The Economic Journal

Drazen,, Allan

doi: 10.1111/j.1468-0297.2007.02031_1.xpmid: N/A

Review I Democratisation as Commitment What are the economic determinants of the transition from authoritarian to democratic rule and democratic consolidation (the process by which, in the oft‐quoted phrase of Linz and Stepan (1996, p. 5), ‘democracy has become ‘‘the only game in town’’… when no significant political groups seriously attempt to overthrow the democratic regime’.)? How do the size of the middle class, external influences (such as ‘globalisation’), or even resource endowments affect these processes? These questions have long interested political scientists and have become central to current policy debates. However, there has been surprisingly little use of formal political economy models to study them, nor even much attention given to these questions by economists. In this book, Acemoglu and Robinson go a long way toward filling this gap. They present a truly path‐breaking analysis of the democratisation, democratic consolidation and reversion to authoritarian rule. What is striking about the book is not simply that they have tackled a set of critically important questions using formal models, but also how they have done it. They present an integrated technical framework of analysis that can (and no doubt will) serve as the starting point for future work on these issues. Moreover, while mathematical formalisation requires simplifying complicated phenomena, the book is rich with conceptual insight and a wealth of evidence on democratic experiences and transitions. This is in the form of case studies, cross‐sectional evidence and historical narratives. The main purpose is to motivate their modelling approach and assumptions, but anyone wanting a good empirical overview of the issues would do well to read this material, even if they had no intention of continuing on to the formal modelling which is the heart of the book. In fact, as the authors point out, much of the literature actually doubts the usefulness of trying to study democratisation and democratic consolidation using a general model. Many readers may think of Huntington’s (1991),The Third Wave, and the wide variety of factors he suggested are associated with the emergence of democracy, suggesting the near impossibility of models of the sort these authors present. Or, consider the argument of Linz and Stepan (1978, p. xi) that ‘the historicity of macro‐political processes precludes the highly abstract generalising of ahistorical social scientific models’. Hence, Acemoglu and Robinson’s construction of a general formal framework in fact provides a challenge to a common approach to studying these questions. To economists who view the approach of modelling as a highly useful tool to organising thought, the challenge is welcome. The Approach and Framework of Analysis After presenting some modelling preliminaries, they proceed with setting out their formal framework. The cliché that ‘mere words cannot do justice’ to their formal analysis is especially apt, for after all, the whole point is the construction of a formal mathematical model. It is not an exercise in simply representing ideas in symbols rather than words, but use of sophisticated game‐theoretic models to illustrate the dynamics of institutional change. (Lest this sound too forbidding, the book should be accessible to any graduate student in economics or political science with standard microeconomic training.) However, page constraints limit me to concentrating on a verbal and conceptual representation. Their approach and basic argument may be summarised as follows. They begin with optimising actors who have well‐defined preferences over outcomes or the consequences of their actions, and who evaluate different options according to their assessments of the economic and social consequences. They argue that deriving support for or opposition to democracy from underlying preferences over primitives (such as income) is preferable to simply assuming that groups have different ideological preferences for democratic versus autocratic forms of government. This sounds quite sensible if ideological preferences were exogenous but in fact they may develop over time. Moreover, the process of building support for democracy per se rather than conditional on the economic outcomes it delivers is itself a central part of the process of democratic consolidation. They quote Diamond (1993, p. 35) that ‘democracy becomes truly stable only when people come to value it widely not solely for its economic and social performance but intrinsically for its political attributes’. I return below to the issue of ‘becoming convinced’ of the value of democracy and its importance. Under the approach of judging a system by the results it generates, the key observation being that almost all policy choices have distributional implications and therefore inherently create conflict. More specifically, they concentrate on fiscal policy and the distributional conflict engendered by choice of alternative tax and transfer schemes. Equilibrium policy favours the group that has political power, that is, the ability to have society’s policy reflect its preferences goals rather than those of another group. Hence, as in any analysis of policy choices when there is a heterogeneity of interests (as there always is), the distribution of political power is crucial in determining the equilibrium policy outcome. Acemoglu and Robinson’s analysis turns on their distinction between two types of political power and the relation of this distinction to political institutions. De facto political power can be thought of, very simply, as power deriving from use of force, reflecting for example more and better weaponry, control of the army, or perhaps even economic power. In almost all societies, however, societal decisions reflect not brute force and the will of the physically or economically strongest, but the operation of institutions of decision making and power allocation, in short, political institutions. They term such power allocated by formal political institutions de jure political power. The key observation that will define their approach to institutional change is that while political influence at a point in time may reflect de facto political power, such power is often transitory (and recognised as such by those who hold it). Hence, current de facto power is no guarantee of future political influence, which will depend on how (presumably more durable) institutions allocate de jure power. I return to how this works shortly. To model the central political conflicts, Acemoglu and Robinson consider a society composed of two groups – minority ‘elites’ who want a non‐democratic regime (as it favours elite interests) and citizens, the majority of the population, who want democracy (as it favours their interests). It is the balance of political power between the two groups that determines whether a society makes the transition from non‐democracy to democracy and whether it stays democratic or reverts to authoritarianism. For much of the analysis, elites are basically analogous to the rich (where their empirical overview is meant to convince readers, among other things, that ‘there is often a close association between what nondemocratic regimes do and what the rich want’.) However, they do not limit class divisions to those along income lines and subsequently show how their framework may be applied when divisions other than between socioeconomic classes, say ethnic identity. What is crucial in applying their framework is the assumption that within‐group decisions (and hence ultimate equilibrium outcomes) depend on economic situation of agents, rather than group identity. This is consistent with their fundamental assumption that the choices of individual agents reflect economic self‐interest. Most economists might agree with the fundamental assumption, but it does leave open the question of how different would political dynamics and outcomes be if not only conflicting groups were defined along non‐economic lines, but also group dynamics reflected considerations other than individual economic self‐interest. At this point the reader might be tempted to say that the argument is straightforward. Once citizens have de facto power in a nondemocracy, they use it for institutional change to ‘cement’ their power. That is, institutional change – democratisation, in this case – results from a shift of de facto power from elites to citizens, who use this power to gain de jure political power and thus continued political influence. Though this formulation summarises their basic line of reasoning, it is too simple to represent their argument, for a number of reasons. First, before the citizenry accumulates enough de facto power to democratise and displace the elites, the elites may (and often do) attempt to repress the citizenry (that is, pro‐democratic forces in order to remain in power. Why do the elites not always use their power under authoritarianism to repress democracy? The basic answer of course is that it is sometimes too costly to repress pressure for democratisation relative to the cost of making concessions. One interesting thing that an analysis based on economic self‐interest delivers here is how the cost of concessions depends on income inequality, this cost being lower the more equal is income distribution in society. Hence, economic development increases the probability of democratisation not simply because attitudes towards nondemocracy change, but also because the elites find it less costly to make concessions. Second, Acemoglu and Robinson point out that in practice, transition to democracy does not occur simply because the citizens have de facto power but often takes place when the elite controlling the existing regime extend voting rights. Why do the elites not give the citizenry the policies they want (but keep power), instead of giving away power? In the absence of moves to satisfy the citizenry, it is assumed the masses will overthrow the elites via a (socially costly) revolution. Hence, satisfying the citizenry requires concessions that are significant not only for today but in the future as well. The elite must make a therefore make a credible commitment to future pro‐majority policies. Simple promises of future citizen‐friendly policies by the elite may not be credible, since citizenry may not have de facto power tomorrow. Hence, they have to transfer political power to the citizens. That is precisely what a transition to democracy does: it shifts future political power away from the elite to the citizens, thereby creating a credible commitment to future pro‐majority policies. Though the argument is logical and tight at this point, some readers might question how widely applicable it is. They consider many examples where democratisation reflects the dynamics of a conflict between the masses and domestic elites (who are concerned about their position under alternative regimes). However, democratisation often reflects the triumph of the citizenry over a foreign (that is, colonial) elite, who do not make institutional concessions to avoid an even worse fate if they did not. The colonial elite go home, often after being forced out by one means or another. Concessions are not chosen over repression because the elites see it as the wiser path; repression was tried and failed. In short, I think their theory of democratisation may be more limited than they suggest but it is a powerful model in many cases. Institutions as a Commitment Device To summarise: political institutions – specifically, democratisation – are the only credible commitment device the elites have not to use de facto power in the future against the citizens. Of course, this raises a third question, namely. For institutions to act as a credible commitment device on future policy, the institutions themselves must be durable. What then makes it credible that democracy as an institution will persist? One might answer that there is a ‘democracy hysteresis’ effect – democracy, once created, becomes difficult to reverse. But this is unconvincing without a more precise argument of why this is so, especially since reversion to authoritarian rule is all too common, a phenomenon studied by Acemoglu and Robinson. Moreover, the question of the persistence of institutions, central to their work on the policy commitment problem being central to political economy, is more general (and even more important, one might say) than the specific question of the durability of democracy. I found their argument here incomplete, since durability of institutions is largely taken as given. They argue that ‘[f]or us, the main difference between policies and institutions is their ‘‘durability’’… Policies are much easier to reverse, whereas institutions are more durable’. (p. 177). It is certainly true that political institutions are more durable and persistent than policies. But, are they perceived to be persistent enough to make them a credible commitment device? The perception that a reversion to the previous institutional arrangement is possible (not to mention perhaps likely) weakens the institutions‐as‐commitment‐device argument. If this perception is strong enough, it would call into question the basic story they are telling. One logical response is that changes in political institutions, by affecting the incentives of both politicians and citizens, may create constituencies that support their continuation. For example, as they wrote in the recent Handbook of Economic Growth, Acemoglu et al. (2005, p. 392), ‘those who hold political power influence the evolution of political institutions, and they will generally opt to maintain the political institutions that give them political power’. On the other hand, in recent work (subsequent to this book), they pursue the question of the persistence of institutions and suggest it is economic rather than political institutions that are persistent. For example, in Acemoglu and Robinson (2006) and companion work, they find that when elites have power, equilibrium changes in political institutions favouring the masses will induce offsetting changes in the distribution of de facto political power, so that economic outcomes may remain largely invariant. In short, here the elites may still have disproportionate influence in politics in spite of democratisation. As they write Political institutions can change from nondemocracy to democracy, changing the distribution of de jure political power. But this may have little effect on (the equilibrium distribution of) economic institutions because now the elite invest more in their de facto political power … (p. 328) In short, it seems to me that it is still an unresolved question of whether political institutions can be thought of as sufficiently persistent in general to ‘solve’ the commitment problem that Acemoglu and Robinson view as key to policy credibility. This is obviously the main question on the agenda, for their model of democratisation rests critically on the answer being positive. Democratic Consolidation One way to think about durability of democracy more specifically is to study the determinants of democratic consolidation. They do this by considering a society where democracy has been created, and policy is chosen by majority vote (so the preferences of the median voter determine fiscal policy). They use the two‐group model developed in previous chapters, associating the elite with the rich and the citizens with the poor, implying that the median voter in democracy will be a poor agent. Democracy is taken to be fully consolidated when there is never any effective coup threat. As with transition to democracy, elites are worse off under democracy and hence anti‐democratic, while citizens are assumed unambiguously pro‐democratic. Hence, the survival of democracy becomes a question simply of whether elites perceive an expected benefit from mounting an anti‐democratic coup, which depends on the probability of success, the costs of mounting a coup, and the (further) costs of failure. Since it is elites who pose the threat to democracy, a central question is the constraints that the possibility of reversion to non‐democracy puts on policy making in a democracy, specifically on the setting of fiscal policy. The greater the threat, the more they must be ‘bought off’. (Given the actual role of the military in coup attempts, it is presumed that they represent the interests of the elite more than those of the citizens, and placating them is part of the same story.) Here I found unwarranted both the basic assumption that after transition to democracy the masses are fully committed to democracy and its implication that they can therefore essentially be taken for granted. In fact, there is considerable evidence that in many new democracies a significant fraction of citizens are not initially convinced democrats or citizens. As Linz and Stepan (1996, p. 144) stress: … the overwhelming majority of consolidated democracies did not actually begin their transition to democracy with a majority of members of the polity or even many of the key agents of the transition being either convinced democrats or citizens who rejected everything about the past regime. Rather, a democratic majority emerges when elites and ordinary citizens alike begin to evaluate, for the societal problems they then face and the overall world within which they then live, that democratic procedures of conflict regulation are better or less dangerous than any other form of governance. Hence, positive public attitudes about democracy evolve over time, implying that a key challenge for policy makers is to engender these attitudes not only among elites but also among the public. The history of the Weimar Republic shows not only that public support of democracy is not automatic after the transition from non‐democracy but also the dire consequences of ignoring the need to build mass support. Determinants of Democracy Having set out the framework for studying transitions from authoritarianism to democracy and vice versa, Acemoglu and Robinson then use their basic framework to examine standard arguments about the factors that make the emergence and consolidation of democracy more or less likely. What makes this approach different is that an explicit model makes more precise how various factors may help or hinder democratisation. Consider, for example, economic development, often seen as a key indicator of the likelihood of transition to democracy in an authoritarian state. Many analyses often seem incomplete because they lack discussion of explicit mechanisms on how higher income or less income inequality lead to the adoption and consolidation of democracy. Development is seen as changing the balance of power between socio‐economic classes, but this is not an explanation of how democracy comes to be adopted. Moreover, an explicit choice‐theoretic model reveals influences that conceptual approaches may miss. A first often‐discussed factor is ‘civil society’, the array of voluntary civic and social organisations and institutions found in politically advanced countries. The link that the authors stress runs through the collective action problem that the citizenry must solve if they are to use their de facto power to force institutional change. As groups that have ‘solved’ the problem of uncoerced collective action, civil society may provide the necessary organisation for the citizenry. Second is the influence of shocks and crises, that is, how they might make democratisation more likely. First, as in many models, crises make it easier to solve collective action problems. These crises are inherent, but lead to large short‐term fluctuations in de facto political power. Conversely, crises may also give anti‐democratic forces temporary de facto power so that anti‐democratic coups are also more likely. A third, less‐discussed influence is the sources of income and the composition of wealth. A major focus is whether the wealthy elite are owners of land or of capital (either physical or human). They suggest that for a number of reasons, landowners may see themselves as having more to lose under democracy than capital owners (for example, land may be easier to tax than physical and human capital) and more willing to use force to prevent democratisation (as capital owners require a larger degree of cooperation from workers to produce). Hence they argue that democratisation will be more likely in a more industrialised society than a more agricultural one, a democratic consolidation more difficult. A higher degree of inequality between the masses and the elites is often argued to increase the probability of violent overthrow of the elites, which their model supports. Larger inter‐group inequality means that once the old regime is overthrown, the extra income the citizenry will share (net of what is destroyed by revolution) will be larger. Of course, larger inter‐group inequality means democracy is more costly for the elite, so that repression becomes more attractive. Hence, they argue that there is an inverted U‐shaped relationship between inter‐group inequality and the likelihood of transition to democracy. Many studies of democratisation assign a central role to the middle class. Though their basic theory only distinguishes between two groups, the elite and the citizens, they present an extended version of the model to take the role of a class between the elite and the great mass of citizens into account. Due to a higher degree of education, the middle class may be better able to organise the use of de facto against the elite, as is common in many analyses. Moreover, if the existence of a middle class limits the degree of redistribution the masses can achieve under democracy (their higher incomes mean they will support fiscal policies closer to those preferred by the elite), the elites are more likely to accept rather than repress democracy. Hence, a strong middle class may make democratisation more likely, independent of any role it may play as the vanguard of institutional change. Globalisation has a number of potential effects on democratisation and consolidation. The increase in wages after a trade opening in an authoritarian, developing economy reduces the gap between the incomes of labour and capital, reducing pressure for but also resistance to democratisation. In the opposite direction, international financial integration makes capital flight easier, reducing the extent of redistribution from elites after the transition to democracy. Increased international trade increases the cost of the disruption of economic activity induced by repression (but perhaps also the cost of revolution). In short, given the various effects, their framework does not imply an unambiguous effect of globalisation on pressure for democratisation. Political Institutions Another set of issues concerns the nature of political institutions themselves. The nature of democratic political institutions determine how much power the elite can exercise under democracy and hence may be crucial for explaining why some societies democratise but others do not. One way they can do this is through the design of democratic institutions. Explicit consideration of institutions may appear standard in political economy nowadays, but it contrasts with the approach often found in analyses of democratisation. For example, many studies stress the role of class conflict but provide no explicit mechanism by which a democratic transition is induced. Their approach allows a far more careful consideration of how the mediation of conflicting interests via political institutions affects outcomes. To take one example, consider the provisions in the constitution of a new democracy concerning the role of former elites. Constitutions often protect the military in the transition to democracy (Chile, Korea, Turkey), or the position of previously privileged groups (whites in South Africa). Acemoglu and Robinson point out that such provisions that make elites feel less threatened by democracy may not only enable a process of transition to democracy but also (perhaps ironically) have helped democratic consolidation. Too Much of a Good Thing? This is an extremely impressive book and will no doubt shape both the research agenda of people working on democratisation and democratic consolidation and the modelling choices they make. So impressive in fact that one hopes it does not crowd out alternative modelling based on different assumptions and perspectives. Perhaps it sounds funny to worry about an approach unintentionally stifling intellectual competition, but such things happen. Such a concern illustrates just how influential this book is likely to be. References Acemoglu , Daron , Johnson , Simon and Robinson , James ( 2005 ) ‘Institutions as a fundamental cause of long‐run growth’, in ( Philippe Aghion and Steven Durlauf, eds.), Handbook of Economic Growth , Volume 1A. pp. 386 – 472 , Amsterdam : Elsevier. Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Acemoglu , Daron and Robinson , James ( 2006 ), ‘De facto political power and institutional persistence’ , AER Papers and Proceedings , vol. 96 , pp. 325 – 9 . OpenURL Placeholder Text WorldCat Diamond , Larry J. ( 1993 ) ‘Economic development and democracy reconsidered’ , American Behavioral Scientist , vol. 35 , pp. 450 – 99 . Google Scholar Crossref Search ADS WorldCat Huntington , Samuel P. ( 1991 ) The Third Wave: Democratization in the Late Twentieth Century , Norman OK: University of Oklahoma Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Linz , Juan J. and Stepan , Alfred ( 1978 ) The Breakdown of Democratic Regimes , Baltimore MD: Johns Hopkins University Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Linz , Juan J. and Stepan , Alfred ( 1996 ), Problems of Democratic Transition and Consolidation: Southern Europe, South America, and Post‐communist Europe , Baltimore MD: Johns Hopkins University Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC © The Author(s). Journal compilation © Royal Economic Society 2007

Cull,, Robert;Demirgu¨ç‐Kunt,, Asli;Morduch,, Jonathan

doi: 10.1111/j.1468-0297.2007.02017.xpmid: N/A

Abstract Microfinance promises to reduce poverty by employing profit‐making banking practices in low‐income communities. Many microfinance institutions have secured high loan repayment rates but, so far, relatively few earn profits. We examine why this promise remains unmet. We explore patterns of profitability, loan repayment, and cost reduction with unusually high‐quality data on 124 institutions in 49 countries. The evidence shows the possibility of earning profits while serving the poor, but a trade‐off emerges between profitability and serving the poorest. Raising fees to very high levels does not ensure greater profitability and the benefits of cost‐cutting diminish when serving better‐off customers. Running banks in low‐income communities is not easy. One of the great accomplishments of the economics of information, after all, has been to show how information asymmetries undermine credit markets in places where potential customers have few assets to offer as collateral (Besley, 1995). Microfinance providers, though, have specialised in making uncollateralised loans in low‐income communities. Through innovative contracts and new microfinance management practices, institutions are generating high loan repayment rates in contexts as diverse as the slums of Dhaka, war‐torn Bosnia, and rural Senegal. In doing so, microfinance providers have forced economic theorists to re‐think pessimistic views of the scope for improving credit markets.1 But microfinance would be a grand failure if securing high repayment rates was all there was to it. Meeting the full promise of microfinance – to reduce poverty without ongoing subsidies – requires translating high repayment rates into profits, a challenge that remains for most microbanks.2 The overall equation linking capital and labour inputs into profits and social change still proves difficult to master. We take a close look at this equation with unusually high‐quality financial information on 124 institutions in 49 countries; the institutions are united by claiming strong commitments to achieving financial self‐sufficiency and a willingness to open their accounts to careful scrutiny.3 The institutions thus represent some of the best hopes for achieving poverty reduction with profit (or at least without ongoing subsidy). Still, the average share of funding (total liabilities plus total equity) made up of subsidy exceeds 20% in this sample. The data do not allow us to answer the big (and controversial) question: can such ongoing subsidy be justified? Answering that would require reliable data on social impacts, and the evidence is scant. The data, though, allow us to illuminate other important questions for the first time in a large comparative survey. Does raising interest rates exacerbate agency problems as detected by lower loan repayment rates and less profitability? Is there evidence of a trade‐off between the depth of outreach to the poor and the pursuit of profitability? Has ‘mission drift’ occurred – i.e., have microbanks moved away from serving their poorer clients in pursuit of commercial viability? The questions are at the heart of debates within academic economics, as well as being of immediate relevance for policymakers and practitioners. As with other cross‐country analyses, the aim is to describe patterns in the data. There is insufficient exogenous variation in key variables to reliably estimate causal impacts, so we focus on associations that help to illuminate and frame key debates. Since the institutions in the survey are more focused on financial performance than typical microbanks, we expect that the trade‐offs described below are even starker for institutions that did not participate in the survey. Our results bring some good news for microfinance advocates. First, over half of the institutions in the survey were profitable after accounting adjustments were made (although the average return on assets is negative overall). Others are approaching profitability and should be able to soon achieve financial self‐sufficiency. Second, simple correlations show little evidence of agency problems, outreach‐profit trade‐offs or mission drift. The correlations thus attest to the possibility of raising interest rates without undermining repayment rates, achieving both profit and substantial outreach to poorer populations, and staying true to initial social missions even when aggressively pursuing commercial goals. Disaggregating by lending type, though, uncovers trade‐offs and tensions, even among these leading institutions. The patterns of profitability and the nature of customers vary considerably with the design of the institutions and their contracts. Microfinance lenders use a variety of approaches to lending, and we focus on three main categories. The best‐known approach is ‘group lending’, made popular within microfinance by the Grameen Bank of Bangladesh and BancoSol in Bolivia. The method uses self‐formed groups of customers that assume joint liability for the repayment of loans given to group members. The joint liability contract can, in principle, mitigate moral hazard and adverse selection by harnessing local information and enforcement possibilities and putting them to use for the bank. (Ahlin and Townsend, (2007) and Cassar et al. (2007), provide theoretical perspectives and empirical tests of group mechanisms in this Feature.) Another method is village banking, based on larger groups but a similar notion of joint liability – the focus of Karlan (2007), also in this Feature. The third main method is ‘individual‐based lending’, which draws on traditional banking practices and involves a standard bilateral relationship between the bank and customer – and, in the absence of other interventions, is most vulnerable to problems imposed by information asymmetries and weak enforcement capacities. The data set contains institutions representative of each approach: 20 institutions based on village banks, 56 individual‐based lenders, and 48 group‐based lenders. Our findings on the latter two groups are generally robust across specifications. We find some institutions that have both achieved profitability and meaningful outreach to the poor but disaggregation by lending‐type reveals trade‐offs between the two objectives. Individual‐based lenders as a group have the highest average profit levels but they perform least well on measures of outreach. Taking average loan size as a proxy for the poverty level of customers (smaller loans indicate poorer customers), individual‐based institutions lend with an average size of $1,220, while village banks (the most subsidy‐dependent category of institutions) lend with an average loan size of $148. Village banks also serve a much larger fraction of women (88%) relative to individual‐based lenders (46%). The economics of information yields a series of predictions that are explored here. First is the hypothesis that raising interest rates will undermine portfolio quality due to adverse selection (Stiglitz and Weiss, 1981) and moral hazard. Evidence consistent with the hypothesis emerges (up to a point) in the sub‐sample of banks that do not use group‐based methods to address information problems. Specifically, the fraction of the loan ‘portfolio at risk’ rises with interest rates for most of the institutions that employ individual‐based lending methods. With respect to profitability, raising interest rates helps – but only up to a point. Beyond an interest rate equivalent to 60% per year, higher rates are no longer associated with higher profits for individual‐based lenders. The evidence is consistent with falling demand (and thus reduced scale economies) at higher rates, coupled with limits on the ability to leverage assets. Neither of these relationships holds in this sample for group lenders. These results generate a preliminary puzzle: why then is the individual‐based approach ever favoured over either of the group‐based methods? The puzzle is sharpened by knowledge that two pioneers of group‐based lending (the Grameen Bank of Bangladesh and BancoSol of Bolivia) have now switched to individual‐based models. One clue is that the individual‐based lenders here provide substantially larger‐sized loans, as noted above. Taking loan size as a proxy for poverty levels, the evidence suggests that the group methods become cumbersome for customers who are less poor and who are willing and able to invest in larger businesses. Working with customers able to use larger loans can be an important path to financial self‐sufficiency for lenders. Taking this path veers from the traditional focus of microfinance (with its emphasis on making smaller loans at as wide a scale as possible), but the shift could improve overall welfare: it is not just the poorest that demand and can take advantage of better access to finance. We find some evidence of ‘reverse mission drift’ for individual‐based lenders as a class: i.e., once an institution is established, pursuing higher profits and focusing on poorer customers can go hand in hand. At the same time, the data show that larger microbanks on average have lower measures of outreach. This last finding is consistent with an important trade‐off between the breadth of outreach (scale) and the depth of outreach (reaching the poor). The question remains open as to whether larger institutions serve an absolutely greater number of the very poor – a question that can only be answered with disaggregated data. 1. Data and Empirical Approach Empirical progress on understanding the trade‐offs in microfinance has been held back by the lack of variation in prices and programme elements necessary for identification of key parameters. Most financial institutions offer their clients a uniform set of products and they seldom change the product mix, price, or design – or institutions change policies in ways that make it difficult for researchers to disentangle patterns of product changes versus other contemporaneous changes.4 The cross‐country data here, however, provide substantial variation in contractual types, prices, institutional sizes and locations, and target markets. The variation provides a means to describe the nature and trade‐offs of lending relationships. The data on 124 microfinance institutions (MFIs) in 49 developing countries were collected by the Microfinance Information Exchange (or the MIX), a not‐for‐profit private organisation that aims to promote information exchange in the microfinance industry. The database contains one observation per institution from 1999 to 2002; 70% of the observations are from 2002. These data, collected for publication in the MicroBanking Bulletin (MBB), have been adjusted to help ensure comparability across institutions. The adjustments, which are summarised in the Appendix, include an inflation adjustment, a reclassification of some long‐term liabilities as equity, an adjustment for the cost of subsidised funding, an adjustment for current‐year cash donations to cover operating expenses, an in‐kind subsidy adjustment for donated goods and services, loan loss reserve and provisioning adjustments, some adjustments for write‐offs, and the reversal of any interest income accrued on non‐performing loans. The institutions were selected based in large part on the quality and extent of their data. The data set is thus not representative of all microfinance institutions. They do, however, collectively serve a large fraction of microfinance customers worldwide. A sense of the skewed size distribution of microfinance is given by a recent analysis of data provided by the Microcredit Summit organisation, a data set whose top end largely overlaps the data here. Honohan (2004, p. 3) finds that ‘the largest 30 microfinance firms account between them for more than 90 per cent of the clients served worldwide by the 234 top firms (and hence for more than three‐quarters of those served by all of the 2,572 firms reporting to the Microcredit Summit)’. While we cannot make a similar comparison here, Honohan’s evidence suggests that during the sample period the banks here served over half of all microfinance customers worldwide. An important feature of our data is qualitative information on the lending style employed by the MFI, the range of the services it offers, its profit status, ownership structure, and sources of funds. These detailed data enable us to offer a more complete analysis of MFI performance by lending type than has been possible before. Lending methods vary across regions, as shown in Table 3 (classifications and acronyms are those employed by the World Bank). There are no village banks in East Asia in the sample, for example. Individual‐based lending predominates in East Asia and the Pacific, while institutions in South Asia and Sub‐Saharan Africa tend to lend through group mechanisms. Institutions in Eastern Europe and North Africa do not strongly favour either individual‐based or group lending. Summary statistics at the bottom of Table 1 indicate that, with the possible exception of Latin America and the Caribbean (LAC), our sample is reasonably balanced across regions. 17% of the institutions come from Eastern Europe and Central Asia (ECA), another 17% from Sub‐Saharan Africa (AFR). South Asian (SA) institutions comprise 10% of the sample, while institutions from East Asia and the Pacific (EAP) and the Middle East and North Africa (MENA) comprise 9% and 7%, respectively. Institutions from Latin America and the Caribbean comprise 40% of the sample. Table 1 Variable Description and Summary Statistics Variable Name . Definition . Mean . Median . Minimum . Maximum . Financial self‐sufficiency Adjusted operating revenue / Adjusted (financial expense + loan loss provision expense + operating expense) 1.035 1.016 0.146 2.183 Operational self‐sufficiency Operating revenue / (Financial expense + loan loss provision expense + operating expense) 1.165 1.115 0.157 3.872 Return on assets adjusted Adjusted net operating income after taxes / Average total assets −0.027 0.002 −1.541 0.280 Average loan size to GNP per capita 0.676 0.376 0.025 5.831 Age Age of the MFI in years 9.300 8 2 40 Size of MFI indicator Size of the loan portfolio, which is 1 for small, 2 for medium and 3 for large. 2.025 2 1 3 For‐profit status For‐profit is 1, non‐profit is 0. 0.237 0 0 1 Village bank lender The MFI does village bank style lending (as opposed to MFI who do individual lending or solidarity lending). 0.165 0 0 1 Solidarity lender The MFI does some solidarity style lending (as opposed to MFI who do only individual lending or do village bank lending). 0.397 0 0 1 Real gross portfolio yield [Yield on gross portfolio (nominal) – Inflation rate] / (1 + Inflation rate) 0.354 0.305 0.051 1.059 Capital costs to Assets (Rent + transportion + depreciation + office + other) / total assets 0.089 0.062 0.01 0.725 Labour costs to Assets Personnel expenses/total assets 0.1 0.078 0 0.794 Loans to Assets Gross loan portfolio/total assets 0.689 0.726 0.077 0.987 Donations to loan portfolio Donations for financial services/gross loan portfolio 0.122 0.005 0 2.081 Average loan size to GNP per capita of the poorest 20% 2.983 1.324 0.108 19.511 Average loan size In US dollars. 715.698 360.500 36.000 5131.231 Women borrowers Percentage of borrowers who are women. 0.649 0.615 0.150 1 Eastern Europe and Central Asia 0.169 0 0 1 Africa 0.169 0 0 1 Middle East and North Africa 0.073 0 0 1 South Asia 0.097 0 0 1 East Asia and the Pacific 0.089 0 0 1 Variable Name . Definition . Mean . Median . Minimum . Maximum . Financial self‐sufficiency Adjusted operating revenue / Adjusted (financial expense + loan loss provision expense + operating expense) 1.035 1.016 0.146 2.183 Operational self‐sufficiency Operating revenue / (Financial expense + loan loss provision expense + operating expense) 1.165 1.115 0.157 3.872 Return on assets adjusted Adjusted net operating income after taxes / Average total assets −0.027 0.002 −1.541 0.280 Average loan size to GNP per capita 0.676 0.376 0.025 5.831 Age Age of the MFI in years 9.300 8 2 40 Size of MFI indicator Size of the loan portfolio, which is 1 for small, 2 for medium and 3 for large. 2.025 2 1 3 For‐profit status For‐profit is 1, non‐profit is 0. 0.237 0 0 1 Village bank lender The MFI does village bank style lending (as opposed to MFI who do individual lending or solidarity lending). 0.165 0 0 1 Solidarity lender The MFI does some solidarity style lending (as opposed to MFI who do only individual lending or do village bank lending). 0.397 0 0 1 Real gross portfolio yield [Yield on gross portfolio (nominal) – Inflation rate] / (1 + Inflation rate) 0.354 0.305 0.051 1.059 Capital costs to Assets (Rent + transportion + depreciation + office + other) / total assets 0.089 0.062 0.01 0.725 Labour costs to Assets Personnel expenses/total assets 0.1 0.078 0 0.794 Loans to Assets Gross loan portfolio/total assets 0.689 0.726 0.077 0.987 Donations to loan portfolio Donations for financial services/gross loan portfolio 0.122 0.005 0 2.081 Average loan size to GNP per capita of the poorest 20% 2.983 1.324 0.108 19.511 Average loan size In US dollars. 715.698 360.500 36.000 5131.231 Women borrowers Percentage of borrowers who are women. 0.649 0.615 0.150 1 Eastern Europe and Central Asia 0.169 0 0 1 Africa 0.169 0 0 1 Middle East and North Africa 0.073 0 0 1 South Asia 0.097 0 0 1 East Asia and the Pacific 0.089 0 0 1 Source. Authors’ calculations, based on data from the Microfinance Information eXchange, Inc. * indicates significance at the 5% level. Open in new tab Table 1 Variable Description and Summary Statistics Variable Name . Definition . Mean . Median . Minimum . Maximum . Financial self‐sufficiency Adjusted operating revenue / Adjusted (financial expense + loan loss provision expense + operating expense) 1.035 1.016 0.146 2.183 Operational self‐sufficiency Operating revenue / (Financial expense + loan loss provision expense + operating expense) 1.165 1.115 0.157 3.872 Return on assets adjusted Adjusted net operating income after taxes / Average total assets −0.027 0.002 −1.541 0.280 Average loan size to GNP per capita 0.676 0.376 0.025 5.831 Age Age of the MFI in years 9.300 8 2 40 Size of MFI indicator Size of the loan portfolio, which is 1 for small, 2 for medium and 3 for large. 2.025 2 1 3 For‐profit status For‐profit is 1, non‐profit is 0. 0.237 0 0 1 Village bank lender The MFI does village bank style lending (as opposed to MFI who do individual lending or solidarity lending). 0.165 0 0 1 Solidarity lender The MFI does some solidarity style lending (as opposed to MFI who do only individual lending or do village bank lending). 0.397 0 0 1 Real gross portfolio yield [Yield on gross portfolio (nominal) – Inflation rate] / (1 + Inflation rate) 0.354 0.305 0.051 1.059 Capital costs to Assets (Rent + transportion + depreciation + office + other) / total assets 0.089 0.062 0.01 0.725 Labour costs to Assets Personnel expenses/total assets 0.1 0.078 0 0.794 Loans to Assets Gross loan portfolio/total assets 0.689 0.726 0.077 0.987 Donations to loan portfolio Donations for financial services/gross loan portfolio 0.122 0.005 0 2.081 Average loan size to GNP per capita of the poorest 20% 2.983 1.324 0.108 19.511 Average loan size In US dollars. 715.698 360.500 36.000 5131.231 Women borrowers Percentage of borrowers who are women. 0.649 0.615 0.150 1 Eastern Europe and Central Asia 0.169 0 0 1 Africa 0.169 0 0 1 Middle East and North Africa 0.073 0 0 1 South Asia 0.097 0 0 1 East Asia and the Pacific 0.089 0 0 1 Variable Name . Definition . Mean . Median . Minimum . Maximum . Financial self‐sufficiency Adjusted operating revenue / Adjusted (financial expense + loan loss provision expense + operating expense) 1.035 1.016 0.146 2.183 Operational self‐sufficiency Operating revenue / (Financial expense + loan loss provision expense + operating expense) 1.165 1.115 0.157 3.872 Return on assets adjusted Adjusted net operating income after taxes / Average total assets −0.027 0.002 −1.541 0.280 Average loan size to GNP per capita 0.676 0.376 0.025 5.831 Age Age of the MFI in years 9.300 8 2 40 Size of MFI indicator Size of the loan portfolio, which is 1 for small, 2 for medium and 3 for large. 2.025 2 1 3 For‐profit status For‐profit is 1, non‐profit is 0. 0.237 0 0 1 Village bank lender The MFI does village bank style lending (as opposed to MFI who do individual lending or solidarity lending). 0.165 0 0 1 Solidarity lender The MFI does some solidarity style lending (as opposed to MFI who do only individual lending or do village bank lending). 0.397 0 0 1 Real gross portfolio yield [Yield on gross portfolio (nominal) – Inflation rate] / (1 + Inflation rate) 0.354 0.305 0.051 1.059 Capital costs to Assets (Rent + transportion + depreciation + office + other) / total assets 0.089 0.062 0.01 0.725 Labour costs to Assets Personnel expenses/total assets 0.1 0.078 0 0.794 Loans to Assets Gross loan portfolio/total assets 0.689 0.726 0.077 0.987 Donations to loan portfolio Donations for financial services/gross loan portfolio 0.122 0.005 0 2.081 Average loan size to GNP per capita of the poorest 20% 2.983 1.324 0.108 19.511 Average loan size In US dollars. 715.698 360.500 36.000 5131.231 Women borrowers Percentage of borrowers who are women. 0.649 0.615 0.150 1 Eastern Europe and Central Asia 0.169 0 0 1 Africa 0.169 0 0 1 Middle East and North Africa 0.073 0 0 1 South Asia 0.097 0 0 1 East Asia and the Pacific 0.089 0 0 1 Source. Authors’ calculations, based on data from the Microfinance Information eXchange, Inc. * indicates significance at the 5% level. Open in new tab Although we include regional dummy variables in the regressions that follow, the regional preferences for certain lending types should be kept in mind when interpreting results. For example, cultural factors could tip countries or regions in the direction of one lending type over another, and it could be these social factors that are ultimately driving the relationships we find rather than lending methods. To address this concern we have re‐run our models using data from institutions located in LAC, the region with the greatest number of institutions to study. Of the fifty MFIs from LAC in our sample, there are 32 individual‐based lenders, 10 solidarity group lenders and eight village banks. The base results that follow on financial performance (Table 7), operating costs/loan size trade‐offs (Table 10) and mission drift (Table 11) are quite similar to those for the LAC sub‐sample. To conserve space, we do not report the LAC results below but the results are available from the authors. 1.1. Dependent Variables The key dependent variable in our analysis of profitability is the financial self‐sufficiency (FSS) ratio, a measure of an institution’s ability to generate sufficient revenue to cover its costs.5 Values below one indicate that it is not doing so. The financial self‐sufficiency ratio bests other measures of financial performance because the data are adjusted as described above and because it offers a more complete summary of inputs and outputs than standard financial ratios such as return on assets or equity. For robustness, however, we also use as dependent variables an unadjusted measure of operation self‐sufficiency (OSS) and a measure of adjusted return on assets (ROA).6 Table 2 shows that the correlations between financial outcomes (FSS, OSS, and ROA) are positive and significant, but not perfect (ranging from 0.59 to 0.90). The three measures are also significantly positively correlated with the age and size of institutions. Regression analyses allow us to investigate the strength of those correlations after controlling for region, lending type, and other relevant covariates. Table 2 Correlations . Financial Self‐ Suffiency . Operational Self‐ Suffiency . Return on Assets adjusted . Average Loan Size to GNPPC . Age . Size of MFI . For‐Profit Status . Village Bank Lender . Solidarity Lender . Real Gross Portfolio Yield . Capital Costs to Assets . Labor Costs to Assets . Loans to Assets . Donations to Loan Portfolio . ALS to GNPPC of the poorest . Operational Self‐Suffiency 0.8940* 1 123 124 Return on Assets adjusted 0.6922* 0.5943* 1 122 123 123 ALS to GNPPC 0.0621 0.0828 0.1174 1 116 117 117 117 Age 0.2562* 0.1858* 0.1899* 0.1521 1 119 120 119 113 120 Size 0.3525* 0.2962* 0.3517* 0.3408* 0.2285* 1 121 122 121 115 118 122 For‐Profit Status 0.0607 −0.0141 0.0929 0.2073* −0.0427 0.2664* 1 117 118 118 112 114 117 118 Village Bank −0.1197 −0.0738 −0.1323 −0.2594* −0.1568 −0.2744* 0.2487* 1 120 121 120 114 117 117 121 120 Solidarity Lender −0.1007 −0.0926 −0.0886 −0.1374 −0.0917 −0.038 0.0575 −0.3608* 1 120 121 120 114 117 120 117 121 121 Portfolio Yield −0.0444 −0.1439 −0.0402 −0.4064* −0.2196* 0.2960* 0.0305 0.3805* −0.0938 1 120 121 121 115 117 119 116 119 118 122 Capital Costs to Assets −0.4218* −0.4048* −0.5896* −0.2380* −0.2343* −0.2648 −0.0933 0.3634* −0.0544 0.4544* 1 123 124 123 117 120 122 118 121 121 121 124 Labor Costs to Assets −0.3779* −0.4006* −0.7095* −0.2669* −0.2364* −0.3226* −0.1567 0.2801* 0.1312 0.4359* 0.6482* 1 123 124 123 117 120 122 118 121 121 121 124 124 Loans to Assets 0.2673* 0.1383 0.2648* 0.1701 0.0161 0.3284* 0.0267 −0.2011* 0.1067 −0.2770* −0.0785 −0.0893 1 123 124 123 117 120 122 118 121 121 121 124 124 124 Donations to Loans −0.4684* −0.4382* −0.6707* −0.2052* −0.2001* −0.4045* −0.178 0.2294* 0.1023 0.2148* 0.3672* 0.5264* −0.4439* 1 123 124 123 117 120 122 118 121 121 121 124 124 124 124 ALS to GNPPC of the Poorest 20% 0.0012 −0.0077 0.0891 0.8653* 0.3922* 0.3098* 0.0768 −0.2508* −0.2632* −0.4471* −0.2379* −0.2603* 0.1157 −0.1682 1 102 102 102 99 98 100 97 99 99 102 102 102 102 102 102 Women Borrowers −0.1605 −0.2277* −0.2092* −0.3567* −0.1124 −0.3404* −0.2563* 0.3760* 0.3259* 0.2901* 0.2464* 0.3092* −0.0091* 0.2873* −0.3778* 113 114 114 108 110 112 109 111 111 113 114 114 114 114 94 . Financial Self‐ Suffiency . Operational Self‐ Suffiency . Return on Assets adjusted . Average Loan Size to GNPPC . Age . Size of MFI . For‐Profit Status . Village Bank Lender . Solidarity Lender . Real Gross Portfolio Yield . Capital Costs to Assets . Labor Costs to Assets . Loans to Assets . Donations to Loan Portfolio . ALS to GNPPC of the poorest . Operational Self‐Suffiency 0.8940* 1 123 124 Return on Assets adjusted 0.6922* 0.5943* 1 122 123 123 ALS to GNPPC 0.0621 0.0828 0.1174 1 116 117 117 117 Age 0.2562* 0.1858* 0.1899* 0.1521 1 119 120 119 113 120 Size 0.3525* 0.2962* 0.3517* 0.3408* 0.2285* 1 121 122 121 115 118 122 For‐Profit Status 0.0607 −0.0141 0.0929 0.2073* −0.0427 0.2664* 1 117 118 118 112 114 117 118 Village Bank −0.1197 −0.0738 −0.1323 −0.2594* −0.1568 −0.2744* 0.2487* 1 120 121 120 114 117 117 121 120 Solidarity Lender −0.1007 −0.0926 −0.0886 −0.1374 −0.0917 −0.038 0.0575 −0.3608* 1 120 121 120 114 117 120 117 121 121 Portfolio Yield −0.0444 −0.1439 −0.0402 −0.4064* −0.2196* 0.2960* 0.0305 0.3805* −0.0938 1 120 121 121 115 117 119 116 119 118 122 Capital Costs to Assets −0.4218* −0.4048* −0.5896* −0.2380* −0.2343* −0.2648 −0.0933 0.3634* −0.0544 0.4544* 1 123 124 123 117 120 122 118 121 121 121 124 Labor Costs to Assets −0.3779* −0.4006* −0.7095* −0.2669* −0.2364* −0.3226* −0.1567 0.2801* 0.1312 0.4359* 0.6482* 1 123 124 123 117 120 122 118 121 121 121 124 124 Loans to Assets 0.2673* 0.1383 0.2648* 0.1701 0.0161 0.3284* 0.0267 −0.2011* 0.1067 −0.2770* −0.0785 −0.0893 1 123 124 123 117 120 122 118 121 121 121 124 124 124 Donations to Loans −0.4684* −0.4382* −0.6707* −0.2052* −0.2001* −0.4045* −0.178 0.2294* 0.1023 0.2148* 0.3672* 0.5264* −0.4439* 1 123 124 123 117 120 122 118 121 121 121 124 124 124 124 ALS to GNPPC of the Poorest 20% 0.0012 −0.0077 0.0891 0.8653* 0.3922* 0.3098* 0.0768 −0.2508* −0.2632* −0.4471* −0.2379* −0.2603* 0.1157 −0.1682 1 102 102 102 99 98 100 97 99 99 102 102 102 102 102 102 Women Borrowers −0.1605 −0.2277* −0.2092* −0.3567* −0.1124 −0.3404* −0.2563* 0.3760* 0.3259* 0.2901* 0.2464* 0.3092* −0.0091* 0.2873* −0.3778* 113 114 114 108 110 112 109 111 111 113 114 114 114 114 94 Open in new tab Table 2 Correlations . Financial Self‐ Suffiency . Operational Self‐ Suffiency . Return on Assets adjusted . Average Loan Size to GNPPC . Age . Size of MFI . For‐Profit Status . Village Bank Lender . Solidarity Lender . Real Gross Portfolio Yield . Capital Costs to Assets . Labor Costs to Assets . Loans to Assets . Donations to Loan Portfolio . ALS to GNPPC of the poorest . Operational Self‐Suffiency 0.8940* 1 123 124 Return on Assets adjusted 0.6922* 0.5943* 1 122 123 123 ALS to GNPPC 0.0621 0.0828 0.1174 1 116 117 117 117 Age 0.2562* 0.1858* 0.1899* 0.1521 1 119 120 119 113 120 Size 0.3525* 0.2962* 0.3517* 0.3408* 0.2285* 1 121 122 121 115 118 122 For‐Profit Status 0.0607 −0.0141 0.0929 0.2073* −0.0427 0.2664* 1 117 118 118 112 114 117 118 Village Bank −0.1197 −0.0738 −0.1323 −0.2594* −0.1568 −0.2744* 0.2487* 1 120 121 120 114 117 117 121 120 Solidarity Lender −0.1007 −0.0926 −0.0886 −0.1374 −0.0917 −0.038 0.0575 −0.3608* 1 120 121 120 114 117 120 117 121 121 Portfolio Yield −0.0444 −0.1439 −0.0402 −0.4064* −0.2196* 0.2960* 0.0305 0.3805* −0.0938 1 120 121 121 115 117 119 116 119 118 122 Capital Costs to Assets −0.4218* −0.4048* −0.5896* −0.2380* −0.2343* −0.2648 −0.0933 0.3634* −0.0544 0.4544* 1 123 124 123 117 120 122 118 121 121 121 124 Labor Costs to Assets −0.3779* −0.4006* −0.7095* −0.2669* −0.2364* −0.3226* −0.1567 0.2801* 0.1312 0.4359* 0.6482* 1 123 124 123 117 120 122 118 121 121 121 124 124 Loans to Assets 0.2673* 0.1383 0.2648* 0.1701 0.0161 0.3284* 0.0267 −0.2011* 0.1067 −0.2770* −0.0785 −0.0893 1 123 124 123 117 120 122 118 121 121 121 124 124 124 Donations to Loans −0.4684* −0.4382* −0.6707* −0.2052* −0.2001* −0.4045* −0.178 0.2294* 0.1023 0.2148* 0.3672* 0.5264* −0.4439* 1 123 124 123 117 120 122 118 121 121 121 124 124 124 124 ALS to GNPPC of the Poorest 20% 0.0012 −0.0077 0.0891 0.8653* 0.3922* 0.3098* 0.0768 −0.2508* −0.2632* −0.4471* −0.2379* −0.2603* 0.1157 −0.1682 1 102 102 102 99 98 100 97 99 99 102 102 102 102 102 102 Women Borrowers −0.1605 −0.2277* −0.2092* −0.3567* −0.1124 −0.3404* −0.2563* 0.3760* 0.3259* 0.2901* 0.2464* 0.3092* −0.0091* 0.2873* −0.3778* 113 114 114 108 110 112 109 111 111 113 114 114 114 114 94 . Financial Self‐ Suffiency . Operational Self‐ Suffiency . Return on Assets adjusted . Average Loan Size to GNPPC . Age . Size of MFI . For‐Profit Status . Village Bank Lender . Solidarity Lender . Real Gross Portfolio Yield . Capital Costs to Assets . Labor Costs to Assets . Loans to Assets . Donations to Loan Portfolio . ALS to GNPPC of the poorest . Operational Self‐Suffiency 0.8940* 1 123 124 Return on Assets adjusted 0.6922* 0.5943* 1 122 123 123 ALS to GNPPC 0.0621 0.0828 0.1174 1 116 117 117 117 Age 0.2562* 0.1858* 0.1899* 0.1521 1 119 120 119 113 120 Size 0.3525* 0.2962* 0.3517* 0.3408* 0.2285* 1 121 122 121 115 118 122 For‐Profit Status 0.0607 −0.0141 0.0929 0.2073* −0.0427 0.2664* 1 117 118 118 112 114 117 118 Village Bank −0.1197 −0.0738 −0.1323 −0.2594* −0.1568 −0.2744* 0.2487* 1 120 121 120 114 117 117 121 120 Solidarity Lender −0.1007 −0.0926 −0.0886 −0.1374 −0.0917 −0.038 0.0575 −0.3608* 1 120 121 120 114 117 120 117 121 121 Portfolio Yield −0.0444 −0.1439 −0.0402 −0.4064* −0.2196* 0.2960* 0.0305 0.3805* −0.0938 1 120 121 121 115 117 119 116 119 118 122 Capital Costs to Assets −0.4218* −0.4048* −0.5896* −0.2380* −0.2343* −0.2648 −0.0933 0.3634* −0.0544 0.4544* 1 123 124 123 117 120 122 118 121 121 121 124 Labor Costs to Assets −0.3779* −0.4006* −0.7095* −0.2669* −0.2364* −0.3226* −0.1567 0.2801* 0.1312 0.4359* 0.6482* 1 123 124 123 117 120 122 118 121 121 121 124 124 Loans to Assets 0.2673* 0.1383 0.2648* 0.1701 0.0161 0.3284* 0.0267 −0.2011* 0.1067 −0.2770* −0.0785 −0.0893 1 123 124 123 117 120 122 118 121 121 121 124 124 124 Donations to Loans −0.4684* −0.4382* −0.6707* −0.2052* −0.2001* −0.4045* −0.178 0.2294* 0.1023 0.2148* 0.3672* 0.5264* −0.4439* 1 123 124 123 117 120 122 118 121 121 121 124 124 124 124 ALS to GNPPC of the Poorest 20% 0.0012 −0.0077 0.0891 0.8653* 0.3922* 0.3098* 0.0768 −0.2508* −0.2632* −0.4471* −0.2379* −0.2603* 0.1157 −0.1682 1 102 102 102 99 98 100 97 99 99 102 102 102 102 102 102 Women Borrowers −0.1605 −0.2277* −0.2092* −0.3567* −0.1124 −0.3404* −0.2563* 0.3760* 0.3259* 0.2901* 0.2464* 0.3092* −0.0091* 0.2873* −0.3778* 113 114 114 108 110 112 109 111 111 113 114 114 114 114 94 Open in new tab Table 3 MFI Lending Style by Region . Individual . Solidarity . Village Bank . Total . East Asia and Pacific 7 4 0 11 Eastern Europe and Central Asia 8 11 2 21 Latin America 32 10 8 50 Middle East and North Africa 3 3 3 9 South Asia 1 9 2 12 Sub‐Saharan Africa 5 11 5 21 Total 56 48 20 124 . Individual . Solidarity . Village Bank . Total . East Asia and Pacific 7 4 0 11 Eastern Europe and Central Asia 8 11 2 21 Latin America 32 10 8 50 Middle East and North Africa 3 3 3 9 South Asia 1 9 2 12 Sub‐Saharan Africa 5 11 5 21 Total 56 48 20 124 Source. Authors’ calculations, based on data from the Microfinance Information eXchange, Inc, 37. Open in new tab Table 3 MFI Lending Style by Region . Individual . Solidarity . Village Bank . Total . East Asia and Pacific 7 4 0 11 Eastern Europe and Central Asia 8 11 2 21 Latin America 32 10 8 50 Middle East and North Africa 3 3 3 9 South Asia 1 9 2 12 Sub‐Saharan Africa 5 11 5 21 Total 56 48 20 124 . Individual . Solidarity . Village Bank . Total . East Asia and Pacific 7 4 0 11 Eastern Europe and Central Asia 8 11 2 21 Latin America 32 10 8 50 Middle East and North Africa 3 3 3 9 South Asia 1 9 2 12 Sub‐Saharan Africa 5 11 5 21 Total 56 48 20 124 Source. Authors’ calculations, based on data from the Microfinance Information eXchange, Inc, 37. Open in new tab 2. The Microfinance Landscape Figure 1 shows basic patterns across the three main institutional types identified in the survey: Fig. 1. Open in new tabDownload slide Profitability, Portfolio Yield, and Expenses by Lending Type Source. Calculated from data in MicroBanking Bulletin, July 2003 (n = 124) Fig. 1. Open in new tabDownload slide Profitability, Portfolio Yield, and Expenses by Lending Type Source. Calculated from data in MicroBanking Bulletin, July 2003 (n = 124) ‘Individual‐based lenders’: institutions that use standard bilateral lending contracts between a lender and a single borrower. Liability for repaying the loan rests with the individual borrower only, although in some cases another individual might serve as a guarantor; ‘Solidarity group lenders’: institutions that employ contracts based on joint liability implemented with ‘solidarity groups’ (in the spirit of contracts used initially at the Grameen Bank in Bangladesh and at BancoSol in Bolivia). Loans are made to individuals but the group, which has between 3 and 10 members depending on the institution and location, shoulders responsibility for a loan if a member cannot repay, and ‘Village banks’, where each branch forms a single, large group and is given a degree of self‐governance (this kind of arrangement was pioneered by FINCA and is now employed by organisations like Pro Mujer and Freedom from Hunger). Figure 1 shows that patterns of average revenues and costs vary systematically by lending type. The village banks in the survey charge the highest average interest rates and face the highest average costs. The measure of interest rates we use, the real gross portfolio yield, captures both direct interest charges and any additional fees charged by lenders. The total expense ratio gives the ratio of total expenses (including labour and capital costs) to assets. Costs outweigh interest revenues, though, and the result is that the average return on assets for village banks is negative (−0.08). The microbanks using solidarity groups charge lower interest rates and face lower costs but again costs exceed revenues and the average return on assets is −0.05. Only for the individual‐based lenders in the survey is the average return on assets positive, though small (0.01). These patterns reflect differences in social mission, target customers and location as much as management strategies. The summary statistics suggest, for example, that one reason that costs are so much higher for village banks and group lenders (relative to individual‐based lenders) is that they make smaller‐sized loans and serve poorer populations. The data in Table 4 show that village banks, the least profitable lending type as a class, serve the poorest customers (as proxied by loan size) and their clients are more likely to be women. The customers of village banks and group lenders, for example, are largely women: 88% and 75%, respectively. In comparison, just under half of the customers of individual‐based lenders are women (46%). Table 4 Summary Statistics by Lending Type . Individual lenders . Solidarity lenders . Village bank lenders . Mean . Stndrd. Dev. . Mean . Stndrd. Dev. . Mean . Stndrd. Dev. . Financial self‐suffiency 1.11 0.29 0.98 0.32 0.95 0.47 Operational self‐suffiency 1.23 0.28 1.12 0.35 1.09 0.75 Return on assets adjusted 0.01 0.08 −0.05 0.24 −0.08 0.22 Average loan size to GNP per capita 1.01 1.10 0.54 0.52 0.20 0.17 Age 11.12 8.67 8.60 5.85 6.95 3.71 Size of MFI indicator 2.23 0.67 2.00 0.72 1.60 0.60 For‐profit status 0.29 0.46 0.26 0.44 0.00 0.00 Real gross portfolio yield 0.31 0.16 0.33 0.14 0.54 0.31 Capital costs to Assets 0.07 0.06 0.08 0.07 0.17 0.16 Labor costs to Assets 0.07 0.05 0.12 0.12 0.15 0.11 Loans to Assets 0.70 0.17 0.71 0.20 0.60 0.17 Donations to Loan portfolio 0.02 0.06 0.17 0.43 0.30 0.47 Average loan size to GNP per capita of the poorest 20% 4.80 4.92 1.63 1.97 0.63 0.39 Average loan size (USD) 1220.23 1184.51 430.98 499.56 148.69 126.61 Women borrowers 0.46 0.16 0.75 0.24 0.88 0.21 . Individual lenders . Solidarity lenders . Village bank lenders . Mean . Stndrd. Dev. . Mean . Stndrd. Dev. . Mean . Stndrd. Dev. . Financial self‐suffiency 1.11 0.29 0.98 0.32 0.95 0.47 Operational self‐suffiency 1.23 0.28 1.12 0.35 1.09 0.75 Return on assets adjusted 0.01 0.08 −0.05 0.24 −0.08 0.22 Average loan size to GNP per capita 1.01 1.10 0.54 0.52 0.20 0.17 Age 11.12 8.67 8.60 5.85 6.95 3.71 Size of MFI indicator 2.23 0.67 2.00 0.72 1.60 0.60 For‐profit status 0.29 0.46 0.26 0.44 0.00 0.00 Real gross portfolio yield 0.31 0.16 0.33 0.14 0.54 0.31 Capital costs to Assets 0.07 0.06 0.08 0.07 0.17 0.16 Labor costs to Assets 0.07 0.05 0.12 0.12 0.15 0.11 Loans to Assets 0.70 0.17 0.71 0.20 0.60 0.17 Donations to Loan portfolio 0.02 0.06 0.17 0.43 0.30 0.47 Average loan size to GNP per capita of the poorest 20% 4.80 4.92 1.63 1.97 0.63 0.39 Average loan size (USD) 1220.23 1184.51 430.98 499.56 148.69 126.61 Women borrowers 0.46 0.16 0.75 0.24 0.88 0.21 Source. Authors’ calculations, based on data from the Microfinance Information eXchange, Inc. Open in new tab Table 4 Summary Statistics by Lending Type . Individual lenders . Solidarity lenders . Village bank lenders . Mean . Stndrd. Dev. . Mean . Stndrd. Dev. . Mean . Stndrd. Dev. . Financial self‐suffiency 1.11 0.29 0.98 0.32 0.95 0.47 Operational self‐suffiency 1.23 0.28 1.12 0.35 1.09 0.75 Return on assets adjusted 0.01 0.08 −0.05 0.24 −0.08 0.22 Average loan size to GNP per capita 1.01 1.10 0.54 0.52 0.20 0.17 Age 11.12 8.67 8.60 5.85 6.95 3.71 Size of MFI indicator 2.23 0.67 2.00 0.72 1.60 0.60 For‐profit status 0.29 0.46 0.26 0.44 0.00 0.00 Real gross portfolio yield 0.31 0.16 0.33 0.14 0.54 0.31 Capital costs to Assets 0.07 0.06 0.08 0.07 0.17 0.16 Labor costs to Assets 0.07 0.05 0.12 0.12 0.15 0.11 Loans to Assets 0.70 0.17 0.71 0.20 0.60 0.17 Donations to Loan portfolio 0.02 0.06 0.17 0.43 0.30 0.47 Average loan size to GNP per capita of the poorest 20% 4.80 4.92 1.63 1.97 0.63 0.39 Average loan size (USD) 1220.23 1184.51 430.98 499.56 148.69 126.61 Women borrowers 0.46 0.16 0.75 0.24 0.88 0.21 . Individual lenders . Solidarity lenders . Village bank lenders . Mean . Stndrd. Dev. . Mean . Stndrd. Dev. . Mean . Stndrd. Dev. . Financial self‐suffiency 1.11 0.29 0.98 0.32 0.95 0.47 Operational self‐suffiency 1.23 0.28 1.12 0.35 1.09 0.75 Return on assets adjusted 0.01 0.08 −0.05 0.24 −0.08 0.22 Average loan size to GNP per capita 1.01 1.10 0.54 0.52 0.20 0.17 Age 11.12 8.67 8.60 5.85 6.95 3.71 Size of MFI indicator 2.23 0.67 2.00 0.72 1.60 0.60 For‐profit status 0.29 0.46 0.26 0.44 0.00 0.00 Real gross portfolio yield 0.31 0.16 0.33 0.14 0.54 0.31 Capital costs to Assets 0.07 0.06 0.08 0.07 0.17 0.16 Labor costs to Assets 0.07 0.05 0.12 0.12 0.15 0.11 Loans to Assets 0.70 0.17 0.71 0.20 0.60 0.17 Donations to Loan portfolio 0.02 0.06 0.17 0.43 0.30 0.47 Average loan size to GNP per capita of the poorest 20% 4.80 4.92 1.63 1.97 0.63 0.39 Average loan size (USD) 1220.23 1184.51 430.98 499.56 148.69 126.61 Women borrowers 0.46 0.16 0.75 0.24 0.88 0.21 Source. Authors’ calculations, based on data from the Microfinance Information eXchange, Inc. Open in new tab The village banks in the survey also make the smallest‐sized loans ($149 on average), followed by group lenders ($431). Individual‐based lenders made far larger average loans on average ($1,220). Average loan size is often taken to be a proxy for the poverty of customers, and these results are in line with anecdotal evidence about the depth of outreach across lending types. The loan size comparisons are made at official exchange rates, though, which can substantially distort the purchasing power of a given amount of money in local currency. Patterns are broadly similar, however, even if the average loan sizes are deflated by gross national product per capita (a metric often preferred by microfinance donors) or deflated by the average income per capita of the bottom 20% in the country. For average loan size to GNP per capita, the ratio for village banks: solidarity group lenders: individual‐based group lenders is 0.20: 0.54: 1.01. Where the deflator is the average income per capita of the bottom 20% in the country, the ratios are 0.63: 1.63: 4.80. These basic distinctions by lending type play out in important ways in the regression analyses below. If predicted revenues fall short of costs, lenders are likely to lean on subsidies. It is not surprising that village banks as a class take most advantage of subsidies. Table 5 shows that the average fraction of subsidies in funding (total liabilities plus total equity) is over one‐third. For solidarity group‐based institutions, the fraction is 28%, and for individual‐based lenders the subsidized share of funding is just 11%. Table 5 Subsidised Share of Funding . Mean . Standard deviation . Sample average 21.4% 29.3% By Lending type  Individual‐based (n = 56) 11.0 17.9  Solidarity group (n = 48) 27.7 37.3  Village banks (n = 20) 35.5 23.6 By Profit status  For‐profit (n = 28) 6.6 14.9  Not‐for‐profit (n = 90) 26.2 31.6 . Mean . Standard deviation . Sample average 21.4% 29.3% By Lending type  Individual‐based (n = 56) 11.0 17.9  Solidarity group (n = 48) 27.7 37.3  Village banks (n = 20) 35.5 23.6 By Profit status  For‐profit (n = 28) 6.6 14.9  Not‐for‐profit (n = 90) 26.2 31.6 Notes. Subsidised share of funding is equal to (subsidised costs of funds adjustment + in‐kind subsidy adjustment + donated equity)/(total liabilities + total equity). ‘Profit status’ refers to the institution’s official designation and is independent of actual profitability. Source. Authors’ calculations, based on data from the Microfinance Information eXchange, Inc. Open in new tab Table 5 Subsidised Share of Funding . Mean . Standard deviation . Sample average 21.4% 29.3% By Lending type  Individual‐based (n = 56) 11.0 17.9  Solidarity group (n = 48) 27.7 37.3  Village banks (n = 20) 35.5 23.6 By Profit status  For‐profit (n = 28) 6.6 14.9  Not‐for‐profit (n = 90) 26.2 31.6 . Mean . Standard deviation . Sample average 21.4% 29.3% By Lending type  Individual‐based (n = 56) 11.0 17.9  Solidarity group (n = 48) 27.7 37.3  Village banks (n = 20) 35.5 23.6 By Profit status  For‐profit (n = 28) 6.6 14.9  Not‐for‐profit (n = 90) 26.2 31.6 Notes. Subsidised share of funding is equal to (subsidised costs of funds adjustment + in‐kind subsidy adjustment + donated equity)/(total liabilities + total equity). ‘Profit status’ refers to the institution’s official designation and is independent of actual profitability. Source. Authors’ calculations, based on data from the Microfinance Information eXchange, Inc. Open in new tab Table 6 shows the correlation of subsidies and costs. The only statistically significant correlations are the positive coefficients with respect to the costs faced by institutions using solidarity groups (0.75 for capital costs, 0.66 for labour costs). The same correlation for village banks and individual‐based lenders are small and weak, suggesting a diversity of rationales for subsidisation. The most striking result holds with respect to portfolio at risk, and it is a ‘non‐result’: in contrast to expert belief that increased subsidisation weakens incentives to maintain high portfolio quality, no such evidence emerges in these correlations. Table 6 Correlations of Subsidised Share of Funding . Correlations with: . Capital costs/Assets . Labour costs/Assets . Port‐folio at risk . Sample average 0.26* 0.59 −0.08 By lending type: Individual‐based 0.08 0.20 −0.21 Solidarity Group 0.75* 0.66* −0.02 Village Bank −0.34 −0.01 0.03 . Correlations with: . Capital costs/Assets . Labour costs/Assets . Port‐folio at risk . Sample average 0.26* 0.59 −0.08 By lending type: Individual‐based 0.08 0.20 −0.21 Solidarity Group 0.75* 0.66* −0.02 Village Bank −0.34 −0.01 0.03 Notes. Subsidised Share of Funding is equal to (subsidised costs of funds adjustment + in‐kind subsidy adjustment + donated equity)/(total liabilities + total equity). FSS is financial self‐sufficiency, OSS is operational self‐sufficiency, and portfolio at risk is the share of loans delinquent at least thirty days. There is no variation in profit status for village banks, and thus no correlation can be calculated for that variable and our subsidy measures for that group. Source. Authors’ calculations, based on data from the Microfinance Information eXchange, Inc. Open in new tab Table 6 Correlations of Subsidised Share of Funding . Correlations with: . Capital costs/Assets . Labour costs/Assets . Port‐folio at risk . Sample average 0.26* 0.59 −0.08 By lending type: Individual‐based 0.08 0.20 −0.21 Solidarity Group 0.75* 0.66* −0.02 Village Bank −0.34 −0.01 0.03 . Correlations with: . Capital costs/Assets . Labour costs/Assets . Port‐folio at risk . Sample average 0.26* 0.59 −0.08 By lending type: Individual‐based 0.08 0.20 −0.21 Solidarity Group 0.75* 0.66* −0.02 Village Bank −0.34 −0.01 0.03 Notes. Subsidised Share of Funding is equal to (subsidised costs of funds adjustment + in‐kind subsidy adjustment + donated equity)/(total liabilities + total equity). FSS is financial self‐sufficiency, OSS is operational self‐sufficiency, and portfolio at risk is the share of loans delinquent at least thirty days. There is no variation in profit status for village banks, and thus no correlation can be calculated for that variable and our subsidy measures for that group. Source. Authors’ calculations, based on data from the Microfinance Information eXchange, Inc. Open in new tab Table 7 Profitability Regressions . Financial Self‐Sufficiency . Operational Self‐Sufficiency . Return on Assets . (1) (2) (3) Real Yield 0.964 0.678 0.374 [5.04]*** [3.24]*** [4.28]*** Real Yield (Village bank) −0.435 −0.514 0.013 [0.97] [0.77] [0.08] Real Yield (Solidarity) −0.011 −0.025 0.405 [0.03] [0.05] [3.14]*** Capital Costs to Assets −1.882 −0.924 −0.727 [5.80]*** [1.88]* [3.90]*** Capital Costs to Assets (Village bank) 1.053 −0.064 −0.020 [1.41] [0.06] [0.06] Capital Costs to Assets (solidarity) −1.877 −2.762 −0.713 [1.42] [1.50] [1.57] Labour Costs to Assets 0.840 0.097 −0.099 [0.91] [0.06] [0.38] Labour Costs to Assets (Village bank) −0.221 −1.649 −0.528 [0.18] [0.94] [0.87] Labour Costs to Assets (Solidarity) 1.648 0.642 −1.144 [1.44] [0.36] [3.42]*** Village bank 0.119 0.452 0.038 [0.31] [0.70] [0.48] Solidarity −0.105 0.028 0.024 [0.58] [0.12] [0.57] Size Indicator 0.158 0.202 0.037 [3.30]*** [2.80]*** [2.89]*** Log of age 0.116 0.148 0.058 [1.94]* [1.52] [3.29]*** Average Loan Size to GNP per capita −0.016 −0.008 0.004 [0.57] [0.19] [0.50] Loans to assets ratio 0.539 0.135 0.224 [2.05]** [0.30] [3.81]*** For‐profit dummy −0.018 −0.101 −0.004 [0.35] [1.27] [0.22] Eastern Europe and Central Asia 0.138 0.338 0.052 [1.33] [2.12]** [1.77]* Sub‐Saharan Africa 0.151 0.243 0.053 [1.66] [2.20]** [2.12]** Middle East and N. Africa 0.110 0.121 0.063 [1.26] [0.94] [2.37]** South Asia 0.212 0.363 −0.024 [1.34] [1.39] [0.64] East Asia −0.083 −0.036 −0.023 [1.34] [0.49] [0.91] Constant −0.031 0.129 −0.440 [0.17] [0.53] [5.51]*** Observations 104 104 104 R‐squared 0.55 0.41 0.87 . Financial Self‐Sufficiency . Operational Self‐Sufficiency . Return on Assets . (1) (2) (3) Real Yield 0.964 0.678 0.374 [5.04]*** [3.24]*** [4.28]*** Real Yield (Village bank) −0.435 −0.514 0.013 [0.97] [0.77] [0.08] Real Yield (Solidarity) −0.011 −0.025 0.405 [0.03] [0.05] [3.14]*** Capital Costs to Assets −1.882 −0.924 −0.727 [5.80]*** [1.88]* [3.90]*** Capital Costs to Assets (Village bank) 1.053 −0.064 −0.020 [1.41] [0.06] [0.06] Capital Costs to Assets (solidarity) −1.877 −2.762 −0.713 [1.42] [1.50] [1.57] Labour Costs to Assets 0.840 0.097 −0.099 [0.91] [0.06] [0.38] Labour Costs to Assets (Village bank) −0.221 −1.649 −0.528 [0.18] [0.94] [0.87] Labour Costs to Assets (Solidarity) 1.648 0.642 −1.144 [1.44] [0.36] [3.42]*** Village bank 0.119 0.452 0.038 [0.31] [0.70] [0.48] Solidarity −0.105 0.028 0.024 [0.58] [0.12] [0.57] Size Indicator 0.158 0.202 0.037 [3.30]*** [2.80]*** [2.89]*** Log of age 0.116 0.148 0.058 [1.94]* [1.52] [3.29]*** Average Loan Size to GNP per capita −0.016 −0.008 0.004 [0.57] [0.19] [0.50] Loans to assets ratio 0.539 0.135 0.224 [2.05]** [0.30] [3.81]*** For‐profit dummy −0.018 −0.101 −0.004 [0.35] [1.27] [0.22] Eastern Europe and Central Asia 0.138 0.338 0.052 [1.33] [2.12]** [1.77]* Sub‐Saharan Africa 0.151 0.243 0.053 [1.66] [2.20]** [2.12]** Middle East and N. Africa 0.110 0.121 0.063 [1.26] [0.94] [2.37]** South Asia 0.212 0.363 −0.024 [1.34] [1.39] [0.64] East Asia −0.083 −0.036 −0.023 [1.34] [0.49] [0.91] Constant −0.031 0.129 −0.440 [0.17] [0.53] [5.51]*** Observations 104 104 104 R‐squared 0.55 0.41 0.87 *significant at 10%; ** significant at 5%; *** significant at 1%. Source: Authors’ calculations, based on data from the Microfinance Information eXchange, Inc. Open in new tab Table 7 Profitability Regressions . Financial Self‐Sufficiency . Operational Self‐Sufficiency . Return on Assets . (1) (2) (3) Real Yield 0.964 0.678 0.374 [5.04]*** [3.24]*** [4.28]*** Real Yield (Village bank) −0.435 −0.514 0.013 [0.97] [0.77] [0.08] Real Yield (Solidarity) −0.011 −0.025 0.405 [0.03] [0.05] [3.14]*** Capital Costs to Assets −1.882 −0.924 −0.727 [5.80]*** [1.88]* [3.90]*** Capital Costs to Assets (Village bank) 1.053 −0.064 −0.020 [1.41] [0.06] [0.06] Capital Costs to Assets (solidarity) −1.877 −2.762 −0.713 [1.42] [1.50] [1.57] Labour Costs to Assets 0.840 0.097 −0.099 [0.91] [0.06] [0.38] Labour Costs to Assets (Village bank) −0.221 −1.649 −0.528 [0.18] [0.94] [0.87] Labour Costs to Assets (Solidarity) 1.648 0.642 −1.144 [1.44] [0.36] [3.42]*** Village bank 0.119 0.452 0.038 [0.31] [0.70] [0.48] Solidarity −0.105 0.028 0.024 [0.58] [0.12] [0.57] Size Indicator 0.158 0.202 0.037 [3.30]*** [2.80]*** [2.89]*** Log of age 0.116 0.148 0.058 [1.94]* [1.52] [3.29]*** Average Loan Size to GNP per capita −0.016 −0.008 0.004 [0.57] [0.19] [0.50] Loans to assets ratio 0.539 0.135 0.224 [2.05]** [0.30] [3.81]*** For‐profit dummy −0.018 −0.101 −0.004 [0.35] [1.27] [0.22] Eastern Europe and Central Asia 0.138 0.338 0.052 [1.33] [2.12]** [1.77]* Sub‐Saharan Africa 0.151 0.243 0.053 [1.66] [2.20]** [2.12]** Middle East and N. Africa 0.110 0.121 0.063 [1.26] [0.94] [2.37]** South Asia 0.212 0.363 −0.024 [1.34] [1.39] [0.64] East Asia −0.083 −0.036 −0.023 [1.34] [0.49] [0.91] Constant −0.031 0.129 −0.440 [0.17] [0.53] [5.51]*** Observations 104 104 104 R‐squared 0.55 0.41 0.87 . Financial Self‐Sufficiency . Operational Self‐Sufficiency . Return on Assets . (1) (2) (3) Real Yield 0.964 0.678 0.374 [5.04]*** [3.24]*** [4.28]*** Real Yield (Village bank) −0.435 −0.514 0.013 [0.97] [0.77] [0.08] Real Yield (Solidarity) −0.011 −0.025 0.405 [0.03] [0.05] [3.14]*** Capital Costs to Assets −1.882 −0.924 −0.727 [5.80]*** [1.88]* [3.90]*** Capital Costs to Assets (Village bank) 1.053 −0.064 −0.020 [1.41] [0.06] [0.06] Capital Costs to Assets (solidarity) −1.877 −2.762 −0.713 [1.42] [1.50] [1.57] Labour Costs to Assets 0.840 0.097 −0.099 [0.91] [0.06] [0.38] Labour Costs to Assets (Village bank) −0.221 −1.649 −0.528 [0.18] [0.94] [0.87] Labour Costs to Assets (Solidarity) 1.648 0.642 −1.144 [1.44] [0.36] [3.42]*** Village bank 0.119 0.452 0.038 [0.31] [0.70] [0.48] Solidarity −0.105 0.028 0.024 [0.58] [0.12] [0.57] Size Indicator 0.158 0.202 0.037 [3.30]*** [2.80]*** [2.89]*** Log of age 0.116 0.148 0.058 [1.94]* [1.52] [3.29]*** Average Loan Size to GNP per capita −0.016 −0.008 0.004 [0.57] [0.19] [0.50] Loans to assets ratio 0.539 0.135 0.224 [2.05]** [0.30] [3.81]*** For‐profit dummy −0.018 −0.101 −0.004 [0.35] [1.27] [0.22] Eastern Europe and Central Asia 0.138 0.338 0.052 [1.33] [2.12]** [1.77]* Sub‐Saharan Africa 0.151 0.243 0.053 [1.66] [2.20]** [2.12]** Middle East and N. Africa 0.110 0.121 0.063 [1.26] [0.94] [2.37]** South Asia 0.212 0.363 −0.024 [1.34] [1.39] [0.64] East Asia −0.083 −0.036 −0.023 [1.34] [0.49] [0.91] Constant −0.031 0.129 −0.440 [0.17] [0.53] [5.51]*** Observations 104 104 104 R‐squared 0.55 0.41 0.87 *significant at 10%; ** significant at 5%; *** significant at 1%. Source: Authors’ calculations, based on data from the Microfinance Information eXchange, Inc. Open in new tab 3. Regression Approach The aim of the benchmark regressions is to understand why some microbanks are more profitable than others. The base regressions describe the correlates of profitability, focusing particularly on the roles of costs and interest rates charged on loans. We allow these factors to vary by lending type using the following reduced‐form equation: (1) where FSS is the financial self‐sufficiency ratio of microfinance institution i. As noted, we also use OSS and ROA as dependent variables. The construction of these measures and their summary statistics appear in Table 1. The means and medians for FSS, OSS, and ROA are all within the expected range but the minimum and maximum values suggest a wide range for each variable, prompting the use of robust regression methods as a check on robustness to outliers. Yield is the real gross portfolio yield, a measure of interest charges faced by customers described in Table 1. Because loan losses are not netted out of the revenues, this measure is intended to capture the ex ante interest rate charged by the lender rather than the ex post interest rate realised on the portfolio. The coefficient matrix β2 includes coefficients that show how the effects of Yield vary by lending type, described in greater detail below. In the results that follow, the omitted category is ‘individual‐based lenders’. Thus, there is one Yield coefficient for solidarity group lenders and another for village banks. Each of those coefficients measures the difference between that lending type and individual‐based lenders with regard to the effect of yields. The coefficient vector β1thus summarises the effect of yields on financial self‐sufficiency for individual‐based lenders. The coefficient matrix β4 shows how the effects of labour costs vary across lending types, while β6 does the same for capital costs. The coefficient vectors β3 and β5 therefore summarise the effects of labour costs and capital costs on financial self‐sufficiency for individual‐based lenders. The lending type variables also enter the specification independently. Because they again are the omitted category, individual‐based lenders do not have their own coefficient.7 The matrix MFI history includes two variables, one for age and the other for size (as measured by total assets). The matrix Orientation contains three variables that describe the microfinance institution’s business practices: the ratio of loans to assets, the average loan size (relative to GNP per capita), and a dummy variable indicating the institution’s formal profit status (equal to one if the organisation is for‐profit). Finally, region is a matrix of dummy variables for each main region of the developing world, with ‘Latin America and the Caribbean’ as the omitted category. Having summarised the correlates of profitability, the next set of regressions explores the relationship between interest rates and profitability for each lending type. Here, the interest is in evidence of declining profitability as interest rates rise to high levels. We first introduce a quadratic term for the gross portfolio yield variable in the profitability equations, allowing the quadratic effect to differ across lending types. The quadratic form can generate U‐shaped patterns consistent with the prediction that agency problems become so severe that overall profitability eventually falls as interest rates rise. This result is also consistent with falling demand for credit (and thus diminishing scale economies) at high interest rates. To shed further light on the specific hypothesis from agency theory, we then replace the profit measures with the share of the portfolio that is delinquent (portfolio at risk) to test directly whether high interest rates are associated with higher rates of non‐repayment – and find some evidence that they are, but only for individual‐based lenders. Moreover, according to one specification, individual‐based lenders charging the highest interest rates in our sample enjoy better repayment performance than those charging intermediate rates. Yet, their lending volumes are substantially lower, a finding that is more consistent with falling demand for credit as rates push past threshold values than with predictions from agency theory. 4. Results 4.1. Financial Sustainability Table 7 gives the results from the estimation of (1) above. The results show that raising interest rates is associated with improved financial performance for individual‐based lenders. The coefficient for real gross portfolio yield (the measure of average interest rates on loans to customers) is positive and significant across all three profitability indicators (financial self‐sufficiency, operational sustainability, and return on assets), indicating that individual‐based lenders tend to be more profitable when their average interest rates are higher. The result, in itself, is not surprising. In addition, we cannot reject the hypothesis that the effects are similar for village banks and group lenders, since their coefficients are not statistically significant in either column (1) or (2). However, when we sum the coefficients for yield and the yield interactions, we also cannot reject the hypothesis that the effect is zero for village banks. In that sense, for that type of group lender there is not a pronounced significant relationship between interest rates and profitability, even after controlling for costs. Capital costs are asssociated with reduced profitability for individual‐based lenders for all three of our profitability measures. The capital costs coefficients for solidarity group lenders are also negative (though insignificant) indicating that containing those costs is a key to their profitability.8 The capital costs coefficient for village banks is positive in the FSS specification and its sum with the simple capital costs coefficient is not statistically distinguishable from zero. However, for OSS and return on assets, the coefficient for village banks is small and insignificant. Thus capital costs are also negatively associated with profitability for village banks for two of our measures. Labour costs tend not to be significant for any of the profitability measures in Table 7. This could be because they are so highly correlated with capital costs(0.65 in Table 2). The exception to this pattern is the negative, significant coefficient for labour costs for solidarity group lenders in the reurn on assets specification. In all, cost containment (especially capital costs) plays an important role in determining the profitability of all three lending types. Note that neither the village bank nor the solidarity group dummy variable is significant in Table 7, indicating that once the effects of costs and yields are permitted to vary by lending type, those types explain no additional variation in financial performance. The regional dummy variables do explain some variation in financial performance. Institutions from Eastern Europe and Central Asia and those from Sub‐Saharan Africa out‐performed those from other regions in terms of operational self‐sufficiency and return on assets. An institution’s age and size are significantly positively linked to financial performance across indicators.9 Finally, neither the indicator for being constituted formally as a for‐profit bank nor the average loan size variable is significantly linked to the financial performance indicators. The latter result shows that, even after controlling for region and other covariates, institutions that make smaller loans are not less profitable on average. The basic pattern of results also holds when we control for regional variation in different ways. For example, in unreported specifications we allowed for correlation between observations from the same country using clustered standard errors. In another set of unreported specifications, we allowed for random effects at the country level. Given the small size of our dataset, we were not able to incorporate country fixed effects in our models. 4.2. Interest Rates We next extend the results to examine the implications of agency theory. Specifically, when lenders face informational asymmetry and borrowers lack collateral, charging interest above a certain threshold could aggravate problems of adverse selection and moral hazard. At high enough rates, only low‐quality borrowers that do not expect to be able to repay would find it in their interest to borrow. If the conjecture is true, microbanks charging relatively high interest rates should expect to face lower repayment rates and profitability. The relationship with regard to profitability (but not portfolio quality) could also arise from demand forces: overly high prices may reduce demand and hence profits. We begin by establishing the basic patterns in the data by including the square of portfolio yield in our base specifications. As in previous specifications, we allow the association between the squared yield variable and financial performance to vary by lending type. We have a relatively small dataset, so introducing the squared yield terms makes it difficult to separate labour and capital costs variables for each lending type. Therefore, those costs variables enter the specifications in Table 8 without lending type interactions. When the costs variables are collapsed in this way, the overall fit of these regressions is almost identical to that for the base profitability regressions in Table 7. Table 8 Profitability Regressions – Allowing Nonlinear Effects of Interest Rates . Financial Self‐Sufficiency . Operational Self‐Sufficiency . Return on assets . (1) (2) (3) Real Yield 1.562 1.695 0.964 [2.01]** [1.82]* [3.00]*** Real Yield squared −0.809 −1.137 −0.640 [0.94] [1.11] [1.75]* Real Yield (Village bank) −0.649 −2.461 −0.334 [0.33] [0.71] [0.70] Real Yield (Village bank) squared 0.396 1.919 0.531 [0.22] [0.65] [1.12] Real Yield (Solidarity) −3.503 −3.639 −1.164 [2.07]** [1.81]* [2.27]** Real Yield (Solidarity) squared 4.451 4.402 1.725 [2.19]** [1.89]* [2.80]*** Capital Costs to Assets −1.385 −1.266 −0.705 [2.97]*** [2.24]** [3.00]*** Labour Costs to Assets −0.253 −0.464 −1.271 [0.66] [1.01] [3.89]*** Village bank 0.199 0.653 0.043 [0.38] [0.68] [0.37] Solidarity 0.493 0.539 0.160 [1.52] [1.40] [1.78]* Size Indicator 0.153 0.181 0.011 [3.17]*** [2.61]** [0.70] Log of age 0.152 0.149 0.065 [2.84]*** [1.89]* [2.77]*** Average Loan Size to GNP per capita −0.009 −0.004 0.004 [0.29] [0.10] [0.56] Loans to assets ratio 0.453 0.091 0.289 [2.03]** [0.23] [4.72]*** For‐profit dummy −0.057 −0.133 −0.014 [0.75] [1.58] [0.68] Eastern Europe and Central Asia 0.115 0.315 0.063 [1.21] [2.22]** [1.69]* Sub‐Saharan Africa 0.124 0.186 0.075 [1.35] [1.72]* [2.46]** Middle East and N. Africa 0.103 0.089 0.074 [1.32] [0.78] [2.08]** South Asia 0.248 0.356 0.013 [1.77]* [1.56] [0.34] East Asia −0.062 −0.030 0.001 [0.96] [0.38] [0.02] Constant −0.195 0.104 −0.468 [0.94] [0.43] [4.55]*** Observations 104 104 104 R‐squared 0.54 0.41 0.83 . Financial Self‐Sufficiency . Operational Self‐Sufficiency . Return on assets . (1) (2) (3) Real Yield 1.562 1.695 0.964 [2.01]** [1.82]* [3.00]*** Real Yield squared −0.809 −1.137 −0.640 [0.94] [1.11] [1.75]* Real Yield (Village bank) −0.649 −2.461 −0.334 [0.33] [0.71] [0.70] Real Yield (Village bank) squared 0.396 1.919 0.531 [0.22] [0.65] [1.12] Real Yield (Solidarity) −3.503 −3.639 −1.164 [2.07]** [1.81]* [2.27]** Real Yield (Solidarity) squared 4.451 4.402 1.725 [2.19]** [1.89]* [2.80]*** Capital Costs to Assets −1.385 −1.266 −0.705 [2.97]*** [2.24]** [3.00]*** Labour Costs to Assets −0.253 −0.464 −1.271 [0.66] [1.01] [3.89]*** Village bank 0.199 0.653 0.043 [0.38] [0.68] [0.37] Solidarity 0.493 0.539 0.160 [1.52] [1.40] [1.78]* Size Indicator 0.153 0.181 0.011 [3.17]*** [2.61]** [0.70] Log of age 0.152 0.149 0.065 [2.84]*** [1.89]* [2.77]*** Average Loan Size to GNP per capita −0.009 −0.004 0.004 [0.29] [0.10] [0.56] Loans to assets ratio 0.453 0.091 0.289 [2.03]** [0.23] [4.72]*** For‐profit dummy −0.057 −0.133 −0.014 [0.75] [1.58] [0.68] Eastern Europe and Central Asia 0.115 0.315 0.063 [1.21] [2.22]** [1.69]* Sub‐Saharan Africa 0.124 0.186 0.075 [1.35] [1.72]* [2.46]** Middle East and N. Africa 0.103 0.089 0.074 [1.32] [0.78] [2.08]** South Asia 0.248 0.356 0.013 [1.77]* [1.56] [0.34] East Asia −0.062 −0.030 0.001 [0.96] [0.38] [0.02] Constant −0.195 0.104 −0.468 [0.94] [0.43] [4.55]*** Observations 104 104 104 R‐squared 0.54 0.41 0.83 All models estimated via OLS, with White’s Heteroskedasticity consistent standard errors. * significant at 10%; ** significant at 5%; *** significant at 1%. Source: Authors’ calculations, based on data from the Microfinance Information eXchange, Inc. Open in new tab Table 8 Profitability Regressions – Allowing Nonlinear Effects of Interest Rates . Financial Self‐Sufficiency . Operational Self‐Sufficiency . Return on assets . (1) (2) (3) Real Yield 1.562 1.695 0.964 [2.01]** [1.82]* [3.00]*** Real Yield squared −0.809 −1.137 −0.640 [0.94] [1.11] [1.75]* Real Yield (Village bank) −0.649 −2.461 −0.334 [0.33] [0.71] [0.70] Real Yield (Village bank) squared 0.396 1.919 0.531 [0.22] [0.65] [1.12] Real Yield (Solidarity) −3.503 −3.639 −1.164 [2.07]** [1.81]* [2.27]** Real Yield (Solidarity) squared 4.451 4.402 1.725 [2.19]** [1.89]* [2.80]*** Capital Costs to Assets −1.385 −1.266 −0.705 [2.97]*** [2.24]** [3.00]*** Labour Costs to Assets −0.253 −0.464 −1.271 [0.66] [1.01] [3.89]*** Village bank 0.199 0.653 0.043 [0.38] [0.68] [0.37] Solidarity 0.493 0.539 0.160 [1.52] [1.40] [1.78]* Size Indicator 0.153 0.181 0.011 [3.17]*** [2.61]** [0.70] Log of age 0.152 0.149 0.065 [2.84]*** [1.89]* [2.77]*** Average Loan Size to GNP per capita −0.009 −0.004 0.004 [0.29] [0.10] [0.56] Loans to assets ratio 0.453 0.091 0.289 [2.03]** [0.23] [4.72]*** For‐profit dummy −0.057 −0.133 −0.014 [0.75] [1.58] [0.68] Eastern Europe and Central Asia 0.115 0.315 0.063 [1.21] [2.22]** [1.69]* Sub‐Saharan Africa 0.124 0.186 0.075 [1.35] [1.72]* [2.46]** Middle East and N. Africa 0.103 0.089 0.074 [1.32] [0.78] [2.08]** South Asia 0.248 0.356 0.013 [1.77]* [1.56] [0.34] East Asia −0.062 −0.030 0.001 [0.96] [0.38] [0.02] Constant −0.195 0.104 −0.468 [0.94] [0.43] [4.55]*** Observations 104 104 104 R‐squared 0.54 0.41 0.83 . Financial Self‐Sufficiency . Operational Self‐Sufficiency . Return on assets . (1) (2) (3) Real Yield 1.562 1.695 0.964 [2.01]** [1.82]* [3.00]*** Real Yield squared −0.809 −1.137 −0.640 [0.94] [1.11] [1.75]* Real Yield (Village bank) −0.649 −2.461 −0.334 [0.33] [0.71] [0.70] Real Yield (Village bank) squared 0.396 1.919 0.531 [0.22] [0.65] [1.12] Real Yield (Solidarity) −3.503 −3.639 −1.164 [2.07]** [1.81]* [2.27]** Real Yield (Solidarity) squared 4.451 4.402 1.725 [2.19]** [1.89]* [2.80]*** Capital Costs to Assets −1.385 −1.266 −0.705 [2.97]*** [2.24]** [3.00]*** Labour Costs to Assets −0.253 −0.464 −1.271 [0.66] [1.01] [3.89]*** Village bank 0.199 0.653 0.043 [0.38] [0.68] [0.37] Solidarity 0.493 0.539 0.160 [1.52] [1.40] [1.78]* Size Indicator 0.153 0.181 0.011 [3.17]*** [2.61]** [0.70] Log of age 0.152 0.149 0.065 [2.84]*** [1.89]* [2.77]*** Average Loan Size to GNP per capita −0.009 −0.004 0.004 [0.29] [0.10] [0.56] Loans to assets ratio 0.453 0.091 0.289 [2.03]** [0.23] [4.72]*** For‐profit dummy −0.057 −0.133 −0.014 [0.75] [1.58] [0.68] Eastern Europe and Central Asia 0.115 0.315 0.063 [1.21] [2.22]** [1.69]* Sub‐Saharan Africa 0.124 0.186 0.075 [1.35] [1.72]* [2.46]** Middle East and N. Africa 0.103 0.089 0.074 [1.32] [0.78] [2.08]** South Asia 0.248 0.356 0.013 [1.77]* [1.56] [0.34] East Asia −0.062 −0.030 0.001 [0.96] [0.38] [0.02] Constant −0.195 0.104 −0.468 [0.94] [0.43] [4.55]*** Observations 104 104 104 R‐squared 0.54 0.41 0.83 All models estimated via OLS, with White’s Heteroskedasticity consistent standard errors. * significant at 10%; ** significant at 5%; *** significant at 1%. Source: Authors’ calculations, based on data from the Microfinance Information eXchange, Inc. Open in new tab The results follow the theoretical predictions for individual‐based lenders. The main finding from Table 8 is that for individual‐based lenders, financial self‐sufficiency, operational sustainability and return on assets are increasing in portfolio yield but only up to the point at which the negative quadratic yield coefficient outweighs the positive linear coefficient.10 Both the linear and the squared yield variables are significant in the return on assets specification. Figure 2 plots the estimated relationship between the financial self‐sufficiency ratio and the yield ratio based on Table 8, column 1, for an individual‐based lender assigned the median value for all other variables that enter the regression. As hypothesised, financial self‐sufficiency is increasing in yield up to a point. Near 60% per annum, the financial self‐sufficiency curve flattens, though there are only a few observations beyond that break point (Figure 2). The paucity of rates above 60% is consistent with individual‐based lenders adjusting in order to avoid potential incentive problems (or to avoid a loss of demand), and thus opting not to push interest rates beyond threshold values. Fig. 2. Open in new tabDownload slide Financial Self‐Sufficiency and Portfolio Yield Source. Authors’ calculations, based on data from the Microfinance Information eXchange, Inc. Fig. 2. Open in new tabDownload slide Financial Self‐Sufficiency and Portfolio Yield Source. Authors’ calculations, based on data from the Microfinance Information eXchange, Inc. For village banks, coefficients on both the linear and squared yield variables are not statistically significant, but similar in magnitude to, and the opposite sign of, the simple yield variables (corresponding to effects for individual‐based lenders). Summing the respective squared and linear variables, the relationship between yields and our profitability indicators for village banks is not significantly different from zero. Though again, because the coefficients for the village bank interactions are insignificant, we also cannot reject the hypothesis that the yield relationships are equal to those for individual‐based lenders. For group lenders, the coefficients for yield and yield squared are also the opposite sign of those for individual‐based lenders but they are significant and much larger in magnitude. When the respective coefficients are summed, the yield coefficients are negative and nearly significant, while the coefficients for yield squared are positive and significant. The result is the U‐shaped curve in Figure 2, which gives the relationship between yields and financial self‐sufficiency for the median group lending institution. For portfolio yields under 40% per annum, which characterises the majority of solidarity group lenders in our sample, the relationship is negative. Had we not imposed the non‐linearity by including a separate squared yield term for group lenders, the simple linear relation between financial self‐sufficiency and yield would have been negative (though insignificant). Overall, the results in Table 8 suggest that individual‐based lenders that charge higher interest rates are more profitable than others but only up to a point. For most solidarity group lenders, an opposite pattern holds. The results for individual‐based lenders are consistent with agency problems or demand forces that reduce scale at high interest rates but, while the specifications control for costs and geographic variables, we note that the result could also be driven by omitted customer characteristics or reverse causation. Reducing interest rates (and thus lowering profits) might be especially likely when the institution is driven by social objectives or if it seeks to maximise profits but faces potential competition. With the existing data, the competing explanations cannot be distinguished. 4.3. Portfolio at Risk We push further to ask whether higher interest rates are also associated with rising loan delinquency. Throughout most of our sample range, loan delinquency is more common for individual‐based lenders that charge higher yields (as predicted by theory). Direct insight into agency problems comes from analysing determinants of loan delinquencies and their relation to interest rates. We start with a specification that does not control for average loan size, and find a statistically significant inverted‐U shaped pattern for individual‐based lenders (Table 9, column 1). The specification replaces the dependent variables in the Table 8 regressions with the share of the portfolio that is at risk, defined as the share of loans that are delinquent for at least thirty days. Summing the respective squared and linear yield variables, there is no significant relationship between yields and portfolio risk for group lenders or village banks. Table 9, column 2 shows that including average loan size reduces the yield coefficients such that they lose significance but the quadratic pattern remains similar. Table 9 Portfolio at Risk Regressions . Portfolio at Risk . (1) . (2) . Real Yield 0.262 0.219 [2.11]** [1.18] Real Yield squared −0.337 −0.294 [2.41]** [1.58] Real Yield (Villagebank) −0.155 −0.106 [0.91] [0.48] Real Yield (Villagebank) squared 0.205 0.156 [1.09] [0.69] Real Yield (Solidarity) −0.451 −0.431 [2.01]** [1.76]* Real Yield (Solidarity) squared 0.445 0.431 [1.71]* [1.58] Capital Costs to Assets 0.087 0.086 [1.39] [1.37] Labour Costs to Assets −0.058 −0.050 [1.02] [0.82] Village bank 0.019 0.011 [0.62] [0.27] Solidarity 0.098 0.096 [2.31]** [1.91]* Size Indicator −0.016 −0.017 [1.63] [1.62] Log of age 0.017 0.019 [1.87]* [1.92]* Average Loan Size to GNP per capita 0.001 [0.15] Loans to assets ratio −0.072 −0.072 [2.15]** [2.11]** For‐profit dummy 0.011 0.013 [1.03] [0.99] Eastern Europe and Central Asia −0.010 −0.012 [0.79] [0.87] Sub‐Saharan Africa −0.016 −0.017 [1.31] [1.29] Middle East and N. Africa −0.001 −0.0003 [0.07] [0.02] South Asia −0.031 −0.032 [2.14]** [2.05]** East Asia −0.012 −0.012 [0.54] [0.52] Constant 0.043 0.046 [1.19] [1.05] Observations 107 101 R‐squared 0.25 0.26 . Portfolio at Risk . (1) . (2) . Real Yield 0.262 0.219 [2.11]** [1.18] Real Yield squared −0.337 −0.294 [2.41]** [1.58] Real Yield (Villagebank) −0.155 −0.106 [0.91] [0.48] Real Yield (Villagebank) squared 0.205 0.156 [1.09] [0.69] Real Yield (Solidarity) −0.451 −0.431 [2.01]** [1.76]* Real Yield (Solidarity) squared 0.445 0.431 [1.71]* [1.58] Capital Costs to Assets 0.087 0.086 [1.39] [1.37] Labour Costs to Assets −0.058 −0.050 [1.02] [0.82] Village bank 0.019 0.011 [0.62] [0.27] Solidarity 0.098 0.096 [2.31]** [1.91]* Size Indicator −0.016 −0.017 [1.63] [1.62] Log of age 0.017 0.019 [1.87]* [1.92]* Average Loan Size to GNP per capita 0.001 [0.15] Loans to assets ratio −0.072 −0.072 [2.15]** [2.11]** For‐profit dummy 0.011 0.013 [1.03] [0.99] Eastern Europe and Central Asia −0.010 −0.012 [0.79] [0.87] Sub‐Saharan Africa −0.016 −0.017 [1.31] [1.29] Middle East and N. Africa −0.001 −0.0003 [0.07] [0.02] South Asia −0.031 −0.032 [2.14]** [2.05]** East Asia −0.012 −0.012 [0.54] [0.52] Constant 0.043 0.046 [1.19] [1.05] Observations 107 101 R‐squared 0.25 0.26 All models estimated via OLS, with White’s Heteroskedasticity consistent standard errors. *significant at 10%; ** significant at 5%; *** significant at 1%. Source: Authors’ calculations, based on data from the Microfinance Information eXchange, Inc. Open in new tab Table 9 Portfolio at Risk Regressions . Portfolio at Risk . (1) . (2) . Real Yield 0.262 0.219 [2.11]** [1.18] Real Yield squared −0.337 −0.294 [2.41]** [1.58] Real Yield (Villagebank) −0.155 −0.106 [0.91] [0.48] Real Yield (Villagebank) squared 0.205 0.156 [1.09] [0.69] Real Yield (Solidarity) −0.451 −0.431 [2.01]** [1.76]* Real Yield (Solidarity) squared 0.445 0.431 [1.71]* [1.58] Capital Costs to Assets 0.087 0.086 [1.39] [1.37] Labour Costs to Assets −0.058 −0.050 [1.02] [0.82] Village bank 0.019 0.011 [0.62] [0.27] Solidarity 0.098 0.096 [2.31]** [1.91]* Size Indicator −0.016 −0.017 [1.63] [1.62] Log of age 0.017 0.019 [1.87]* [1.92]* Average Loan Size to GNP per capita 0.001 [0.15] Loans to assets ratio −0.072 −0.072 [2.15]** [2.11]** For‐profit dummy 0.011 0.013 [1.03] [0.99] Eastern Europe and Central Asia −0.010 −0.012 [0.79] [0.87] Sub‐Saharan Africa −0.016 −0.017 [1.31] [1.29] Middle East and N. Africa −0.001 −0.0003 [0.07] [0.02] South Asia −0.031 −0.032 [2.14]** [2.05]** East Asia −0.012 −0.012 [0.54] [0.52] Constant 0.043 0.046 [1.19] [1.05] Observations 107 101 R‐squared 0.25 0.26 . Portfolio at Risk . (1) . (2) . Real Yield 0.262 0.219 [2.11]** [1.18] Real Yield squared −0.337 −0.294 [2.41]** [1.58] Real Yield (Villagebank) −0.155 −0.106 [0.91] [0.48] Real Yield (Villagebank) squared 0.205 0.156 [1.09] [0.69] Real Yield (Solidarity) −0.451 −0.431 [2.01]** [1.76]* Real Yield (Solidarity) squared 0.445 0.431 [1.71]* [1.58] Capital Costs to Assets 0.087 0.086 [1.39] [1.37] Labour Costs to Assets −0.058 −0.050 [1.02] [0.82] Village bank 0.019 0.011 [0.62] [0.27] Solidarity 0.098 0.096 [2.31]** [1.91]* Size Indicator −0.016 −0.017 [1.63] [1.62] Log of age 0.017 0.019 [1.87]* [1.92]* Average Loan Size to GNP per capita 0.001 [0.15] Loans to assets ratio −0.072 −0.072 [2.15]** [2.11]** For‐profit dummy 0.011 0.013 [1.03] [0.99] Eastern Europe and Central Asia −0.010 −0.012 [0.79] [0.87] Sub‐Saharan Africa −0.016 −0.017 [1.31] [1.29] Middle East and N. Africa −0.001 −0.0003 [0.07] [0.02] South Asia −0.031 −0.032 [2.14]** [2.05]** East Asia −0.012 −0.012 [0.54] [0.52] Constant 0.043 0.046 [1.19] [1.05] Observations 107 101 R‐squared 0.25 0.26 All models estimated via OLS, with White’s Heteroskedasticity consistent standard errors. *significant at 10%; ** significant at 5%; *** significant at 1%. Source: Authors’ calculations, based on data from the Microfinance Information eXchange, Inc. Open in new tab When we plot the relation between yields and portfolio at risk for individual‐based lenders from column 1 (see Figure 3), there is a positive relation up to real yields of about 40%. Beyond that point, however, the share of loans at risk is declining in portfolio yield, and 7 to 9 individual‐based lenders have yields that high. Reconciliation with the earlier results (showing a flattening in profitability only when real yields surpass about 60%) comes from taking demand also into account. From Table 2 we know that the portfolio yield variable is significantly negatively correlated with size of an institution (i.e., total assets) and the ratio of loans to assets, so there is a negative association between charging higher interest rates and having a large customer base. By the same token, there is also a significant negative correlation between real portfolio yields and average loan size (relative to GNP per capita), which indicates that lenders charging high interest rates tend to make small loans, another possible reason for seeing the downward pressure on profitability at very high fees. Note, though, that the benchmark regressions show no general association of average loan sizes and financial sustainability. Fig. 3. Open in new tabDownload slide Portfolio Yield and Portfolio Risk, Individual Lenders Source. Authors’ calculations, based on data from the Microfinance Information eXchange, Inc. Fig. 3. Open in new tabDownload slide Portfolio Yield and Portfolio Risk, Individual Lenders Source. Authors’ calculations, based on data from the Microfinance Information eXchange, Inc. 4.4. Reducing Costs Table 10 offers regressions that relate the total cost per dollar lent to the microbank’s average loan size.11 We also include the square of average loan size to capture potential non‐linearities. The question is the degree to which expanding loan sizes improves profitability by lowering average costs. We find that larger loans are associated with lower average costs – but only up to a point. The loan size coefficient is negative and significant in both OLS and robust regressions, while loan size squared is positive and significant. The two coefficients imply a U‐shaped relationship between costs per dollar lent and average loan size for individual‐based lenders that reaches its minimum for loans two to three times per capita GNP (Figure 4). Note also that only a handful of individual‐based lenders exceed the minimum. Fig. 4. Open in new tabDownload slide Loan Size and Costs Source. Authors’ calculations, based on data from the Microfinance Information eXchange, Inc. Fig. 4. Open in new tabDownload slide Loan Size and Costs Source. Authors’ calculations, based on data from the Microfinance Information eXchange, Inc. Similar findings hold for group lenders, although they appear to be less able to exploit scale economies. For solidarity group lenders, coefficients for the loan size variables are significant and of the same sign as those for individual‐based lenders, which also implies a U‐shaped relationship between costs per dollar lent and loan size. However, the magnitudes of those coefficients imply a minimum slightly above the level of GDP per capita. Based on the respective minima for the two groups, individual‐based lenders seem better able to exploit these scale economies. The patterns for village banks are not robust to the estimation technique. Results for village banks are, in general, estimated with less precision than those for the other types of lenders. In the OLS regressions in Table 10, the loan size variables tend to share the same signs as those for individual‐based lenders but are insignificant. In the robust regressions in columns 3 and 4, the loan size coefficients are significant, large in magnitude, and of the opposite sign of those for individual‐based lenders. Future work with a larger data set may lead to more robust estimates for village banks, but the present data do not provide a reliable guide to patterns. We therefore do not present a figure for those banks. 4.5. Mission Drift Mission drift is a concern for socially‐driven microbanks. As clients mature and develop their businesses they should be able to increase loan sizes and their incomes should rise. A successful microbank will thus find that, over time, their clients receive larger loans and will be less poor. The bank’s mission and practices may well need to shift with these changes, but the result is not ‘mission drift’ as the term is generally understood. Mission drift, instead, is a shift in the composition of new clients, or a re‐orientation from poorer to wealthier clients among existing clients. The evidence above shows that the concern cannot be brushed away easily. In particular, tensions between outreach and sustainability emerged when results were disaggregated by lending type. Results from the Section above suggest that individual‐based lenders (and to a lesser extent group lenders) find it cost‐effective to increase their average loan size. In pursuing profit, microbanks would then naturally ask whether it can make sense to shift focus to wealthier borrowers who can absorb larger loans, even at the sacrifice of outreach to the poorest segments in a community. Table 10 Cost/Loan Size Trade‐offs . White’s standard errors . Robust regressions . −1 . −2 . −1 . −2 . Loan size indicator −0.223 [3.49]*** −0.218 [3.41]*** −0.219 [2.45]** −0.205 [2.60]** Loan size indicator squared 0.054 [3.26]** 0.052 [2.87]** 0.052 [2.09]** 0.049 [2.22]** Loan size × Village bank −0.238 [0.19] −0.369 [0.37] 4.823 [3.09]*** 4.792 [3.43]*** Loan size squared × Village bank 0.072 [0.05] 0.608 [0.52] −11.832 [3.23]*** −11.516 [3.49]*** Loan size × Solidarity −0.736 [3.31]*** −0.507 [2.83]*** −0.573 [3.10]*** −0.461 [2.80]*** Loan size squared × Solidarity 0.337 [3.35]*** 0.225 [2.79]*** 0.253 [2.73]*** 0.201 [2.44]** Village bank dummy 0.186 [1.10] 0.094 [0.68] −0.171 [1.19] −0.211 [1.66]* Solidarity dummy 0.228 [2.24]** 0.141 [1.84]* 0.189 [2.41]** 0.156 [2.23]** Size indicator −0.127 [3.47]*** −0.052 [2.01]** −0.087 [3.22]*** −0.062 [2.50]** Age −0.011 [3.04]*** −0.011 [3.74]*** −0.010 [3.14]*** −0.009 [3.23]*** Donations over Loan portfolio 0.507 [5.99]*** 0.408 [7.90]*** E. Eur. and Ctrl Asia 0.021 [0.23] −0.051 [0.76] −0.007 [0.12] −0.041 [0.79] Sub. Africa 0.139 [1.64] 0.101 [1.49] 0.146 [2.71]*** 0.126 [2.62]*** Middle East and N. Africa −0.142 [1.25] −0.212 [2.71]*** −0.187 [2.92]*** −0.189 [3.33]*** South Asia −0.166 [1.41] −0.186 [2.57]** −0.178 [2.78]*** −0.169 [3.01]*** East Asia 0.035 [0.45] 0.040 [0.73] 0.046 [0.68] 0.058 [0.97] Constant 0.786 [7.85]*** 0.627 [7.05]*** 0.672 [8.44]*** 0.590 [8.19]*** Observations 106 106 105 105 R‐squared 0.52 0.73 0.66 0.81 . White’s standard errors . Robust regressions . −1 . −2 . −1 . −2 . Loan size indicator −0.223 [3.49]*** −0.218 [3.41]*** −0.219 [2.45]** −0.205 [2.60]** Loan size indicator squared 0.054 [3.26]** 0.052 [2.87]** 0.052 [2.09]** 0.049 [2.22]** Loan size × Village bank −0.238 [0.19] −0.369 [0.37] 4.823 [3.09]*** 4.792 [3.43]*** Loan size squared × Village bank 0.072 [0.05] 0.608 [0.52] −11.832 [3.23]*** −11.516 [3.49]*** Loan size × Solidarity −0.736 [3.31]*** −0.507 [2.83]*** −0.573 [3.10]*** −0.461 [2.80]*** Loan size squared × Solidarity 0.337 [3.35]*** 0.225 [2.79]*** 0.253 [2.73]*** 0.201 [2.44]** Village bank dummy 0.186 [1.10] 0.094 [0.68] −0.171 [1.19] −0.211 [1.66]* Solidarity dummy 0.228 [2.24]** 0.141 [1.84]* 0.189 [2.41]** 0.156 [2.23]** Size indicator −0.127 [3.47]*** −0.052 [2.01]** −0.087 [3.22]*** −0.062 [2.50]** Age −0.011 [3.04]*** −0.011 [3.74]*** −0.010 [3.14]*** −0.009 [3.23]*** Donations over Loan portfolio 0.507 [5.99]*** 0.408 [7.90]*** E. Eur. and Ctrl Asia 0.021 [0.23] −0.051 [0.76] −0.007 [0.12] −0.041 [0.79] Sub. Africa 0.139 [1.64] 0.101 [1.49] 0.146 [2.71]*** 0.126 [2.62]*** Middle East and N. Africa −0.142 [1.25] −0.212 [2.71]*** −0.187 [2.92]*** −0.189 [3.33]*** South Asia −0.166 [1.41] −0.186 [2.57]** −0.178 [2.78]*** −0.169 [3.01]*** East Asia 0.035 [0.45] 0.040 [0.73] 0.046 [0.68] 0.058 [0.97] Constant 0.786 [7.85]*** 0.627 [7.05]*** 0.672 [8.44]*** 0.590 [8.19]*** Observations 106 106 105 105 R‐squared 0.52 0.73 0.66 0.81 Robust t‐statistics in brackets. *significant at 10%; **significant at 5%; ***significant at 1%. Source. Authors’ calculations, based on data from the Microfinance Information eXchange, Inc. Open in new tab Table 10 Cost/Loan Size Trade‐offs . White’s standard errors . Robust regressions . −1 . −2 . −1 . −2 . Loan size indicator −0.223 [3.49]*** −0.218 [3.41]*** −0.219 [2.45]** −0.205 [2.60]** Loan size indicator squared 0.054 [3.26]** 0.052 [2.87]** 0.052 [2.09]** 0.049 [2.22]** Loan size × Village bank −0.238 [0.19] −0.369 [0.37] 4.823 [3.09]*** 4.792 [3.43]*** Loan size squared × Village bank 0.072 [0.05] 0.608 [0.52] −11.832 [3.23]*** −11.516 [3.49]*** Loan size × Solidarity −0.736 [3.31]*** −0.507 [2.83]*** −0.573 [3.10]*** −0.461 [2.80]*** Loan size squared × Solidarity 0.337 [3.35]*** 0.225 [2.79]*** 0.253 [2.73]*** 0.201 [2.44]** Village bank dummy 0.186 [1.10] 0.094 [0.68] −0.171 [1.19] −0.211 [1.66]* Solidarity dummy 0.228 [2.24]** 0.141 [1.84]* 0.189 [2.41]** 0.156 [2.23]** Size indicator −0.127 [3.47]*** −0.052 [2.01]** −0.087 [3.22]*** −0.062 [2.50]** Age −0.011 [3.04]*** −0.011 [3.74]*** −0.010 [3.14]*** −0.009 [3.23]*** Donations over Loan portfolio 0.507 [5.99]*** 0.408 [7.90]*** E. Eur. and Ctrl Asia 0.021 [0.23] −0.051 [0.76] −0.007 [0.12] −0.041 [0.79] Sub. Africa 0.139 [1.64] 0.101 [1.49] 0.146 [2.71]*** 0.126 [2.62]*** Middle East and N. Africa −0.142 [1.25] −0.212 [2.71]*** −0.187 [2.92]*** −0.189 [3.33]*** South Asia −0.166 [1.41] −0.186 [2.57]** −0.178 [2.78]*** −0.169 [3.01]*** East Asia 0.035 [0.45] 0.040 [0.73] 0.046 [0.68] 0.058 [0.97] Constant 0.786 [7.85]*** 0.627 [7.05]*** 0.672 [8.44]*** 0.590 [8.19]*** Observations 106 106 105 105 R‐squared 0.52 0.73 0.66 0.81 . White’s standard errors . Robust regressions . −1 . −2 . −1 . −2 . Loan size indicator −0.223 [3.49]*** −0.218 [3.41]*** −0.219 [2.45]** −0.205 [2.60]** Loan size indicator squared 0.054 [3.26]** 0.052 [2.87]** 0.052 [2.09]** 0.049 [2.22]** Loan size × Village bank −0.238 [0.19] −0.369 [0.37] 4.823 [3.09]*** 4.792 [3.43]*** Loan size squared × Village bank 0.072 [0.05] 0.608 [0.52] −11.832 [3.23]*** −11.516 [3.49]*** Loan size × Solidarity −0.736 [3.31]*** −0.507 [2.83]*** −0.573 [3.10]*** −0.461 [2.80]*** Loan size squared × Solidarity 0.337 [3.35]*** 0.225 [2.79]*** 0.253 [2.73]*** 0.201 [2.44]** Village bank dummy 0.186 [1.10] 0.094 [0.68] −0.171 [1.19] −0.211 [1.66]* Solidarity dummy 0.228 [2.24]** 0.141 [1.84]* 0.189 [2.41]** 0.156 [2.23]** Size indicator −0.127 [3.47]*** −0.052 [2.01]** −0.087 [3.22]*** −0.062 [2.50]** Age −0.011 [3.04]*** −0.011 [3.74]*** −0.010 [3.14]*** −0.009 [3.23]*** Donations over Loan portfolio 0.507 [5.99]*** 0.408 [7.90]*** E. Eur. and Ctrl Asia 0.021 [0.23] −0.051 [0.76] −0.007 [0.12] −0.041 [0.79] Sub. Africa 0.139 [1.64] 0.101 [1.49] 0.146 [2.71]*** 0.126 [2.62]*** Middle East and N. Africa −0.142 [1.25] −0.212 [2.71]*** −0.187 [2.92]*** −0.189 [3.33]*** South Asia −0.166 [1.41] −0.186 [2.57]** −0.178 [2.78]*** −0.169 [3.01]*** East Asia 0.035 [0.45] 0.040 [0.73] 0.046 [0.68] 0.058 [0.97] Constant 0.786 [7.85]*** 0.627 [7.05]*** 0.672 [8.44]*** 0.590 [8.19]*** Observations 106 106 105 105 R‐squared 0.52 0.73 0.66 0.81 Robust t‐statistics in brackets. *significant at 10%; **significant at 5%; ***significant at 1%. Source. Authors’ calculations, based on data from the Microfinance Information eXchange, Inc. Open in new tab Table 11 Mission Drift . Average Loan Size over GNP per capita . Avg. Loan Size over GNP p.c. poorest 20% . Percentage Women Borrowers . (1) . (2) . (3) . Financial self‐sufficiency −1.618 −6.457 0.27 [2.44]** [1.17] [2.13]** Village bank −0.917 −3.953 0.623 [1.72]* [1.36] [3.67]*** Solidarity −1.143 −5.096 0.458 [1.96]* [1.73]* [2.63]** Self‐sufficiency × Village bank 1.876 7.331 −0.355 [2.68]*** [1.29] [2.08]** Self‐sufficiency × Solidarity 1.586 5.985 −0.033 [2.26]** [1.04] [0.19] Age 0.043 0.245 −0.001 [1.50] [2.05]** [0.59] Age × Village bank −0.075 −0.365 −0.026 [1.98]* [2.34]** [1.44] Age × Solidarity −0.064 −0.284 0.007 [2.09]** [2.17]** [1.55] Size Indicator 0.624 2.155 −0.113 [1.72]* [1.30] [2.31]** Size × Village bank −0.556 −1.949 0.176 [1.60] [1.13] [1.74]* Size × Solidarity −0.203 −0.523 −0.102 [0.57] [0.30] [1.48] Eastern Europe and Central Asia 0.3 −0.141 −0.096 [1.26] [0.12] [1.58] Sub‐Saharan Africa 0.418 0.161 −0.05 [2.13]** [0.17] [0.85] Middle East and N. Africa −0.189 −2.155 −0.033 [1.00] [2.34]** [0.50] South Asia 0.49 0.117 0.027 [0.93] [0.09] [0.30] East Asia −0.066 −1.915 0.075 [0.34] [1.94]* [1.07] Constant 0.832 4.534 0.447 [1.58] [1.55] [3.65]*** Observations 108 94 105 R‐squared 0.35 0.45 0.6 . Average Loan Size over GNP per capita . Avg. Loan Size over GNP p.c. poorest 20% . Percentage Women Borrowers . (1) . (2) . (3) . Financial self‐sufficiency −1.618 −6.457 0.27 [2.44]** [1.17] [2.13]** Village bank −0.917 −3.953 0.623 [1.72]* [1.36] [3.67]*** Solidarity −1.143 −5.096 0.458 [1.96]* [1.73]* [2.63]** Self‐sufficiency × Village bank 1.876 7.331 −0.355 [2.68]*** [1.29] [2.08]** Self‐sufficiency × Solidarity 1.586 5.985 −0.033 [2.26]** [1.04] [0.19] Age 0.043 0.245 −0.001 [1.50] [2.05]** [0.59] Age × Village bank −0.075 −0.365 −0.026 [1.98]* [2.34]** [1.44] Age × Solidarity −0.064 −0.284 0.007 [2.09]** [2.17]** [1.55] Size Indicator 0.624 2.155 −0.113 [1.72]* [1.30] [2.31]** Size × Village bank −0.556 −1.949 0.176 [1.60] [1.13] [1.74]* Size × Solidarity −0.203 −0.523 −0.102 [0.57] [0.30] [1.48] Eastern Europe and Central Asia 0.3 −0.141 −0.096 [1.26] [0.12] [1.58] Sub‐Saharan Africa 0.418 0.161 −0.05 [2.13]** [0.17] [0.85] Middle East and N. Africa −0.189 −2.155 −0.033 [1.00] [2.34]** [0.50] South Asia 0.49 0.117 0.027 [0.93] [0.09] [0.30] East Asia −0.066 −1.915 0.075 [0.34] [1.94]* [1.07] Constant 0.832 4.534 0.447 [1.58] [1.55] [3.65]*** Observations 108 94 105 R‐squared 0.35 0.45 0.6 All models estimated via OLS, with White’s heteroscedasticity consistent standard errors. * significant at 10%; ** significant at 5%; *** significant at 1% Source: Authors’ calculations, based on data from the Microfinance Information eXchange, Inc. Open in new tab Table 11 Mission Drift . Average Loan Size over GNP per capita . Avg. Loan Size over GNP p.c. poorest 20% . Percentage Women Borrowers . (1) . (2) . (3) . Financial self‐sufficiency −1.618 −6.457 0.27 [2.44]** [1.17] [2.13]** Village bank −0.917 −3.953 0.623 [1.72]* [1.36] [3.67]*** Solidarity −1.143 −5.096 0.458 [1.96]* [1.73]* [2.63]** Self‐sufficiency × Village bank 1.876 7.331 −0.355 [2.68]*** [1.29] [2.08]** Self‐sufficiency × Solidarity 1.586 5.985 −0.033 [2.26]** [1.04] [0.19] Age 0.043 0.245 −0.001 [1.50] [2.05]** [0.59] Age × Village bank −0.075 −0.365 −0.026 [1.98]* [2.34]** [1.44] Age × Solidarity −0.064 −0.284 0.007 [2.09]** [2.17]** [1.55] Size Indicator 0.624 2.155 −0.113 [1.72]* [1.30] [2.31]** Size × Village bank −0.556 −1.949 0.176 [1.60] [1.13] [1.74]* Size × Solidarity −0.203 −0.523 −0.102 [0.57] [0.30] [1.48] Eastern Europe and Central Asia 0.3 −0.141 −0.096 [1.26] [0.12] [1.58] Sub‐Saharan Africa 0.418 0.161 −0.05 [2.13]** [0.17] [0.85] Middle East and N. Africa −0.189 −2.155 −0.033 [1.00] [2.34]** [0.50] South Asia 0.49 0.117 0.027 [0.93] [0.09] [0.30] East Asia −0.066 −1.915 0.075 [0.34] [1.94]* [1.07] Constant 0.832 4.534 0.447 [1.58] [1.55] [3.65]*** Observations 108 94 105 R‐squared 0.35 0.45 0.6 . Average Loan Size over GNP per capita . Avg. Loan Size over GNP p.c. poorest 20% . Percentage Women Borrowers . (1) . (2) . (3) . Financial self‐sufficiency −1.618 −6.457 0.27 [2.44]** [1.17] [2.13]** Village bank −0.917 −3.953 0.623 [1.72]* [1.36] [3.67]*** Solidarity −1.143 −5.096 0.458 [1.96]* [1.73]* [2.63]** Self‐sufficiency × Village bank 1.876 7.331 −0.355 [2.68]*** [1.29] [2.08]** Self‐sufficiency × Solidarity 1.586 5.985 −0.033 [2.26]** [1.04] [0.19] Age 0.043 0.245 −0.001 [1.50] [2.05]** [0.59] Age × Village bank −0.075 −0.365 −0.026 [1.98]* [2.34]** [1.44] Age × Solidarity −0.064 −0.284 0.007 [2.09]** [2.17]** [1.55] Size Indicator 0.624 2.155 −0.113 [1.72]* [1.30] [2.31]** Size × Village bank −0.556 −1.949 0.176 [1.60] [1.13] [1.74]* Size × Solidarity −0.203 −0.523 −0.102 [0.57] [0.30] [1.48] Eastern Europe and Central Asia 0.3 −0.141 −0.096 [1.26] [0.12] [1.58] Sub‐Saharan Africa 0.418 0.161 −0.05 [2.13]** [0.17] [0.85] Middle East and N. Africa −0.189 −2.155 −0.033 [1.00] [2.34]** [0.50] South Asia 0.49 0.117 0.027 [0.93] [0.09] [0.30] East Asia −0.066 −1.915 0.075 [0.34] [1.94]* [1.07] Constant 0.832 4.534 0.447 [1.58] [1.55] [3.65]*** Observations 108 94 105 R‐squared 0.35 0.45 0.6 All models estimated via OLS, with White’s heteroscedasticity consistent standard errors. * significant at 10%; ** significant at 5%; *** significant at 1% Source: Authors’ calculations, based on data from the Microfinance Information eXchange, Inc. Open in new tab The cross‐sectional data here are not ideal for addressing mission drift since the issues inherently involve adaptation over time. We focus instead on the relationship between outreach and profitability, using a variety of outreach measures as dependent variables. Table 11 gives results on the relationship between profitability and three common measures of outreach: average loan size/GDP per capita, average loan size/GDP per capita of the poorest 20% of the population, and the share of loans extended to women. Smaller average loan size is taken as an indication of better outreach to the poor. Deflating by GDP per capita both normalises the loan size variable so that it is no longer in terms of local currency and provides an adjustment for the overall wealth of a country. In high‐inequality countries, GDP per capita is a poor reflection of typical resources for households, so normalising instead by the income accruing to the bottom 20% should be a better denominator. It turns out, though, that the results are comparable across measures. While the simple correlations show that financial self‐sufficiency is not significantly linked to any of the outreach measures, the relationship between financial self‐sufficiency and outreach becomes apparent when we disaggregate by lending type. In column 1 of Table 11, the coefficient for financial self‐sufficiency (corresponding to individual‐based lenders) is negative and significant for the average loan size variable. That coefficient is also positive and significant in column 3, where the percentage of women borrowers is the dependent variable. This suggests that individual‐based lenders that are financially self‐sustainable tend to be more focused on the poor and women. In column 1, the interaction between FSS and lending type is positive and significant for both village banks and group lenders. This does not necessarily indicate that village banks and group lenders with relatively high profitability lend less to the poor. When we sum the coefficients for the FSS variable and the respective interaction terms, we find that FSS is not significantly linked to the average loan size indicator for either type of lender (i.e., there is no evidence of mission drift). The significant coefficients on the interaction terms do however indicate that the relationship is different than that for individual‐based lenders. Countervailing trends emerge, though, when we push further by investigating the role of institutions’ age and size. The significant positive coefficient for institution size in the average loan size specification and the significant negative coefficient in the specification on gender indicate that larger individual‐based lenders do relatively poorly in terms of outreach. For village banks, the interaction with size produces coefficients of the opposite sign of those for the simple size variable. Because the magnitudes of the two sets of coefficients are similar, size is not significantly associated with outreach for village banks. For group lenders, the coefficient on institution size in the average loan size specification takes the same sign as that for village banks but the magnitude is substantially smaller. The net effect of summing the coefficients for the size variable and the Group lending × Size interaction term is significantly greater than zero, indicating that large group lenders have larger average loan sizes. Similarly, when those coefficients are summed in the ‘women borrowers’ specification, the total effect is significantly less than zero, indicating that large group lenders lend less to women. Controlling for financial self‐sufficiency, age and size by type of lending, village banks and group lenders have much smaller average loan sizes and extend a higher share of their loans to women (based on the coefficients for the simple dummy variables for those two groups). However, the interactions between lending type and age, size and financial self‐sufficiency reveal more complicated relationships than those dummy variables would suggest. The significant positive coefficient for age in the specification for average loan size divided by the GNP per capita of the poorest 20% provides some evidence of mission drift over time for individual‐based lenders. For village banks and group lenders, age appears to have less association with outreach. For example, in the women borrower specification, neither the age variable nor the Village bank × Age interaction term is significant. In the loan size specifications, the age coefficient is positive (and significant in column 2), while the interaction terms are negative and significant. The net effect of the two coefficients is never significantly different from zero for either group lenders or village banks. In sum, outreach appears to be driven by two countervailing influences for individual‐based lenders (Table 12). Size (and to a lesser extent age) is associated with less outreach, while profitability is associated with more. On balance, the evidence is consistent with the hypothesis that, as they grow larger, individual‐based lenders are more susceptible to mission drift than village banks. Outreach indicators for village banks and group lenders tend not to be significantly negatively associated with age, size, or financial self‐sufficiency. For them, mission drift would appear to be a less severe concern, although large group lenders do have worse outreach than smaller ones. Table 12 Summary of Mission Drift Results . Association with Size of Loans (significance) . Association with Proportion of Loans to Women (significance) . Individual‐Based Lenders Increases in:  Age of firm Larger (5%) No significant relation  Size of firm Larger (10%) Lower (5%)  Financial Self‐Sufficiency Smaller (5%) Higher (5%) Solidarity Group Lenders Increases in:  Age of firm No significant relation No significant relation  Size of firm Larger (1%) Lower (1%)  Financial Self‐Sufficiency No significant relation Higher (5%) Village Banks Increases in:  Age of firm No significant relation No significant relation  Size of firm No significant relation No significant relation  Financial Self‐Sufficiency No significant relation No significant relation . Association with Size of Loans (significance) . Association with Proportion of Loans to Women (significance) . Individual‐Based Lenders Increases in:  Age of firm Larger (5%) No significant relation  Size of firm Larger (10%) Lower (5%)  Financial Self‐Sufficiency Smaller (5%) Higher (5%) Solidarity Group Lenders Increases in:  Age of firm No significant relation No significant relation  Size of firm Larger (1%) Lower (1%)  Financial Self‐Sufficiency No significant relation Higher (5%) Village Banks Increases in:  Age of firm No significant relation No significant relation  Size of firm No significant relation No significant relation  Financial Self‐Sufficiency No significant relation No significant relation Notes. Significance levels in parenthesis. A significant result for loan size implies that the coefficient was significant in either model 1 of Table 11 (with dependent variable average loan size over GNP per capita) or model 2 (with dependent variable average loan size over the GNP per capita of the poorest 20% of the population), or both. Open in new tab Table 12 Summary of Mission Drift Results . Association with Size of Loans (significance) . Association with Proportion of Loans to Women (significance) . Individual‐Based Lenders Increases in:  Age of firm Larger (5%) No significant relation  Size of firm Larger (10%) Lower (5%)  Financial Self‐Sufficiency Smaller (5%) Higher (5%) Solidarity Group Lenders Increases in:  Age of firm No significant relation No significant relation  Size of firm Larger (1%) Lower (1%)  Financial Self‐Sufficiency No significant relation Higher (5%) Village Banks Increases in:  Age of firm No significant relation No significant relation  Size of firm No significant relation No significant relation  Financial Self‐Sufficiency No significant relation No significant relation . Association with Size of Loans (significance) . Association with Proportion of Loans to Women (significance) . Individual‐Based Lenders Increases in:  Age of firm Larger (5%) No significant relation  Size of firm Larger (10%) Lower (5%)  Financial Self‐Sufficiency Smaller (5%) Higher (5%) Solidarity Group Lenders Increases in:  Age of firm No significant relation No significant relation  Size of firm Larger (1%) Lower (1%)  Financial Self‐Sufficiency No significant relation Higher (5%) Village Banks Increases in:  Age of firm No significant relation No significant relation  Size of firm No significant relation No significant relation  Financial Self‐Sufficiency No significant relation No significant relation Notes. Significance levels in parenthesis. A significant result for loan size implies that the coefficient was significant in either model 1 of Table 11 (with dependent variable average loan size over GNP per capita) or model 2 (with dependent variable average loan size over the GNP per capita of the poorest 20% of the population), or both. Open in new tab 5. Conclusion At the outset of this article, we sought to address three questions. Does raising interest rates exacerbate agency problems as detected by lower repayment rates and less profitability? Is there evidence of a trade‐off between the depth of outreach to the poor and the pursuit of profitability? Has ‘mission drift’ occurred – i.e., have microbanks moved away from serving their poorer clients in pursuit of commercial viability? Based on a high‐quality survey of 124 microfinance institutions, we find that the answers to our questions depend on an institution’s lending method. For example, we find that individual‐based lenders that charge higher interest rates are more profitable than others but only up to a point. Beyond threshold interest rates, profitability tends to be lower. The patterns are consistent with greater loan delinquency (following predictions from agency theory) and, at the highest rates, to falling demand for credit. In contrast, for solidarity group lenders, financial performance tends not to improve (or even worsens in models with quadratic terms) as yields increase throughout most of our sample range. We acknowledge the possibility of alternative interpretations. For example, the social objectives of some MFIs might compel them to charge lower interest rates and thus earn lower profits. Those institutions might require substantial subsidies to operate, consistent with the negative correlations between subsidies and profitability in Table 6. However, this would not explain the trade‐offs we find for MFIs charging relatively high yields. On our second question, regarding trade‐offs between outreach to the poor and profitability, the simple relationship between profitability and average loan size is insignificant in our base regressions. Controlling for other relevant factors, institutions that make smaller loans are not necessarily less profitable. But we do find that larger loan sizes are associated with lower average costs for both individual‐based lenders and solidarity group lenders. Since larger loan size is often taken to imply less outreach to the poor, the result could have negative implications. For individual‐based lenders, the pattern of results are consistent with disincentives for depth of outreach – i.e., the personnel expenses devoted to identifying borrowers worthy of larger loans could deter institutions from serving the poorest segments of society. At the same time, we note that it is not just the poorest that demand and can take advantage of better access to finance. We also find some positive results for individual‐based lenders regarding mission drift, the third issue we sought to address. Financially self‐sustaining individual‐based lenders tend to have smaller average loan size and lend more to women, suggesting that pursuit of profit and outreach to the poor can go hand in hand. There are however countervailing influences: larger individual‐based and group‐based lenders tend to extend larger loans and lend less frequently to women. Older individual‐based lenders also do worse on outreach measures than younger ones. While this is not evidence of mission drift in the strict sense (i.e., that pursuit of improved financial performance reduces focus on the poor), the results for larger and older microbanks are consistent with the idea that as institutions mature and grow, they focus increasingly on clients that can absorb larger loans. On the whole, our results suggest that institutional design and orientation matters importantly in considering trade‐offs in microfinance. The trade‐offs can be stark: village banks, which focus on the poorest borrowers, face the highest average costs and the highest subsidy levels. By the same token, individual‐based lenders earn the highest average profits but do least well on indicators of outreach to the very poor. At the same time, we find examples of institutions that have managed to achieve profitability together with notable outreach to the poor – achieving the ultimate promise of microfinance. But they are, so far, the exceptions. Footnotes 1 " A great deal has been written on microfinance theory within the past fifteen years (Stiglitz, 1990; Banerjee et al., 1994; Besley and Coate, 1995; Conning, 1999; Ghatak and Guinnane, 1999; Laffont and Rey, 2003; Rai and Sjöström, 2004). Armendáriz de Aghion and Morduch (2005) provide a critical guide to the economics literature on microfinance, and Ahlin and Townsend (2007) test leading models with Thai data. 2 " We take this goal on face value, although we recognise the case for subsidised microfinance when social benefits sufficiently outweigh social costs and subsidies do not undercut non‐subsidised firms. The goal of profit‐making microfinance is discussed by Robinson (2001). Armendáriz de Aghion and Morduch (2005) discuss subsidy and sustainability in their chapter 9. 3 " The Microfinance Information eXchange, Inc. (MIX) kindly provided the data (with confidentiality safeguards) through an agreement with the World Bank Research Department. The MIX is a non‐profit company dedicated to improving the information infrastructure of the microfinance industry by promoting standards of financial and operational reporting and providing data. The data we use were collected as part of the MIX’s MicroBanking Bulletin project. Summary statistics on the institutions are available in the Bulletin (http://www.mixmarket.org). 4 " New work using field experiments (Karlan and Zinman, 2005) or natural experiments (Dehejia et al., 2005) shows promise in ways either to exploit the variation that exists or to create variation as part of a research programme. 5 " The financial self‐sufficiency ratio is adjusted financial revenue divided by the sum of adjusted financial expenses, adjusted net loan loss provision expenses, and adjusted operating expenses. It indicates the institution’s ability to operate without ongoing subsidy, including soft loans and grants. The definition is from MicroBanking Bulletin (2005), p. 57. 6 " The operational sustainability ratio is financial revenue divided by the sum of financial expenses, net loan loss provision expenses and operating expenses. Unlike the financial self‐sufficiency ratio, OSS is not adjusted. Return on assets is measured as adjusted net operating income (net of taxes) divided by adjusted average total assets. Definitions are from MicroBanking Bulletin (2005), p. 57. 7 " Within the group of individual‐based lenders there are also sources of variation that we would ideally like to capture. For example, such lenders vary in the extent to which they require collateral to secure loans. Unfortunately, we lack the data necessary to capture finer distinctions between institutions that use the same lending type. 8 " The sum of the capital costs and capital costs × solidarity lender coefficients is significantly less than zero for all specifications in Table 7. 9 " In addition to controlling for age in the base regressions, we also ran models on subsets of MFIs of similar vintage (5–20 years old). Because the performance indicators for young MFIs are widely dispersed, our results are at least as strong when we restrict the sample in this way. See Appendix for reasons why data from young MFIs might be most in need of adjustment. 10 " Note that the models in Table 8 are run via OLS, with White’s standard errors. Similar qualitative results were obtained for robust regressions, although significance levels were lower. Because we are trying to illustrate the effect of relatively extreme portfolio yields, the OLS models were more appropriate than robust techniques that were likely to downweight such observations. We thank David Roodman for pointing out that strong correlations between linear and quadratic terms of the same series can spuriously generate the kinds of patterns here. While we acknowledge the point, the base regressions in Table 7 do not contain the quadratic term, so we feel confident that the positive linear relation between portfolio yield and financial performance is not spurious. We include specifications with the quadratic yield term as a simple test of whether the linear relation becomes less steep as interest rates climb. And again, there are strong theoretical reasons to expect that this might be the case. 11 " We exclude from the regressions institutions with operating costs less than 5% (one observation) or greater than 150% (two observations) of total loans, as these seemed implausible. Two of those observations were already not part of our sample in the base profitability models due to missing data for some variables. References Ahlin , C. and Townsend , R. ( 2007 ). ‘Using repayment data to test across models of joint liability lending’ , Economic Journal , vol. 117, pp. F11 – F51 . OpenURL Placeholder Text WorldCat Armendáriz de Aghion , B. ( 1999 ). ‘On the design of a credit agreement with peer monitoring’ , Journal of Development Economics , vol. 60 (October), pp. 79 – 104 . Google Scholar Crossref Search ADS WorldCat Armendáriz de Aghion , B. and Morduch , J. ( 2005 ). The Economics of Microfinance , Cambridge, MA: MIT Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Banerjee , A. , Besley , T. and Guinnane , T. ( 1994 ). ‘Thy neighbor’s keeper: the design of a credit cooperative with theory and a test’ , Quarterly Journal of Economics , vol. 109 (May), pp. 491 – 515 . Google Scholar Crossref Search ADS WorldCat Besley , T. ( 1995 ). ‘Savings, credit, and insurance’, in ( J. Behrman and T.N. Sarinivasan, eds.), Handbook of Development Economics , Volume 3A , pp. 2123 – 207 , Amsterdam: North‐Holland . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Besley and Coate ( 1995 ). ‘Group lending, repayment incentives, and social collateral’, Journal of Developing Economics , vol. 46 (February), pp. 1 – 18 . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Cassar , A. , Crowley , L. and Wydick , B. ( 2006 ). ‘The effect of social capital on group loan repayment: evidence from field experiments’ , Economic Journal , vol. 117, pp. F85 – F106 . OpenURL Placeholder Text WorldCat Conning , J. ( 1999 ). ‘Outreach, sustainability and leverage in monitored and peer‐monitored lending’ , Journal of Development Economics , vol. 60 (October), pp. 51 – 77 . Google Scholar Crossref Search ADS WorldCat Dehejia , R. , Montgomery , H. and Morduch , J. ( 2005 ). ‘Do interest rates matter? Evidence from the Dhaka slums’ , Columbia University: Department of Economics and NYU Wagner School working paper . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Ghatak , M. ( 2000 ). ‘Screening by the company you keep: joint liability lending and the peer selection effect’ , Economic Journal , vol. 110 (July), pp. 601 – 31 . Google Scholar Crossref Search ADS WorldCat Ghatak , M . Guinnane , T. ( 1999 ). ‘The economics of lending with joint liability: theory and practice’, Journal of Development Economics , vol. 60 (October), pp. 195 – 228 . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Honohan , P. ( 2004 ). ‘Financial sector policy and the poor: selected findings and issues’ , World Bank Working Paper No. 43, Washington, DC: World Bank . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Karlan , D. ( 2006 ). ‘Social connections and group banking’ , Economic Journal , vol. 117, pp. F52 – F84 . OpenURL Placeholder Text WorldCat Karlan , D. and Zinman , J. ( 2005 ). ‘Observing unobservables: identifying information asymmetries with a consumer credit field experiment’ , Princeton University , Woodrow Wilson School, mimeo. Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Laffont , J. and Rey , P. ( 2003 ). ‘Collusion and group lending with moral hazard’ , IDEI, Toulouse and University of Southern California , mimeo. Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Microbanking Bulletin ( 2005 ). ‘Trend lines’ , Issue 10 ( 5 ), March. Rai , A. and Sjöström , T. ( 2004 ). ‘Is Grameen lending efficient? Repayment incentives and insurance in village economies’ , Review of Economic Studies , vol. 71 (January), pp. 217 – 34 . Google Scholar Crossref Search ADS WorldCat Robinson , M. ( 2001 ). The Microfinance Revolution: Sustainable Banking for the Poor , Washington, DC: The World Bank . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Stiglitz , J. ( 1990 ). ‘Peer monitoring and credit markets’ , World Bank Economic Review , vol. 4 (September), pp. 351 – 66 . Google Scholar Crossref Search ADS WorldCat Stiglitz , J. and Weiss , A. ( 1981 ). ‘Credit rationing in markets with imperfect information’ , American Economic Review , vol. 71 (June), pp. 393 – 410 . OpenURL Placeholder Text WorldCat Appendix. Financial Statement Adjustments and their Effects Adjustment . Effect on Financial Statements . Type of Institution Most Affected by Adjustment . Inflation adjustment of equity (minus net fixed assets) Increases financial expense accounts on income statement, to some degree offset by inflation income account for revaluation of fixed assets. Generates a reserve in the balance sheet’s equity account, reflecting that portion of the MFI’s retained earnings that has been consumed by the effects of inflation. Decreases profitability and ‘real’ retained earnings. MFIs funded more by equity than by liabilities will be hardest hit, especially in high‐inflation countries. Reclassification of certain long‐term liabilities into equity, and subsequent inflation adjustment Decreases concessionary loan account and increases equity account; increases inflation adjustment on income statement and balance sheet. NGOs that have long‐term low‐interest ‘loans’ from international agencies that function more as donations than loans. Subsidised cost of funds adjustment. Increases financial expense on income statement to the extent that the MFI’s liabilities carry a below‐market rate of interest. Decreases net income and increases subsidy adjustment account on balance sheet. MFIs with heavily subsidised loans (i.e., large lines of credit from governments or international agencies at highly subsidised rates). Subsidy adjustment: current‐year cash donations to cover operating expenses Reduces operating expense on income statement (if the MFI records donations as operating income). Increases subsidy adjustment account on balance sheet. NGOs during their start‐up phase. The adjustment is relatively less important for mature institutions. In‐kind subsidy adjustment (e.g. donation of goods or services: line staff paid for by technical assistance providers) Increases operating expense on income statement to the extent that the MFI is receiving subsidised or donated goods or services. Decreases net income, increases subsidy adjustment on balance sheet. MFIs using goods or services for which they are not paying a market‐based cost (i.e., MFIs during their start‐up phase). Loan loss reserve and provision expense adjustment Usually increases loan loss provision expense on income statement and loan loss reserve on balance sheet. MFIs that have unrealistic loan loss provisioning policies. Write‐off adjustment On balance sheet, reduces gross loan portfolio and loan loss reserve by an equal amount, so that neither the net loan portfolio nor the income statement is affected. Improves (lowers) portfolio‐at ‐risk ratio. MFIs that do not write off non‐performing loans aggressively enough. Reversal of interest income accrued on non‐performing loans Reduces financial income and net profit on the income statement, and equity on the balance sheet. MFIs that continue accruing income on delinquent loans past the point where collection becomes unlikely, or that fail to reverse previously accrued income on such loans. Adjustment . Effect on Financial Statements . Type of Institution Most Affected by Adjustment . Inflation adjustment of equity (minus net fixed assets) Increases financial expense accounts on income statement, to some degree offset by inflation income account for revaluation of fixed assets. Generates a reserve in the balance sheet’s equity account, reflecting that portion of the MFI’s retained earnings that has been consumed by the effects of inflation. Decreases profitability and ‘real’ retained earnings. MFIs funded more by equity than by liabilities will be hardest hit, especially in high‐inflation countries. Reclassification of certain long‐term liabilities into equity, and subsequent inflation adjustment Decreases concessionary loan account and increases equity account; increases inflation adjustment on income statement and balance sheet. NGOs that have long‐term low‐interest ‘loans’ from international agencies that function more as donations than loans. Subsidised cost of funds adjustment. Increases financial expense on income statement to the extent that the MFI’s liabilities carry a below‐market rate of interest. Decreases net income and increases subsidy adjustment account on balance sheet. MFIs with heavily subsidised loans (i.e., large lines of credit from governments or international agencies at highly subsidised rates). Subsidy adjustment: current‐year cash donations to cover operating expenses Reduces operating expense on income statement (if the MFI records donations as operating income). Increases subsidy adjustment account on balance sheet. NGOs during their start‐up phase. The adjustment is relatively less important for mature institutions. In‐kind subsidy adjustment (e.g. donation of goods or services: line staff paid for by technical assistance providers) Increases operating expense on income statement to the extent that the MFI is receiving subsidised or donated goods or services. Decreases net income, increases subsidy adjustment on balance sheet. MFIs using goods or services for which they are not paying a market‐based cost (i.e., MFIs during their start‐up phase). Loan loss reserve and provision expense adjustment Usually increases loan loss provision expense on income statement and loan loss reserve on balance sheet. MFIs that have unrealistic loan loss provisioning policies. Write‐off adjustment On balance sheet, reduces gross loan portfolio and loan loss reserve by an equal amount, so that neither the net loan portfolio nor the income statement is affected. Improves (lowers) portfolio‐at ‐risk ratio. MFIs that do not write off non‐performing loans aggressively enough. Reversal of interest income accrued on non‐performing loans Reduces financial income and net profit on the income statement, and equity on the balance sheet. MFIs that continue accruing income on delinquent loans past the point where collection becomes unlikely, or that fail to reverse previously accrued income on such loans. 1 Source. The Microbanking Bulletin, ‘Our Methodology’ (http://www.mixmbb.org/en/company/our_methodology.html) Adjustment . Effect on Financial Statements . Type of Institution Most Affected by Adjustment . Inflation adjustment of equity (minus net fixed assets) Increases financial expense accounts on income statement, to some degree offset by inflation income account for revaluation of fixed assets. Generates a reserve in the balance sheet’s equity account, reflecting that portion of the MFI’s retained earnings that has been consumed by the effects of inflation. Decreases profitability and ‘real’ retained earnings. MFIs funded more by equity than by liabilities will be hardest hit, especially in high‐inflation countries. Reclassification of certain long‐term liabilities into equity, and subsequent inflation adjustment Decreases concessionary loan account and increases equity account; increases inflation adjustment on income statement and balance sheet. NGOs that have long‐term low‐interest ‘loans’ from international agencies that function more as donations than loans. Subsidised cost of funds adjustment. Increases financial expense on income statement to the extent that the MFI’s liabilities carry a below‐market rate of interest. Decreases net income and increases subsidy adjustment account on balance sheet. MFIs with heavily subsidised loans (i.e., large lines of credit from governments or international agencies at highly subsidised rates). Subsidy adjustment: current‐year cash donations to cover operating expenses Reduces operating expense on income statement (if the MFI records donations as operating income). Increases subsidy adjustment account on balance sheet. NGOs during their start‐up phase. The adjustment is relatively less important for mature institutions. In‐kind subsidy adjustment (e.g. donation of goods or services: line staff paid for by technical assistance providers) Increases operating expense on income statement to the extent that the MFI is receiving subsidised or donated goods or services. Decreases net income, increases subsidy adjustment on balance sheet. MFIs using goods or services for which they are not paying a market‐based cost (i.e., MFIs during their start‐up phase). Loan loss reserve and provision expense adjustment Usually increases loan loss provision expense on income statement and loan loss reserve on balance sheet. MFIs that have unrealistic loan loss provisioning policies. Write‐off adjustment On balance sheet, reduces gross loan portfolio and loan loss reserve by an equal amount, so that neither the net loan portfolio nor the income statement is affected. Improves (lowers) portfolio‐at ‐risk ratio. MFIs that do not write off non‐performing loans aggressively enough. Reversal of interest income accrued on non‐performing loans Reduces financial income and net profit on the income statement, and equity on the balance sheet. MFIs that continue accruing income on delinquent loans past the point where collection becomes unlikely, or that fail to reverse previously accrued income on such loans. Adjustment . Effect on Financial Statements . Type of Institution Most Affected by Adjustment . Inflation adjustment of equity (minus net fixed assets) Increases financial expense accounts on income statement, to some degree offset by inflation income account for revaluation of fixed assets. Generates a reserve in the balance sheet’s equity account, reflecting that portion of the MFI’s retained earnings that has been consumed by the effects of inflation. Decreases profitability and ‘real’ retained earnings. MFIs funded more by equity than by liabilities will be hardest hit, especially in high‐inflation countries. Reclassification of certain long‐term liabilities into equity, and subsequent inflation adjustment Decreases concessionary loan account and increases equity account; increases inflation adjustment on income statement and balance sheet. NGOs that have long‐term low‐interest ‘loans’ from international agencies that function more as donations than loans. Subsidised cost of funds adjustment. Increases financial expense on income statement to the extent that the MFI’s liabilities carry a below‐market rate of interest. Decreases net income and increases subsidy adjustment account on balance sheet. MFIs with heavily subsidised loans (i.e., large lines of credit from governments or international agencies at highly subsidised rates). Subsidy adjustment: current‐year cash donations to cover operating expenses Reduces operating expense on income statement (if the MFI records donations as operating income). Increases subsidy adjustment account on balance sheet. NGOs during their start‐up phase. The adjustment is relatively less important for mature institutions. In‐kind subsidy adjustment (e.g. donation of goods or services: line staff paid for by technical assistance providers) Increases operating expense on income statement to the extent that the MFI is receiving subsidised or donated goods or services. Decreases net income, increases subsidy adjustment on balance sheet. MFIs using goods or services for which they are not paying a market‐based cost (i.e., MFIs during their start‐up phase). Loan loss reserve and provision expense adjustment Usually increases loan loss provision expense on income statement and loan loss reserve on balance sheet. MFIs that have unrealistic loan loss provisioning policies. Write‐off adjustment On balance sheet, reduces gross loan portfolio and loan loss reserve by an equal amount, so that neither the net loan portfolio nor the income statement is affected. Improves (lowers) portfolio‐at ‐risk ratio. MFIs that do not write off non‐performing loans aggressively enough. Reversal of interest income accrued on non‐performing loans Reduces financial income and net profit on the income statement, and equity on the balance sheet. MFIs that continue accruing income on delinquent loans past the point where collection becomes unlikely, or that fail to reverse previously accrued income on such loans. 1 Source. The Microbanking Bulletin, ‘Our Methodology’ (http://www.mixmbb.org/en/company/our_methodology.html) Author notes " The views are those of the authors and not necessarily those of the World Bank or its affiliate institutions. The Microfinance Information eXchange, Inc. (MIX) provided the data through an agreement with the World Bank Research Department. Confidentiality of institution‐level data has been maintained. We thank Isabelle Barres, Joao Fonseca, Didier Thys and Peter Wall of the Microfinance Information Exchange (MIX) for their substantial efforts in assembling both the adjusted data and the qualitative information on MFIs for us. We have benefited from comments from Thorsten Beck, Patrick Honohan, Stijn Claessens, Bert Sholtens and participants at the Groningen conference. Sarojini Hirshleifer and Varun Kshirsagar provided expert assistance with the research. Any errors are ours only. © The Author(s). Journal compilation © Royal Economic Society 2007

Hermes,, Niels;Lensink,, Robert

doi: 10.1111/j.1468-0297.2007.02013.xpmid: N/A

Abstract Microfinance has received a lot of attention recently, both from policy makers as well as in academic circles. Two of the main topics that have been hotly debated are explaining joint liability group lending and its implications for reducing information asymmetries, and the trade‐off between the financial sustainability and outreach of microfinance programmes. This Feature contains three novel empirical contributions providing new insights with respect to why and how joint liability group lending works. It also contains the first large‐scale systematic analysis of the trade‐off between financial performance and outreach of microfinance institutions. Lack of access to credit is generally seen as one of the main reasons why many people in developing economies remain poor. Usually, the poor have no access to loans from the banking system, because they cannot put up acceptable collateral and/or because the costs for banks of screening and monitoring the activities of the poor, and of enforcing their contracts, are too high to make lending to this group profitable. Since the late 1970s, however, the poor in developing economies have increasingly gained access to small loans with the help of so‐called microfinance programmes. Especially during the past ten years, these programmes have been introduced in many developing economies. Well‐known examples are the Grameen Bank in Bangladesh, Banco Sol in Bolivia and Bank Rakyat in Indonesia. The Grameen Bank system of group lending (established in 1976 by Mohammad Yunus, a Bengal banker and economist), in particular, has been widely copied in other developing countries. Between December 1997 and December 2005 the number of microfinance institutions increased from 618 to 3,133. The number of people who received credit from these institutions rose from 13.5 million to 113.3 million (84% of them being women) during the same period (Daley‐Harris, 2006). According to the United Nations (UN), in 2002 almost one fifth of the world population (i.e. 1.3 billion people) were living in extreme poverty, earning less than one dollar a day. In recent public debates microfinance has been mentioned as an important instrument to combat extreme poverty. To support this view the UN declared 2005 to be the International Year of Microcredit. According to the UN, microfinance can contribute significantly to the achievement of the United Nations Millennium Development Goals, as agreed upon by world leaders at the UN Millennium Summit in September 2000, and which aim at halving extreme poverty by 2015. In October 2006, the attention for microfinance and its role in reducing poverty was further increased when Mohammad Yunus received the Nobel peace prize. According to the Nobel Committee microfinance can help people to break out of poverty, which in turn is seen as an important prerequisite to establish long lasting peace (Norwegian Nobel Committee, 2006). Next to the growing attention from policy makers, the academic world has also shown increased interest in microfinance, especially during the last ten or so years. Several questions have been addressed in the literature. One major strand of literature focuses on explaining how and why microfinance works from a theoretical perspective. In this context, most models focus on explaining so‐called joint liability group lending and its implications for reducing information asymmetries. Yet, there are only a few empirical studies investigating whether and how microfinance helps to reduce existing information asymmetries. A second important and related issue discussed in the literature deals with the trade‐off between the financial sustainability and outreach of microfinance programmes. Although this issue is the subject of a heated debate, there is a lack of systematic empirical analyses on the nature and determinants of the trade‐off. This Feature aims to provide new empirical evidence on several important questions related to microfinance. In particular, the feature contains four novel empirical contributions to the literature. Three of them (Ahlin and Townsend, 2007; Karlan, 2007 and Cassar et al., 2007) deal with joint liability group lending. They provide new insights with respect to why and how this type of lending works in enhancing repayment rates, which may contribute to improving the sustainability of these programmes. The fourth article addresses the issues of financial performance and outreach of microfinance programmes (Cull et al., 2007). The remainder of this introduction consists of a short review of the existing literature on the two topics to which the articles in this Feature are related. 1. The Economics of Joint Liability Group Lending Generally speaking, microfinance programmes provide credit to the poor, either through joint liability group lending or through individual‐based lending. While the latter comes close to traditional banking, involving a direct relationship between the programme and an individual, the joint liability lending approach uses groups of borrowers to which loans are made. Currently, the majority of microfinance borrowers have access to loans through group lending programmes. According to one recent survey of a sample of microfinance programmes, only 16% of these made use of so‐called group lending to provide credit to the poor; yet, they served more than two thirds of all borrowers from the microfinance programmes included in the survey (Lapenu and Zeller, 2001). With joint liability lending the group of borrowers is made responsible for the repayment of the loan, i.e. all group members are jointly liable. Thus, if one group member does not repay her loan, others may have to contribute so as to ensure repayment. Non‐repayment by the group means that all group members will be denied future access to loans from the programme. In this way, group lending creates incentives for individual group members to screen and monitor other members of the group and to enforce repayment in order to reduce the risk of having to contribute to the repayment of loans of others and to ensure access to future loans. Thus, joint liability group lending stimulates screening, monitoring and enforcement of contracts among borrowers, reducing or erasing the agency costs of the lender. Moreover, the group lending structure is also expected to be more effective in providing such activities as compared to the lender, because group members usually live close to each other and/or have social ties (also referred to as social capital in the existing literature). They are therefore better informed about each other’s activities. Since joint liability group lending stimulates screening, monitoring and enforcement within the group, and since it improves the effectiveness of these activities due to geographical proximity and close social ties, repayment performance of group loans is expected to be high. Several theoretical models confirm that joint liability group lending leads to more and more effective screening, monitoring and enforcement among group members. Some of these models explicitly focus on the properties of joint liability lending related to mitigating information asymmetries. For example, models by Stiglitz (1990) and Varian (1990), Banerjee et al. (1994), Armendáriz de Aghion (1999) and Chowdury (2005) explicitly deal with moral hazard and monitoring problems, showing how joint liability may help to solve these problems. Ghatak (1999; 2000) and Gangopadhyay et al. (2005), among others, provide models focusing on adverse selection and screening. Some other models specifically discuss the role of social ties within group lending in improving repayment performance of groups. The work of Besley and Coate (1995) and Wydick (2001) fall into this category of models. In spite of the abundance of theoretical literature, there has been surprisingly little empirical evidence of whether and how microfinance actually helps to reduce existing information asymmetries. This is, at least partly, due to the difficulty of obtaining reliable data on the working of these programmes and the behaviour of their participants. Most of the available empirical studies address the general question of whether joint liability group lending improves repayment performance of groups, using different types of proxies for screening, monitoring and enforcement behaviour taking place within groups. Wenner (1995) provides one of the first empirical studies on the determinants of repayment of groups, using information of 25 groups from a lending programme in Costa Rica. His analysis indicates that repayment performance of groups improves when groups have written (formal) rules stating how members should behave. This variable implicitly measures screening, monitoring and enforcement activities that take place within the groups. Another variable that is found to determine repayment is the location of groups: if groups are located in remote areas this reduces their possibilities for access to alternative sources of credit, which stimulates them to ensure group repayment as much as possible in order to have future access to loans. Sharma and Zeller (1997), using data of 128 groups from four group lending programmes in Bangladesh, show that repayment problems increase when there are more relatives in the same group. This supports the hypothesis that screening, monitoring and enforcement among relatives does not take place or at least is less effective, since relatives may more easily collude against the programme and delay repayment. Second, the results indicate that if borrowers are more credit rationed this increases repayment performance. This result can be taken as evidence for the fact that group members have more incentives to screen, monitor and enforce if they have no alternative credit sources. Third, Sharma and Zeller (1997) find that groups that were formed using a self‐selection (screening) process show a better repayment performance. Zeller (1998), based on information from 146 groups in Madagascar, focuses on the role of social ties and finds evidence that groups with stronger ties show higher repayment rates. Moreover, he shows that groups with internal rules and regulations demonstrate better repayment rates, a result that was also reported in Wenner (1995). An influential study is carried out by Wydick (1999), who uses data of 137 groups from a group‐based lending programme in Guatemala. This paper uses the most extensive list of variables to measure screening, monitoring and enforcement within groups. Wydick finds evidence for the fact that the average distance between group members negatively influences repayment performance, whereas the knowledge one member has of the weekly sales of other members is positively related to repayment performance. Both variables are assumed to measure monitoring activities within groups. However, he also finds evidence that social ties within groups reduces the pressure members put on each other to repay loans. Paxton et al. (2000) use data of 140 groups from a group‐based lending programme in Burkina Faso. They show that the homogeneity of the group in terms of their ethnicity, occupation, income etc., reduces its repayment performance. This may indicate that if members are more homogeneous they have lower incentives to screen, monitor and enforce each other and/or may start to collude against the programme. They also show that social pressure within groups is positively related to repayment performance. Finally, they find that the quality of the group leader in running the group is positively related to repayment performance, which may be seen as evidence for the fact that the group leader plays a prominent role in screening, monitoring and enforcement within the group. Hermes et al. (2005) elaborate on this last result and investigate the role of the group leader in reducing moral hazard behaviour, using data of 102 groups from two Eritrean group lending programmes. They find evidence that monitoring and social ties of the group leader reduce moral hazard behaviour of group members. This result is not found for the other group members. In a related paper they also find evidence that the role of the group leader is most important in improving repayment performance of the group (Hermes et al., 2006). The empirical studies mentioned above present interesting results on how and why joint liability group lending works. However, they also suffer from a number of potential weaknesses. First, in most papers the link between theory and empirics is rather implicit. Many of the variables used to measure group member behaviour in terms of screening, monitoring and enforcement are only indirectly related to the contents of these concepts from a theoretical perspective. Moreover, in several cases crude, or at least one‐dimensional, measures are used to proxy for complex constructs such as social ties. Finally, the empirical analyses may suffer from endogeneity problems. This may be especially problematic for studies investigating the role of social ties in mitigating information asymmetries and improving repayment rates (Karlan, 2007). Three of the four articles in this Feature address the potential weaknesses of previous empirical work. Ahlin and Townsend (2007) explicitly derive direct empirical tests from four well‐known theoretical models of adverse selection, moral hazard and social sanctions. Karlan focuses on the role of social ties in group lending and uses an empirical setting, which allows the solving of the endogeneity problem other papers suffer from when investigating this issue. Cassar et al. (2007) take a novel approach by carrying out microfinance experiments. In this way, they are able to analyse several different components of social ties and their influence on the working of groups. All three articles provide important contributions to a better understanding of how joint liability group lending works. In their article, Ahlin and Townsend (2007) focus on the empirical implications of four well‐known theoretical models of joint liability group lending. In particular, they take the models Stiglitz (1990) and Banerjee et al. (1994) that explain how joint liability may solve moral hazard problems; they use the Besley and Coate (1995) model which describes how group lending may solve problems of limited contract enforcement by using social sanctions; and they use the Ghatak (1999) explaining how joint liability contracts help to solve adverse selection problems. Based on these models they generate theoretical predictions regarding the determinants of the repayment performance of groups. Since the models assume different economic environments and focus on different types of problems joint liability group lending should solve, predictions regarding determinants of repayment performance may differ between models, and this is what they indeed find. In particular, they show that conflicting predictions can be found for the role of cooperation (or social cohesion) between group members, the correlation between borrower returns and the degree of joint liability in explaining repayment performance. Using a very rich dataset containing detailed information on 262 groups of the Bank for Agricultural Copperatives (BAAC) in Thailand, Ahlin and Townsend (2007) empirically test the predictions of the four different models. They find empirical support for the fact that repayment performance is negatively associated with higher levels of relatedness and sharing within groups and with higher levels of joint liability. Their results also support the suggestion that repayment performance is positively associated with the strength of local sanctions and with higher correlations between borrower returns. Their most interesting result is that social ties between group members are not necessarily positive in promoting group repayment, which contrasts the generally accepted view in the literature. Karlan’s article investigates the role of social ties, or social connections in his terminology, in group lending by explicitly testing whether groups with stronger connections outperform those with weaker connections. As was mentioned above, most of the earlier studies on the role of social connections in group lending suffer from an endogeneity problem. It may well be that the nature of social connections correlates with other economic or social characteristics that may independently influence repayment performance. If this is the case, one cannot draw conclusions on the causal nature of the relationship between higher repayment performance and stronger social connections. Karlan (2007) is able to circumvent this endogeneity problem by making use of a natural experiment, which allows him to rule out the possibility that the nature social connections correlates with other group characteristics influencing repayment. The empirical setting of the article focuses on the microfinance organisation FINCA‐Peru. This organisation randomly creates groups: if a person wants to obtain a loan, she is put on a list, without taking into account where she lives or whether she knows the other persons who are already on the list. Once the list contains 30 persons the group can start. This process of group formation exogenously creates groups with different levels of initial social ties, which enables the actual measurement of the impact of these social ties on monitoring and enforcement efforts within the group. The empirical analysis is based on a large dataset containing information of over 2,000 individual group members. The most important empirical result is that individual group members who have stronger social connections to other group members are more likely to repay their loans and to save more. Karlan shows that this is due to the fact that these members are better able to monitor each other and to enforce each other’s repayment. He also shows that members with stronger connections are better able to distinguish between strategic default and default due to negative external shocks, as well to distinguish between who should and who should not be punished for her behaviour. The results of the article strongly support the view that monitoring and enforcement are positively related to group performance and that social connections are important in assisting monitoring and enforcement efforts within groups. Cassar et al. (2007) also focus on the importance of social ties (the authors use the term socialcapital) in explaining repayment performance of groups. Yet, they take an innovative approach to analyse this issue by using microfinance experiments. The main advantage of this approach vis‐à‐vis other approaches in the literature is that it permits the disentangling of different aspects of social capital within groups and their effects on group performance. The authors argue that repayment by individual members depends on their belief that other members will do the same, since this will determine whether or not credit will be available to them in the next loan cycle. This belief depends, at least partially, on the existence of social capital within the group. Social capital may consist of aspects such as general trust of individual group members in the society as a whole, specific trust of one individual towards one or more group members, acquaintanceship among group members, and trust based on (positive) experiences with other group members in the past related to repayment of loans. Cassar et al. (2007) use a microfinance game at two different locations: Nyanga, South Africa and Berd, Armenia. Their total sample consists of 36 microfinance groups, which include 498 individual group members. The results of their experiments provide clear evidence for the fact that different aspects of social capital have a different impact on group performance. Most importantly, they find that specific trust between group members is more important for group performance than trust in society as a whole. Moreover, social and cultural homogeneity of group members improves performance. They also find that past (positive) experience with other members helping an individual to repay her loan provides incentives to this individual to help others repaying their loans in the future. Finally, the fact that people merely know each other does not help to improve group performance. These results clearly indicate that it is really important to disentangle different aspects of social capital when explaining group repayment performance. 2. Financial Performance and Outreach A second important issue raised in the literature on microfinance deals with the sustainability of microfinance programmes. Providing microfinance is a costly business due to high transaction and information costs. At present, a large number of microfinance programmes still depend on donor subsidies to meet the high costs, i.e. they are not financially sustainable. In the 1990s, the importance of financial sustainability of microfinance institutions gave rise to an important debate between the financial systems approach and the poverty lending approach (Robinson, 2001). If both approaches agree on the ultimate goal, which is to serve as many poor people as possible in a sustainable way, the means by which these goals should be reached differ fundamentally. The financial systems approach, on the one hand, emphasises the importance of financial sustainable microfinance programmess. On the other hand, the poverty lending approach concentrates on using credit to help overcome poverty, primarily by providing credit with subsidised interest rates. Ultimately, the debate comes down to the question whether subsidising interest rates is justified. The advocates of the poverty reduction approach would argue that the poor cannot afford higher interest rates; hence that financial sustainability ultimately goes against the aim of serving large groups of poor borrowers. The financial services camp, however, claims that empirical evidence neither shows that the poor cannot afford higher interest rates nor that there is a negative correlation between the financial sustainability of the institution and the poverty level of the clients. The debate between the two approaches has not been concluded yet, although the most recent microfinance paradigm seems to favour the financial systems approach. The main argument to support this view is that large‐scale outreach to the poor on a long‐term basis cannot be guaranteed if microfinance institutions are incapable of standing on their own feet. Nonetheless, there remains a huge variety in microfinance institutions, some of which can be characterised as subsidised credit institutions, whereas others are becoming sustainable commercial financial institutions. This new microfinance paradigm has stimulated research on financial performance and financial efficiency of microfinance institutions. Hulme and Mosley (1996), for instance, provide alternative measures of financial performance of some microfinance institutions. By using the Subsidy Dependence Index (SDI) devised in Yaron (1992), indicating how much higher the interest rates charged to borrowers would have to be in order for the institution to cover all operating costs, Hulme and Mosley show that almost all institutes in their sample are still subsidy dependent. Morduch (1999a) provides a similar calculation for the Grameen Bank. He shows that, in order to become subsidy independent, the Grameen Bank would have needed to increase the lending rates by some 75% between 1985 and 1996. Calculations of the SDI to determine financial sustainability are useful. Yet, there are also some major drawbacks. First, the SDI assumes that a rise in lending rates automatically leads to higher profits. This, however, need not be the case since higher lending rates could lead to lower profits of banks in case of adverse selection and moral hazard effects. Cull et al. (2007) in this Feature explicitly deal with this possibility (see below). A more general problem with focusing on SDIs is that it puts too much emphasis on financial sustainability of microfinance institutions (Morduch, 1999a). SDIs do not indicate to what extent subsidies are justified. A more accurate assessment of the microfinance institutions would have to compare the costs and benefits of subsidies. Unfortunately, there are only a few studies that attempt to do this. Examples of such studies are Townsend and Yaron (2001) for the BAAC in Thailand, and Khandker (2005) for the Grameen Bank in Bangladesh. These studies, although based on some far reaching assumptions, suggest that the social benefits of these microfinance institutions exceed the costs. The greater emphasis on financial sustainability and the trend toward commercialisation of microfinance has raised concerns about the effects of this shift on outreach, or more specifically on the number (breadth) and socioeconomic level (depth) of the clients that are served by microfinance institutions. There is some discussion in the literature on the outreach of microfinance programmes. For an overall survey of recent evidence on this issue, see Goldberg (2005). Useful overviews are also given by Weiss and Montgomery (2004), who summarise the evidence for the microfinance industry in Asia and Latin America, and Lafourcade et al. (2005) who focus on microfinance institutions in Africa. This literature provides mixed evidence, especially regarding depth of outreach. Some studies indicate that it is the ‘better off’ poor rather than the ‘starkly’ poor who stand to benefit most. Evidence for this is given in e.g. Hulme and Mosley (1996) and Copestake et al. (2005). Other studies, e.g. Khandker (2005) and EDA Rural Systems (2004), find that the extremely poor benefit more from microfinance than the moderately poor. However, most of the evidence on the depth of outreach of microfinance institutions suffers from being anecdotal and case study driven. The existing studies do not systematically explain differences in depth of outreach of microfinance institutions, nor do they explicitly explore whether there is a trade‐off between the depth of outreach versus the strife for financial sustainability. The study by Cull et al. (2007) provides a new dimension to the existing literature on financial performance of microfinance institutions. This study attempts to examine financial performance and outreach systematically for the first time in a large comparative study based on a new extensive data set of 124 microfinance institutions in 49 countries. The authors explicitly explore whether there is empirical evidence for a trade‐off between the depth of outreach and profitability. They examine this issue by examining whether more profitability is associated with a lower depth of outreach to the poor, and whether there is a deliberate move away from serving poor clients to wealthier clients in order to achieve higher financial sustainability (mission drift). They also test whether a rise in lending rates causes a deterioration of the loan portfolio due to adverse selection and moral hazard. A special feature of the study by Cull et al. (2007) is that an explicit distinction has been made between three types of microfinance institutions, i.e. group lending systems, village banking, and individual‐based lending. Their dataset contains 56 individual‐based lenders, 48 group‐based lenders and 20 village banks. This enables them to examine the relevance of institutional design with respect to the trade‐off between financial performance and depth of outreach of microfinance institutions. The existing literature on microfinance focuses almost entirely on group lending, while hardly paying attention to other approaches to microfinance lending, e.g. individual‐based lending. In the light of the current move to individual‐based lending systems (even the most well‐known examples of group‐based lending, the Grameen Bank of Bangladesh and BancoSol of Bolivia now use individual‐based models) this is a bit surprising. There is a general descriptive discussion in the literature on the advantages of group loans over individual loans (Conning, 1999; Morduch, 1999b). Some authors prefer individual loans because they are assumed to be more flexible, whereas others are in favour of group loans. However, until now there has been no systematic and rigorous comparison of group‐based versus individual‐based microfinance institutions. Cull et al. (2007) are the first to provide such a systematic comparison. The results of the analyses are extremely interesting and highly policy relevant. Individual‐based microfinance institutions seem to perform better in terms of profitability, but the fraction of poor borrowers and female borrowers in the loan portfolio is lower than for group‐based institutions. The study also shows that a rise in interest rates, above a certain threshold, leads to a worsening of portfolio quality in case of individual‐based lending, whereas this relation does not exist for the group‐based microfinance institutions. This confirms the hypothesis that screening and monitoring by peers in group‐based systems helps to overcome problems of moral hazard and adverse selection. The study also suggests that individual‐based microfinance institutions, especially if they grow larger, focus increasingly on wealthier clients (mission drift), whereas this is less so for the group‐based microfinance institutions. Most importantly, the study strongly underlines the importance of institutional design in considering trade‐offs in microfinance. References Ahlin , C. and Townsend , R. ( 2007 ). ‘Using repayment data to test across models of joint liability lending’ , Economic Journal , vol. 117, pp. F11 – F51 . OpenURL Placeholder Text WorldCat Armendáriz De Aghion , B. ( 1999 ). ‘On the design of a credit agreement with peer monitoring’ , Journal of Development Economics , vol. 60 ( 1 ), pp. 79 – 104 . Google Scholar Crossref Search ADS WorldCat Banerjee , A. , Besley , T. and Guinnane , T. ( 1994 ). ‘Thy neighbor’s keeper: the design of a credit cooperative with theory and a test’ , Quarterly Journal of Economics , vol. 109 ( 2 ), pp. 491 – 515 . Google Scholar Crossref Search ADS WorldCat Besley , T. and Coate , S. ( 1995 ). ‘Group lending, repayment incentives and social collateral’ , Journal of Development Economics , vol. 46 ( 1 ), pp. 1 – 18 . Google Scholar Crossref Search ADS WorldCat Cassar , A. , Crowley , L. and Wydick , B. ( 2007 ). ‘The effect of social capital on group loan repayment: evidence from field experiments’ , Economic Journal , vol. 117, pp. F85 – F106 . OpenURL Placeholder Text WorldCat Chowdury , P.R. ( 2005 ). ‘Group lending: sequential financing, lender monitoring and joint liability’ , Journal of Development Economics , vol. 77 ( 2 ), pp. 415 – 39 . Google Scholar Crossref Search ADS WorldCat Conning , J. ( 1999 ). ‘Outreach, sustainability and leverage in monitored and peer‐monitored lending’ , Journal of Development Economics , vol. 60 ( 1 ), pp. 51 – 77 . Google Scholar Crossref Search ADS WorldCat Copestake , J. , Dawson , P., Fanning , J.P., McKay , A. and Wright‐Revolledo , K. ( 2005 ). ‘Monitoring the diversity of the poverty outreach and impact of microfinance: a comparison of methods using data from Peru’ , Development Policy Review , vol. 23 ( 6 ), pp. 703 – 23 . Google Scholar Crossref Search ADS WorldCat Cull , R. , Demirgüç‐Kunt , A. and Morduch , J. ( 2007 ). ‘Financial performance and outreach: a global analysis of leading microbanks’ , Economic Journal , vol. 117, pp. F107 – F133 . OpenURL Placeholder Text WorldCat Daley‐Harris , S. ( 2006 ). ‘State of the microcredit summit campaign report 2006’ , Washington, DC: Microcredit summit campaign . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC EDA Rural Systems ( 2004 ). The Maturing of Indian Microfinance: a Longitudinal Study , New Delhi: India, EDA Rural Systems . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Gangopadhyay , S. , Ghatak , M. and Lensink , R. ( 2005 ). ‘Joint liability lending and the peer selection effect’ , Economic Journal , vol. 115 ( 506 ), pp. 1005 – 15 . Google Scholar Crossref Search ADS WorldCat Ghatak , M. ( 1999 ). ‘Group lending, local information and peer selection’ , Journal of Development Economics , vol. 60 ( 1 ), pp. 27 – 50 . Google Scholar Crossref Search ADS WorldCat Ghatak , M. ( 2000 ). ‘Screening by the company you keep: joint liability lending and the peer selection effect’ , Economic Journal , vol. 110 ( 465 ), pp. 601 – 31 . Google Scholar Crossref Search ADS WorldCat Goldberg , N. ( 2005 ). ‘Measuring the impact of microfinance: taking stock of what we know’ , Grameen Foundation USA Weblink: http://www.grameenfoundation.org. OpenURL Placeholder Text WorldCat Hermes , N. , Lensink , R. and Mehrteab , H.T. ( 2005 ). ‘Peer monitoring, social ties and moral hazard in group lending programmes: evidence from Eritrea’ , World Development , vol. 33 ( 1 ), pp. 149 – 69 . Google Scholar Crossref Search ADS WorldCat Hermes , N. , Lensink , R. and Mehrteab , H.T. ( 2006 ). ‘Does the group leader matter: the impact of monitoring activities and social ties of group leaders on the repayment performance of group‐based lending in Eritrea’ , African Development Review , vol. 18 ( 1 ), pp. 72 – 97 . Google Scholar Crossref Search ADS WorldCat Hulme , D. and Mosley , P. ( 1996 ). Finance Against Poverty , London: Routledge . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Karlan , D. ( 2007 ). ‘Social connections and group banking’ , Economic Journal , vol. 117, pp. F52 – F84 . OpenURL Placeholder Text WorldCat Khandker , S.R. ( 2005 ). ‘Microfinance and poverty: evidence using panel data from Bangladesh’ , The World Bank Economic Review , vol. 19 ( 2 ), pp. 263 – 86 . Google Scholar Crossref Search ADS WorldCat Lafourcade , A.C. , Isern , J., Mwangi , P. and Brown , M. ( 2005 ). ‘Overview of the outreach and financial performance of microfinance institutions in Africa’ , Microfinance Information Exchange (MIX) April 2005. weblink: http://www.mixmarket.org. OpenURL Placeholder Text WorldCat Lapenu , C. and Zeller , M. ( 2001 ). ‘Distribution, growth and performance of microfinance institutions in Africa, Asia and Latin America’ , FCND discussion paper 114, Washington, DC: International Food Policy Research Institute (IFPRI) . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Morduch , J. ( 1999a ). ‘The role of subsidies in microfinance: evidence from the Grameen bank’ , Journal of Development Economics , vol. 60 ( 1 ), pp. 229 – 48 . Google Scholar Crossref Search ADS WorldCat Morduch , J. ( 1999b ). ‘The microfinance promise’ , Journal of Economic Literature , vol. 37 ( 4 ), pp. 1569 – 614 . Google Scholar Crossref Search ADS WorldCat Norwegian Nobel Committee ( 2006 ). ‘The Nobel Peace Prize 2006’ , Press release , Oslo, October. weblink: http://nobelprize.org/nobel_prizes/peace/laureates/2006/press.html. Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Paxton , J. , Graham , D. and Thraen , C. ( 2000 ), ‘Modeling group loan repayment behaviour: new insights from Burkina Faso’ , Economic Development & Cultural Change , vol. 48 ( 3 ), pp. 639 – 55 . Google Scholar Crossref Search ADS WorldCat Robinson , M. ( 2001 ). The Microfinance Revolution: Sustainable Banking for the Poor , Washington, DC: The World Bank . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Sharma , M. and Zeller , M. ( 1997 ). ‘Repayment performance in group‐based credit programs in Bangladesh: an empirical analysis’ , World Development , vol. 25 ( 10 ), pp. 1731 – 42 . Google Scholar Crossref Search ADS WorldCat Stiglitz , J. ( 1990 ). ‘Peer monitoring and credit markets’ , World Bank Economic Review , vol. 4 ( 3 ), pp. 351 – 66 . Google Scholar Crossref Search ADS WorldCat Townsend , R.M. and Yaron , J. ( 2001 ). ‘The credit risk‐contingency system of an Asian development bank’ , Federal Reserve Bank of Chicago Economic Perspectives , Third Quarter, pp. 31 – 48 . OpenURL Placeholder Text WorldCat Varian , H. ( 1990 ). ‘Monitoring agents with other agents’ , Journal of Institutional and Theoretical Economics , vol. 146 ( 2 ), pp. 153 – 74 . OpenURL Placeholder Text WorldCat Weiss , J. and Montgomery , J. ( 2004 ). ‘Great expectations: microfinance and poverty reduction in Asia and Latin America’ , ADB Institute discussion paper No. 14, The Asian Development Bank Institute . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Wenner , M. ( 1995 ). ‘Group credit: a means to improve information transfer and loan repayment performance’ , Journal of Development Studies , vol. 32 ( 2 ), pp. 263 – 81 . Google Scholar Crossref Search ADS WorldCat Wydick , B. ( 1999 ). ‘Can social cohesion be harnessed to repair market failures? Evidence from group‐based lending in Guatemala’ , Economic Journal , vol. 109 ( 457 ), pp. 463 – 75 . Google Scholar Crossref Search ADS WorldCat Wydick , B. ( 2001 ). ‘Group lending under dynamic incentives as a borrower discipline device’ , Review of Development Economics , vol. 5 ( 3 ), pp. 406 – 20 . Google Scholar Crossref Search ADS WorldCat Yaron , J. ( 1992 ). ‘Successful rural finance institutions’ , World Bank discussion paper No. 150, Washington, DC: The World Bank . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Zeller , M. ( 1998 ). ‘Determinants of repayment performance in credit groups: the role of program design, intra‐group risk pooling and social cohesion’ , Economic Development and Cultural Change , vol. 46 ( 1 ), pp. 599 – 620 . Google Scholar Crossref Search ADS WorldCat © The Author(s). Journal compilation © Royal Economic Society 2007

Easterly,, William

doi: 10.1111/j.1468-0297.2007.02031_2.xpmid: N/A

Review II Daron Acemoglu and James Robinson have made a huge contribution to the literature on democracy in their fascinating new book. To cut to the argument in brief, Acemoglu and Robinson see democracy emerging out of a strategic face‐off between the rich minority and the poor majority. The rich prefer not to have democracy because of the threat of redistribution. However, an even worse threat to the elite is total revolution by the poor, which will destroy the elite altogether. The poor can threaten revolution to try to extract democratic concessions from the rich. Often there is only a temporary revolutionary window of opportunity, such as during a war or a major economic crisis. (Although Acemoglu and Johnson at first seem to have in mind a traditional elite, they also make clear the rich minority could just as well be a recently created group of political insiders who feed off state revenues.) Why do the rich not just defuse the temporary crisis by promising some redistribution towards the poor, instead of agreeing to democracy? Or why do they not just repress the poor with military force? Acemoglu and Johnson show the first option does not work because the poor are not stupid – they know that the autocratic elite can reverse the redistributive policies after the revolutionary crisis passes. Only a permanent institutional change towards democracy assures the poor that they will remain in charge and permanently benefit from some redistribution. Democracy is a way to solve a commitment problem. Repression could work with a poor disorganised population but gets more and more costly (and less likely to succeed) as the majority gets more politically active. Under these circumstances, the elite agree to a transition to democracy. Acemoglu and Robinson cite the gradual movement towards universal suffrage in Britain in the nineteenth century as an example. They give an illustrative quote from Prime Minister Earl Grey in 1831: There is no‐one more decided against annual parliaments, universal suffrage and the ballot, than I am. My object is not to favour, but to put an end to such hopes and projects … The principle of my reform is, to prevent the necessity of revolution … reforming to preserve and not to overthrow. The authors give similar examples of democratic transitions amidst revolutionary threats by the populace in France, Germany, Sweden, Argentina, Colombia, and more recently in Africa (including especially South Africa) and Eastern Europe. The rich gave in more easily to democracy in Britain and America because the design of the new democratic system had some checks against the redistributive powers of the majority. A two‐chamber legislature had the upper chamber less under the sway of the majority. A system of winner‐take‐all elections for legislative representatives (as opposed to holding plebiscites on how much to tax the rich) made the more radical redistributionists unelectable. The rich may also find it reassuring that they can spend their money lobbying against redistribution. You need just the right amount of protection for the rich under democracy: too little and the elite will not want to agree to democracy, too much and the poor will go ahead and have the revolution anyway. A more recent example is the Chilean military oligarchy agreeing to democracy in 1990, conditional on giving the military enough remaining power to protect free market and private property reforms they had introduced during their bloody tenure from 1973 to 1990. The authors have an important comparative prediction for when this process will work itself out. It is less likely when the elite are landowners. Land is immobile and visible, and so much easier to tax than machinery or human skills. The rich thus have much more to lose from a democratic majority deciding on taxes. If the elite’s assets are mainly physical and human capital, they have more to lose from the disruptions of a revolution, while under the rule of a democratic majority, human and physical capital are more mobile than land and thus less taxable. Another important prediction of the book is that democracy is less likely with higher inequality between the elite and the majority. The poor majority will voter for a higher tax rate the higher inequality is (they have more to gain from redistribution if the gap between rich and poor is large, and they have less of a stake in future income to lose if tax rates penalise income growth). For the same reason, the poor are also likely to engage in attempted revolutions in non‐democracies with high inequality. On the elite side, the elite have much more to fear from democracy (and revolution) and so repression is more attractive to the elite with high inequality. Poor peasants are also much easier to repress with force than richer industrial workers. Thus oligarchy is more likely in unequal agrarian societies than in more equal industrial societies, with periodic attempts at revolution. Hence, in Latin America, we have seen successful violent revolutions such as Mexico early in the twentieth century, Bolivia (at least an incomplete revolution), Cuba, and Nicaragua and attempted revolutions such as in El Salvador, Guatemala and Colombia. The big successful Communist revolutions occurred in poor agrarian societies – Russia in 1917 or China in 1949 – not the industrialised societies as Marx had predicted. Democracy in unequal agrarian societies tends not to last, as it alternates between populist demagogues attempting redistribution and the rich striking back with military coups, as Latin America’s history abundantly demonstrates. Argentina is Acemoglu and Johnson’s classic example of an unstable democracy in a society that throughout its history has had a traditional land‐based elite. An important exception to the general inequality story is the persistence of autocracy in relatively equal countries in East Asia, such as South Korea and Taiwan (where democracy was late to arrive) and Singapore (still not a democracy). Acemoglu and Johnson argue that the majority has little incentive to engage in revolution when they are already sharing a lot of the wealth anyway, and so autocracy can be stable. Hence, they predict a nonlinear relationship between inequality and democracy, with democracy most likely at intermediate levels of inequality. Another twist is to examine the game when there are three players – the elite, the middle class and the poor. The middle class will demand less redistribution than the poor, hence a democracy in which the middle class is the pivotal player is less of a threat to the elite than a democracy with no middle class. The prediction is that a large middle class will make democracy more likely, and more likely to persist if it emerges. Acemoglu and Johnson illustrate the point with the classic example of middle class Costa Rica versus the other countries of Central America that lack a middle class, such as Guatemala and El Salvador. Acemoglu and Johnson have a rich story of the varieties of situations that lead to perpetual dictatorship. Another kind of elite desperate to stave off democracy at all costs is a minority ethnic group that manages to get into power. Acemoglu and Johnson’s classic example of this is apartheid South Africa. Other examples are Tutsis in Burundi (and more recently in Rwanda), Americo‐Liberians, the mulatto elite in Haiti and, perhaps, the European elite in some Latin American countries. Unfortunately for these elites, the grievances of inequality are even worse when combined with the toxic force of ethnic animosity, and so rebellion against the elite is likely (as seen in many of the above cases). Acemoglu and Johnson point out, however, that if ethnic cleavages cut different ways than class cleavages, they tend to weaken the effect of class cleavages on the democratic outcome. A pluralistic society with many ethnic groups thus could have democracy emerge more easily, which might explain the success of democracy in the US with its many immigrant groups (although African countries with many ethnic groups and little democracy seem like counter‐examples). Acemoglu and Johnson provide a great model for tying together some of the findings of empirical researchers on democracy. Cross‐country studies have indeed found democracy to be higher in more societies with a higher share of income going to the middle class, even addressing possible reverse causality from democracy to the size of the middle class; see Easterly (2001) for a discussion, and Figure 1 for the correlation. What determined different size middle classes in different countries? Is the relationship causal? Answering the first question helps answer the second. Many authors have pointed to natural resource endowments as a determinant of structural inequality. Economic historians Sokoloff and Engerman (2000) have highlighted the role of sugar plantations and silver mines in contributing to Latin America and the Caribbean’s high inequality. The plantations and mines had to be operated at a large scale, and wound up in the hands of a few, and the planters relied on slavery to work the sugar plantations – an extreme form of inequality. You could not grow sugar in North America – wheat was the crop of choice. Wheat could be produced on a small scale, hence a middle class formed made up of family farmers in the US and Canada. The relative proportions of wheat and sugar are a natural instrument for inequality, which has then been shown to have a negative causal effect on quality of institutions – including whether a society is democratic (Easterly, 2006). This confirms the general tenor of the Acemoglu‐Johnson prediction about inequality (although only linearly, I am not aware of any nonlinear tests). A natural extension to Acemoglu and Johnson’s story of a land‐based elite is an oil‐based elite. Oil is infamous for undermining or preventing democracy. Oil revenues are very easy to redistribute, so the Acemoglu‐Johnson model would suggest that wealthy and well‐connected insiders who benefit from oil controlled by a dictatorship have a lot to lose from a democracy that will surely result in redistribution. Hence we get oil societies desperate to prevent democracy, ranging from the oil‐rich Middle East to Africa. Jensen and Wantchekon (2004) documented systematically the association of resource wealth with autocracy in Africa, as others did using worldwide patterns (Ross, 2001; Collier and Hoeffler, 2005). Jensen and Wantchekon shows that new democracies have succeeded in Africa mainly in resource‐poor places like Benin, Mali, and Madagascar, while oil‐rich states like Algeria, Libya, Gabon and Cameroon still have dictators. Worldwide, oil producers are on average in the worst fourth of the world’s countries in democracy in 2004, as democracy was recently measured by Kaufmann et al. (2005). The negative effect of oil on democracy could be one of the main mechanisms for the ‘natural resource curse’, in which windfalls of natural resources, even though boosting income directly, have a negative effect on subsequent economic growth. Isham et al. (2005) find that ‘point‐source’ natural resources are more inimical to institutions than ‘diffuse’ natural resources, which may again get at Acemoglu and Johnson’s emphasis on just how taxable is the income of the elite if democracy triumphs. Another variation on the book’s story on land‐based elites could be a foreign aid‐based elite. Foreign aid can act a lot like natural resources in its effect on incentives for democracy through Acemoglu‐Johnson’s mechanisms. Indeed, Djankov et al. (2006) find that high aid caused setbacks to democracy over 1960–99. They found aid’s effect on democracy to be worse than the effect of oil on democracy. Although the predictions of the book about inequality and about the source of income for the elite generally conform to the consensus of the empirical literature, the central ‘fear of revolution’ mechanisms of the model are more difficult to test (they test it mainly with some suggestive case studies). The Engerman and Sokoloff (2002, 2005), Sokoloff and Engerman (2000) story is simpler (although not necessarily contradicting the A‐J story) for the relationship of elites and democracy – if initial inequality is too high (because of factor endowments), the elite will do things to perpetuate inequality such as restricting suffrage and restricting education, while the elite will accede to broad suffrage and mass literacy in an initially more equal society. (One important difference in predictions between this book and the rest of the literature is the A‐J non‐linear relationship between inequality and democracy, where relatively equal societies have benevolent autocracies.) An alternative line of research in the literature emphasises a human capital mechanism driving democracy. As the elite accumulates physical capital in industry, they have more demand for skilled labour and are more likely to support public education for the majority (Galor and Moav, 2006). Education is correlated with political participation and so, by giving in to mass education, the elite make mass democracy inevitable as well (Bourguignon and Verdier, 2000). Glaeser et al. (2006) argue that education drives democracy empirically, and provide a story for why this is so – that benefits of democracy are individually modest but widely distributed, while benefits of dictatorship are high for a restricted few. The positive effect of education on civic activity makes it possible to solve the coordination problem to attain, and then retain, democracy. Acemoglu et al. (2005) dispute this empirical correlation between education and democracy (in an article coming out after this book, but the result is mentioned in passing), pointing out it disappears once you control for country fixed effects. These debates will likely continue for a while. For Acemoglu and Johnson’s specific story, further work is needed to provide a more complete empirical test of their mechanism that the elite are motivated by fear of revolution in a strategic game with the majority Still, the story is plausible and appealing, has very rich theoretical predictions, and is largely borne out by reduced form stylised facts and case studies. Nobody working in political economy can afford not to read Economic Origins of Dictatorship and Democracy, which is one of the most important contributions to the literature on the economics of democracy in a very long time. References Acemoglu , Daron , Johnson , Simon, Robinson , James A. and Yared , Pierre ( 2005 ). ‘From education to democracy?’ , American Economic Review , vol. 95 ( 2 ), (May), pp. 44 – 9 . Google Scholar Crossref Search ADS WorldCat Bourguignon , Francois and Verdier , Thierry ( 2000 ). ‘Oligarchy, democracy, inequality and growth’ , Journal of Development Economics , vol. 62 ( 2 ), (August), pp. 285 – 313 . Google Scholar Crossref Search ADS WorldCat Collier , Paul and Hoeffler , Anke ( 2005 ), ‘Democracy and resource rents’ , Department of Economics , University of Oxford, April 26. Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Djankov , Simeon , Montalvo , Jose G. and Reynal‐Querol , Marta ( 2006 ). ‘The curse of aid’ , SSRN Working Paper, March. Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Easterly , William ( 2001 ). ‘The middle class consensus and economic development’ , Journal of Economic Growth , vol. 6 ( 4 ), (December), pp. 317 – 36 . Google Scholar Crossref Search ADS WorldCat Easterly , William ( 2006 ). ‘Inequality does cause underdevelopment: insights from a new instrument’ , Journal of Development Economics , forthcoming. OpenURL Placeholder Text WorldCat Engerman , Stanley L. and Sokoloff , Kenneth L. ( 2005 ). ‘The evolution of suffrage institutions in the new world’ , Journal of Economic History , vol. 65 ( 4 ), (December), pp. 891 – 921 . Google Scholar Crossref Search ADS WorldCat Engerman , Stanley L. and Sokoloff , Kenneth L. ( 2002 ). ‘Factor endowments, inequality, and paths of development among new world economies’ , Economia: Journal of the Latin American and Caribbean Economic Association , vol. 3 ( 1 ), (Fall), pp. 41 – 88 . OpenURL Placeholder Text WorldCat Galor , Oded and Moav , Omer ( 2006 ). ‘Das Human‐Kapital: a theory of the demise of the class structure’ , Review of Economic Studies , vol. 73 ( 1 ), (January), pp. 85 – 117 . Google Scholar Crossref Search ADS WorldCat Glaeser , Edward , Ponzetto , Giacomo and Shleifer , Andrei ( 2006 ). ‘Why does democracy need education?’ , NBER Working Paper No. 12128, April. Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Isham , Jonathan , Woolcock , Michael, Pritchett , Lant and Busby , Gwen ( 2005 ). ‘The varieties of resource experience: natural resource export structures and the political economy of economic growth’ , The World Bank Economic Review , vol. 19 , pp. 141 – 74 . Google Scholar Crossref Search ADS WorldCat Jensen , Nathan and Wantchekon , Leonard ( 2004 ). ‘Resource wealth and political regimes in Africa’ , Comparative Political Studies , vol. 37 ( 7 ), pp. 816 – 41 . Google Scholar Crossref Search ADS WorldCat Kaufmann , Daniel , Kraay , Aart and Mastruzzi , Massimo ( 2005 ). ‘Governance matters IV: governance indicators for 1996–2004’ , mimeo, World Bank, May. Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Ross , Michael ( 2001 ). ‘Does oil hinder democracy?’ , World Politics , vol. 53 (April), pp. 25 – 61 . Google Scholar Crossref Search ADS WorldCat Sokoloff , Kenneth L. and Engerman , Stanley L. ( 2000 ). ‘Institutions, factor endowments, and paths of development in the new world’ , Journal of Economic Perspectives , vol. 14 ( 3 ), pp. 217 – 32 . Google Scholar Crossref Search ADS WorldCat © The Author(s). Journal compilation © Royal Economic Society 2007

Ahlin,, Christian;Townsend, Robert, M.

doi: 10.1111/j.1468-0297.2007.02014.xpmid: N/A

Abstract Theories rationalising joint liability lending are rich in implications for repayment rates. We exploit this fact to test four diverse models. We show that the models’ repayment implications do not always coincide. For example, higher correlation of output and borrowers’ ability to act cooperatively can raise or lower repayment, depending on the model. Data from Thai borrowing groups suggest that repayment is affected negatively by the joint liability rate (ceteris paribus) and social ties, and positively by the strength of local sanctions and correlated returns. Further, the relative fit of the adverse selection versus informal sanctions models varies by region. A number of theoretical papers explore the key mechanism that gives group loans an advantage over individual loans. But these models take different stands on the underlying economic environment and the problem which groups are imagined to try to overcome. Specifically, we use as springboards to the data four widely cited papers representative of the literature. Two of these papers highlight moral hazard problems which joint liability lending and monitoring can mitigate: Stiglitz (l990) and Banerjee et al. (l994) hence forward BBG. One focuses on an environment of limited contract enforcement and the remedy of village sanctions: Besley and Coate (l995) hence forward BC. The fourth shows how adverse selection of borrowers can be partially overcome by joint liability contracts: Ghatak (l999).1 We take these models as emphasising the problems of moral hazard, limited commitment, and adverse selection typically thought of as obstacles to trade and the cause of limited (or non‐existent) lending by formal financial intermediaries. The empirical side of research in joint liability lending has lagged relatively far behind. There is little empirical work that views the data through the lens of theory.2 The point of our article is to attempt to bridge the gap between theoretical and empirical work. All four models are found to be rich in predictions regarding the determinants of the group repayment rate. We examine the predictions both of variables already included in the models as published, and especially of new variables we can introduce in a general way. See Table 1 of Section 1 for a summary of theoretical predictions. All predictions are derived holding all else constant; in particular, we discard the lender’s zero‐profit condition imposed in some of the models. The reason has to do with the lending institution in our data, discussed in more detail in Section 2.2. Table 1 Repayment Implications An entry marked with a ‘‘‡’’ corresponds to a variable not included in the original model. . Variable . Effect on Repayment . Stiglitz . BBG . BC . Ghatak . liability payment q ↓a ↑ ↓ positive correlation ↑‡b ↓‡b ↑‡b cooperative behavior ↑‡ ↓‡c ↓‡d cost of monitoring ↓ official penalties ↑ unofficial penalties ↑ screening ↑ productivity H ↑‡ ↑‡ ↑‡ ↑‡ interest rate r ↓ ↓ ↓ ↓ loan size L ↓ ↓‡ ↗↘‡ An entry marked with a ‘‘‡’’ corresponds to a variable not included in the original model. . Variable . Effect on Repayment . Stiglitz . BBG . BC . Ghatak . liability payment q ↓a ↑ ↓ positive correlation ↑‡b ↓‡b ↑‡b cooperative behavior ↑‡ ↓‡c ↓‡d cost of monitoring ↓ official penalties ↑ unofficial penalties ↑ screening ↑ productivity H ↑‡ ↑‡ ↑‡ ↑‡ interest rate r ↓ ↓ ↓ ↓ loan size L ↓ ↓‡ ↗↘‡ Under assumption A2, section 2.1.1. All correlation results rely on general, symmetric parametrizations of the correlation. If the marginal cost of penalizing is less than one. If unofficial penalties are larger than the loss to a borrower due to his partner’s default. Open in new tab Table 1 Repayment Implications An entry marked with a ‘‘‡’’ corresponds to a variable not included in the original model. . Variable . Effect on Repayment . Stiglitz . BBG . BC . Ghatak . liability payment q ↓a ↑ ↓ positive correlation ↑‡b ↓‡b ↑‡b cooperative behavior ↑‡ ↓‡c ↓‡d cost of monitoring ↓ official penalties ↑ unofficial penalties ↑ screening ↑ productivity H ↑‡ ↑‡ ↑‡ ↑‡ interest rate r ↓ ↓ ↓ ↓ loan size L ↓ ↓‡ ↗↘‡ An entry marked with a ‘‘‡’’ corresponds to a variable not included in the original model. . Variable . Effect on Repayment . Stiglitz . BBG . BC . Ghatak . liability payment q ↓a ↑ ↓ positive correlation ↑‡b ↓‡b ↑‡b cooperative behavior ↑‡ ↓‡c ↓‡d cost of monitoring ↓ official penalties ↑ unofficial penalties ↑ screening ↑ productivity H ↑‡ ↑‡ ↑‡ ↑‡ interest rate r ↓ ↓ ↓ ↓ loan size L ↓ ↓‡ ↗↘‡ Under assumption A2, section 2.1.1. All correlation results rely on general, symmetric parametrizations of the correlation. If the marginal cost of penalizing is less than one. If unofficial penalties are larger than the loss to a borrower due to his partner’s default. Open in new tab Some of the more interesting theoretical predictions overturn conventional wisdom and expose conflict between the models. Take cooperation, modelled in this article as the ability to commit costlessly to a set of actions that is Pareto optimal within the borrowing group. In one moral hazard model, this ability to act cooperatively leads to less risk‐taking by eliminating a borrower’s ability to free‐ride on his partner’s safe behaviour. Cooperation thus raises the repayment rate. However, in the other moral hazard model and in the limited enforcement model, both of which introduce informal penalties, cooperation can lower repayment by making possible binding agreements not to use excessively harsh penalties. Social capital can thus lower repayment and promote collusion. Consider next correlation between borrower returns, which none of the original models addresses. We find that under plausible assumptions higher correlation lowers repayment in the limited enforcement model. In that environment, default happens when both borrowers realize low returns; this event is more likely when returns are more correlated. On the other hand, two models illustrate that correlation can raise repayment, for different reasons. In one moral hazard model, higher correlation raises the safe project payoff more than the risky project payoff, because saving the joint liability fee (which correlation raises the odds of) is more valuable to risk‐averse borrowers when realised returns are moderate, as they are with the safe project. In the adverse selection model, higher correlation similarly raises borrowing payoffs by making execution of joint liability rarer; the effect is to draw in marginal borrowers, who are safer than the average, and raise the expected repayment rate. Correlation may thus have counter‐intuitive, positive effects on repayment. We also find that a higher degree of joint liability can lower repayment, ceteris paribus. One might expect a higher degree of joint liability to increase incentives for repayment‐promoting group behaviour, whether it be monitoring, penalising, or screening. This is indeed the result that emerges in one of the moral hazard models. But another moral hazard model shows that joint liability is itself an additional repayment burden and thus may skew incentives toward risky projects; and the adverse selection model shows that higher joint liability pushes marginal, safer borrowers out of the market and lowers the expected repayment rate. These results may seem to counter the original papers’ results on the efficiency of group lending; however, they hold with the interest rate fixed, while previous results incorporate joint liability with a simultaneous decline in the interest rate. Empirical results confirm some of these and other theoretical predictions. We find that cooperation, measured by the degree of intra‐group sharing, is negatively associated with repayment. A relatively direct measure of covariance of income shocks is positively associated with repayment. Joint liability within the group, measured by the percent of the group that is landless, is a negative predictor of repayment. We also find that the strength of social sanctions, measured by the likelihood of a village‐wide lending shutdown to a defaulter, positively predicts repayment. The theoretical and empirical results here should also serve to challenge, or refine, the notion that groups succeed because of their ability to access and make use of social collateral. They point to the fact that strong social cohesion may lead to weak incentives to repay, and lack of social cohesion may lead to excessive penalties that actually promote high repayment rates. Evidently, care must be taken in interpreting the interaction between borrower welfare and high repayment rates. Wydick (1999) and Ghatak and Guinnane (1999) make similar points about the ambiguous relationship between social ties and repayment, the latter in the context of several historical and contemporary examples; Rahman (1999) is also related. Here we formalise the theory and provide empirical tests. We do find that no model matches all the full‐sample empirical findings, though each contributes substantially to explaining some of them. Breaking the sample up by region causes some better overall fits to emerge. The limited enforcement model of Besley and Coate does best in the poorer, low‐infrastructure Northeast in its prediction that village penalties go with higher repayment. The screening model of Ghatak does best in the wealthier, central region in its predictions that the degree of joint liability payment will decrease repayment and that the loan size has a positive, then negative effect on repayment. Interestingly, follow‐up results indicate the regional divergence in results is due less to wealth per se than to whatever factors make commercial banks rarer, perhaps physical and legal infrastructure. The patterns we find suggest that strategic default may be a more prevalent problem in low‐infrastructure areas, while information problems (in particular adverse selection) may be more prevalent in more developed areas.3 Most of the theoretical literature focuses on conditions under which joint liability contracts are optimal relative to individual liability contracts, typically with an endogenous, market‐clearing interest rate. For example, in Stiglitz (1990) and in Ghatak (1999), an increase in the degree of joint liability allows profit‐maximising lenders to lower the interest rate and induce safer project choice or draw in safer borrowers. In a competitive equilibrium, lenders still make zero profits but borrower welfare is enhanced through joint liability. Direct tests involve relating the prevalence of joint liability versus individual liability contracts or other outside options against the covariates suggested by the models. In a companion paper, Ahlin and Townsend (2007), we use this direct approach to test predictions of Holmström and Milgrom (1990), the related Prescott and Townsend (2002), and Ghatak (1999). We find, consistent with the theory, that intra‐village wealth heterogeneity predicts group, joint liability contracts and; further, that there is a U‐shaped relationship of group borrowing with household wealth. We also find strong evidence for adverse selection, and distinguish this finding from one for moral hazard. In this article we adopt an indirect, but equally telling, approach. Specifically, we test the models’ implications for borrower repayment rates, given the joint liability contract is in use. We are thus not attempting to assess directly the effect of group lending on repayment.4 Instead, we are finding determinants of repayment in a group lending context and using this as evidence for or against group lending models. The probability of repayment plays a fundamental role in each model’s setup and results, and is therefore a useful key for unlocking and examining the mechanics of each of the models. We therefore turn our attention from contract choice per se to those internal mechanics. The exception in this article is our addition to the model of Ghatak, in which both loan size (borrow less than offered) and whether to enter into a group (borrow at all under joint liability) are two individual selections which determine how observed repayment should vary with observed loan amounts. 1. Theories and Implications We discuss next the specific setups of the four models, focusing on the mechanics and intution behind their repayment implications. That is, if p is the group probability of repayment and X is a key determinant of p, we attempt to sign ∂p/∂X is as general a way as possible. These are not, for the most part, the theoretical results of the published papers, which focus on comparing the efficiency of joint liability and individual liability; we are therefore in uncharted territory. All repayment results are essentially partial derivatives, i.e. are derived without imposing a zero‐profit condition on the bank, a decision we discuss in Section 2.2. For purposes of comparing the models and fully using our data to address key questions, we introduce new variables when this is possible in a relatively general way and with a minimum of assumptions. The implications of the various models, some of which we derive in this Section and the rest of which can be found in Ahlin and Townsend (2002), but all of which will be tested empirically, are summarised in Table 1. The models’ predictions agree along some dimensions and disagree along others. We do not claim that a prediction by a model here is general to all models in its genre (moral hazard, adverse selection, etc.), nor that our way of modelling a particular phenomenon is the only way (except when generality is claimed). Further, we do not claim these models are the best possible choices among all alternatives to fit the data. Still, we view the results as indicative about how leading representatives of each genre work. In the following analysis, we omit some details and adopt common notation wherever possible. All the models analyse groups of size two in a static setting of limited liability. Both members of a group face the same contract terms.5 Three of the models restrict attention to binomial output distributions, while the fourth allows for more general distributions. Omitted proof details can be found in Ahlin and Townsend (2002). 1.1. Moral Hazard: Stiglitz (1990) When liability is limited, borrowers may prefer project outcome distributions with greater probability weight in the tails. This is because the lower tail outcomes are effectively subsidised by the lender. In essence, limited liability raises the incentives to gamble. Stiglitz addresses this kind of moral hazard and shows how joint liability decreases group incentives to gamble by giving each borrower a stake in the success of his partner. Each borrower receives a loan L and chooses a risky or safe project, producing output Y(pR,L) with probability pR or Y(pS,L) with probability pS > pR, respectively. The complementary probabilities result in zero output. The safe project gives higher expected output, but lower output when successful:6 (A1) Limited liability can skew incentives toward the less efficient, risky project. Joint liability contracts take the following form. The lender gets nothing from a borrower who fails (due to limited liability), rL from a borrower who succeeds, and an additional payment qL from a borrower who succeeds and whose partner fails. Expected utility of a borrower who chooses technology i and whose partner chooses technology j, call it Vij, can be written, under independent returns and standard utility function U(·), (1) The first term represents the expected payoff from both borrowers succeeding, the second from borrower i succeeding and borrower j failing. The article assumes cooperative behaviour and restricts attention to symmetric choices. Thus a group makes whichever symmetric project choice gives each member higher utility: (2) where 1[·] represents the indicator function. In this context, the impact of some variable X on the repayment rate p is determined by whether a change in X tilts incentives more toward the safe or risky project. For example, if ∂VSS/∂X > ∂VRR/∂X, then the repayment rate p is increasing in X, in the sense that there may exist a cutoff value in the range of X values above (below) which safe (risky) projects are chosen. Stiglitz shows that in this sense the repayment rate is declining in r and L. An increase in r makes success more onerous and raises the implicit, relative subsidy to failure, encouraging gambling. An increase in L raises the risky payoff relative to the safe one, by a direct assumption Stiglitz makes. These results are represented graphically in Figure 1. The ‘Switch Line’ gives combinations of (r,L) that leave the borrowers indifferent between projects: VSS = VRR. Above it (higher r) or to the right (higher L), risky projects are preferred; below it, safe projects. Fig. 1. Open in new tabDownload slide The Switch Line. Note. To the left of the solid line, safe projects are chosen; to the right, risky ones. The dashed line is the Switch Line for groups acting non‐cooperatively, discussed in Section 1.1.2. Fig. 1. Open in new tabDownload slide The Switch Line. Note. To the left of the solid line, safe projects are chosen; to the right, risky ones. The dashed line is the Switch Line for groups acting non‐cooperatively, discussed in Section 1.1.2. 1.1.1. Checking q To examine the effect of q on repayment we use a substantive assumption: (A2) This ensures that asymmetric borrower outcomes, which involve a payment of the joint liability fee, occur more often with safe than with risky projects. Proposition 1.Under assumptions A1 and A2, the group repayment rate is lower for groups with higher q. Proof. We compare ∂VSS/∂q with ∂VRR/∂q. Higher joint liability lowers the payoff of either project (importantly, with r fixed). Which payoff it hurts more depends on two factors: (1) in which project is qL paid more often and (2) under which project is the utility loss of paying qL greater, conditional on paying it. Assumption A2 guarantees that the safe project involves paying qL more often. Assumption A1 guarantees that the payment costs more in utility terms with a safe project, since output is lower there (and utility is concave); in other words, after gambling and winning, it is less painful to make an extra payment. In short, the joint liability payment is paid more often and during times of lower income under the safe project choice, so an increase in q tilts incentives toward risky projects.7 Graphically, the result on q would be represented by shifting the Switch Line left as q increases. 1.1.2. Subtracting cooperation In a departure from the Stiglitz model, assume borrowers play non‐cooperatively: both choose safe projects only if neither can gain by deviation to a risky project. The condition for both to choose safe projects is VSS ≥ VRS, which is stronger than VSS ≥ VRR of the cooperative case, since (using (1)) (3) Proposition 2.The group repayment rate is higher for groups acting cooperatively. With non‐cooperative behaviour, the temptation to free‐ride on one’s partner’s safe behaviour sometimes derails the symmetric safe‐project equilibrium. Non‐cooperative groups choose risky projects more often. Graphically, the Switch Line for non‐cooperative groups is shifted down as compared to the cooperative Switch Line, as in Figure 1. 1.1.3. Adding correlation A final modification to the model introduces correlation in borrower output realisations.8 Let the joint probability distribution for the returns of a borrowing group choosing projects that succeed with probability pi and pj be . j Succeeds (pj) . j Fails (1 − pj) . i Succeeds (pi) pipj + ε pi(1 − pj) − ε i Fails (1 − pi) (1 − pi)pj − ε (1 − pi)(1 − pj) + ε . j Succeeds (pj) . j Fails (1 − pj) . i Succeeds (pi) pipj + ε pi(1 − pj) − ε i Fails (1 − pi) (1 − pi)pj − ε (1 − pi)(1 − pj) + ε . j Succeeds (pj) . j Fails (1 − pj) . i Succeeds (pi) pipj + ε pi(1 − pj) − ε i Fails (1 − pi) (1 − pi)pj − ε (1 − pi)(1 − pj) + ε . j Succeeds (pj) . j Fails (1 − pj) . i Succeeds (pi) pipj + ε pi(1 − pj) − ε i Fails (1 − pi) (1 − pi)pj − ε (1 − pi)(1 − pj) + ε A positive ε adds probability to the symmetric events (both‐succeed and both‐fail) and subtracts it from the asymmetric events; and vice versa for a negative ε. The zero‐correlation case assumed above has ε = 0, while ε > 0 (ε < 0) implies positive (negative) correlation. Any joint distribution that preserves pi and pj as the unconditional probabilities of success must take this form.9 However, for each (pi,pj) combination, ε could in theory be different. Restricting attention to symmetric project choices as before, there are two potentially different correlation structures: εR ≡ ε(pR,pR) and εS ≡ ε(pS,pS). These capture the degrees of covariance across safe and risky projects, respectively. We focus on two cases in which covariance structure is in some sense independent of project risk. The first variation on this assumption is (A3) The second variation ensures that both the safe and risky project distributions have the same correlation coefficient, equal to ρ, as straightforward calculation verifies: (A4) Under either assumption, the probability of success is as in (2), with payoffs equal to (4) Proposition 3.Under Assumptions A1 and A3, or under Assumptions A1, A2, and A4, the group repayment rate is higher for groups with higher project return correlation. Correlation shifts probability weight from the state in which a borrower is successful and his partner fails to the state in which both borrowers are successful. This shift is more valuable with the safe project, since output is lower and thus the utility gain from not paying qL is higher there. This somewhat surprising result contrasts with the common assumption of the empirical literature that positive correlation is bad for repayment. This model makes clear it can differentially raise payoffs of safe projects, lowering the temptation to gamble. 1.2. Moral Hazard: Banerjee et al. (1994) BBG focus on the same moral hazard problem as Stiglitz: the temptation of limited liability borrowers to gamble with riskier projects. The key differences are in agents’ risk neutrality, role asymmetry and in the introduction of monitoring backed by punishment capability.10 Combined with joint liability, threatened punishments can reduce incentives for risk‐taking. Groups are asymmetric: they consist of one member who borrows and a cosigner who monitors.11 The borrower receives one unit of capital and chooses a project indexed by p ∈ [p,1], where p > 0. As in Stiglitz, the project return is Y(p) with probability p and zero otherwise. The lender collects r from the borrower when his project is successful and q from the cosigner otherwise (loan size is normalised to one). The risk‐neutral borrower’s payoff (gross of any penalties) is thus p[Y(p) − r] = E(p) − pr, where E(p) ≡ pY(p) is expected output. By assumption, expected output is increasing in p, as in Stiglitz; p = 1 is thus socially optimal. Also assumed is that the interest rate and loan size are such that given limited liability, the borrower prefers riskier projects to safer ones, and left alone will gamble with p = p. Specifically: (A5) implying that expected output is decreasing in risk but the borrower’s expected payoff is increasing in risk. The monitor can penalise the borrower based on his project choice. A penalty that costs the borrower c to bear costs the monitor M(c) to impose, with M increasing and convex. The minimum penalty needed to enforce project p exactly outweighs the borrower’s gain from deviating to the riskiest project p = p. It is (5) The monitor then chooses p to maximise payoff which includes the joint liability fee q, paid with probability (1 − p), and the monitoring cost of implementing p. The first‐order condition is (6) We call this the Monitoring Equation. The benefit of more monitoring (left‐hand side) is saving the joint liability fee q more often. The cost of more monitoring (right‐hand side) is proportional to the marginal cost of penalising, M ′(c), and the size of the additional penalty needed to lower risk, cp(p,r), which captures the severity of the moral hazard problem. Costs and benefits are graphed against p in Figure 2. Fig. 2. Open in new tabDownload slide Determination of p Under Costly Monitoring (solid line) and Costless Cooperation (dashed line) Note. Discussed in Section 1.2.3. Fig. 2. Open in new tabDownload slide Determination of p Under Costly Monitoring (solid line) and Costless Cooperation (dashed line) Note. Discussed in Section 1.2.3. Repayment effects of some variable X, say, come through the effects of X on the costs and benefits of monitoring. They can be derived by totally differentiating the Monitoring Equation with respect to p and X to obtain ∂p/∂X. 1.2.1. Checking q An increase in joint liability q raises the benefit of monitoring without affecting the cost. The unambiguous result is more monitoring and less risk‐taking by the borrower. Proposition 4.Under Assumption A5, the group repayment rate p is higher for groups with higher joint liability payment q. Interestingly, the prediction for q is opposite that of the Stiglitz model. In BBG, greater liability creates incentives for more intense group pressure to perform well. Stiglitz incorporates not only this effect, but also the idea that the joint liability payment acts like an additional tax on success, since only a successful borrower pays it. 1.2.2. Adding L Here we let loan size L vary, as in Stiglitz. Assuming separability, we write (A6) Usual assumptions are made on F, including (weak) concavity. We let E(p) ≡ pY(p,1), so that expected output equals E(p)F(L). The amounts due from the borrower upon success and the monitor upon failure, respectively, are rL and qL. The analogue to assumption A5, ensuring a moral hazard problem, is (A7) Proposition 5.Under Assumptions A6 and A7, the group repayment rate is lower for groups with higher L. Proof: See Appendix A. A larger loan from the lender raises the cost of monitoring for two reasons. First, under diminishing returns it increases the interest payment (rL) relatively faster than the gross returns (F(L));12 since risky projects avoid repaying the loan more often, they become relatively more attractive to the borrower, and the monitor must threaten stiffer penalties. Second, even under constant returns, a larger loan scales up required penalties and therefore makes the marginal penalty more costly, since the cost of imposing penalties is convex. 1.2.3. Adding cooperation Assume the monitor and borrower can enforce any joint agreement on project choice costlessly, as in Stiglitz. They will thus maximise the sum of payoffs. The problem becomes to choose p in order to maximise the sum of net payoffs of the borrower and monitor. The first order condition can be written (7) Comparison with non‐cooperative Monitoring Equation (6) reveals as the critical condition whether M ′(c) is greater or less than one. If less, the non‐cooperative sloped line in Figure 2 is lower than the cooperative one, and the resulting p is lower under cooperation. Proposition 6.Under assumption A5, the group repayment rate is lower for groups that can cooperate and enforce side‐contracts if M ′(c) < 1 and higher if M ′(c) > 1. The cooperative setup is isomorphic to the non‐cooperative case where M(c) = c, where the monitor can apply penalties that affect the borrower’s payoff at a one‐for‐one cost to the monitor’s own payoff (or equivalently, pay a bonus contingent on project choice). If the marginal cost of penalising is less than (greater than) one, then the monitor enforces a safer project (riskier project) than in the group‐surplus maximising case, and non‐cooperation results in a higher (lower) repayment rate. This prediction (when M ′(c) < 1) is counter to that of the Stiglitz model, where cooperation enables the group to circumvent free‐riding. Here, non‐cooperative behaviour facilitates the monitor’s use of cheap penalties to enforce a higher probability of repayment than is optimal from the group’s perspective. The common idea that social capital leads to better‐behaving groups may thus be turned on its head. Lenders may even prefer groups with less social capital if this translates into less ability to collude. 1.3. Strategic Default: Besley and Coate (1995) In BC, project choice is fixed; the game begins after project outcomes have been realised as borrowers decide whether or not to repay. The cost of repayment is the gross interest rate r (loan size is normalised to one). The benefit of repayment is avoiding penalties imposed by the lender and, under joint liability, penalties imposed by the group or community. Joint liability is found to raise repayment rates if these informal sanctions are strong enough. The two borrowers’ returns are drawn independently from distribution F (Y), with support [0,Ymax]. Repayment decisions are then made non‐cooperatively.13 Joint liability here implies that if the lender does not receive the full repayment amount from the group, 2r, he imposes an official penalty on each borrower. The penalty on borrower i depends on borrower i’s output, so we write it as co(Yi). It is increasing in Yi but always less than Yi, by assumption. In other words, the lender penalises more severely when output is higher but never as severely as outright confiscation. Since penalties depend positively on output, borrowers who realise high returns (low returns) will choose to repay (default). One can define a cutoff function (8) By construction, when weighing repaying r against incurring penalties co(Y), repayment is more attractive if Y ≥ Y(r) and default is more attractive if Y < Y(r). Above Y(r), official penalties are greater than r and vice versa. It is now possible to classify most outcomes by whether the group will repay or not. First, if both borrowers realise returns Yi,Yj < Y(r), the group will default. Official penalties are not strong enough to give incentives for either borrower to pay r. This outcome corresponds to box A in Figure 3. Second, if both borrowers realise returns Y(r) ≤ Yi,Yj < Y(2r), the group will repay. Both borrowers prefer repaying r to incurring official penalties.14 This outcome corresponds to box B in Figure 3. Third, if either borrower realises return Y ≥ Y(2r), the group will repay. This is because the more successful borrower will bail out the group if he has to, since paying 2r is better than incurring official penalties when returns are this high. This corresponds to the unlabelled part of the unit square in Figure 3. Fig. 3. Open in new tabDownload slide The Default Region. Note. Here a ≡ Y(r), b ≡ Y(2r), and Ymax is normalised to one. Default in the non‐cooperative game occurs if joint output realisations fall in box A, or in boxes AB below the curve. The dashed (dotted) curve demarcates a Default Region with relatively weak (strong) unofficial penalties. The Default Region under cooperation (see Section 1.3.2) and linear official penalties is demarcated by the dash‐dotted line. Fig. 3. Open in new tabDownload slide The Default Region. Note. Here a ≡ Y(r), b ≡ Y(2r), and Ymax is normalised to one. Default in the non‐cooperative game occurs if joint output realisations fall in box A, or in boxes AB below the curve. The dashed (dotted) curve demarcates a Default Region with relatively weak (strong) unofficial penalties. The Default Region under cooperation (see Section 1.3.2) and linear official penalties is demarcated by the dash‐dotted line. The remaining case is the one in which there is disagreement, but neither borrower is willing to bail out the group: Yi < Y(r) and Y(r) ≤ Yj < Y(2r), say. This corresponds to boxes AB in Figure 3. Here, borrower i prefers to default while borrower j prefers to repay, his own share at least, but not pay for both. With no further assumption, the group defaults. However, BC introduce informal penalties that are imposed on a borrower i, say, who decides to default when his partner j would want to repay. The effect of the unofficial penalty is to increase the willingness to repay of the low‐output borrower in these situations of disagreement. If informal penalties are arbitrarily severe, nearly all of these situations result in group repayment, and vice versa if they are arbitrarily weak. By assumption, the unofficial penalty, cu(Yi, Λj), depends on two things. One is the delinquent borrower i’s ability to repay, Yi. The second is his partner j’s desire to repay, proportional to his gain from repayment relative to default, Λj ≡ co(Yj)−r. More generally, one can define a new cutoff output level above which repayment is optimal, accounting for both official and unofficial penalties. Call this cutoff for borrower i, say.15 It depends on the partner’s output Yj in the following way: the higher Yj, the stronger are the partner’s desire to repay and thus the higher unofficial penalties, so the lower is . In summary, default occurs in two circumstances only: when both borrowers’ realisations are below Y(r), and when one realises output Y(r) ≤ Yj < Y(2r) and the other . The repayment rate p is thus (10) In Figure 3, the set of joint output realisations leading to default, the Default Region, consists of box A and some parts of the AB boxes. In particular, there is a curve running through the AB boxes and the point (a,a), below which repayment does not happen (the dashed curve here). This curve represents , and is thus lower when unofficial penalties are stronger. In this context, the effect of any variable on repayment will come either as it changes the boundaries of the Default Region, via a change in costs or benefits of repayment, or as it changes the probability of falling into the Default Region. 1.3.1. Checking official and unofficial penalties Stiffer penalties raise the cost of default and do not affect the cost of repayment. Interestingly, stiffer official penalties also can raise unofficial penalties, since they raise the non‐delinquent borrower’s desire to repay. Graphically, stronger penalties shrink the Default Region in Figure 3 via a lowering of a, b, and the curves. Proposition 7.The repayment rate is higher for groups with stronger official or unofficial penalties. 1.3.2. Adding cooperation Assume borrowers can costlessly enforce agreements among themselves. Since utility is transferable, they will maximise the sum of payoffs and repay if and only if the sum of official penalties is greater than the group’s total debt: (11) An indifference curve in joint output space can be defined, below which the group defaults. Note that every point of indifference occurs in exactly the situations of disagreement discussed above (the AB boxes in Figure 3), where one borrower realises low output, Y < Y(r), and the other moderate output, Y(r) ≤ Y < Y(2r). If both realise low output (box A), penalties are clearly too low to encourage repayment; if both realise moderate output (box B) or one realises high output (the unlabelled region), penalties are sufficient. The repayment condition (11) at equality, using the definition of Y(·) as (co)−1, is Yi = Y[2r − co(Yj)]. Thus the indifference curve is decreasing and goes through (0,Y(2r)), (Y(r),Y(r)), and (Y(2r),0) (respectively, (0,b), (a,a), and (b,0) in Figure 3). An example (under linear official penalties) is the dash‐dotted line in Figure 3. Below this line is the cooperative Default Region. The cooperative repayment rate can be written (12) Comparison with (10) reveals that the severity of unofficial penalties, which cooperation renders unused, determines the effect on the repayment rate. If unofficial penalties are severe, the cutoff is low and the non‐cooperative repayment rate is higher; and vice versa. Graphically, non‐cooperative Default Regions under weak and strong unofficial penalties are demarcated, respectively, by the dashed and dotted curves in Figure 3. Proposition 8.The repayment rate is lower (higher) for groups acting cooperatively if unofficial penalties are greater than (less than) the non‐defaulting borrower’s loss from default. 16 The cooperative setup is isomorphic to the non‐cooperative case where cu(Y, Λ) = Λ, under which the unofficial punishment exactly fits the ‘crime’, i.e. the cost to the partner of default. If unofficial penalties are more severe, there are output realisations where official penalties on the group would be less than 2r, yet the low‐output partner repays to avoid the unofficial penalties. If acting cooperatively, the low‐output borrower could instead compensate his partner directly for his loss Λ, leaving some surplus to be split as the group defaults. As with BBG and for similar reasons, the common idea that social capital leads to better‐behaving groups may thus be turned on its head. 1.3.3 Adding correlation Here we introduce correlation between project returns. Unlike the previous two exercises, this leaves the Default Region unchanged but alters the probability of falling into it. No perfectly general results are available; it is clearly possible to increase overall correlation while lowering the chance of falling into the Default Region, and the reverse. Our aim is to analyse a symmetric, parametric yet general correlation structure, as in the earlier Section 1.1.3. The basic idea in our parameterisation is to raise the probability of similar outcomes (Yi near Yj) and lower the probability of dissimilar outcomes (Yi far from Yj). Accordingly, we add or subtract probability mass relative to (a transformation of) an arbitrary monotonic polynomial function of the absolute difference |Yi −Yj|. For details, see the discussion leading up to Assumption A9 in the Appendix as well as Ahlin and Townsend (2002). Proposition 9.Under Appendix Assumption A9, and if unofficial penalties are severe enough, the repayment rate is lower for groups with higher covariance of returns. Proof: see Appendix 4. As discussed above, default occurs in two types of situations. If both returns are low, i.e. less than Y(r) (box A in Figure 3), borrowers unanimously default. Higher correlation raises the probability of this event since borrower outcomes are relatively similar (both low). Default also occurs when one borrower wants to repay but only for himself, Y(r) ≤ Yj < Y(2r) say, and the other borrower wants to default despite official and unofficial penalties, (boxes AB below the Default Region boundary). This event is arbitrarily rare as unofficial penalties get stronger. Thus, while higher correlation may raise or lower the probability of this event,17 the effect is arbitrarily small relative to correlation’s effect on the prevalence of unanimous default. Graphically, strong unofficial penalties eliminate arbitrarily much of the AB boxes from the Default Region, guaranteeing that the probability mass added or subtracted within box A is the dominant effect of correlation.18 1.4. Adverse Selection: Ghatak (1999) Here agents’ project types are fixed and they repay whenever possible; their only decisions are whether and with whom to borrow. Due to the same kind of limited liability as in Stiglitz and BBG, borrowing is more attractive to agents with riskier projects. Safe borrowers may thus be excluded from the market. Joint liability, Ghatak shows, can be used to take advantage of information borrowers have about each other’s types to draw into the market borrowers who would otherwise be excluded.19 Agents weigh the outside, non‐borrowing option payoff u > 0 against that of undertaking their endowed project using capital borrowed in a joint liability group. Agents differ in the riskiness of their endowed projects; there is a density g (p) > 0 of borrowers at each project type p ∈ [p,1]. Project type is observable among borrowers but not to the outside lender. As in Stiglitz and BBG, an agent carrying out a project of type p realises output Y(p) with probability p and zero otherwise. As in Stiglitz, a borrower pays gross interest rate r if he succeeds and an additional joint liability payment q if he succeeds and his partner fails (loan size is normalised to one). Thus, a borrower of type p who pairs with one of type p ′ has expected payoff (13) This incorporates risk neutrality but is otherwise identical to Stiglitz (1). The first decision centres on group formation: who pairs with whom? Ghatak shows that groups form homogeneously in risk‐type, p. While everyone would prefer a safer partner, safe borrowers prefer them more strongly, since they succeed more often and thus are in the position of being potentially liable for their partner more often. Thus p ′ = p in payoff 6. The second decision is whether or not to borrow. In contrast to Stiglitz and BBG, benefits of borrowing, pY(p), are assumed not to vary with project risk: (A8) Thus all borrowers have equally worthwhile projects in an expected value sense, but differ only in second‐order risk (higher p implies lower risk, e.g. variance). However, costs of borrowing, i.e. expected repayment pr + p(1 − p)q, do vary: they are higher for safer borrowers (for q≤r), since payments to the bank are made only upon success. Thus, assuming that not everyone borrows, only borrowers riskier than some cutoff risk‐type will borrow, while safer borrowers will take the outside option. This marginal type, , solves the following Selection Equation: (14) which sets the borrowing payoff of the marginal type equal to the outside option. Only agents with borrow. Expected repayment burdens are too high for safer borrowers. Unlike the others, this model does not produce a probability of repayment p as a function of key variables (r, q, and so on). Rather, it delivers a range for the probability of repayment, , where is a function of key variables through the Selection Equation. Observing r, q, and so on, our best guess for the repayment rate is then ; is thus the analog to the p’s of the other models. Since varies positively and monotonically with (), we can focus on rather than . Thus the Selection Equation is the key to understanding repayment determinants. In general, any change that makes borrowing more attractive draws in more borrowers; and since a larger borrowing pool is a less risky one (the marginal borrowers are always safer than the average borrowers), this raises the expected repayment rate. 1.4.1 Checking q A higher joint liability payment makes borrowing relatively less attractive. Thus the higher a group’s q, the smaller and more risky the pool from which it is drawn. Proposition 10.Under Assumption A8, the expected group repayment rate is lower for groups with higher q. Proof. By total differentiation of the Selection Equation, for q ≤ r. 1.4.2. Subtracting screening ability What happens when homogeneous matching is replaced with random matching? In particular, assume that borrowers do not know each other’s types, but only the distribution of borrowing types. Matching is random and each borrower expects to match with a partner of average risk within the borrowing pool. Compared with homogeneous matching, safe borrowers are worse off and risky borrowers better off. Since safe borrowers are the marginal ones, they are driven out of the market and the residual borrowing pool is more risky. Specifically, the expected repayment rate is , where is the new cutoff risk‐type, defined as the value20 for p satisfying a modified Selection Equation (15) The left‐hand side involves the expectation over all potential partners, which is just when the marginal type is . For (), the payoff is strictly larger (smaller) than u. Thus is the equilibrium cutoff risk‐type. Proposition 11.Under Assumption A8, the expected group repayment rate is higher for groups with the ability to screen. Proof. Equation (14), (15), and the fact that , respectively, give that (16) Since E − pr − p(1 − p)q is decreasing in p, inequality 16 implies that . 1.4.3 Adding loan size Here we take a simple version of loan size determination: the lender makes loan offers that are random across groups but equal for both borrowing partners within a group. Borrowers then choose to borrow any amount up to the lender’s offer, or take their outside option. Assumptions A6 and A8 give the borrower payoff, under loan size L and homogeneous matching, as (17) We assume some properties of F in the Appendix, including strict concavity and Inada conditions. The effect of a change in loan size on expected repayment is non‐monotonic. First, assume loan sizes are small, such that all borrowers would prefer larger loans. Observing a higher loan size means borrowing is more attractive relative to the outside option. Hence the pool of borrowers from which the group was drawn is larger and safer, and the expected repayment rate is higher. Second, assume loan sizes are large. In this context and under diminishing returns to capital, observing a larger loan implies that (the upper bound on) the cost of capital of the borrower must be lower. This is because the marginal product of a larger loan is smaller, yet still must be above the borrower’s cost of capital since the borrower did not revise his loan size downward. Since cost of capital declines with risk in this limited liability setting, the borrower is drawn from a riskier pool and the expected repayment rate is lower. Proposition 12.Under Assumptions A6, A8, and A10 (in the Appendix), there exists ansuch that for (), the expected group repayment rate is higher (coarsely lower) for groups with higher L. Details and proof are in Appendix A and Ahlin and Townsend (2002). In empirical tests we will allow for a non‐monotonic relationship between expected repayment and loan size. 1.4.4. Adding correlation Here we use the same parameterisations of correlation as in our modification of the Stiglitz model; see Section 1.1.3 for details.21 Given that homogeneous matching still obtains,22 the Selection Equation becomes (18) Proposition 13.Under Assumptions A8 and either A3 or A4, the expected group repayment rate is higher for groups with higher project return correlation (higher ε or ρ). Higher correlation implies that if a borrower is successful, his partner is more likely to be. Hence, the borrower’s chances of having to make the joint liability payment are lower. Higher‐correlation groups thus have higher borrowing payoffs, and are drawn from a larger and safer pool. 2. Empirical Results In this Section we discuss our results from data on Thai joint liability borrowing groups and the villages where they are located. We discuss our methodology in Section 2.1. Data, sampling, and variables are described in Section 2.2, with Appendix B and Table 2 giving greater detail on construction of the variables. In Section 2.3 and Table 3, we report the results and discuss how they fit the theoretical predictions. Robustness checks and empirical concerns are discussed in Section 2.4. Table 2 Variables Variables marked with an asterisk are taken or constructed from the household‐level survey, HH. All others are from the group‐level survey, BAAC. . Variable . Description . Mean . (σ) . Dependent Variable:  Repayment Outcome Did BAAC ever raise interest rate as penalty for late payment? 0.27 (0.44) Joint liability:  Degree of joint liability q Percent landless in the group 0.06 (0.15) Covariance:  COVARIABILITY* Measure of coincidence of economically bad years across villagers 0.28 (0.16)  HOMOGENEOUS_ OCCUPATIONSa Measure of occupational homogeneity within the group 0.87 (0.24) Cooperation:  SHARING_RELATIVES Measure of sharing among closely related group members 2.1 (1.6)  SHARING_NON‐RELATIVES Measure of sharing among unrelated group members 1.5 (1.4)  BEST_COOPERATION* Percent in tambon naming this village best in the tambon for “cooperation among villagers’’ 0.25 (0.11)  JOINT_DECISIONS Number of decisions made collectively 0.37 (0.91) Cost of monitoring:  IN_VILLAGE Percent of group living in the same village 0.88 (0.22)  RELATEDNESSb Percent of group members having a close relative in the group 0.58 (0.36) Screening:  SCREEN Do some want to join this group but cannot? 0.39 (0.49)  KNOW_TYPE Do group members know the quality of each other’s work? 0.94 (0.24) Penalties for default:  BEST_INSTITUTIONS* Percent in tambon naming this village best in the tambon for ‘‘availability and quality of institutions’’ 0.27 (0.19)  SANCTIONS* Percent of village loans where default is punishable by informal sanctions 0.10 (0.11) Productivity:  AVERAGE_LAND Average landholdings of group members (rai) 23.6 (15.7)  AVERAGE_EDUCATION Index of group average education levels 3.1c (0.32) Contract terms:  Interest rate r Average interest rate faced by the group 10.9 (2.0)  Loan size L Average loan size borrowed by the group (thousand 97 Thai baht) 18.7 (18.3) Control:  LN(GROUP_AGE) Number of years group has existed (Log) 11.4d (8.5)  GROUP_SIZE Number of members in the group 12.3 (5.1)  VILLAGE_RISK* Village average coefficient of variation for next year’s expected income 0.30 (0.09)  VILLAGE_WEALTH* Village average wealth (million 97 Thai baht) 1.1 (2.1)  PCG_MEMBERSHIP*e Percent in village claiming Production Credit Group membership 0.08 (0.16)  BANK_MEMBERSHIP* Percent in village claiming to be clients of a commercial bank 0.28 (0.18) Variables marked with an asterisk are taken or constructed from the household‐level survey, HH. All others are from the group‐level survey, BAAC. . Variable . Description . Mean . (σ) . Dependent Variable:  Repayment Outcome Did BAAC ever raise interest rate as penalty for late payment? 0.27 (0.44) Joint liability:  Degree of joint liability q Percent landless in the group 0.06 (0.15) Covariance:  COVARIABILITY* Measure of coincidence of economically bad years across villagers 0.28 (0.16)  HOMOGENEOUS_ OCCUPATIONSa Measure of occupational homogeneity within the group 0.87 (0.24) Cooperation:  SHARING_RELATIVES Measure of sharing among closely related group members 2.1 (1.6)  SHARING_NON‐RELATIVES Measure of sharing among unrelated group members 1.5 (1.4)  BEST_COOPERATION* Percent in tambon naming this village best in the tambon for “cooperation among villagers’’ 0.25 (0.11)  JOINT_DECISIONS Number of decisions made collectively 0.37 (0.91) Cost of monitoring:  IN_VILLAGE Percent of group living in the same village 0.88 (0.22)  RELATEDNESSb Percent of group members having a close relative in the group 0.58 (0.36) Screening:  SCREEN Do some want to join this group but cannot? 0.39 (0.49)  KNOW_TYPE Do group members know the quality of each other’s work? 0.94 (0.24) Penalties for default:  BEST_INSTITUTIONS* Percent in tambon naming this village best in the tambon for ‘‘availability and quality of institutions’’ 0.27 (0.19)  SANCTIONS* Percent of village loans where default is punishable by informal sanctions 0.10 (0.11) Productivity:  AVERAGE_LAND Average landholdings of group members (rai) 23.6 (15.7)  AVERAGE_EDUCATION Index of group average education levels 3.1c (0.32) Contract terms:  Interest rate r Average interest rate faced by the group 10.9 (2.0)  Loan size L Average loan size borrowed by the group (thousand 97 Thai baht) 18.7 (18.3) Control:  LN(GROUP_AGE) Number of years group has existed (Log) 11.4d (8.5)  GROUP_SIZE Number of members in the group 12.3 (5.1)  VILLAGE_RISK* Village average coefficient of variation for next year’s expected income 0.30 (0.09)  VILLAGE_WEALTH* Village average wealth (million 97 Thai baht) 1.1 (2.1)  PCG_MEMBERSHIP*e Percent in village claiming Production Credit Group membership 0.08 (0.16)  BANK_MEMBERSHIP* Percent in village claiming to be clients of a commercial bank 0.28 (0.18) a Could also measure cost of monitoring. b Could also measure cooperation. c See text for the interpretation of this education index. d Here the mean and standard deviation are for age, not log of age. e Could also measure cooperation. Open in new tab Table 2 Variables Variables marked with an asterisk are taken or constructed from the household‐level survey, HH. All others are from the group‐level survey, BAAC. . Variable . Description . Mean . (σ) . Dependent Variable:  Repayment Outcome Did BAAC ever raise interest rate as penalty for late payment? 0.27 (0.44) Joint liability:  Degree of joint liability q Percent landless in the group 0.06 (0.15) Covariance:  COVARIABILITY* Measure of coincidence of economically bad years across villagers 0.28 (0.16)  HOMOGENEOUS_ OCCUPATIONSa Measure of occupational homogeneity within the group 0.87 (0.24) Cooperation:  SHARING_RELATIVES Measure of sharing among closely related group members 2.1 (1.6)  SHARING_NON‐RELATIVES Measure of sharing among unrelated group members 1.5 (1.4)  BEST_COOPERATION* Percent in tambon naming this village best in the tambon for “cooperation among villagers’’ 0.25 (0.11)  JOINT_DECISIONS Number of decisions made collectively 0.37 (0.91) Cost of monitoring:  IN_VILLAGE Percent of group living in the same village 0.88 (0.22)  RELATEDNESSb Percent of group members having a close relative in the group 0.58 (0.36) Screening:  SCREEN Do some want to join this group but cannot? 0.39 (0.49)  KNOW_TYPE Do group members know the quality of each other’s work? 0.94 (0.24) Penalties for default:  BEST_INSTITUTIONS* Percent in tambon naming this village best in the tambon for ‘‘availability and quality of institutions’’ 0.27 (0.19)  SANCTIONS* Percent of village loans where default is punishable by informal sanctions 0.10 (0.11) Productivity:  AVERAGE_LAND Average landholdings of group members (rai) 23.6 (15.7)  AVERAGE_EDUCATION Index of group average education levels 3.1c (0.32) Contract terms:  Interest rate r Average interest rate faced by the group 10.9 (2.0)  Loan size L Average loan size borrowed by the group (thousand 97 Thai baht) 18.7 (18.3) Control:  LN(GROUP_AGE) Number of years group has existed (Log) 11.4d (8.5)  GROUP_SIZE Number of members in the group 12.3 (5.1)  VILLAGE_RISK* Village average coefficient of variation for next year’s expected income 0.30 (0.09)  VILLAGE_WEALTH* Village average wealth (million 97 Thai baht) 1.1 (2.1)  PCG_MEMBERSHIP*e Percent in village claiming Production Credit Group membership 0.08 (0.16)  BANK_MEMBERSHIP* Percent in village claiming to be clients of a commercial bank 0.28 (0.18) Variables marked with an asterisk are taken or constructed from the household‐level survey, HH. All others are from the group‐level survey, BAAC. . Variable . Description . Mean . (σ) . Dependent Variable:  Repayment Outcome Did BAAC ever raise interest rate as penalty for late payment? 0.27 (0.44) Joint liability:  Degree of joint liability q Percent landless in the group 0.06 (0.15) Covariance:  COVARIABILITY* Measure of coincidence of economically bad years across villagers 0.28 (0.16)  HOMOGENEOUS_ OCCUPATIONSa Measure of occupational homogeneity within the group 0.87 (0.24) Cooperation:  SHARING_RELATIVES Measure of sharing among closely related group members 2.1 (1.6)  SHARING_NON‐RELATIVES Measure of sharing among unrelated group members 1.5 (1.4)  BEST_COOPERATION* Percent in tambon naming this village best in the tambon for “cooperation among villagers’’ 0.25 (0.11)  JOINT_DECISIONS Number of decisions made collectively 0.37 (0.91) Cost of monitoring:  IN_VILLAGE Percent of group living in the same village 0.88 (0.22)  RELATEDNESSb Percent of group members having a close relative in the group 0.58 (0.36) Screening:  SCREEN Do some want to join this group but cannot? 0.39 (0.49)  KNOW_TYPE Do group members know the quality of each other’s work? 0.94 (0.24) Penalties for default:  BEST_INSTITUTIONS* Percent in tambon naming this village best in the tambon for ‘‘availability and quality of institutions’’ 0.27 (0.19)  SANCTIONS* Percent of village loans where default is punishable by informal sanctions 0.10 (0.11) Productivity:  AVERAGE_LAND Average landholdings of group members (rai) 23.6 (15.7)  AVERAGE_EDUCATION Index of group average education levels 3.1c (0.32) Contract terms:  Interest rate r Average interest rate faced by the group 10.9 (2.0)  Loan size L Average loan size borrowed by the group (thousand 97 Thai baht) 18.7 (18.3) Control:  LN(GROUP_AGE) Number of years group has existed (Log) 11.4d (8.5)  GROUP_SIZE Number of members in the group 12.3 (5.1)  VILLAGE_RISK* Village average coefficient of variation for next year’s expected income 0.30 (0.09)  VILLAGE_WEALTH* Village average wealth (million 97 Thai baht) 1.1 (2.1)  PCG_MEMBERSHIP*e Percent in village claiming Production Credit Group membership 0.08 (0.16)  BANK_MEMBERSHIP* Percent in village claiming to be clients of a commercial bank 0.28 (0.18) a Could also measure cost of monitoring. b Could also measure cooperation. c See text for the interpretation of this education index. d Here the mean and standard deviation are for age, not log of age. e Could also measure cooperation. Open in new tab Table 3 Logit Results Dependent Variable = 1 if BAAC has never raised the interest rate as a penalty, 0 if it has. . Standard errors in parentheses; significance at 5, 10 and 15% denoted by ***, **, and *, respectively. . . All groups . All groups, w/village effects . Northeast . Central . N = 219 . N = 219 . N = 130 . N = 89 . Joint liability:  Degree of joint liability q −3.65 (1.51)*** −16.7 (7.22)*** −4.69 (6.76) −7.69 (3.31)*** Covariance:  COVARIABILITY 2.05 (1.39)* 33.9 (12.5)*** 1.66 (2.13) 1.82 (4.03)  HOMOGENEOUS_OCCUPATIONSa 0.220 (.858) −0.065 (2.82) 1.39 (1.45) 0.061 (1.69) Cooperation:  SHARING_RELATIVES 0.382 (.250)* 0.433 (.599) 0.491 (.417) 0.375 (.487)  SHARING_NON−RELATIVES −0.553 (.266)*** −1.75 (.762)*** −0.497 (.410) −0.586 (.558)  BEST_COOPERATION −2.30 (2.40) 3.05 (8.41) −6.65 (3.53)** −5.12 (5.97)  JOINT_DECISIONS 0.499 (.265)** 0.327 (.829) 0.317 (.358) 1.58 (.765)*** Cost of Monitoring:  IN_VILLAGE 0.879 (.831) 8.41 (3.85)*** −0.694 (1.84) 1.01 (1.26)  RELATEDNESSb −0.590 (.573) −6.84 (3.15)*** −1.29 (.925) −0.574 (1.21) Screening:  SCREEN −0.364 (.402) −1.68 (1.42) −0.950 (.624)* 1.17 (.876)  KNOW_TYPE −0.139 (.773) −4.71 (3.12)* −1.38 (1.23) 2.24 (2.12) Penalties for default:  BEST_INSTITUTIONS 2.10 (1.36)* 16.7 (8.19)*** 3.42 (1.91)** 5.24 (3.81)  SANCTIONS 3.18 (1.95)* 24.9 (13.1)** 12.1 (4.34)*** −1.04 (3.55) Productivity:  AVERAGE_LAND −0.006 (.013) 0.0002 (.037) −0.013 (.026) −0.007 (.023)  AVERAGE_EDUCATION 1.28 (.698)** 5.43 (2.13)*** 1.88 (.935)*** 0.949 (1.25) Contract terms:  Interest rate r −0.119 (.101) −0.799 (.421)** −0.056 (.142) −0.385 (.299)  Loan size L 32.9 (30.8) 278.8 (165.2)** 20.4 (48.3) 187.4 (99.8)**  Loan size squared L2 −0.463 (.335) −4.60 (2.45)** −0.319 (.409) −2.39 (1.42)** Control:  LN(GROUP_AGE) −0.958 (.282)*** −1.64 (.886)** −1.54 (.488)*** −1.61 (.701)***  GROUP_SIZE 0.034 (.047) 0.145 (.148) −0.014 (.089) 0.113 (.079)  VILLAGE_RISK −3.47 (2.66) −57.1 (22.1)*** 1.41 (4.40) −10.1 (5.82)**  VILLAGE_WEALTH 0.026 (.083) −0.66 (.43)* 1.00 (1.29) 0.125 (.117)  PCG_MEMBERSHIPc −3.81 (1.18)*** −29.1 (10.7)*** −6.56 (1.98)*** −2.98 (3.42)  BANK_MEMBERSHIP 0.288 (1.21) −23.8 (11.0)*** −2.07 (2.39) 0.206 (2.34) Dependent Variable = 1 if BAAC has never raised the interest rate as a penalty, 0 if it has. . Standard errors in parentheses; significance at 5, 10 and 15% denoted by ***, **, and *, respectively. . . All groups . All groups, w/village effects . Northeast . Central . N = 219 . N = 219 . N = 130 . N = 89 . Joint liability:  Degree of joint liability q −3.65 (1.51)*** −16.7 (7.22)*** −4.69 (6.76) −7.69 (3.31)*** Covariance:  COVARIABILITY 2.05 (1.39)* 33.9 (12.5)*** 1.66 (2.13) 1.82 (4.03)  HOMOGENEOUS_OCCUPATIONSa 0.220 (.858) −0.065 (2.82) 1.39 (1.45) 0.061 (1.69) Cooperation:  SHARING_RELATIVES 0.382 (.250)* 0.433 (.599) 0.491 (.417) 0.375 (.487)  SHARING_NON−RELATIVES −0.553 (.266)*** −1.75 (.762)*** −0.497 (.410) −0.586 (.558)  BEST_COOPERATION −2.30 (2.40) 3.05 (8.41) −6.65 (3.53)** −5.12 (5.97)  JOINT_DECISIONS 0.499 (.265)** 0.327 (.829) 0.317 (.358) 1.58 (.765)*** Cost of Monitoring:  IN_VILLAGE 0.879 (.831) 8.41 (3.85)*** −0.694 (1.84) 1.01 (1.26)  RELATEDNESSb −0.590 (.573) −6.84 (3.15)*** −1.29 (.925) −0.574 (1.21) Screening:  SCREEN −0.364 (.402) −1.68 (1.42) −0.950 (.624)* 1.17 (.876)  KNOW_TYPE −0.139 (.773) −4.71 (3.12)* −1.38 (1.23) 2.24 (2.12) Penalties for default:  BEST_INSTITUTIONS 2.10 (1.36)* 16.7 (8.19)*** 3.42 (1.91)** 5.24 (3.81)  SANCTIONS 3.18 (1.95)* 24.9 (13.1)** 12.1 (4.34)*** −1.04 (3.55) Productivity:  AVERAGE_LAND −0.006 (.013) 0.0002 (.037) −0.013 (.026) −0.007 (.023)  AVERAGE_EDUCATION 1.28 (.698)** 5.43 (2.13)*** 1.88 (.935)*** 0.949 (1.25) Contract terms:  Interest rate r −0.119 (.101) −0.799 (.421)** −0.056 (.142) −0.385 (.299)  Loan size L 32.9 (30.8) 278.8 (165.2)** 20.4 (48.3) 187.4 (99.8)**  Loan size squared L2 −0.463 (.335) −4.60 (2.45)** −0.319 (.409) −2.39 (1.42)** Control:  LN(GROUP_AGE) −0.958 (.282)*** −1.64 (.886)** −1.54 (.488)*** −1.61 (.701)***  GROUP_SIZE 0.034 (.047) 0.145 (.148) −0.014 (.089) 0.113 (.079)  VILLAGE_RISK −3.47 (2.66) −57.1 (22.1)*** 1.41 (4.40) −10.1 (5.82)**  VILLAGE_WEALTH 0.026 (.083) −0.66 (.43)* 1.00 (1.29) 0.125 (.117)  PCG_MEMBERSHIPc −3.81 (1.18)*** −29.1 (10.7)*** −6.56 (1.98)*** −2.98 (3.42)  BANK_MEMBERSHIP 0.288 (1.21) −23.8 (11.0)*** −2.07 (2.39) 0.206 (2.34) a Could also measure cost of monitoring. b Could also measure cooperation. c Could also measure cooperation. Open in new tab Table 3 Logit Results Dependent Variable = 1 if BAAC has never raised the interest rate as a penalty, 0 if it has. . Standard errors in parentheses; significance at 5, 10 and 15% denoted by ***, **, and *, respectively. . . All groups . All groups, w/village effects . Northeast . Central . N = 219 . N = 219 . N = 130 . N = 89 . Joint liability:  Degree of joint liability q −3.65 (1.51)*** −16.7 (7.22)*** −4.69 (6.76) −7.69 (3.31)*** Covariance:  COVARIABILITY 2.05 (1.39)* 33.9 (12.5)*** 1.66 (2.13) 1.82 (4.03)  HOMOGENEOUS_OCCUPATIONSa 0.220 (.858) −0.065 (2.82) 1.39 (1.45) 0.061 (1.69) Cooperation:  SHARING_RELATIVES 0.382 (.250)* 0.433 (.599) 0.491 (.417) 0.375 (.487)  SHARING_NON−RELATIVES −0.553 (.266)*** −1.75 (.762)*** −0.497 (.410) −0.586 (.558)  BEST_COOPERATION −2.30 (2.40) 3.05 (8.41) −6.65 (3.53)** −5.12 (5.97)  JOINT_DECISIONS 0.499 (.265)** 0.327 (.829) 0.317 (.358) 1.58 (.765)*** Cost of Monitoring:  IN_VILLAGE 0.879 (.831) 8.41 (3.85)*** −0.694 (1.84) 1.01 (1.26)  RELATEDNESSb −0.590 (.573) −6.84 (3.15)*** −1.29 (.925) −0.574 (1.21) Screening:  SCREEN −0.364 (.402) −1.68 (1.42) −0.950 (.624)* 1.17 (.876)  KNOW_TYPE −0.139 (.773) −4.71 (3.12)* −1.38 (1.23) 2.24 (2.12) Penalties for default:  BEST_INSTITUTIONS 2.10 (1.36)* 16.7 (8.19)*** 3.42 (1.91)** 5.24 (3.81)  SANCTIONS 3.18 (1.95)* 24.9 (13.1)** 12.1 (4.34)*** −1.04 (3.55) Productivity:  AVERAGE_LAND −0.006 (.013) 0.0002 (.037) −0.013 (.026) −0.007 (.023)  AVERAGE_EDUCATION 1.28 (.698)** 5.43 (2.13)*** 1.88 (.935)*** 0.949 (1.25) Contract terms:  Interest rate r −0.119 (.101) −0.799 (.421)** −0.056 (.142) −0.385 (.299)  Loan size L 32.9 (30.8) 278.8 (165.2)** 20.4 (48.3) 187.4 (99.8)**  Loan size squared L2 −0.463 (.335) −4.60 (2.45)** −0.319 (.409) −2.39 (1.42)** Control:  LN(GROUP_AGE) −0.958 (.282)*** −1.64 (.886)** −1.54 (.488)*** −1.61 (.701)***  GROUP_SIZE 0.034 (.047) 0.145 (.148) −0.014 (.089) 0.113 (.079)  VILLAGE_RISK −3.47 (2.66) −57.1 (22.1)*** 1.41 (4.40) −10.1 (5.82)**  VILLAGE_WEALTH 0.026 (.083) −0.66 (.43)* 1.00 (1.29) 0.125 (.117)  PCG_MEMBERSHIPc −3.81 (1.18)*** −29.1 (10.7)*** −6.56 (1.98)*** −2.98 (3.42)  BANK_MEMBERSHIP 0.288 (1.21) −23.8 (11.0)*** −2.07 (2.39) 0.206 (2.34) Dependent Variable = 1 if BAAC has never raised the interest rate as a penalty, 0 if it has. . Standard errors in parentheses; significance at 5, 10 and 15% denoted by ***, **, and *, respectively. . . All groups . All groups, w/village effects . Northeast . Central . N = 219 . N = 219 . N = 130 . N = 89 . Joint liability:  Degree of joint liability q −3.65 (1.51)*** −16.7 (7.22)*** −4.69 (6.76) −7.69 (3.31)*** Covariance:  COVARIABILITY 2.05 (1.39)* 33.9 (12.5)*** 1.66 (2.13) 1.82 (4.03)  HOMOGENEOUS_OCCUPATIONSa 0.220 (.858) −0.065 (2.82) 1.39 (1.45) 0.061 (1.69) Cooperation:  SHARING_RELATIVES 0.382 (.250)* 0.433 (.599) 0.491 (.417) 0.375 (.487)  SHARING_NON−RELATIVES −0.553 (.266)*** −1.75 (.762)*** −0.497 (.410) −0.586 (.558)  BEST_COOPERATION −2.30 (2.40) 3.05 (8.41) −6.65 (3.53)** −5.12 (5.97)  JOINT_DECISIONS 0.499 (.265)** 0.327 (.829) 0.317 (.358) 1.58 (.765)*** Cost of Monitoring:  IN_VILLAGE 0.879 (.831) 8.41 (3.85)*** −0.694 (1.84) 1.01 (1.26)  RELATEDNESSb −0.590 (.573) −6.84 (3.15)*** −1.29 (.925) −0.574 (1.21) Screening:  SCREEN −0.364 (.402) −1.68 (1.42) −0.950 (.624)* 1.17 (.876)  KNOW_TYPE −0.139 (.773) −4.71 (3.12)* −1.38 (1.23) 2.24 (2.12) Penalties for default:  BEST_INSTITUTIONS 2.10 (1.36)* 16.7 (8.19)*** 3.42 (1.91)** 5.24 (3.81)  SANCTIONS 3.18 (1.95)* 24.9 (13.1)** 12.1 (4.34)*** −1.04 (3.55) Productivity:  AVERAGE_LAND −0.006 (.013) 0.0002 (.037) −0.013 (.026) −0.007 (.023)  AVERAGE_EDUCATION 1.28 (.698)** 5.43 (2.13)*** 1.88 (.935)*** 0.949 (1.25) Contract terms:  Interest rate r −0.119 (.101) −0.799 (.421)** −0.056 (.142) −0.385 (.299)  Loan size L 32.9 (30.8) 278.8 (165.2)** 20.4 (48.3) 187.4 (99.8)**  Loan size squared L2 −0.463 (.335) −4.60 (2.45)** −0.319 (.409) −2.39 (1.42)** Control:  LN(GROUP_AGE) −0.958 (.282)*** −1.64 (.886)** −1.54 (.488)*** −1.61 (.701)***  GROUP_SIZE 0.034 (.047) 0.145 (.148) −0.014 (.089) 0.113 (.079)  VILLAGE_RISK −3.47 (2.66) −57.1 (22.1)*** 1.41 (4.40) −10.1 (5.82)**  VILLAGE_WEALTH 0.026 (.083) −0.66 (.43)* 1.00 (1.29) 0.125 (.117)  PCG_MEMBERSHIPc −3.81 (1.18)*** −29.1 (10.7)*** −6.56 (1.98)*** −2.98 (3.42)  BANK_MEMBERSHIP 0.288 (1.21) −23.8 (11.0)*** −2.07 (2.39) 0.206 (2.34) a Could also measure cost of monitoring. b Could also measure cooperation. c Could also measure cooperation. Open in new tab 2.1. Methodology Our empirical goal given cross sectional data on group repayment R (a binary variable) and characteristics X = (X1, …, XM) is to see how the frequency of repayment R varies across groups with different characteristics X. Estimating the partial derivatives in this way will be the analogue to the analytical partial derivatives and comparative statics of Table 1 and Section 1. One of our more interesting if challenging goals is to try to distinguish the models. To clarify, it seems useful to adopt the definitions of the econometric literature: Two models f and h are said to be completely nested if for each parameter θ under model f, for all R and X, with repayment Pf(R|X,θ), we can find in model h a parameter ϕ with an equivalent repayment: Ph(R |X,ϕ)=Pf(R |X,θ). That is, the two models could not be distinguished in the data because we could always rationalise the results from one, the relationship among observables (R, X), by a configuration of the other. Two models are partially overlapping if this happens some (but not all) of the time. On the other hand, if the derivative under f were ∂Pf(R |X,θ)/∂Xm > 0 for all X and θ and ∂Ph(R |X,ϕ)/∂Xm < 0 for all X and ϕ, then the two models must be completely non‐nested because the derivatives have opposite signs over the entire (relevant) range of (R,X). Thus in principle the models can be distinguished by the sign of the derivative in the data, subject to statistical tests. Inspection of Table 1 reveals that the models’ predictions about some variables are unanimous; other variables are uniquely featured by just one of the models; and the models give conflicting predictions about a third set of variables. While all predictions will be tested, it is the second and especially the third category of variable that allow for distinguishing the models. If we were to parameterise the models, for example, specifying θ and ϕ in the above discussion, then we could proceed by maximum likelihood methods, comparing as in Vuong (1989) across non‐nested models by examining (adjusted) likelihood ratios. In that way the sign restrictions inherent to each model would be loaded automatically into the probability and in a sense forced onto the data. But our goal is to be explicit about the consistency or inconsistency of a model with the data by looking more deeply into the determination of the likelihoods, at the signs of the derivatives. Moreover we seek to do this in a relatively non‐parametric way, for example specifying that agents are risk averse in Stiglitz without pinning down the exact degree of curvature of the utility function, i.e. the parameter θ. On the other hand, though it is possible to determine the shape of the entire probability of repayment surface P(R |X) in each of the theories, it is not possible with data, especially with limited sample size, to reliably plot the non‐parametric version of the corresponding multi‐dimensional histograms. Thus we focus on first derivatives and make some simplifying approximations. We note in particular that the cross partial derivatives ∂2P(R |X)/∂Xm∂Xn are determined in each of the models, and typically many of these cross partial terms are not zero. But we do not have enough data to estimate these.23 Indeed it is difficult to estimate the direct partial, ∂P(R |X)/∂Xm without making some simplifying, approximation assumptions about how the Xn, n ≠ m, enter. The picture grows more complicated if agents can select into ‘models’ based on characteristics X and unobservables correlated with error terms influencing repayment rates. In this case, varying a particular Xm would not trace out the function P of one model, but segments of the P functions from several models. Most likely, the predominant model in the data would then determine the sign. In theory, however, cases could arise in which the sign reflects not any one model’s partial, but the effect of switching between models. One potential way to solve this issue would be to embed selection into the models where it is not already endogenous (i.e. all but Ghatak), and look for identifying restrictions that ensure one model or another is or is not in force. At this point identifying restrictions seem hard to come by. Our main approach to testing the repayment predictions is the most structural. It involves making two simplifying assumptions on the models themselves. First we assume that for each model, P(R g = 1|X g) can be written as a function P(β′X g), where β is an M × 1 vector of parameters and Xg is an M × 1 vector containing group g’s values for the M covariates, g = 1, …, G. This restricts covariates to enter repayment probabilities as a linear combination while leaving the function P unrestricted. This is the single‐index model, studied by Ichimura (1993) among others, and potentially computationally complex to estimate. Our second assumption is that P = Λ, that is the probability function is logistic. This is the logit model, easily estimated by maximum likelihood, as in (19): (19) This approach forces all models into the same structure of the function P(R | X), but allows the data to determine the signs of the coefficients. We also used two bivariate, non‐parametric approaches. One simply tested for mean repayment differences across high and low values of each Xm. For robustness, we varied the cutoff value defining ‘high’ and ‘low’ in a systematic way. The second used locally linear non‐parametric regressions (Cleveland, 1979; Fan, 1992). These regressions calculated an expected repayment rate at each value of the covariate Xm using only the 80% of the sample closest to , in a weighted least squares regression with the tri‐cube weighting function (Cleveland, 1979). Standard errors were obtained from bootstrap techniques. The two univariate approaches gave results consistent with the multivariate logits in many cases, but not all. To sort these cases out, our final approach involved the same locally linear regressions, but with linear multivariate controls. That is, we assume a partially linear model – where all regressors but one affect R linearly, and the remaining regressor’s effect can take any smooth shape. We estimate this model using Yatchew’s (1998) differencing method for estimating and removing the linear regressors’ effects, then using the local linear regression to plot the residual relationship. Almost without exception, these tests confirm the results of the multivariate, logit specification. Hence, for brevity we focus almost exclusively on results from the logits in this article; the one exception will be to examine the effect of loan size in a partially linear model.24 2.2. Data The data used in this article are from the Townsend Thai data base, in particular from a large cross section of l92 villages, conducted in May l997. The survey covers two contrasting regions of Thailand. The central region is relatively close to Bangkok and enjoys a degree of industrialisation, as in the province of Chachoengsao, and also fertile land for farming, as in the province of Lopburi. The Northeast region is poorer and semi‐arid, with the province of Srisaket regarded as one of the poorest in the entire country and the province of Buriram offering a transition as one moves back west toward Bangkok. Within each province, twelve subcounties, or tambons, were chosen. Within each tambon, a cluster of four villages was selected, and within each village fifteen households were administered a Household instrument. There are thus 2,875 households in the household data base. We call this instrument the HH survey. Of key importance for the paper here, in each village as many borrowing groups of the Bank for Agriculture and Agricultural Cooperatives (BAAC) as possible were interviewed, up to two. In all we have data on 262 groups, 62 of which are the only groups in their respective village. We call this instrument the BAAC survey. Each group designates an official leader, and the leader responded to questions on behalf of the group. The BAAC is a government‐operated development bank in Thailand. It was established in l966 and is the primary formal financial institution serving rural households. By its own estimates, it serves 4.88 million farm families, in a country with just over sixty million inhabitants, about eighty percent of which live in rural areas. In the data here, BAAC loans constitute 34.3% of the total number of loans, but we include in this total loans and reciprocal gift giving from friends, relatives, and moneylenders (Kaboski and Townsend, 1998). Indeed, commercial banks in the sample here have only 3.4% of total loans, and provide loans to only about 6% of the household sample. Occasionally a village will have established a local financial institution but typically these are small and constitute on average only 12.8% of total loans. Informal loans, though 39.4% of the total, are also small in size. The BAAC requires some kind of collateral for all loans but it allows smaller loans to be backed with social collateral in the form of joint liability. Thus loans underwritten by a BAAC group do not in principle require land or other physical collateral, only the promise that individual members be jointly liable. Loans larger than 50,000 baht must be backed by an asset such as land. Any particular loan is classified as a group‐guaranteed or individual loan, and the appropriate collateral box checked off on the loan form. The nature of BAAC lending justifies our decision not to impose a zero‐profit constraint in our theoretical work. For one, the BAAC receives a non‐trivial government subsidy. Its subsidy dependency index, the amount that would be necessary to raise the average on‐lending rate in order to break even, has been estimated at 35% (Townsend and Yaron, 2001). Under its charter the BAAC is responsible for the well‐being of farmers and those in rural areas, and it carries out that responsibility by charging a lower interest rate to small clients. Even a subsidised bank may tailor interest rates to borrowers. The BAAC appears to do so only very broadly. There is an exogenous, pre‐specified, unified national schedule mapping loan size into interest rates. For example in l997, at the time the data used here were collected, all loans under 60,000 baht carried a 9% interest rate, while loans between 60,000 baht and 1,000,000 baht charged 12.25% interest rates. Thus, except for the highest loan amounts and any exceptions to the policy, we should see virtually no variation. Observed variation may be due in part to measurement error, as respondents do not distinguish clearly between the part of repayment which is principal and the part which is interest. We are thus dealing with a bank that does not attempt to break even by adjusting interest rates based on risk or other group and location specifics. We turn briefly to descriptions of variables, which are described in greater detail in Appendix B and summarised (including statistically) in Table 2. Our measure of default is a binary dummy from the BAAC survey, which equals zero if the BAAC has ever, in the history of the group, raised the interest rate as a penalty for late payment, and one otherwise. Twenty seven percent of the groups responded affirmatively. This relatively high figure should not be taken as a mark against the BAAC lending programme. Annual default rates are much lower, whereas this measures default over the entire history of the group (median group age is ten years). Further, imposing an interest rate penalty is one of the first remedial actions in a dynamic process the BAAC uses with delinquent group‐guaranteed borrowers; repayment ultimately may have occurred.25 We include several control variables that are not featured in any of the four models: log‐age of group; size of group; a measure of village‐wide risk; a measure of village‐average household wealth; and two measures of village‐wide non‐BAAC credit options, measuring village‐wide prevalence of commercial bank membership and production credit group (PCG) membership, respectively. The age of group is particularly important to control since our measure of default applies to default at any time during the history of the group. Others of these control variables could fruitfully be added explicitly to the theory, but we do not do so here. Our proxy for the degree of joint liability q is the fraction of the group that is landless. This has validity because, in case of default, the BAAC has the option of taking legal action to seize assets, often land, of a borrower or his guarantors. The more borrowers are landless, the more likely guarantors will end up liable. While the BAAC rarely takes legal action, there have been such cases and certainly instances of group members being pressured to repay for a delinquent borrower in their group. All told, the threat seems to carry some credibility. Of course, if used alone, this variable might capture group wealth more closely than joint liability. However, we control for group average land‐holdings, so the partial effect of the landlessness variable is to capture the lopsidedness of group landholdings conditional on the mean. Average group landholdings, along with average education in the group, are included as productivity shifters. Two BAAC survey dummy variables proxy screening, one reflecting whether group members know the quality of each other’s work (a key assumption in the Ghatak model), the other reflecting whether there are households who would like to join but are screened out of a given group. Cost of monitoring is measured by the percentage of the group living in the same village and the percentage of group members who have a close relative in the group, both from the BAAC survey. However, BBG inextricably ties monitoring to imposing penalties, making the degree of relatedness a mixed signal. It can also be thought of as a measure for cooperation. Cooperation is further captured by two measures of sharing and cooperation within groups, among related and unrelated group members, respectively. These measures are based on questions about whether there has been sharing of free labour or coordination to procure inputs, for example, within the group in the past six months. We also use a village‐wide measure based on a poll of nearby villages in the HH survey that ranks villages based on the amount of cooperation. Finally, we use an index of joint decision‐making within the group regarding production. We proxy covariance by the degree of occupational homogeneity within the group and by a village‐level measure that captures the degree of agreement in the village about which year of the past five was worst for income. Official penalties are proxied by a poll of nearby villages that ranks villages based on availability and quality of institutions. This captures to some degree the legal infrastructure, which is related to the official penalties the BAAC can impose on borrowers. Unofficial penalties are reflected in a village‐wide measure reflecting the frequency of village‐wide denial of credit to, or loss of reputation of, a borrower who defaults on a loan. This captures very directly a form of unofficial penalties – widespread exclusion from future credit transactions. Finally, data on groups’ loan sizes and interest rates come from the BAAC survey which asks about maximum and minimum loan sizes and interest rates within the group; in each case we take a weighted average of the two. As noted above, however, variation in the interest rate should be rare due to a standardised national policy. We discuss endogeneity issues, which may be particularly acute with r and L, in Section 2.4. 2.3. Results Logit results on all variables simultaneously are listed in Table 3. There are 219 groups included in the regression incorporating both regions; 43 are excluded for missing data. Of these 130 observations are in the northeast region, and 89 in the central region. To focus on the within‐village variation, we include a specification with village dummies, of course only for the villages with two groups represented in the data (of which there are 75). In analysing the results, we focus primarily on whole‐sample results, since they contain the most data. Of the control variables, the log‐age of group exhibits a consistent significantly negative correlation with repayment. This is almost certainly because the dependent variable involves default at any time in the history of the group, which is more likely for older groups. There is some evidence for village income variability predicting lower repayment, and a slight amount for village wealth. Outside credit options, particularly the informal village‐based production credit groups, also are associated with lower repayment to the BAAC. There are three kinds of variables that shed light on the theory derived in this article (see Table 1). The first type of variable is the focal variable in exactly one of the models. This includes screening in the Ghatak model, cost of monitoring in BBG, and penalties in BC. BC is confirmed along this dimension, with official and unofficial penalties being good predictors of repayment, especially in the northeast sample. We seem to be the first to document the effect of unofficial penalties on repayment; to our knowledge, other research that examined informal sanctions found little effect (Wydick, 1999). Our measure for unofficial penalties – the exclusion of a delinquent borrower from future credit in his village – also appears to be unique. Some evidence in favour of BBG is present with percentage of group living in village positively predicting repayment in one specification.26 The percentage of members with a relative in the group is negatively associated with repayment in one specification. This contradicts the cost of monitoring prediction of the BBG model. However, that model equates monitoring with the ability to impose penalties; it may be that imposing penalties is harder among relatives. The evidence for Ghatak is not strong along this dimension of our data, if anything slightly negative (though outside conventional significance levels). However, we do not place significant weight on this result since our proxies are dummy variables, one with very little variation (KNOW_TYPE), and may not fully capture the phenomenon of screening. A second kind of comparison involves variables about which the models disagree –L, q, covariance, and cooperation – and leads to the possibility of rejecting one model in favour of another. The evidence on q and L favours Ghatak, especially in the central region, at the expense of Stiglitz and BBG. Joint liability q, as proxied by the landless fraction of the group, robustly predicts lower repayment. (Note that this result holds controlling for group average land.) This is consistent with the Stiglitz story of higher q raising the repayment burden and encouraging gambling and with the Ghatak story of higher q driving out marginal, safer borrowers. It may seem an interesting result, perhaps paradoxical, result given the popularity of these types of contract. But recall, the main idea of the authors’ models is that increasing q allows a decrease in r, while here we (and the BAAC) hold r fixed and vary q. The results on L are not overwhelming, and should be downplayed due to potential endogeneity. However, they best fit the Ghatak model’s non‐monotonic prediction, that larger loans draw in marginal, safer borrowers when loans are small but signal riskier borrowers with lower expected repayment costs when loans are large. Nonparametric estimates of the relationship, where the other covariates are controlled for linearly, as discussed above and in Appendix C, are presented in Figure 4. Especially in the central region the inverted‐U is pronounced in the function estimates. The results are not conclusive, though, due to lack of data at high loan sizes. However, the upward‐sloping segment of the relationship is well supported and this portion of the curve is enough to favour the Ghatak model. Fig. 4. Open in new tabDownload slide Partially Linear Regressions – Loan size against Repayment. Note. Lack of large loans in the data means the downward‐sloping portions are estimated imprecisely and suggests that most borrowers are in the credit‐constrained, upward‐sloping region. Fig. 4. Open in new tabDownload slide Partially Linear Regressions – Loan size against Repayment. Note. Lack of large loans in the data means the downward‐sloping portions are estimated imprecisely and suggests that most borrowers are in the credit‐constrained, upward‐sloping region. The evidence on covariability is weak but favours Stiglitz and Ghatak at the expense of BC. The direct measure of village covariance of output is a significant predictor of good repayment, in the whole‐sample logits at least. This result is consistent with the Stiglitz story that higher covariation of income can differentially increase safe project payoffs and the Ghatak story of higher covariation drawing in marginal, safer borrowers. It is at odds with BC under certain assumptions as well as the empirical literature, which assumes high correlation should lead to lower repayment. Finally, the evidence on cooperation is rich but seems to turn in favour of BBG and BC at the expense of Stiglitz, pointing toward a negative relationship between cooperation and repayment rates (though not necessarily borrower welfare). Most prominently, the degree of relatedness and the amount of sharing among non‐relatives in the group show up as significant negative predictors of repayment. The village cooperation poll also registers as negative in the northeast region. These results are consistent with the BBG and BC stories that groups behaving cooperatively may choose not to repay rather than pressure each other more than is optimal. A minor exception is sharing among group relatives but this coefficient turns negative (but insignificant) if the two sharing indices are combined or sharing among group non‐relatives is excluded.27 A potential exception to this conclusion arises in the index for the number of production decisions made cooperatively, which is a positive predictor of repayment in several specifications. However, JOINT_DECISIONS may not measure cooperation in the sense of being able costlessly to enforce agreements, but rather may reflect a transfer of knowledge and expertise.28 A second possibility is that the ability to cooperate in project choice is different from the ability to cooperate in punishment behaviour. Since the Stiglitz model focuses on cooperation in project choice, and predicts a positive effect, while (our extensions of) BBG and BC focus on cooperation in punishment behaviour, and predict a negative effect, this would be an interesting reconciliation of the results.29 The negative association between cooperation and repayment seems to conflict with conventional wisdom and with empirical results of Karlan (2007) (this issue).30 Karlan explores a group lending programme in which groups are formed randomly by the lender, so all social ties must be exogenous; he finds some measures of social ties positively predict repayment. One key difference between our results seems to be the range of social ties that we observe. In a programme of random group formation, the range of social ties in the realised groups is likely to be low – certainly not comparable to our case where the mean group has nearly 60% relatives. Thus, one might view Karlan’s results as indicating positive effects of social ties in a neighbourhood of zero, while ours indicate negative effects as social ties vary across more substantial levels. While the theory in this article does not point to such a reconciliation, future work potentially could. The third kind of variable elicits unanimous predictions from multiple models. This includes r and productivity. In both cases there is some evidence to support the model’s unanimous predictions. The negative results on r are at best suggestive, due to a combination of little variation and potential endogeneity. The positive and significant result on education is to our knowledge new; the measure of human capital used in previous studies was literacy and the coefficient was not significantly different from zero (Zeller, 1998). In summary, the unanimous predictions of the models receive some support; the unique predictions are not strongly upheld with the exception of BC’s focus on unofficial penalties; and the conflicting predictions go both for and against almost every model, depending on the prediction, with only the Ghatak model not being contradicted. Pooling all the evidence by region, it can be said that BC is upheld in the northeast region, in particular its focus on unofficial penalties, and not contradicted in any. In the central region, however, Ghatak does well in matching the results on q and L. It is quite possible, and not surprising, that different mechanisms are at work in the different regions, with joint liability potentially solving more of a selection problem in the central region and an enforcement problem in the northeast. What is it about the regions that leads to different results? Two dimensions along which the regions differ significantly are wealth (the northeast being poorer) and financial access (the northeast having less access). This is seen in our data in significant regional differences in VILLAGE_WEALTH and BANK_MEMBERSHIP. Accordingly, we stratify by these two variables, running the same logits on the respective above‐median and below‐median samples. Though the overlap in both cases is quite high, it is the low‐financial‐access sample that parallels very closely the northeast region results on unofficial penalties while the low‐wealth sample turns up insignificant results for unofficial penalties. The results are suggestive that the lack of basic infrastructure (physical and legal) for a formal financial system in the northeast, not the lack of wealth, makes informal penalties an effective substitute method for guaranteeing loans. 2.4. Robustness Checks and Empirical Concerns How representative are our data? The BAAC is not a universal bank. BAAC clients are more educated and wealthier than the typical rural household for example, particularly so for those taking out individual loans. However, this article does not seek to explain among all rural households of the sample who borrows. Rather, this article follows the models in taking as given the selection of some of the rural population into BAAC groups and then focusing on the potential inner workings of those groups themselves. The major exception is our examination of Ghatak’s adverse selection model, with its implication that group members are more risky than those who take the outside option, for which we find evidence, see Ahlin and Townsend (2007). We also note that the data here do not capture groups that have gone out of existence, perhaps because of a pattern of default. This may lead to bias in estimated coefficient magnitudes; however, our focus is on the sign not the magnitude. Further, this seems to be a relatively unsubstantial tail of the data. Using a survey of village heads from the same villages, we find that there are 469 BAAC groups reported in 192 villages, and 6 BAAC groups reported to have once existed but since disbanded. That is, only 1.3 groups have disappeared for every 100 active groups. Of these, some probably disbanded for non‐default reasons. The turnover is slightly higher, but still not alarming, at the individual level: in the household survey, for every 100 households reporting to be members of a BAAC group there are 4.1 households reporting that they were, but are no longer, members of a BAAC group. Our data thus appear to be broadly representative. Another potential problem with the data is that our repayment proxy reflects the entire history of the group, while our covariates are for the most part contemporaneous measures. We thus implicitly assume that the covariates are stable over time. This is certainly true of some (e.g. education) and may be less true of others (e.g. sharing). However, if the covariates are not stable, our sense is that the most probable bias in estimates would be due to independent measurement error and thus uniformly toward zero. We discuss cases in which default itself may plausibly cause changes in covariates below. At any rate, the ideal case would involve both dynamic models and dynamic data. For robustness, we experiment with several different controls and specifications. We use the coefficient of variation of village wealth, both in addition to and as a substitute for village wealth. Entered together, they are both insignificant. By itself village wealth inequality shows the same behaviour as village wealth – negative and (marginally) significant in the fixed‐effect specification only. We also add a regional dummy to the logit regressions and find it insignificant in the baseline but significant in the fixed‐effect specification, predicting lower repayment in the northeast. It does not change other results appreciably. One might be concerned that using several proxies for a given factor, such as both landholdings and education for productivity, may confound the result for any given proxy. Accordingly, we run the baseline logit specification using one proxy at a time for each factor, rather than all simultaneously, and find no appreciable difference in results. Relatedly, each model, as studied here, makes predictions over a different set of variables. It might then be preferable to use model‐specific sets of variables in the regressions rather than the same set for all models. However, to say whether one variable does or does not fit into a given model is somewhat arbitrary. In this article we choose to carry out some extensions based on simplicity of assumptions needed, but more could be done if one is willing to make stronger assumptions. For example, L and q are not inherently absent from the BC model, they are just not introduced here. Thus defining the appropriate set of variables for a model is not straightforward. That said, we do run the logits using for each model only the variables with predictions marked in Table 1. The results are not appreciably different. It is also a departure from the theory to enter the variables additively. The models make predictions about the own and cross‐partials of the repayment rate. We therefore add in various interactive terms suggested by the theories. However, the insertion of these interactive effects failed to uncover significant terms and in some cases undercut the significance of coefficients on the variables entered in levels. We attribute this to lack of data. Further, the logit model builds in non‐zero cross‐partials even when terms are entered additively. Though we have not established causality beyond doubt, we view most of the evidence as strongly suggestive of causal relationships between key variables and repayment. There are exceptions. In principle, group‐level variables may themselves be functions of repayment history, or may be correlated with unobserved determinants of repayment (due to a group selection or other effect). This appears somewhat likely for interest rate r (to the extent it deviates from the national schedule) and loan size L, though not as likely as if the BAAC were satisfying a zero‐profit constraint. Loan size in particular varies for a number of reasons including borrower seniority, borrower’s estimated revenue, government targeting of specific regions, but also perhaps based on assessment of the client’s reliability in repayment. Accordingly, though some exogenous variation seems to exist, we are cautious about inferring a direction of causation. We also note that the results on other variables are robust to a more cautious approach that excludes r and L from the logits. Other group‐level variables are plausibly exogenous to repayment history and do not appear correlated with unobserved repayment determinants. Average landholdings in the group and our proxy for q, the landless fraction of the group, are likely to be stable with respect to repayment history given the rarity of full default and consequent seizure of land by the BAAC (which of course does not imply the threat is non‐credible). Education levels are in general predetermined with respect to repayment history. Other characteristics of group members such as relatedness, location, and occupation are relatively fixed over time. One might worry that groups re‐sort after a default experience to change the composition of these characteristics, perhaps expelling non‐relatives or borrowers from other villages. But the relatively low turnover of borrowers cited previously makes this less of a concern. It is also possible that the sharing variables may respond to group repayment history. If sharing decreases after repayment problems, this would push toward a positive correlation between repayment and sharing; we see the opposite. On the other hand, intra‐group transfers may be higher after default within the group, if one member is repaying another. However, the sharing question is worded so as to imply sharing per se rather than activities with an explicit quid pro quo. More importantly, groups were asked whether one or more members have ever repaid for another in the history of the group. About 10% responded affirmatively. Excluding these groups from the logits still turned up a negative and significant coefficient on sharing among group non‐relatives; this suggests that the negative correlation between repayment and sharing is not being driven by those groups that have experienced internal bailouts. Indeed, the correlation between sharing within the group and past internal bailouts is not significantly different from zero. A related concern is that the result is due to sharing being correlated with some unobserved determinant of repayment, income risk for example. While this cannot be completely ruled out, we note that risk is controlled for at the village level by VILLAGE_RISK and at the group level by OCCUPATIONAL_HOMOGENEITY and (in the unreported robustness check) by existence of an intra‐group bailout. Village‐level variables are unlikely to be functions of the group’s repayment history, since a group makes up only a fraction of the village population. They may, however, be correlated with other unobserved village‐level characteristics that are the actual determinants of repayment. For example, the negative results on outside lenders might be questioned if the prevalence of PCGs indicated the presence of a village attribute that caused it to be shunned by institutional lenders. In this case, it would not be outside lenders per se, but the negative village attribute leading to low repayment. However, this does not appear to be the case, as there is a positive, not negative, correlation between village borrowing from the BAAC and PCG prevalence. We calculate a correlation coefficient not statistically different from zero – 0.072 – between percentage of villagers’ loans that come from the BAAC with percentage of villagers who are PCG members. One might think the BAAC and PCGs are both partly driven by social missions that lead them to difficult areas. But, the correlation between PCG prevalence and commercial bank prevalence is also not statistically different from zero – 0.031 – suggesting that PCGs do not tend to exist primarily in regions of commercial institutional abandonment. One might also wonder if an unobserved village attribute is driving the result on unofficial penalties. In particular, our measure may be correlated with the number of different lenders in the village, since it may be harder to coordinate a shutdown of lending against a delinquent borrower if there are more lenders. We construct a measure of the number of lenders in a village, using data on outstanding loans and their sources from the household survey. We use both the absolute number of lenders represented in this village sample of loans, and the number of lenders normalised by the number of loans in the village sample. Entering either into the baseline regression does not change the sign and significance of the coefficient on unofficial penalties. Our measure of covariability of returns, the coincidence across villagers of bad years within the past five, might be correlated with a village having experienced a negative aggregate shock. For example, even if all villages have the same covariability of returns, those which happened to have a large negative shock recently would register both low repayment and high covariability. This negative relationship between covariability and repayment would be non‐causal. However, this would predict a negative relationship between repayment and covariability, while the data show a positive one. Further, the mean and standard deviation of this variable indicates that aggregate shocks were not common (see Table 2), at least not in the five years prior to data collection in early 1997. Group education levels may also reflect a non‐education related village‐level phenomenon correlated with well‐functioning schools. This does not appear to be the case; the result on group education levels continues to hold even when we include the village‐average education level in the logits. Our interpretation of landlessness as reflecting joint liability, rather than wealth, may be questioned. We do note that the specifications all control for average landholdings in the group, so the partial effect of landlessness should capture the lopsidedness of the distribution conditional on the amount of land. If wealth mattered per se, and linearly (relative to the index), then average landholdings should show up as a positive predictor of repayment. It does not, in the baseline specifications and in specifications that are identical to them but exclude the landlessness variable. Even if wealth mattered positively and non‐linearly, it seems highly likely that average landholdings would show up as a positive predictor of repayment; it does not. One might hypothesise a non‐monotonic relationship between wealth and repayment, where default is worst at either wealth extreme. If this were true, however, one would expect a negative relationship between average landholdings and repayment, controlling for landlessness. This is not what we find; see Table 3. Further evidence is provided in the village‐wealth variable, which is an insignificant, and if anything slightly negative, predictor of repayment. So, while we cannot completely rule out the interpretation of landlessness as capturing a wealth effect, we find it unlikely. Overall, then, all results except on r and L seem plausibly to suggest causal relationships. 3. Conclusion We have compiled and helped construct a theoretical framework through which to view repayment data of joint liability borrowing groups and to test between theories regarding them. We accept for now that the current models have their limitations or shortcomings. For example, the models are static and involve borrowing groups of fixed size (two). Our goal here instead is to evaluate these current, widely‐used theories by a confrontation with the data. Hopefully the insights provided can be used in future research, including the construction of revised models. Using this framework and rich data from Thailand on group characteristics and the villages where they are located, four models were compared. We find that the Besley and Coate model of social sanctions that prevent strategic default performs remarkably well, especially in the low‐infrastructure northeast region. The Ghatak model of peer screening by risk type to overcome adverse selection is supported in the central region, closer to Bangkok. The strongest facts that future modelling should take into account include the negative relationship of the repayment rate with the rate of joint liability (ceteris paribus); its positive relationship with the strength of local sanctions; its potentially positive relationship with correlated returns; and its sometimes negative relationship with more benign social ties such as relatedness and sharing. This is one of the most striking aspects of the results for policy implications:31 strong social ties – measured by sharing among non‐relatives, cooperation, and clustering of relatives, and village‐run savings and loan institutions (PCGs) – having seemingly adverse effects on repayment performance. This result has not been seen in the previous empirical literature, nor focused on in the theoretical models, though Ghatak and Guinnane (1999) provide an insightful discussion using historical and contemporary examples. On the contrary, social ties are typically seen as positive for group lending. This idea must be qualified. Social structures that enable penalties can be helpful for repayment, while those which discourage them can lower repayment. However, a higher repayment rate is not always synonymous with higher welfare. It may merely reflect the use of cheap penalties to enforce repayment when it is not (ex ante) Pareto optimal for the group. Thus joint liability lending may flourish most in areas where social penalties are especially severe, even more severe than the borrowers themselves would prefer.32 Appendix A: Proofs Proof of Proposition 5. The minimum penalty needed to enforce a project choice p is (20) exactly analogous to (5). The new Monitoring Equation is (21) Totally differentiating with respect to p and L gives that (22) The denominator can be shown strictly positive using assumption A7, (20), and the fact that M is increasing and convex. Turning to the second term in the numerator of ∂p/∂L, (20) gives that The concavity of F ensures that F′(L) ≤ F(L)/L. Using this along with the fact that M is increasing and assumption A7 gives that this second term is positive. Turning to the first term, M is convex by assumption and cp is positive by assumption A7. One can show that cL is zero at p = p and is elsewhere positive because it is increasing in p: (23) The strict inequality uses assumption A7 directly, and the weak inequality comes from the concavity of F. Thus the first term of the numerator is positive. Supplement to and Proof of Proposition 9. Let f be the density function associated with F, with [0,1] its support. Under zero correlation, the joint density is f(Yi)f(Yj). Let φ(Yi,Yj) be a generalised joint density: (24) Essentially, κγ(Yi,Yj) is the added (or subtracted) density, relative to the zero‐correlation case, at a point (Yi,Yj). This parameterisation is without loss of generality and allows choice in the structure (γ(Yi,Yj)) and amount (κ) of the added or subtracted density. It does not guarantee that the marginal densities are preserved; necessary and sufficient for this is that γ(Yi,Yj) integrate to zero over Yi and Yj, separately. To guarantee this, we parameterise further: (25) This parameterisation is without loss of generality in the sense that it does not rule out any γ(Yi,Yj) that preserves the marginal densities.33 Its usefulness is in allowing us to choose any integrable function g(Yi,Yj) without worrying about preserving the marginal densities; by construction, the above transformation of g(Yi,Yj), i.e. γ(Yi,Yj), will preserve them. We parameterise g(Yi,Yj), aiming for generality while imposing symmetry on the correlation. Let {α1, α2, …, αN}, {β1, β2, …, βN}, and Q be strictly positive numbers; we assume (A9) This formulation adds mass Q everywhere on the unit square, but also subtracts mass according to an arbitrary, monotonic, polynomial function of the distance between returns. The more disparate the output realisations, the more mass is subtracted. Assumption A9 encompasses simple examples like absolute difference Q − |Yi − Yj| and squared difference Q − (Yi − Yj)2. Intuitively, assumption A9 should lead to positive correlation, and it does.34 Combining it with (24) and (25) and carrying out some detailed integration gives (26) This is strictly positive and linear in κ. Thus κ parameterises the intensity of covariance. We turn now to the proof. Let a ≡ Y(r) and b ≡ Y(2r). Modifying (10) to incorporate the generalised joint density function φ(Yi, Yj; κ) gives (27) From (24), we get that dφ/dκ = γ(Yi,Yj). Using this in (27) gives (28) This equation merely says that the effect of higher correlation on p is inversely related to the amount of mass that the introduced correlation adds to the Default Region. We first show that the first double integral in (28) (corresponding to the probability mass added to box A in Figure 3) is strictly positive. Integration using (25) and assumption A9 gives 𝒜 is strictly positive, since αk,βk > 0, a ∈ (0,1), and [1 − aαk+1 − (1 − a)αk+1] is strictly positive for αk > 0 and a ∈ (0,1).35 We next sketch an argument for why the second double integral in (28) is close enough to zero as unofficial penalties get sufficiently severe; for details see Ahlin and Townsend (2002). First, note that the subtracted mass γ(Yi,Yj) is continuous and so has a maximum and minimum value on [0,1]2. Second, note that can be arbitrarily close to zero over arbitrarily much of the (a,b) interval. In other words, severe enough unofficial penalties remove an arbitrarily large fraction of the AB boxes from the default region. Combining these, even if the maximum or minimum mass is added or subtracted everywhere in the AB boxes, the total amount within the Default Region is arbitrarily small as unofficial penalties get more severe. Thus the second integral can be pinned sufficiently near zero to guarantee the two integrals add to a positive number, which implies ∂p/∂κ is negative. Supplement to and Proof of Proposition 12. We make the following assumptions on F: (A10) The last part of this assumption is the only non‐standard one and requires that F(L) is unbounded and sufficiently concave asymptotically. All parts are satisfied by F(L) = Lα. Observing an agent borrowing L establishes two facts. First, it must be that the borrowing payoff (17) is greater than the outside option u. Defining Z(p) ≡ pr + p(1 − p)q as the (expected) unit borrowing cost of a type‐p agent, this condition is equivalent to (29) Since Z(p) is increasing in p, this implies that , where solves (29) at equality. One can also show that the right‐hand side of inequality (29) first increases, then decreases in L, implying that does the same; see Figure 5. Fig. 5. Open in new tabDownload slide When L is Small Enough – Specifically, Below – then and E(p|L) follows . For , and E(p|L) is a Convex Combination of and . Fig. 5. Open in new tabDownload slide When L is Small Enough – Specifically, Below – then and E(p|L) follows . For , and E(p|L) is a Convex Combination of and . Second, since the borrower can always accept less than the lender’s offer,36 the borrower’s payoff cannot be decreasing in loan size. Otherwise, the borrower could have refused some of the loan and increased his payoff. Applying this to payoff 3 gives, after rearranging, (30) This guarantees that , where solves (30) at equality. The larger L, the tighter the bound of inequality (30), and hence the lower ; see Figure 5. Larger loans signal a lower (expected) cost of capital, which is true of more risky groups. Thus, observing L tells us that . Manipulating (29) and (30) shows that iff (31) Due to strict concavity, Γ′(L) > 0, so Γ(L) can be inverted. Also due to concavity,  limL→0+Γ(L) = 0;37 and due to the last part of assumption A10, Γ(L) is unbounded. Thus there exists an , such that () when (). Assume first that . The expected repayment rate is . Total differentiation of a modified Selection Equation (relation (29) at equality) gives that . Ths is strictly positive since in this range of L, so . Assume next that . The group type is in , but the expected repayment rate is not simply , where the expectation is with respect to density g(p). The reason is that there is a mass point at type corresponding to all groups of type who were offered more than L but only accepted L, their optimal amount.38 We will show that the expected repayment rate is in fact a convex combination of and . Both of these terms are strictly declining in L due to strict concavity of F(L); and both drop (simultaneously) to p for L high enough, due to Inada conditions on F(L). However, the convex combination may be non‐monotonic if the weights shift sufficiently, as pictured in Figure 5. The expected repayment rate is thus coarsely decreasing, in the sense that it is bounded between two strictly decreasing functions that approach p. It remains to show that the expected repayment rate is a convex combination of and . Let h(L),H(L) be the density and distribution functions of lender loan offers. When , there are two categories of borrowers with loan size L. The first includes all agents with types who received a loan offer of exactlyL, of mass . These agents accepted the loan offer without modification because L is less than their desired amount. The second includes all agents of type who received a loan offer greater than L but accepted only L since it is optimal for them, of mass . The probability of observing loan size L is then: (32) This gives rise to a distribution of types conditional on L, call it P(p|L), modified to include a probability mass at . Specifically, using Bayes rule, P(p|L) is zero if ; if ; and g(p)h(L)/P(L) if . E(p|L) is the integral , where the integration treats as a mass point. Carrying out this integration gives the expected group repayment rate as a convex combination of and , where the expectation is with respect to density g(p): (33) Appendix B: Variable Descriptions LN(GROUP_AGE) is the log‐age of the group. If we think of default as having some probability p of occurring each year, then clearly groups with a longer history are more likely to have run into problems. But the effect would be non‐linear in age.39 Results under inclusion of terms for age and age squared, rather than log of age, are similar and not reported here. VILLAGE_RISK is a village‐wide measure of risk, taken from the household survey. Households are asked how much they will earn if next year is a good year (Hi), how much if bad (Lo), and how much they expect to earn (Ex). We assume a distribution of income over two of these mass points, Hi and Lo, as do the models. The coefficient of variation is then equal to This quantity is calculated for each villager in the HH survey, and the village average is used. Thus it is a measure of average riskiness of occupation in a given village. VILLAGE_WEALTH measures average household wealth in the village. Villagers were asked detailed questions about assets of all types – ponds, livestock, appliances, and so on – as well as liabilities. Date of purchase was used to estimate current value after depreciation. These different types of wealth were aggregated for each villager, then averaged across villagers. The unit of measure is one hundred thousand 1997 Thai baht. GROUP_SIZE is the number of members in the group. Groups in our data range in size from five to thirty seven, with eleven being the median. However, each model we consider fixes group size at two. Thus we enter group size as a control variable and leave the introduction of group size into the theory as future work. BANK_MEMBERSHIP and PCG_MEMBERSHIP are measures of outside borrowing opportunities taken from the HH survey. They give the percentage of households surveyed in the group’s village who are members of a commercial bank or production credit group (PCG), respectively. PCGs are village‐run organisations that collect regular savings deposits from members and offer loans after a member has met some threshold requirement involving length of membership, amount deposited, or both. Often the maximum available loans from these institutions are small, possibly one fifth the size of BAAC loans, and the interest rates are similar or slightly higher (Kaboski and Townsend, 1998). There are PCGs large enough to offer loans as large as BAAC loans. Occasionally joint liability is used with these loans. Commercial banks are conventional lenders, requiring collateral. The degree of joint liability q is proxied by a variable constructed from the BAAC survey, the percentage of the group that owns no land. If all group members own land, it is less likely that a guarantor will in the end have to pay rather than the borrower himself. Conversely, if some members of the group are landless, a guarantor will more often have to repay if a landless borrower defaults. Covariance is proxied by two measures. COVARIABILITY is a village‐level measure taken from the HH survey. Villagers were asked which of the previous five years were the best and worst for income, respectively. Our variable is constructed as the probability that two randomly selected respondents from the same village reported the same year as worst. If Nv is the number of villagers in village v and svy is the share of villagers in village v who named year y as the worst, this probability is equal to40 HOMOGENEOUS_OCCUPATIONS is taken from the BAAC survey and equals the probability two randomly chosen group members have the same occupation. It is calculated similarly to COVARIABILITY. Cooperation is measured by four variables. SHARING_RELATIVES, SHARING_NON‐RELATIVES are indices from the BAAC survey. They equal the number of positive responses to five out of six yes/no sharing questions: whether sharing of rice, helping with money, helping with free labour, coordinating to transport crops, coordinating to purchase inputs, and coordinating to sell crops has occurred in the past year. We exclude the sharing of rice, since this may reflect the predominance of rice farming. The same set of questions was asked twice, regarding relatives and non‐relatives, respectively, within the group. These lead to SHARING_RELATIVES and SHARING_NON‐RELATIVES, respectively. BEST_COOPERATION comes from a poll of villagers in the HH survey. Each is asked which village in his tambon (subcounty) enjoys the best cooperation among villagers. The percentage of villagers in the HH survey naming the village in which a group is resident is the measure we use. Finally, JOINT_DECISIONS counts the number of the following three decisions on which some or all group members, as opposed to the individual farmer, have the final say: which crops to grow, pesticide and fertilizer usage, and production techniques. Cost of monitoring is measured in two ways, from the BAAC survey. IN_VILLAGE gives the percentage of the group living in the same village. RELATEDNESS gives the percentage of group members who have a close relative in the group. Official and unofficial penalties are proxied by two village‐level variables from the HH survey. BEST_INSTITUTIONS is a poll similar to BEST_COOPERATION, where the respondent is asked to name the best village in his tambon in terms of availability and quality of institutions. This measure captures to some degree the legal infrastructure, which is related to the official penalties the BAAC can impose on borrowers. SANCTIONS comes from the HH survey and is constructed from a question asking villagers what the penalties for default on their current loans are. We use the percentage of loans in the village that have penalties extending beyond the direct participants in the loan agreement. Specifically, we count loans for which the borrower reports that under default, he cannot borrow again from this lender and other lenders, or that reputation in the village is damaged. Screening is proxied by two dummy variables from the BAAC survey. KNOW_TYPE equals one if the group leader answered that members know the quality of each other’s work. SCREEN equals one if the group leader answered that there are borrowers who would like to join their group but cannot. Productivity shifters include AVERAGE_LAND, the average amount of land per group member, measured in rai,41 and AVERAGE_EDUCATION, average educational attainment in the group. The raw data for education are not years of schooling, but a classification into one of four categories: no schooling; some schooling, but below P4; P4; and higher than P4 schooling. The majority of borrowers have P4 schooling, the minimum level required by the Thai government. Our measure uses the following average: 1(% of group with some schooling, but below P4) + 3(% of group with P4 schooling) + 5(% of group with higher than P4 schooling). The empirical results are robust to various choices of weights. Data on groups’ interest rates r and loan sizes L come from a BAAC survey question asking about the highest and lowest loan size and interest rate experienced by any member of the group over the past year. We take these high (hi) and low (lo) figures and use a weighted average (lo + 0.1hi)/1.1. The high end is only slightly weighted since the upper tail is often quite long and unrepresentative of the group as whole. Appendix C Partially Linear Model Estimation We assume a partially linear model: the repayment rate R is some smooth function of loan size L added to a linear function of the remaining variables, X−L: where β−L is a vector of (M − 1) coefficients and k is a continuous function. The coefficients β−L are estimated by ordering the observations according to L, differencing across nearby observations we use optimal fifth‐order differencing, described in Yatchew (1998) and regressing (the differenced) X−L on (the differenced) R. This produces an estimate . Next we run a non‐parametric, locally linear regression similar to Lowess (Cleveland, 1979); Fan, 1992) where the dependent variable is the residual (and thus not restricted to equal 0 or 1) and the independent variable is L. For each unique value of L, we calculate a fitted value of the residual from a weighted least squares regression on a ‘nearby’ subset of the total sample. Thus the choices are weights for the regression and a bandwidth which determines the subsample. For each unique value l, say, of L, the bandwidth h(l) is set to ensure inclusion of the 80% of the data whose values l g are closest to l.42 The weighting function is the tri‐cube weighting function: This function places more weight on observations located more closely to l. Standard errors at 90% confidence are calculated using the bootstrap method. That is, recreating 1000 samples from the original sample by sampling with replacement, and repeating the entire procedure, for each value l the confidence interval is the 51st and 950th smallest fitted value from these samples. Since our main concern is with the shape (slope) of the functions, we normalise the residuals in each estimate to have mean zero.43 Footnotes 1 " Ghatak (2000), Gangopadhyay et al. (2005), and Armendariz de Aghion and Gollier (2000) also examine joint liability lending in the context of adverse selection. We focus on Ghatak (1999) because its focus on pooling contracts (as opposed to the screening contracts of the first two) and strong informational flows among villagers (as opposed to the weak information flows of the third) best fits the setting of our data. 2 " Wydick (1999) provides an exception. Using self‐collected data from urban and rural borrowing groups Guatemala, he tests the relative importance of social ties, group pressure, and monitoring in explaining repayment performance. He concludes that measures of monitoring are the strongest positive predictors. 3 " The evidence is not direct evidence of a given impediment to trade. Rather, it is evidence about how well a model that features a given impediment to trade does in explaining repayment data. In this context, lack of evidence for a given model may be due to its featured impediment to trade being less important or to its auxiliary assumptions failing to hold. 4 " Using cross‐institution, cross‐country regressions, Cull et al. (2007) (this issue) find that higher interest rates and less direct monitoring are associated with greater portfolio at risk for lenders using individual contracts, but not for lenders using group contracts. They argue this suggests that group lending mitigates information problems directly, lessening the need to rely on lower interest rates or direct monitoring. 5 " In general, the theory we examine does not take into account intra‐group heterogeneity, though this clearly exists in the data. The exception is Ghatak, who allows for intra‐group heterogeneity in risk‐type but shows that it does not exist in equilibrium. 6 " This latter assumption is not explicit in the paper, but if not true, a risky project would have no redeeming feature in comparison with a safe one and would never be chosen. 7 " If assumption A2 did not hold and risky projects involved asymmetric outcomes more often, then the result would depend on how big this difference in frequency of outcomes were, how risk‐averse the borrowers are, and how large risky output is relative to safe output. 8 " A similar modification to a model of strategic default is analysed by Armendariz de Aghion (1999). 9 " This follows from the fact that the entries in row one (column one) must add to pi(pj), and the entries in row two (column two) must add to 1 − pi (1 − pj). 10 " We modify their model by shutting down lending within the group. As BBG show, this departure is appropriate assuming the outside lender has a lower cost of funds. 11 " This is not an a priori unreasonable description of our data, since not all group members take loans every year. 12 " This holds true in the parameter space governed by assumption A7, that is, where there is a moral hazard problem. 13 " For the details of the game, see BC. In short, the borrowers decide simultaneously whether or not to repay r in a first stage. If the decision is not unanimous, the borrower who decided in the first stage to repay can revise his decision in a second stage, paying 0 or 2r. 14 " In this case there can also be a default equilibrium due to group coordination failure. We assume along with BC that the borrowers’ preferred equilibrium (repayment) is played. 15 " is defined implicitly, to satisfy (9) At , it is equally costly to pay r and to suffer default penalties. Above , official and possibly unofficial penalties increase, making it strictly better to pay r; below , the reverse is true. 16 " Specifically, cu(Yi,Λj) > Λj for all Yi ≥ 0 and Λj > 0 implies the cooperative repayment rate is lower and cu(Yi,Λj) < Λj for all Yi ≥ 0 and Λj > 0 implies the cooperative repayment rate is higher. To see this, note that setting cu(Yi,Λj) = Λj =vco(Yj) − r in (9) of the previous footnote gives that for Yj ∈ (Y(r),Y(2r)), which equates the default probabilities. 17 " Note that here borrower outcomes may be relatively similar – for example, Yi,Yj near Y(r) – or dissimilar – for example, Yi near zero and Yj near Y(2r). 18 " Even with our correlation parameterisation, no simple result is available without assuming unofficial penalties are strong. To see this, imagine unofficial penalties are arbitrarily weak, and official penalties are sufficiently weak that no borrower would ever bail out the other but strong enough to induce individual repayment sometimes (Y(2r) > Ymax > Y(r)). In this case, default occurs in every case except when both borrowers are successful, Yi,Yj ≥ Y(r). This non‐default outcome involves relatively similar returns (both high); thus higher correlation raises its probability and makes default less likely. (Graphically, this corresponds to b ≥ 1 in Figure 3 and weak unofficial penalties; the only area leading to repayment would be a single box in the upper‐right corner of (Yi,Yj) space.) This result would require joint assumptions on F(·), r, co, and cu. 19 " Joint liability can be used to screen borrowers, as examined in Ghatak (2000); safer borrowers are more willing to accept higher q since their partners are safer. Here we focus on the pooling case of Ghatak (1999), where joint liability improves efficiency by lowering the subsidy to risk‐taking. The BAAC seems to be pooling, not screening, since they offer just one standard group contract in terms of interest rates and joint liability stipulations (though q may vary for reasons outside the contract, e.g. landholdings). The BAAC does offer individual contracts also, but they require collateral, so the screening there is best thought of as on collateral. 20 " There may be multiple solutions, in which case the highest value can be chosen, giving the upper bound for . Regarding existence, it is possible to show that if (14) has a solution, so does (15). 21 " For technical reasons, we assume that risk‐types are bounded away from one, i.e. that the support of g(p) is for some . 22 " This is verified in Ahlin and Townsend (2002); under assumption A4, we must restrict ρ to be positive. 23 " Indeed the insertion of interactive effects into the logits described below failed to uncover significant terms and in some cases undercut the significance of coefficients on the variables entered in levels. 24 " For more results using this and the non‐parametric techniques, see Ahlin and Townsend (2002). 25 " We are currently trying to get data from the BAAC on other, more severe forms of default. 26 " Hermes et al. (2005) find evidence that stronger monitoring by the group leader, but not by group members, mitigates moral hazard. 27 " Thus the positive result on sharing among group relatives is only due to controlling for sharing among group non‐relatives. This suggests that sharing per se within the group is bad for repayment; but holding fixed sharing among non‐relatives, sharing among relatives is good for repayment. Explaining this result seems to require a theory more precise than a casual invoking of ‘social capital’. 28 " Under this interpretation, JOINT_DECISIONS could fit under the heading of borrower productivity, and a positive sign would thus match our other results and all the models’ predictions. Varian (1990) examines incentives for the transfer of human capital between jointly liable borrowers. 29 " One fact that supports both interpretations is that the correlation of JOINT_DECISIONS with each of the other measures of cooperation is statistically insignificant. Thus it appears to be measuring a different type of cooperation or a different phenomenon entirely. 30 " Cassar et al. (2007) (this issue) also find positive effects of social ties – measured by group homogeneity and survey measures of trust between group members – on repayment, in experimental games performed in South Africa and Armenia. 31 " For more discussion of policy implications of this and related work, see Ahlin (2005). 32 " See Rahman’s (1999) study of Grameen borrowers. Of course, if they still borrow, they are probably better off doing so than not. 33 " To see this, let γ(Yi,Yj) be any function that preserves the marginal densities. Then g(Yi,Yj) can just be set equal to γ(Yi,Yj), and the integrals in (25) are all zero. 34 " Note that though g(Yi,Yj) requires only the distance |Yi − Yj|, the corresponding γ(Yi,Yj) requires the individual values Yi and Yj. This is because adding mass proportional only to the distance |Yi − Yj| would alter the marginal densities. To see this, one can compare the vertical slices of the unit square when Yi = 0 and Yi = 1/2, respectively. The distance |Yi − Yj| on the former slice varies from zero to one, and on the latter slice from zero to one half. Clearly, if the mass added strictly increases with distance |Yi − Yj|, it cannot sum to zero over both of these slices. Thus g(Yi,Yj) is transformed via (25) to add more weight than the distance term alone would imply, near the boundaries of the square. 35 " Considering [1 − aαk+1 − (1 − a)αk+1], note that it equals zero at αk = 0 and is continuous and strictly increasing in αk when a ∈ (0,1). 36 " In keeping with our past treatment of the lender, we assume it does not use the counteroffer to infer the borrower’s risk‐type and adjust contract terms toward some zero‐profit condition. 37 " Note that 0 < LF′(L) < F(L) when L > 0, due to concavity, and F(L) approaches zero. 38 " The mass point is only at , since any borrower who accepts less than offered ends up with his optimal loan size; and by definition, L is optimal for type . When , the mass point does not arise because no one has their optimal loan size. 39 " Specifically, if P(T) is the probability of not having defaulted in T years given an annual probability of not defaulting of p, then P(T) = pT. Further, p ′(T) = ln (p)pT < 0 and p ′′(T) = [ln (p)]2pT > 0. Thus the function is decreasing at a decreasing rate (in absolute value), as is the (negative) natural log. 40 " This is recognisable as the fractionalisation measure, except that here respondents are sampled without replacement. If respondents were sampled with replacement, the measure would be simply . Results would not be affected. 41 " One rai is approximately equal to 0.4 acres. 42 " If there are clusters of observations at the boundary of the bandwidth with the same value for the independent variable, all are included. Thus potentially more than 80% of the sample is used. 43 " Note that the Yatchew procedure identifies the function k(L) up to a constant. De‐meaning the residuals is then essentially a normalisation of each bootstrapped estimate with respect to the constants. Without this normalisation, bootstrap error bands can get large merely because the constants are varying. References Ahlin , C. ( 2005 ). ‘Economic theory meets evidence in rural Thailand: lessons for group lending’ , mimeo, Vanderbilt University. Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Ahlin , C. and Townsend , R. M. ( 2002 ). ‘Using repayment data to test across theories of joint liability lending’ , Vanderbilt University Department of Economics Working Paper 02‐W27, (December). OpenURL Placeholder Text WorldCat Ahlin , C. and Townsend , R. M. ( 2007 ). ‘Selection into and across contracts: theory and field research’ , Journal of Econometrics , vol. 136 ( 2 ), (February), pp. 665 – 98 . Google Scholar Crossref Search ADS WorldCat Armendariz de Aghion , B. ( 1999 ). ‘On the design of a credit agreement with peer monitoring’ , Journal of Development Economics , vol. 60 ( 1 ), (October), pp. 79 – 104 . Google Scholar Crossref Search ADS WorldCat Armendariz de Aghion , B. and Gollier , C. ( 2000 ). ‘Peer group formation in an adverse selection model’ , Economic Journal , vol. 110 ( 465 ), (July), pp. 632 – 43 . Google Scholar Crossref Search ADS WorldCat Banerjee , A. V. , Besley , T. and Guinnane , T. W. ( 1994 ). ‘Thy neighbor’s keeper: the design of a credit cooperative with theory and a test’ , Quarterly Journal of Economics , vol. 109 ( 2 ), (May), pp. 491 – 515 . Google Scholar Crossref Search ADS WorldCat Besley , T. and Coate , S. ( 1995 ). ‘Group lending, repayment incentives and social collateral’ , Journal of Development Economics , vol. 46 ( 1 ), (February), pp. 1 – 18 . Google Scholar Crossref Search ADS WorldCat Cassar , A. , Crowley L. and Wydick , B. ( 2007 ). ‘The effect of social capital on group loan repayment: evidence from field experiments’ , Economic Journal , vol. 117, pp. F85 – F106 . OpenURL Placeholder Text WorldCat Cleveland , W. S. ( 1979 ). ‘Robust locally weighted regression and smoothing scatterplots’ , Journal of the American Statistical Association , vol. 74 , pp. 829 – 36 . Google Scholar Crossref Search ADS WorldCat Cull , R. , Demirguç‐Kunt , A. and Morduch , J. ( 2007 ). ‘Financial performance and outreach: a global analysis of leading microbanks’ , Economic Journal , vol. 117, pp. F107 – F133 . OpenURL Placeholder Text WorldCat Fan , J. ( 1992 ). ‘Design‐adaptive nonparametric regression’ , Journal of the American Statistical Association , vol. 87 ( 420 ), (December), pp. 998 – 1004 . Google Scholar Crossref Search ADS WorldCat Gangopadhyay , S. , Ghatak , M. and Lensink , R. ( 2005 ). ‘On joint liability lending and the peer selection effect’ , Economic Journal , vol. 115 ( 506 ), (October), pp. 1005 – 15 . Google Scholar Crossref Search ADS WorldCat Ghatak , M. , ( 1999 ). ‘Group lending, local information and peer selection’ , Journal of Development Economics , vol. 60 ( 1 ), (October), pp. 27 – 50 . Google Scholar Crossref Search ADS WorldCat Ghatak , M. , ( 2000 ). ‘Screening by the company you keep: joint liability lending and the peer selection effect’ , Economic Journal , vol. 110 ( 465 ), (July), pp. 601 – 31 . Google Scholar Crossref Search ADS WorldCat Ghatak , M. and Guinnane , T. W. ( 1999 ). ‘The economics of lending with joint liability: theory and practice’ , Journal of Development Economics , vol. 60 ( 1 ), (October), pp. 195 – 228 . Google Scholar Crossref Search ADS WorldCat Hermes , N. , Lensink , R. and Mehrteab , H. T. ( 2005 ). ‘Peer monitoring, social ties and moral hazard in group‐based lending programs: evidence from Eritrea’ , World Development , vol. 33 ( 1 ), (January), pp. 149 – 69 . Google Scholar Crossref Search ADS WorldCat Holmström , B. and Milgrom , P. ( 1990 ). ‘Regulating trade among agents’ , Journal of Institutional and Theoretical Economics , vol. 146 ( 1 ), (March), pp. 85 – 105 . OpenURL Placeholder Text WorldCat Ichimura , H. ( 1993 ). ‘Semiparametric least squares (SLS) and weighted SLS estimation of single‐index models’ , Journal of Econometrics , vol. 58 , pp. 71 – 120 . Google Scholar Crossref Search ADS WorldCat Kaboski , J. and Townsend , R. M. ( 1998 ). ‘Borrowing and lending in semi‐urban and rural Thailand’ , mimeo, University of Chicago. Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Karlan , D. S. ( 2007 ). ‘Social connections and group banking’ , Economic Journal , vol. 117, pp. F52 – F84 . OpenURL Placeholder Text WorldCat Prescott , E. S. and Townsend , R. M. ( 2002 ). ‘Collective organizations versus relative performance contracts: inequality, risk sharing, and moral hazard’ , Journal of Economic Theory , vol. 103 ( 2 ), (April), pp. 282 – 310 . Google Scholar Crossref Search ADS WorldCat Rahman , A. ( 1999 ). ‘Micro‐credit initiatives for equitable and sustainable development: who pays’ ? World Development , vol. 27 ( 1 ), (January), pp. 67 – 82 . Google Scholar Crossref Search ADS WorldCat Stiglitz , J. E. ( 1990 ). ‘Peer monitoring and credit markets’ , World Bank Economic Review , vol. 4 ( 3 ), (September), pp. 351 – 66 . Google Scholar Crossref Search ADS WorldCat Townsend , R. M. and Yaron , J. ( 2001 ). ‘The credit risk‐contingency system of an Asian development bank’ , Federal Reserve Bank of Chicago Economic Perspectives , vol. 25 ( 3 ), pp. 31 – 48 . OpenURL Placeholder Text WorldCat Varian , H. R. ( 1990 ). ‘Monitoring agents with other agents’ , Journal of Institutional and Theoretical Economics , vol. 146 ( 1 ), (March), pp. 153 – 74 . OpenURL Placeholder Text WorldCat Vuong , Q. H. ( 1989 ). ‘Likelihood ratio tests for model selection and non‐nested hypotheses’ , Econometrica , vol. 57 ( 2 ), (March), pp. 307 – 33 . Google Scholar Crossref Search ADS WorldCat Wydick , B. ( 1999 ). ‘Can social cohesion be harnessed to repair market failures? Evidence from group lending in Guatemala’ , Economic Journal , vol. 109 ( 457 ), (June), pp. 463 – 75 . Google Scholar Crossref Search ADS WorldCat Yatchew , A. ( 1998 ). ‘Nonparametric regression techniques in economics’ , Journal of Economic Literature , vol. 36 , (June), pp. 669 – 721 . OpenURL Placeholder Text WorldCat Zeller , M. ( 1998 ). ‘Determinants of repayment performance in credit groups: the role of program design, intragroup risk pooling, and social cohesion’ , Economic Development and Cultural Change , vol. 46 ( 3 ), (April), pp. 599 – 620 . Google Scholar Crossref Search ADS WorldCat Author notes " We thank three anonymous referees, Francisco Buera, Jonathan Conning, Hidehiko Ichimura, Steve Levitt, Costas Meghir, Derek Neal, Sombat Sakuntasathien, seminar participants at Chicago, Groningen, London School of Economics, Southern California, University College London, Vanderbilt and at NEUDC 2001, ESSFM 2002, the 2002 European Economic Association meetings, LACEA 2003, and BREAD 2004 for valuable input. Financial support from the National Institutes of Health and the National Science Foundation (including its Graduate Research Fellowship Program) is gratefully acknowledged. All errors are ours. © The Author(s). Journal compilation © Royal Economic Society 2007

Karlan, Dean, S.

doi: 10.1111/j.1468-0297.2007.02015.xpmid: N/A

Abstract Lending to the poor is expensive due to high screening, monitoring and enforcement costs. Group lending advocates believe lenders overcome this by harnessing social connections. Using data from FINCA‐Peru, I exploit a quasi‐random group formation process to find evidence of peers successfully monitoring and enforcing joint‐liability loans. Individuals with stronger social connections to their fellow group members (i.e., either living closer or being of a similar culture) have higher repayment and higher savings. Furthermore, I observe direct evidence that relationships deteriorate after default, and that through successful monitoring, individuals know who to punish and who not to punish after default. Lending to the poor is a difficult task throughout the world, as attested to by the many projects that experience high default rates. Starting with the Grameen Bank in Bangladesh and FINCA village banking in Latin America, development policy makers have embraced group lending as a possible alternative for lenders to provide credit to the poor. Group lending typically links the fate of borrowers by stipulating that if one borrower within a group fails to repay her loan, the others in her group must repay it for her. This potentially works for a few reasons (which all rely on social connections): individuals are able to select creditworthy peers, are able to monitor each others’ use of funds and ability to repay, are able to enforce repayment, or perhaps are more likely to repay merely because of altruism towards those in their group. I test whether groups that are more connected socially perform better and, specifically, whether this is a causal relationship from ex‐post contract monitoring and/or enforcement. The empirical tests employed rule out selection or unobserved dimensions, such as economic opportunities, that coincide with social connections. FINCA‐Peru has a group formation process that generates a natural experiment in which some groups are endowed with stronger social connections than others. I find that stronger social connections of the group lead to higher repayment and savings. While theoretical models have described the potential of group lending to overcome information asymmetries, little empirical evidence has been found to understand if and how group lending actually improves repayment rates (see Banerjee et al., 1994; Besley and Coate, 1995; Ghatak, 2000; Stiglitz, 1990 and Varian, 1990). Cull et al. (2007) (this Feature), comparing institutional profitability of 124 institutions in 49 countries, find positive correlations between interest rate yield and sustainability but at particularly high rates they find default problems begin to occur for individual lending programmes, but not for group lending programmes. This suggests perhaps that the classic models of information asymmetries are indeed salient for individual liability, but that group liability has helped to mitigate the key factors driving the information asymmetry problems. Advocates of group lending not only argue that in fact it does mitigate information asymmetries but typically offer an explanation as to how: by taking advantage of the social networks and relationships. As Varian writes in a 2001New York Times article, ‘Peer pressure can be an immensely strong force, and the Grameen Bank has figured out how to make it work in the cause of economic development.’ In this article, I define social connections as the links and commonalities that bind a group of people together and determine their social interactions.1 Social connections in this context can be thought of as a broader form of social capital, one that encompasses the transaction costs of monitoring successfully, gathering information on each other, and/or punishing in the case of default, or perhaps even just the presence of stronger altruistic motives towards each other. The strength of these connections might merely be a function of living closer to someone else, whereas social capital typically refers to the depth of a given relationship or the level of trust and/or information between individuals. This article analyses the extent to which social connections facilitate the monitoring and enforcement of loans in a group liability arrangement. Typically, showing that higher social connections cause higher loan repayments is a difficult task due to selection and group formation issues. Laboratory experiments in the field provide an alternative approach for examining this question (Cassar et al., 2007; Gine et al., 2006). Such an approach has the advantage of allowing the researcher control over group formation process and specific contract terms, and the disadvantage of being a staged setting in which individuals participate in stylised games albeit with monetary incentives. Using observational data, since most group lending programmes rely on peers to screen each other and form groups, fundamental endogeneity problems exist when analysing the impact of social connections on lending outcomes. For instance, if groups are formed within neighbourhoods, and neighbourhoods with stronger social networks also have more economic opportunities, then empirically one should observe a correlation between the social connections of a group and its likelihood to repay. Indeed prior studies have found correlations between social connections and repayment but have not been able to establish causality.2 If social capital is measured by activities or involvement with others in the community (as is common), then an omitted variable problem persists, in which those with stronger entrepreneurial spirits may also have stronger social capital. Peer‐selection also typically prevents the econometrician from distinguishing between ex‐ante (selection) and ex‐post (monitoring and enforcement) paths through which social connections cause better (or worse) lending outcomes. The analysis here successfully isolates the monitoring and enforcement path. It is important to note that this does not imply anything about the effectiveness (or ineffectiveness) of peer selection on group lending. I collected data from FINCA‐Peru, a group lending organisation, to investigate whether geographic and cultural concentrations make peers more likely to both repay their loans and save more. FINCA‐Peru’s process for assigning individuals to groups creates a natural experiment with quasi‐random group formation. This quasi‐random process provides the strategy for identifying social connections. When lending groups are formed, the initial members neither select each other nor are neighbourhood‐based, as is common in other group‐lending organisations. Instead, when individuals seeking a loan come to FINCA, they are put on a list. Once this list contains thirty names, a group is formed. Group meetings take place in the FINCA office in the city centre and not in the various neighbourhoods, allowing groups to contain members from all over the city. This unique assignment process creates groups with exogenous levels of initial social connections. Since each group has fewer than 30 members, chance alone produces some groups with higher levels of social connections, i.e. they are more geographically or culturally dense. In addition, because individuals do not screen each other beforehand, improved enforcement and monitoring, and not selection, explains the impact of social connections on group outcomes. I find that individuals who live closer to one another and are more culturally similar to others in the group are more likely to repay their loans and save more. There are many reasons to believe that this is a result of their ability to better monitor and enforce the loans. I present direct evidence of monitoring, such as knowledge of each other’s default status, as well as direct evidence of punishment, such as deterioration of relationships. Monitoring and enforcement could lead to improved repayment rates through several mechanisms, such as by increasing the cost of defaulting, by inspiring stronger group solidarity, more diligent work ethics and hence better business outcomes, or by taking advantage of sentiments of altruism or reciprocity within social groups. I also find evidence that better connected individuals are more likely to be forgiven after defaulting, suggesting that their peers were able to distinguish between default due to moral hazard and default due to true negative personal shocks; see Rai and Sjöström (2004) for a theoretical model of this issue. These findings provide important insights into the factors that drive the success of group banking projects. The measures I use (geographic proximity and cultural similarity), although more general than standard measures of social connections (or social capital), have three distinct advantages in this context. First, they are unlikely to be influenced by participation in the credit programme and, hence, are not endogenous with respect to outcomes of the lending group. Second, they can be measured accurately even on a recall basis. Third, they are easily observable, making it simpler to formulate and present policy recommendations.3 However, the broadness of the measures allows for alternative explanations of the lower observed default rates, such as reduced transaction costs for conducting monitoring activities. For this reason, I refer to the measures as measures of social connections rather than social capital as is often done in the group lending literature. Regardless, whereas this broadness limits the ability to interpret the results as evidence of social capital per se influencing lending outcomes, the primary findings provide solid evidence that peer monitoring and enforcement effectively reduce default rates. This is an important finding for policymakers and microfinance institutions, as well as academics interested in testing contract theories. We have observed a plethora of lending schemes to the poor, some with more success than others. Little empirical work has shown why some designs seem to work better than others; the findings here provide valuable insight into these important questions regarding lending to the poor (Banerjee, 2002).4 This article proceeds as follows. Section 1 discusses joint liability mechanisms and FINCA‐Peru, the source of the data for this research. Section 2 discusses the survey procedures and summarises the data collected. Section 3 discusses the identification strategy employed in the analysis. Section 4 presents the central results on the lending and savings outcomes. Section 5 presents results on direct observations of monitoring and enforcement activities. Section 6 concludes. 1. Joint Liability Mechanisms 1.1. FINCA‐Peru FINCA‐Peru uses a village banking lending methodology, first introduced by FINCA International in 1984 and now used by over 80 organisations in 32 countries. A village bank is a group of 30 women who meet weekly at the FINCA office both to borrow and to save, simultaneously. Most members have two loans, one from FINCA (the external loan) and one from their own pool of savings (the internal loan). Interest rates on both external and internal loans are 3% per month. In the case of default on either loan, the group’s savings is used to pay back the loan. Each week the members make an instalment payment on their external loan. In addition to the instalment on their external loan, all members must make a savings deposit such that at the end of the four‐month loan cycle they will have saved at least 20% of the amount borrowed under their external loan. Operationally, the loan instalment and savings deposit are made together, as one payment. Clients also are encouraged to make additional voluntary savings deposits. The savings deposits (both mandatory and voluntary) do not lie idle. Each week, the savings are accumulated and lent out to some of the group members as one month internal loans. At the end of the loan cycle, interest earned on the internal loans is paid out to the members proportionally, by the amount of savings each has amassed.5 FINCA earns the interest on the external loans. The savings and internal loan structure is very similar to a rotating savings and credit association since all members make small weekly deposits, and then each week a small fraction of the members receives large loans from the savings of everyone.6 Empirically, FINCA has perfect repayment on its loan to the group. When there has been default, it always has been on the individual level and fully covered by the individual’s own savings or by the other women’s savings. Regardless, in weekly meetings FINCA employees emphasise to the clients the need to monitor and enforce each other’s loans, even if they are fully collateralised for FINCA. FINCA does this for two reasons. First, although their rate of return is not directly affected by internal default, groups with higher internal default pose a higher risk of eventual default to FINCA. Second, groups with higher internal default have higher dropout rates, and the acquisition of new clients is costly for FINCA, particularly since new clients start out at lower loan sizes than tenured clients. When individuals want to join a bank, they typically arrive on their own or by invitation of a member of another group.7 FINCA does not advertise in the community, nor explicitly ask individuals to seek out new applicants. Most people in the community, in fact, are aware of FINCA and know where the office is. Clients do not come in already formed groups.8 If these individuals meet the basic criteria (have a business, understand the rules, and want a loan), their names are placed on a waiting list. When 30 names are on the list, a group is formed and individuals receive their first loan. This process happens quite quickly, typically in a week or two. FINCA claims to follow this methodology for two reasons, despite its potential drawbacks (e.g. it does not use the peers to help select the best clients). First, they believe that it is the fastest way to create new banks. Asking clients to go out and find others would inevitably take much longer. Thus, individuals do not feel compelled to seek out others in order to speed up the group formation process. Second, FINCA’s mission includes building new social connections, hence they prefer initial group members not to know one another. FINCA hopes that through participation in its programme it not only provides credit to the poor, but also helps the poor develop new relationships, both social and business, and in so doing strengthens the social fabric of the community as a whole. Each week the clients are required to attend a meeting at the FINCA office located in the town centre. Several activities occur at this weekly meeting, including loan payments, savings deposits, issuing new loans, training in group operations and the importance of group solidarity, and monitoring of loan repayment by all members. Attendance at meetings typically exceeds 90%, although poorly performing groups often experience lower attendance.9 For some groups, monitoring activities are very regimented. After all payments are recorded, the group ‘board’ (with supervision by a FINCA employee) reviews all the default situations. It then assigns specific individuals to visit the person in default and to inquire as to the cause of the default or late payment.10 When members leave a group, either voluntarily or involuntarily, their place often is filled by a friend or relative of another group member through direct invitation. FINCA’s operating philosophy encourages clients to develop solidarity or social capital. While this is evident from the meeting hall posters propagating the values of camaraderie, trust and teamwork, it is even more evident in the training materials provided to the employees and clients. In these materials, FINCA emphasises that the clients themselves are responsible for monitoring the group members in order to ensure that loan proceeds are used properly and for enforcing repayment and attendance. 1.2. Why Group Lending? Poor individuals lack formal credit because lenders have little means of screening clients, monitoring the use of funds, or enforcing repayment. In recent years many development organisations have used group lending to deliver credit to poor individuals. Group lending purports to pass off the screening, monitoring and enforcement of the loans to the peers (Banerjee et al. 1994; Diamond, 1984; Ghatak and Guinnane, 1999; Stiglitz, 1990; Varian, 1990). Furthermore, group loans help formal lenders overcome the prohibitively high fixed cost of delivering small loans. Monitoring and enforcement are distinct, although difficult to distinguish empirically. Monitoring itself does not guarantee repayment, but it allows a lending organisation to know whom to punish for not repaying. Although a commercial bank can attempt to monitor business and life outcomes for individuals, it is both difficult and costly to do so. Group lending mechanisms provide incentives to the borrowers to monitor each other to see who can pay and who cannot. Monitoring can take on several forms, such as observing repayment of the loan, visiting another’s business to verify that it is in operation, showing receipts to demonstrate that inventory was purchased with the loan proceeds,11 and talking to others in the community to confirm negative shocks like illness. In these examples, the extent of someone’s social networks is critical and positively related to the ability to monitor or be monitored.12Armendariz de Aghion and Gollier (2000) and Armendariz de Aghion (1999) show theoretically how peer monitoring alone, with random formation of groups, can help overcome adverse selection problems when monitoring is costly for the lending institution itself. Stronger social networks have lower monitoring costs, which results in more credit being extended. To enforce lending contracts, lending institutions typically resort to legal options, such as seizing property of the borrower or garnishing wages directly from the employer. In most poor communities, such punishments fail for one of two reasons, either the legal infrastructure does not support such action, or the borrower has no seizable assets or wages. De Soto (2000) and Besley and Coate (1995) discuss these issues at length. Group lending purports to overcome these failures by using people’s desire to protect their social connections (and social capital) and avoid any possible repercussions. Such repercussions could be economic and result in reduced trading partners for one’s business, social and lead to loss of friends, or psychological and damage one’s self‐esteem. Group lending does not unambiguously facilitate repayment through monitoring and enforcement. Three issues in particular could cause group lending to generate higher default than individual lending and cause groups with higher social connections to have higher default than groups with lower social connections. First, if social connections are strong enough to permit the monitors to distinguish between personal negative shocks and mere reneging, then punishment could be made contingent upon the observations of the monitor. This effectively would be an insurance as well as a lending mechanism and would weaken the incentive to repay after personal negative shocks. Second, Besley and Coate (1995) present a strategic default model: as good individuals observe others defaulting, they themselves default as well since they will not receive a new loan even if they repay and they will suffer no scorn from others for defaulting. If borrowing individually, these individuals might have repaid. In both of these theories, higher social connections should generate higher default.13 Third, the presence of the insurance and possible risk‐sharing arrangements could encourage ex ante moral hazard, or shifts into riskier project choice by the clients. Whereas this may be optimal for the clients, this does pose a greater risk to the lending organisation (which may be compensated in that higher interest could be charged). Hence, the theoretical relationship between social connections and repayment is ambiguous. The existing empirical research on the relationship between social connections and repayments is also inconclusive, partly due to the endogeneity problems discussed earlier. For instance, Sharma and Zeller (1997) using credit groups in Bangladesh, and Ahlin and Townsend (2007) using data from Thailand, find that groups with high levels of family relations have higher default. These findings could be because family members are unable to screen effectively. Ahlin and Townsend (2007) and Wydick (1999) find that groups that report threats of social sanctions for failure to repay have higher repayment; however, why some groups decide to have such policies is not understood, and potentially endogenous (or potentially creating omitted variable problems for drawing causal inferences). Also, such reports do not indicate whether higher levels improve or worsen the ability of social connections to cause better outcomes. Sadoulet and Carpenter (1999) analyse the structure of a Guatemalan peer mechanism and find that by design it lends itself to risk‐sharing as well as enforcement of repayment. Most recently, La Ferrara (2003) studies kin groups in Ghana and finds that punishment is exacted not only on those who default, but also on the kin of those who default, and that the threat of such punishment induces compliance in the short run. These studies demonstrate that the relationship between social connections and group lending outcomes is complicated and worthy of further study. This article builds on that research by using a natural experiment to show that having stronger social connections causes higher repayment and savings by facilitating monitoring and enforcement of group lending contracts. 2. Data This research uses data from participants in the Ayacucho14 programme in Peru from 1998 to 2000. For this study, I divide participants into two groups, those that were invited by a member of their own group and those that were not. The analysis is conducted on the latter, i.e. the uninvited. The primary analysis will regress loan default, savings and attrition on geographic and cultural dispersion.15 The default, savings and attrition data come from FINCA‐Peru’s internal records. These records also contain certain basic demographic information, such as marital status, number of children, and age. For this project, I employed a team of 10 surveyors from January to June 2000 to collect data on cultural identity, social connections amongst group members, method of their arrival to FINCA (i.e., invited or uninvited), location of their home and other demographic information not already collected by FINCA. Three types of surveys were conducted in this phase: group interviews to collect publicly known information (such as who invited whom), individual surveys conducted privately, and individual surveys conducted in the homes or businesses of former members. See the Data Appendix for a description of the data collection process.16 Further data about monitoring and enforcement activities were collected in 2001 and are discussed in Section 5. The primary dataset for this project contains 2,054 individuals over 6,874 loan cycles, or an average of 3.3 loans per individual. For the primary analysis, I restrict the analysis to uninvited individuals (57% of the sample) for reasons explained in more detail below, in the identification section. The dependent variables are the outcome for each uninvited individual’s first loan, and the key independent variables are that person’s connection to the original members of her group. I have data on the selection method (i.e., uninvited or invited) for 1,719 of the 2,054 individuals. Twenty per cent of the uninvited individuals had some default on their first loan, whereas only 16.0% of the invited individuals had some default on their first loan. The average savings deposits made during the 4‐month loan cycle was $59 for both the uninvited and invited. Tables 1 and 2 show the summary statistics for individuals, and Table 3 shows the summary statistics for groups. The summary statistics are shown for the invited as well as the uninvited in order to provide a broad description of the differences across these two sample frames but should not be over‐interpreted since this comparison is not well identified. Table 1 Individual Summary Statistics
Means . Method of arrival to group . Uninvited . Invited . (1) . (2) . Loan data Proportion of loans with default 0.203 0.160 (0.016) (0.011) Default (cond. on default > 0), US$ 69.157 62.867 (8.407) (4.038) Default as proportion of approved FINCA loan (cond. on default >0) 2.797 2.229 (0.222) (0.123) Initial savings, US$ 37.800 39.098 (2.601) (1.342) New savings deposits (both required & voluntary), US$ 59.121 58.690 (3.059) (1.941) Dropout after first loan cycle, proportion 0.240 0.243 (0.017) (0.129) Demographic data Female 0.989 0.996 (0.004) (0.002) Age 34.101 32.102 (0.495) (0.389) Spouse 0.534 0.565 (0.020) (0.149) Completed high school 0.317 0.287 (0.188) (0.136) Individuals Average number of loan cycles per individual 616 1,103 3.13 2.67 . Method of arrival to group . Uninvited . Invited . (1) . (2) . Loan data Proportion of loans with default 0.203 0.160 (0.016) (0.011) Default (cond. on default > 0), US$ 69.157 62.867 (8.407) (4.038) Default as proportion of approved FINCA loan (cond. on default >0) 2.797 2.229 (0.222) (0.123) Initial savings, US$ 37.800 39.098 (2.601) (1.342) New savings deposits (both required & voluntary), US$ 59.121 58.690 (3.059) (1.941) Dropout after first loan cycle, proportion 0.240 0.243 (0.017) (0.129) Demographic data Female 0.989 0.996 (0.004) (0.002) Age 34.101 32.102 (0.495) (0.389) Spouse 0.534 0.565 (0.020) (0.149) Completed high school 0.317 0.287 (0.188) (0.136) Individuals Average number of loan cycles per individual 616 1,103 3.13 2.67 Standard errors of estimated means reported in parentheses. Open in new tab Table 1 Individual Summary Statistics
Means . Method of arrival to group . Uninvited . Invited . (1) . (2) . Loan data Proportion of loans with default 0.203 0.160 (0.016) (0.011) Default (cond. on default > 0), US$ 69.157 62.867 (8.407) (4.038) Default as proportion of approved FINCA loan (cond. on default >0) 2.797 2.229 (0.222) (0.123) Initial savings, US$ 37.800 39.098 (2.601) (1.342) New savings deposits (both required & voluntary), US$ 59.121 58.690 (3.059) (1.941) Dropout after first loan cycle, proportion 0.240 0.243 (0.017) (0.129) Demographic data Female 0.989 0.996 (0.004) (0.002) Age 34.101 32.102 (0.495) (0.389) Spouse 0.534 0.565 (0.020) (0.149) Completed high school 0.317 0.287 (0.188) (0.136) Individuals Average number of loan cycles per individual 616 1,103 3.13 2.67 . Method of arrival to group . Uninvited . Invited . (1) . (2) . Loan data Proportion of loans with default 0.203 0.160 (0.016) (0.011) Default (cond. on default > 0), US$ 69.157 62.867 (8.407) (4.038) Default as proportion of approved FINCA loan (cond. on default >0) 2.797 2.229 (0.222) (0.123) Initial savings, US$ 37.800 39.098 (2.601) (1.342) New savings deposits (both required & voluntary), US$ 59.121 58.690 (3.059) (1.941) Dropout after first loan cycle, proportion 0.240 0.243 (0.017) (0.129) Demographic data Female 0.989 0.996 (0.004) (0.002) Age 34.101 32.102 (0.495) (0.389) Spouse 0.534 0.565 (0.020) (0.149) Completed high school 0.317 0.287 (0.188) (0.136) Individuals Average number of loan cycles per individual 616 1,103 3.13 2.67 Standard errors of estimated means reported in parentheses. Open in new tab To measure social connections, I examine the cultural and geographic proximity of each individual to the original members of the group. Research at both macroeconomic and microeconomic levels suggests that cultural heterogeneity influences the societal norms that dictate how economies and political bodies organise themselves. For instance, Alesina et al. (2004) find evidence for explicit tradeoffs between racial and income heterogeneity and economies of scale in the formation of local jurisdictions. Alesina and La Ferrara (2000) find that cultural heterogeneity negatively influences participation in community and civic activities. Glaeser et al. (2000) discuss the determinants of trust in the US, with strong findings for cultural heterogeneity negatively influencing trust. Most people in Ayacucho, Peru are a blend of indigenous and Western heritage. Individuals of either extreme can be identified easily by their language, dress and hair style. For instance, indigenous individuals wear black hats with large rims, keep their hair in braids and speak only Quechua, whereas Western individuals have short, styled hair, speak only Spanish, and wear jeans and other Western clothing. Using the above characteristics, I create a culture score from zero to eight for each individual. I then calculate the probability that a given individual has the same culture score as a randomly chosen individual from the original group. This is analogous to a standard cultural fragmentation index (Alesina and La Ferrara, 2000) which calculates the probability that two individuals randomly drawn from a group are of the same cultural background. Table 2 Individual Geographic and Cultural Measures
Means and Standard Deviations . Uninvited to group . Invited to group . Mean & std error . Std dev & no. of Obs . Mean & std error . Std dev & no. of Obs . (1) . (2) . Distance data (units in minutes walking) *Distance from current member to original members of group 13.501 9.928 13.106 7.569 (0.400) n = 616 (0.231) n = 1,075 Distance from current member to members of other groups 13.704 9.632 14.109 7.366 (0.388) n = 616 (0.225) n = 1,075 *Prob(Member from original group lives within 10‐minute walk of home) 0.224 0.236 0.205 0.203 (0.009) n = 616 (0.006) n = 1,075 Prob(Person from other group lives within 10‐minute walk of home) 0.211 0.214 0.172 0.170 (0.009) n = 616 (0.005) n = 1,075 Distance to FINCA office (town center) 9.565 10.905 9.858 8.957 (0.439) n = 616 (0.273) n = 1,075 Culture data Culture score (0 = Western, 8 = Indigenous) 2.537 2.224 2.610 2.172 (0.090) n = 616 (0.066) n = 1,075 *Prob(Member from original group is of same culture as individual) 0.201 0.156 0.190 0.140 (0.006) n = 616 (0.004) n = 1,075 Prob(Person from other group is of same culture as individual) 0.185 0.118 0.172 0.098 (0.005) n = 616 (0.003) n = 1,075 Seating arrangements Distance to current members of group 15.795 11.588 15.081 8.984 (0.612) n = 358 (0.357) n = 632 Distance to persons seated next to each other in meeting 14.953 13.395 13.086 10.503 (0.708) n = 358 (0.418) n = 632 Prob(Person from same group is of same culture) 0.231 0.141 0.239 0.134 (0.007) n = 358 (0.005) n = 632 Prob(Person in next seat in meeting is of same culture) (0.262) 0.3016 (0.261) 0.302 (0.016) n = 358 (0.012) n = 632 . Uninvited to group . Invited to group . Mean & std error . Std dev & no. of Obs . Mean & std error . Std dev & no. of Obs . (1) . (2) . Distance data (units in minutes walking) *Distance from current member to original members of group 13.501 9.928 13.106 7.569 (0.400) n = 616 (0.231) n = 1,075 Distance from current member to members of other groups 13.704 9.632 14.109 7.366 (0.388) n = 616 (0.225) n = 1,075 *Prob(Member from original group lives within 10‐minute walk of home) 0.224 0.236 0.205 0.203 (0.009) n = 616 (0.006) n = 1,075 Prob(Person from other group lives within 10‐minute walk of home) 0.211 0.214 0.172 0.170 (0.009) n = 616 (0.005) n = 1,075 Distance to FINCA office (town center) 9.565 10.905 9.858 8.957 (0.439) n = 616 (0.273) n = 1,075 Culture data Culture score (0 = Western, 8 = Indigenous) 2.537 2.224 2.610 2.172 (0.090) n = 616 (0.066) n = 1,075 *Prob(Member from original group is of same culture as individual) 0.201 0.156 0.190 0.140 (0.006) n = 616 (0.004) n = 1,075 Prob(Person from other group is of same culture as individual) 0.185 0.118 0.172 0.098 (0.005) n = 616 (0.003) n = 1,075 Seating arrangements Distance to current members of group 15.795 11.588 15.081 8.984 (0.612) n = 358 (0.357) n = 632 Distance to persons seated next to each other in meeting 14.953 13.395 13.086 10.503 (0.708) n = 358 (0.418) n = 632 Prob(Person from same group is of same culture) 0.231 0.141 0.239 0.134 (0.007) n = 358 (0.005) n = 632 Prob(Person in next seat in meeting is of same culture) (0.262) 0.3016 (0.261) 0.302 (0.016) n = 358 (0.012) n = 632 *Variables with asterisks are the key independent variables used in the specifications in Tables 4, 5, and 6. For each variable, the Table reports the mean, the standard error of the estimate of the mean, the standard deviation and the number of observations. Units for distance measures are in minutes walking distance. Open in new tab Table 2 Individual Geographic and Cultural Measures
Means and Standard Deviations . Uninvited to group . Invited to group . Mean & std error . Std dev & no. of Obs . Mean & std error . Std dev & no. of Obs . (1) . (2) . Distance data (units in minutes walking) *Distance from current member to original members of group 13.501 9.928 13.106 7.569 (0.400) n = 616 (0.231) n = 1,075 Distance from current member to members of other groups 13.704 9.632 14.109 7.366 (0.388) n = 616 (0.225) n = 1,075 *Prob(Member from original group lives within 10‐minute walk of home) 0.224 0.236 0.205 0.203 (0.009) n = 616 (0.006) n = 1,075 Prob(Person from other group lives within 10‐minute walk of home) 0.211 0.214 0.172 0.170 (0.009) n = 616 (0.005) n = 1,075 Distance to FINCA office (town center) 9.565 10.905 9.858 8.957 (0.439) n = 616 (0.273) n = 1,075 Culture data Culture score (0 = Western, 8 = Indigenous) 2.537 2.224 2.610 2.172 (0.090) n = 616 (0.066) n = 1,075 *Prob(Member from original group is of same culture as individual) 0.201 0.156 0.190 0.140 (0.006) n = 616 (0.004) n = 1,075 Prob(Person from other group is of same culture as individual) 0.185 0.118 0.172 0.098 (0.005) n = 616 (0.003) n = 1,075 Seating arrangements Distance to current members of group 15.795 11.588 15.081 8.984 (0.612) n = 358 (0.357) n = 632 Distance to persons seated next to each other in meeting 14.953 13.395 13.086 10.503 (0.708) n = 358 (0.418) n = 632 Prob(Person from same group is of same culture) 0.231 0.141 0.239 0.134 (0.007) n = 358 (0.005) n = 632 Prob(Person in next seat in meeting is of same culture) (0.262) 0.3016 (0.261) 0.302 (0.016) n = 358 (0.012) n = 632 . Uninvited to group . Invited to group . Mean & std error . Std dev & no. of Obs . Mean & std error . Std dev & no. of Obs . (1) . (2) . Distance data (units in minutes walking) *Distance from current member to original members of group 13.501 9.928 13.106 7.569 (0.400) n = 616 (0.231) n = 1,075 Distance from current member to members of other groups 13.704 9.632 14.109 7.366 (0.388) n = 616 (0.225) n = 1,075 *Prob(Member from original group lives within 10‐minute walk of home) 0.224 0.236 0.205 0.203 (0.009) n = 616 (0.006) n = 1,075 Prob(Person from other group lives within 10‐minute walk of home) 0.211 0.214 0.172 0.170 (0.009) n = 616 (0.005) n = 1,075 Distance to FINCA office (town center) 9.565 10.905 9.858 8.957 (0.439) n = 616 (0.273) n = 1,075 Culture data Culture score (0 = Western, 8 = Indigenous) 2.537 2.224 2.610 2.172 (0.090) n = 616 (0.066) n = 1,075 *Prob(Member from original group is of same culture as individual) 0.201 0.156 0.190 0.140 (0.006) n = 616 (0.004) n = 1,075 Prob(Person from other group is of same culture as individual) 0.185 0.118 0.172 0.098 (0.005) n = 616 (0.003) n = 1,075 Seating arrangements Distance to current members of group 15.795 11.588 15.081 8.984 (0.612) n = 358 (0.357) n = 632 Distance to persons seated next to each other in meeting 14.953 13.395 13.086 10.503 (0.708) n = 358 (0.418) n = 632 Prob(Person from same group is of same culture) 0.231 0.141 0.239 0.134 (0.007) n = 358 (0.005) n = 632 Prob(Person in next seat in meeting is of same culture) (0.262) 0.3016 (0.261) 0.302 (0.016) n = 358 (0.012) n = 632 *Variables with asterisks are the key independent variables used in the specifications in Tables 4, 5, and 6. For each variable, the Table reports the mean, the standard error of the estimate of the mean, the standard deviation and the number of observations. Units for distance measures are in minutes walking distance. Open in new tab Geographic distance between members captures social connections for many reasons. Monitoring costs are reduced when individuals live closer to each other. Individuals with more common acquaintances or friends will procure information more easily about each other. Also, the threat of reputation loss is potentially more effective among those who live closer to each other since such individuals will have more frequent future interactions and more acquaintances in common. In order to quantify geographic concentration, I employ two measures: the average distance of an individual’s home to those of the original members, and the percentage of original members who live within a 10‐minute walk of the individual. The first is similar to a metric used by Busch and Reinhardt (1999) to calculate geographic concentration of industries. The second measure recognises that it is costly, perhaps exceedingly so, for everyone to monitor everyone else. Therefore, it is more sensible for individuals to be responsible for monitoring those who live close to them. For reasons discussed in the next section, both measures relate distance to the original, not current, members of the group. For group‐level analysis for both cultural similarity and geographic concentration, I use the average of the individual measures.17 3. Identification Strategy The identification strategy exploits the institutional fact that FINCA‐Ayacucho forms initial groups with little self‐selection. This solves an endogeneity problem fundamental to group lending, that peers select their own group members (Ghatak, 1999; 2000). Peer selection and group formation in this context create two empirical issues: the first issue is about establishing a causal link from social connections and group outcomes and the second issue is about distinguishing between selection and ex post monitoring and enforcement stories. Peer selection might generate omitted variable problems (e.g., individuals assortatively match into groups on characteristics unobservable to the econometrician, yet correlated with both social connections and business success) or simultaneity problems (successful groups help create better social connections). Such omitted variable and simultaneity problems make it difficult to argue that observed correlations between social connections and repayment (or other group outcomes) are causal in nature, rather than spuriously correlative. Second, it prevents the econometrician from identifying the impact of social connections on effective monitoring and enforcement of loans, as distinct from the effective selection of trustworthy individuals. As discussed in Section 1, when individuals want to receive a loan from FINCA, they typically arrive on their own or by invitation of a member in a group without an opening. Their name is then put on a list and once 30 names have been collected, a new group is formed. As individuals leave the group, openings are typically filled by invitation of a member of that group. Out of the 1,078 individuals who came by invitation of a member of the same group, only eight reported coming by invitation of two others, and only one reported coming by invitation of three others. Hence, even when individuals come by invitation, few cases exist of even a small portion of the group forming prior to arrival to FINCA. I divide participants into two groups, invited and uninvited. I claim that the social connection between the current, uninvited members and the original, uninvited members is exogenous (whereas that of the invited is endogenous). I examine this key assumption below. Since the uninvited members can invite members, I want to measure the social connections between each uninvited member and the original, not current, members of the group. This solves another problem as well, that the dropout process may homogenise groups at different rates depending on the prior success of the group. Furthermore, by only analysing the uninvited members, I can eliminate peer selection as a possible explanation of the findings. This issue has been difficult to overcome in prior studies, such as Sharma and Zeller (1997). Table 3 Group Summary Statistics
Means and Standard Deviations . Method of arrival to group . Uninvited to group . Invited to group . Mean & std error . Std dev & no. of obs . Mean & std error . Std dev & no. of obs . Geographic concentration *Average distance to original members from current members (minutes) 12.422 5.342 12.413 4.751 (0.824) n = 42 (0.733) n = 42 *Average % of original members who live within 10 minutes of current member 0.239 0.200 0.243 0.196 (0.031) n = 42 (0.030) n = 42 GD: Geographic concentration 0.147 0.104 0.252 0.174 (0.016) n = 42 (0.027) n = 42 E(GD): Expected geographic concentration 0.127 0.090 0.203 0.168 (0.014) n = 42 (0.026) n = 42 Cultural concentration *Average percent of original members of same culture as current member 0.197 0.098 0.212 0.111 (0.015) n = 42 (0.017) n = 42 CD: Cultural concentration 0.119 0.136 0.184 0.155 (0.021) n = 42 (0.024) n = 42 E(CD): Expected cultural concentration 0.106 0.078 0.167 0.136 (0.012) n = 42 (0.021) n = 42 . Method of arrival to group . Uninvited to group . Invited to group . Mean & std error . Std dev & no. of obs . Mean & std error . Std dev & no. of obs . Geographic concentration *Average distance to original members from current members (minutes) 12.422 5.342 12.413 4.751 (0.824) n = 42 (0.733) n = 42 *Average % of original members who live within 10 minutes of current member 0.239 0.200 0.243 0.196 (0.031) n = 42 (0.030) n = 42 GD: Geographic concentration 0.147 0.104 0.252 0.174 (0.016) n = 42 (0.027) n = 42 E(GD): Expected geographic concentration 0.127 0.090 0.203 0.168 (0.014) n = 42 (0.026) n = 42 Cultural concentration *Average percent of original members of same culture as current member 0.197 0.098 0.212 0.111 (0.015) n = 42 (0.017) n = 42 CD: Cultural concentration 0.119 0.136 0.184 0.155 (0.021) n = 42 (0.024) n = 42 E(CD): Expected cultural concentration 0.106 0.078 0.167 0.136 (0.012) n = 42 (0.021) n = 42 *Variables with asterisks are the key independent variables used in the specifications in Table 7. All results calculated on original group members only. where si is the share of the group from neighbourhood i and xi is the share of the general population from neighbourhood i. CD and E(CD) are constructed identically to GD and E(GD), except by cultural group rather than neighbourhood. The Alesina index for cultural concentration is equal to the sum of squared shares of each cultural group. Open in new tab Table 3 Group Summary Statistics
Means and Standard Deviations . Method of arrival to group . Uninvited to group . Invited to group . Mean & std error . Std dev & no. of obs . Mean & std error . Std dev & no. of obs . Geographic concentration *Average distance to original members from current members (minutes) 12.422 5.342 12.413 4.751 (0.824) n = 42 (0.733) n = 42 *Average % of original members who live within 10 minutes of current member 0.239 0.200 0.243 0.196 (0.031) n = 42 (0.030) n = 42 GD: Geographic concentration 0.147 0.104 0.252 0.174 (0.016) n = 42 (0.027) n = 42 E(GD): Expected geographic concentration 0.127 0.090 0.203 0.168 (0.014) n = 42 (0.026) n = 42 Cultural concentration *Average percent of original members of same culture as current member 0.197 0.098 0.212 0.111 (0.015) n = 42 (0.017) n = 42 CD: Cultural concentration 0.119 0.136 0.184 0.155 (0.021) n = 42 (0.024) n = 42 E(CD): Expected cultural concentration 0.106 0.078 0.167 0.136 (0.012) n = 42 (0.021) n = 42 . Method of arrival to group . Uninvited to group . Invited to group . Mean & std error . Std dev & no. of obs . Mean & std error . Std dev & no. of obs . Geographic concentration *Average distance to original members from current members (minutes) 12.422 5.342 12.413 4.751 (0.824) n = 42 (0.733) n = 42 *Average % of original members who live within 10 minutes of current member 0.239 0.200 0.243 0.196 (0.031) n = 42 (0.030) n = 42 GD: Geographic concentration 0.147 0.104 0.252 0.174 (0.016) n = 42 (0.027) n = 42 E(GD): Expected geographic concentration 0.127 0.090 0.203 0.168 (0.014) n = 42 (0.026) n = 42 Cultural concentration *Average percent of original members of same culture as current member 0.197 0.098 0.212 0.111 (0.015) n = 42 (0.017) n = 42 CD: Cultural concentration 0.119 0.136 0.184 0.155 (0.021) n = 42 (0.024) n = 42 E(CD): Expected cultural concentration 0.106 0.078 0.167 0.136 (0.012) n = 42 (0.021) n = 42 *Variables with asterisks are the key independent variables used in the specifications in Table 7. All results calculated on original group members only. where si is the share of the group from neighbourhood i and xi is the share of the general population from neighbourhood i. CD and E(CD) are constructed identically to GD and E(GD), except by cultural group rather than neighbourhood. The Alesina index for cultural concentration is equal to the sum of squared shares of each cultural group. Open in new tab This analysis then takes advantage of small sample variation. Since each group has on average 15 uninvited individuals, the idiosyncratic variation proves sufficient to conduct an analysis of the impact of social connections on financial outcomes.18Table 3, for instance, shows the means and standard deviations of the group‐level measures of social connections. The basic model I estimate is of the form: (1) where Y is a financial outcome (either default, savings or dropout), X is one of the social connections measures (either geographic proximity or cultural similarity), and Z is a matrix of neighbourhood and cultural dummies and other demographic information. The neighbourhood and cultural dummies are necessary in order to identify the connectedness of an individual to others in the group properly, rather than alternatively the straight effect of living far from town, or being in a minority cultural group. Using invited individuals poses at least two endogeneity problems to the above specification. First, there is an unobservable selection problem. For example, more sophisticated individuals might be more likely to have successful businesses and repay their loan, as well as more likely to invite their peers into the group. Hence, since individuals tend to invite those who live closer to them, geographic proximity would be correlated with repayment, but not because of improved monitoring or enforcement of the loans. Second, a simultaneity problem exists. Most group lending financial institutions claim to provide two key benefits: higher or smoother consumption by resolving credit market failures and greater social cohesion or empowerment. If this second benefit is true, the correlation between social connections and group outcomes easily could be causal from the other direction. I use two tests to confirm that the group members identified as uninvited are placed randomly into groups. These tests show that there was no assignment to groups on observables, but cannot prove this absolutely, as assignment could have been on unobservables. However, interviews with FINCA and the participants support the claim that the uninvited truly were uninvited, and assignment to groups can be considered random. First, I use a test similar to Ellison and Glaeser (1997) to determine whether the observed geographic dispersion is different from what one would expect to arise randomly,19 as if location were chosen using a dartboard. Ellison and Glaeser uses the following measure of geographic dispersion:20 (2) where si is the share of the group from neighbourhood i and xi is the share of the general population from neighbourhood i.21 E(GD), given random selection, is given by: (3) where n is the number of members in each group. The results of this test support the claim that uninvited, but not invited, individuals select into groups randomly. Table 3 shows these results. The mean of GD is not significantly different from the E(GD) for uninvited (0.147 versus 0.127), but is significantly more for invited (0.252 versus 0.203, significant at 95%). This supports the claim that uninvited individuals are grouped together in a random process with respect to geographic concentration and this also supports the omission of invited individuals from the analysis since they do not pass this test. I conduct a parallel test for the cultural dispersion of each group. For both uninvited and invited, the difference between actual and expected cultural concentration is insignificant (0.119 versus 0.106 for uninvited, and 0.184 versus 0.167 for invited).22 This measure of geographical concentration incorporates the dispersion across neighbourhoods but does not take into account distance between neighbourhoods. To capture distance between individuals, I test whether the percentage of individuals in one’s group who live within a 10‐minute walking circle is greater than the percentage of individuals in the entire sample who live within this same 10‐minute circle. Table 2 shows this comparison: 22.4% of fellow group members live within a 10‐minute circle of each uninvited member, whereas 21.1% of the total sample live within these same 10‐minute circles. These are statistically the same. However, for the invited members, the difference is 20.5% versus 17.2%, significant at 99%. This suggests both that the uninvited are randomly located and that this is a powerful test.23 It found that invited individuals are not randomly located, and in fact are more clustered geographically. I conclude that the allocation of uninvited individuals into groups appears random, allowing the idiosyncratic variation to identify the social connections within the groups. 4. Empirical Results on Lending and Savings Outcomes: 4.1. Default Rate The default rate is perhaps the single most focused on outcome for both researchers and practitioners in analysing the effectiveness of a particular mechanism design. To the extent that default, or specifically the risk of default, leads to credit market failures, default is harmful to social welfare. However, over‐monitoring or over‐punishing might yield higher repayment rates but not maximise social welfare. Regardless, microfinance institutions focus intensely on repayment rates as one of the key, if not only, metrics of financial health and sustainability.24 Social connections could facilitate monitoring and enforcement through reduced cost, increased accuracy of information or higher reputation values. In Ayacucho, monitoring means visiting clients or neighbours of delinquent clients to verify their stories. If someone says they have not repaid due to illness or death in the family, a simple house check or conversation with their neighbours typically can confirm this. Hence, group members who are physically close should be better at monitoring one another. Furthermore, with more mutual acquaintances, the information garnered through the monitoring is likely to be more accurate. This may also cause the threat of enforcement to be more effective since reputation matters more among one’s peers. Stories of repossession of assets are rare; most enforcement activities involve moral suasion via frequent visits to the person’s home or place of business. Cultural homogeneity captures the expected level of these social connections between individuals, as well as the likely extent of mutual acquaintances.25 The dependent variable, default as a percentage of potential loan amount,26 is truncated at zero since most individuals fully repay. Furthermore, almost all individuals pay part of their loan. Default typically begins somewhere in the middle of the loan term, at which point the client stops attending meetings and making her weekly loan payments. The estimating model uses a tobit specification as follows: (4) (5) Defaulti is a latent variable for person i’s default, X is either geographic concentration or cultural similarity, Z is a matrix of control variables, including neighbourhood dummies, year and tenure of group, and education.27 I include neighbourhood dummies in order to account for a potential correlation between density of a neighbourhood and business profitability. For similar reasons, I control for distance to the town centre, where the main market and FINCA office are located. Each measure of social connection is included in a separate specification. For bank‐level specifications, the default is calculated as the average default for individuals in that group. Similarly, most control variables are calculated as the average of the group. When examining the impact of geographic dispersion, I control for average distance to town centre and the percentage of the group that lives within 5 minutes of the town centre. This accounts for the possibility that higher‐concentrated groups are closer to the town centre where the most economic activity takes place. When examining the impact of cultural similarity, controls for the percentage of each group that are indigenous and Western are also included. Table 4 shows the results for the specifications with the individual as the unit of observation. Table 7 column 1 shows the results for the specifications with the group as the unit of observation. The Data AppendixTable 2 shows the typical relationships between the control variables and outcomes of interest. Table 4 Individual Default
OLS, Tobit, and Probit . Dependent variable: % of loan in default at end of cycle . 1st Loan Only . All Loans . OLS . Tobit . Probit . OLS . Tobit . Probit . (1) . (2) . (3) . (4) . (5) . (6) . Distance from individual’s home to original members of group 0.019 0.343 0.019 0.049 0.297 0.040 (0.077) (0.342) (0.019) (0.068) (0.024) (0.027) n = 616 n = 616 n = 616 n = 1,801 n = 1,801 n = 1,801 % of original members within 10‐minute walk of individual’s home −1.536*** −6.077*** −0.284*** −1.556*** −3.754*** −0.367*** (0.391) (1.795) (0.079) (0.370) (1.078) (0.134) n = 616 n = 616 n = 616 n = 1,801 n = 1,801 n = 1,801 % of original members with same culture as individual −0.534* −4.230** −0.200*** −0.396 −1.458 −0.177 (0.301) (1.791) (0.069) (0.308) (1.116) (0.111) n = 616 n = 616 n = 616 n = 1,801 n = 1,801 n = 1,801 . Dependent variable: % of loan in default at end of cycle . 1st Loan Only . All Loans . OLS . Tobit . Probit . OLS . Tobit . Probit . (1) . (2) . (3) . (4) . (5) . (6) . Distance from individual’s home to original members of group 0.019 0.343 0.019 0.049 0.297 0.040 (0.077) (0.342) (0.019) (0.068) (0.024) (0.027) n = 616 n = 616 n = 616 n = 1,801 n = 1,801 n = 1,801 % of original members within 10‐minute walk of individual’s home −1.536*** −6.077*** −0.284*** −1.556*** −3.754*** −0.367*** (0.391) (1.795) (0.079) (0.370) (1.078) (0.134) n = 616 n = 616 n = 616 n = 1,801 n = 1,801 n = 1,801 % of original members with same culture as individual −0.534* −4.230** −0.200*** −0.396 −1.458 −0.177 (0.301) (1.791) (0.069) (0.308) (1.116) (0.111) n = 616 n = 616 n = 616 n = 1,801 n = 1,801 n = 1,801 *** 99% significance; ** 95% significance; * 90% significance Each cell is a separate specification. Standard errors corrected for clustering at the group level in all specifications. Individuals weighted evenly ’all loans’ specifications. Individual level specifications include the following control variables (See AppendixTable 2 for results on control variables): Distance to FINCA (town centre), town dummy, neighbourhood dummies, age, education, marital status, siblings, children, no. in household, year, and tenure of group when individual joined. Loan size estimated using approved loan amount, which is savings balance at end of prior cycle. Open in new tab Table 4 Individual Default
OLS, Tobit, and Probit . Dependent variable: % of loan in default at end of cycle . 1st Loan Only . All Loans . OLS . Tobit . Probit . OLS . Tobit . Probit . (1) . (2) . (3) . (4) . (5) . (6) . Distance from individual’s home to original members of group 0.019 0.343 0.019 0.049 0.297 0.040 (0.077) (0.342) (0.019) (0.068) (0.024) (0.027) n = 616 n = 616 n = 616 n = 1,801 n = 1,801 n = 1,801 % of original members within 10‐minute walk of individual’s home −1.536*** −6.077*** −0.284*** −1.556*** −3.754*** −0.367*** (0.391) (1.795) (0.079) (0.370) (1.078) (0.134) n = 616 n = 616 n = 616 n = 1,801 n = 1,801 n = 1,801 % of original members with same culture as individual −0.534* −4.230** −0.200*** −0.396 −1.458 −0.177 (0.301) (1.791) (0.069) (0.308) (1.116) (0.111) n = 616 n = 616 n = 616 n = 1,801 n = 1,801 n = 1,801 . Dependent variable: % of loan in default at end of cycle . 1st Loan Only . All Loans . OLS . Tobit . Probit . OLS . Tobit . Probit . (1) . (2) . (3) . (4) . (5) . (6) . Distance from individual’s home to original members of group 0.019 0.343 0.019 0.049 0.297 0.040 (0.077) (0.342) (0.019) (0.068) (0.024) (0.027) n = 616 n = 616 n = 616 n = 1,801 n = 1,801 n = 1,801 % of original members within 10‐minute walk of individual’s home −1.536*** −6.077*** −0.284*** −1.556*** −3.754*** −0.367*** (0.391) (1.795) (0.079) (0.370) (1.078) (0.134) n = 616 n = 616 n = 616 n = 1,801 n = 1,801 n = 1,801 % of original members with same culture as individual −0.534* −4.230** −0.200*** −0.396 −1.458 −0.177 (0.301) (1.791) (0.069) (0.308) (1.116) (0.111) n = 616 n = 616 n = 616 n = 1,801 n = 1,801 n = 1,801 *** 99% significance; ** 95% significance; * 90% significance Each cell is a separate specification. Standard errors corrected for clustering at the group level in all specifications. Individuals weighted evenly ’all loans’ specifications. Individual level specifications include the following control variables (See AppendixTable 2 for results on control variables): Distance to FINCA (town centre), town dummy, neighbourhood dummies, age, education, marital status, siblings, children, no. in household, year, and tenure of group when individual joined. Loan size estimated using approved loan amount, which is savings balance at end of prior cycle. Open in new tab Of the 616 uninvited individuals in the sample, 125 had defaulted at the end of their first loan. Of the 245 group observations, 44 had individuals with default at some point in the sample. The default only occurred on the internal loans made from the members’ savings. FINCA had perfect repayment on its loans to the groups. For the primary individual‐level analysis, I use the first loan cycle for each client and not the entire history for four reasons: (1) an attrition bias may exist wherein those who leave the programme are more (or less) connected than those who remain, (2) an attenuation bias exists since the connections to the initial members is a noisy (albeit exogenous) measure of connections to current members, (3) the immediacy of the outcome avoids complications from changes in the group dynamics that occur due to new entrants to the group, and (4) a selection effect influences the further loan cycles, since even though the individuals included in the analysis are sorted randomly, not all of the new entrants are. The new entrants may (and often are) chosen by the peers, and may in turn influence the decision of the original member to repay their loan. By restricting to each individual’s first cycle, this problem is avoided. In a secondary analysis, I expand the analysis to each client’s full history with the project and weight each individual equally. Many of those who dropped out without default perhaps were close to default and left because they feared repercussions from failure to repay the next loan or found the pressure exerted from the first loan too unpleasant. This attrition should understate the predictive power of the social connection measures since these are the individuals for whom social connections potentially matter more. For this reason, as well as the reasons stated above, conducting the analysis on the first loan cycle only is a better specification. Furthermore, since the independent variable is a measure of distance (either geographic or cultural) to the original members of the group, attenuation bias suggests that as the group ages, this becomes a noisier measure of the enforcement and monitoring capabilities of the group. To allow for clearer interpretation, each measure of social connection is included in its own specification and is presented separately as a cell in Table 4.28 Columns 1–3 show the OLS, tobit and probit results, respectively. Both cultural similarity and geographic concentration negatively predict default (significance ranges from 99% to marginally insignificant). The second geographic concentration measure, which captures the number of individuals within a 10‐minute walk, is significant statistically and economically. The first measure, average distance to the original members, is signed intuitively but not significant statistically. This is likely due to the irrelevance of the distance of the further individuals for effective monitoring and enforcement. The economic magnitude of these findings is significant: a shift from the 25th percentile (6%) to the 75th percentile (32%) of the second geographic concentration measure suggests a 7.2 percentage point decrease in the probability of default. Similarly, a shift in the cultural dispersion measure from the 25th percentile (8%) to the 75th percentile (28%) decreases the probability of default by 3.9 percentage points. Comparing column 4 to column 1, column 5 to column 2, and column 6 to 3 shows how the attrition and attenuation bias leads to underestimating the impact of social connection, e.g. in the tobit model the coefficient on cultural similarity falls from −4.23 to −1.46 and the coefficient on the second measure of geographic concentration falls from −6.08 to −3.75. The group level specifications in Table 7 show that both measures of geographic concentration predict default, significant to 95% for the average distance of all members and 99% for the percentage that live within a 10‐minute walk. The cultural concentration, although signed intuitively, is not significant statistically. 4.2. Savings All individuals are required to make weekly savings payments such that over one loan term, the individual has saved 20% of their loan from FINCA (e.g., on a $50 loan, at the end of sixteen weeks the client has $10 in savings). In addition, many clients make voluntary savings payments as well. This saving does not lie idle but rather serves as another source of borrowing for these same members. Thus, for each dollar in savings a member typically has access to two dollars of loans: one dollar from FINCA and one dollar from the savings pool. The return on this savings is the same for the entire group, and is calculated as the profits on loans made minus default, divided by total group savings at the end of the loan cycle. Social connection influences each input into this formula. First, as found above, higher social connection leads to lower default, and since defaults are covered by the group’s savings, lower default directly implies a higher return on savings. Second, not all groups lend out all of their savings. Many groups invest their savings if they do not have safe projects. Again, since higher social connections lead to lower default, groups with higher social connections should lend out a higher percentage of their savings. Any savings not lent out remain with the FINCA cashier and do not earn interest. Table 5 shows the results for the specifications with the individual as the unit of observation, and Table 7 shows the results with the group as the unit of observation. Again, to allow for clearer interpretation, each measure of social connection is included in its own OLS specification. Geographic concentration, but not cultural similarity, produces higher savings. Table 5, columns 1 to 3 show the results using three different savings variables: total savings, mandatory savings and voluntary savings. All specifications include the same controls as were included in the default analysis. The results for total savings show that individuals who live further from others in the group save less, significant at 90% in the individual‐level (Table 5, column 1) and insignificant at the group‐level (Table 7, column 2). A shift from the 25th percentile to the 75th percentile in the average distance to others in the group implies an increase of $13.20 in savings per client in their first 4‐month loan cycle, which is significant given that the mean savings is $58.69. As with default, when the analysis uses the entire tenure of each client, the attrition biases the results downward (see Table 5, column 4 versus column 1). Since mandatory savings are paid in the same installment along with weekly loan payment, predictors of loan repayment also predict mandatory savings deposits. As such, the percentage of the group which lives within 5 minutes is a stronger predictor of individual default and is also the stronger predictor of mandatory savings. Furthermore, since voluntary savings should be driven by return on savings, measures that predict group‐wide return on savings should also predict voluntary savings (see Table 7, columns 4 and 5), significant at 95%. Following this logic, Table 7, column 5 shows that as the group is more concentrated, the return on savings rises (significant at 95%). The coefficient of 0.04 suggests that a shift from the 25th percentile to the 75th percentile in geographic concentration would increase the return on savings by 1.31 percentage points per annum. On $100 in savings, such a change in group composition could produce additional interest earnings approximately equal to the daily wage of a poor entrepreneur. Cultural similarity, although influential on default, does not significantly influence the level of savings.29 One possible explanation is that cultural similarity inspires empathy within cultural groups, but where empathy is asymmetric in gains versus losses. In other words, empathy inspires repayment on loans because failure to do so would harm peers; however, empathy does not inspire higher savings since that has a positive and second‐order benefit to the peers. Indeed, no statistically significant relationship is observed between cultural similarity and voluntary savings. Table 5 Individual Savings
OLS . Total savings deposits . Mandatory savings deposits . Voluntary savings deposits . Individual savings deposits . 1st Loan only . 1st Loan only . 1st Loan only . All loans . (1) . (2) . (3) . (4) . Distance from individual’s home to original members of group −9.681* −3.259* −6.428* −7.192** (4.995) (1.749) (3.702) (3.294) n = 616 n = 616 n = 616 n = 1,801 % of original members within 10‐minute walk of individual’s home 15.623 21.154*** −5.531 27.001 (23.565) (5.762) (22.584) (20.304) n = 616 n = 616 n = 616 n = 1,801 % of original members with same culture as individual −11.573 8.335 −19.908 −13.590 (29.700) (7.392) (26.783) (23.721) n = 616 n = 616 n = 616 n = 1,801 . Total savings deposits . Mandatory savings deposits . Voluntary savings deposits . Individual savings deposits . 1st Loan only . 1st Loan only . 1st Loan only . All loans . (1) . (2) . (3) . (4) . Distance from individual’s home to original members of group −9.681* −3.259* −6.428* −7.192** (4.995) (1.749) (3.702) (3.294) n = 616 n = 616 n = 616 n = 1,801 % of original members within 10‐minute walk of individual’s home 15.623 21.154*** −5.531 27.001 (23.565) (5.762) (22.584) (20.304) n = 616 n = 616 n = 616 n = 1,801 % of original members with same culture as individual −11.573 8.335 −19.908 −13.590 (29.700) (7.392) (26.783) (23.721) n = 616 n = 616 n = 616 n = 1,801 *** 99% significance; ** 95% significance; * 90% significance Each cell is a separate specification. Standard errors corrected for clustering at the group level in all specifications. Individuals weighted evenly ‘all loans’ specifications. Individual‐level specifications include the following control variables: Distance to FINCA (town centre), town dummy, neighbourhood dummies, age, education, marital status, siblings, children, no. in household, year and tenure of group when individual joined. Open in new tab Table 5 Individual Savings
OLS . Total savings deposits . Mandatory savings deposits . Voluntary savings deposits . Individual savings deposits . 1st Loan only . 1st Loan only . 1st Loan only . All loans . (1) . (2) . (3) . (4) . Distance from individual’s home to original members of group −9.681* −3.259* −6.428* −7.192** (4.995) (1.749) (3.702) (3.294) n = 616 n = 616 n = 616 n = 1,801 % of original members within 10‐minute walk of individual’s home 15.623 21.154*** −5.531 27.001 (23.565) (5.762) (22.584) (20.304) n = 616 n = 616 n = 616 n = 1,801 % of original members with same culture as individual −11.573 8.335 −19.908 −13.590 (29.700) (7.392) (26.783) (23.721) n = 616 n = 616 n = 616 n = 1,801 . Total savings deposits . Mandatory savings deposits . Voluntary savings deposits . Individual savings deposits . 1st Loan only . 1st Loan only . 1st Loan only . All loans . (1) . (2) . (3) . (4) . Distance from individual’s home to original members of group −9.681* −3.259* −6.428* −7.192** (4.995) (1.749) (3.702) (3.294) n = 616 n = 616 n = 616 n = 1,801 % of original members within 10‐minute walk of individual’s home 15.623 21.154*** −5.531 27.001 (23.565) (5.762) (22.584) (20.304) n = 616 n = 616 n = 616 n = 1,801 % of original members with same culture as individual −11.573 8.335 −19.908 −13.590 (29.700) (7.392) (26.783) (23.721) n = 616 n = 616 n = 616 n = 1,801 *** 99% significance; ** 95% significance; * 90% significance Each cell is a separate specification. Standard errors corrected for clustering at the group level in all specifications. Individuals weighted evenly ‘all loans’ specifications. Individual‐level specifications include the following control variables: Distance to FINCA (town centre), town dummy, neighbourhood dummies, age, education, marital status, siblings, children, no. in household, year and tenure of group when individual joined. Open in new tab 4.3. Attrition Since financial outcomes are highly accurate predictors of retention, an attrition bias must be considered when examining the predictors of default and savings. Those who remain in the project for many years are different in many respects from those who leave. For FINCA, length of participation in a group varies widely, with attrition likelihood initially high and then falling over time. Attrition falls from 24% after the first loan to 16% after one year and 11% after two years. Default is the strongest predictor of attrition: 71% of those with default left while only 13% of those without default left.30 There is neither a firm rule nor a precise process for deciding whether a group member who defaults should remain part of the group. While FINCA influences this decision to some degree, the ultimate judgment lies with the group as a whole. Table 6, column 1 shows a probit model of the dropout decision. The probit model is specified as follows: (6) where Yi = 1 if an individual drops out and Yi = 0 if an individual remains in the group, Xi is one of the three social connection measures, and Zi is a vector of control variables. These specifications are reported in Table 6, Columns 1 – 3. I test two hypotheses. First, I examine whether social connections influence the decision to leave the programme. I do not observe clearly whether an individual leaves by force or voluntarily (the reality is often murky, so this is not just a data issue). Such an effect can be due to lower utility from attending meetings when there are fewer sociable peers at the meeting. Or it could be the fact that those with higher levels of social connections have more to lose in the case of default, and hence might be quicker to leave. Lastly, it could be mechanical through the effect on default, i.e., low social capital leads to higher default, which leads to clients being forced out. Empirically, all three measures of social connections indicate that higher levels of social connections lead to lower dropout rates, although only two of three are significant statistically. To be able to distinguish between idiosyncratic negative shocks and merely reneging, one needs especially good monitoring. Individuals who are particularly close to each other potentially can arrange such a risk‐sharing arrangement.31 Although no anecdotal evidence exists to suggest that such arrangements are made explicitly ex‐ante, both qualitative and quantitative evidence suggest that they take place ex‐post. FINCA reports instances where individuals vouch for delinquent members in order to prevent them from being forced out of the group. I test for this empirically using a multinomial logit model to estimate the joint decision of default and dropout. I find that those with higher levels of social connections are least likely to default and dropout simultaneously. Table 6, Columns 4–6 report three specifications, one for each social capital measure. The omitted category for the dependent variable is being in default and dropping out. I find that higher social capital leads individuals to be more likely to remain in the programme with default, relative to dropping out with default. This is true for all three measures of social capital. Note that this can be interpreted in several ways. Table 6 Dropout
Probit, Multinomial Logit Key Independent variable:Specification . Distance from individual’s home to original members of group Probit . % of original members within 10‐minute walk of individual’s home Probit . % of original members with same culture as individual Probit . Distance from individual’s home to original members of group . % of original members within 10‐minute walk of individual’s home Multinomial Logit . % of original members with same culture as individual . (1) . (2) . (3) . (4) . (5) . (6) . (A) Probit Specification  Dependent variable = 1 if client dropped out 0.038 −0.012** −0.218* (0.032) (0.006) (0.125) (B) Multinomial Logit, categorical dependent variable  Dependent variable = Dropped out with default Omitted Omitted omitted  Dependent variable = Dropped out without default −0.194 0.124** 1.712 (0.359) (0.059) (1.458)  Dependent variable = Remained in with default −0.935* 0.203** 2.397* (0.540) (0.081) (1.295)  Dependent variable = Remained in without default −0.373 0.144** 2.536** (0.312) (0.061) (1.127)  Observations 635 635 635 635 635 635  Observed probability of dropout 0.235 0.235 0.235  Log‐likelihood −234.98 −232.73 −234.59 −428.38 −426.10 −429.65  Pseudo‐Rsquared 0.321 0.327 0.322 0.304 0.308 0.302  Groups 42 42 42 42 42 42 Key Independent variable:Specification . Distance from individual’s home to original members of group Probit . % of original members within 10‐minute walk of individual’s home Probit . % of original members with same culture as individual Probit . Distance from individual’s home to original members of group . % of original members within 10‐minute walk of individual’s home Multinomial Logit . % of original members with same culture as individual . (1) . (2) . (3) . (4) . (5) . (6) . (A) Probit Specification  Dependent variable = 1 if client dropped out 0.038 −0.012** −0.218* (0.032) (0.006) (0.125) (B) Multinomial Logit, categorical dependent variable  Dependent variable = Dropped out with default Omitted Omitted omitted  Dependent variable = Dropped out without default −0.194 0.124** 1.712 (0.359) (0.059) (1.458)  Dependent variable = Remained in with default −0.935* 0.203** 2.397* (0.540) (0.081) (1.295)  Dependent variable = Remained in without default −0.373 0.144** 2.536** (0.312) (0.061) (1.127)  Observations 635 635 635 635 635 635  Observed probability of dropout 0.235 0.235 0.235  Log‐likelihood −234.98 −232.73 −234.59 −428.38 −426.10 −429.65  Pseudo‐Rsquared 0.321 0.327 0.322 0.304 0.308 0.302  Groups 42 42 42 42 42 42 *** 99% significance; ** 95% significance; * 90% significance. Marginal effects of probit reported. Standard errors corrected for clustering at the group level. Control variables not shown for distance to FINCA (town Centre), town dummy, neighbourhood dummies, age, education, marital status, siblings, children, no. in household, year and tenure of group. Open in new tab Table 6 Dropout
Probit, Multinomial Logit Key Independent variable:Specification . Distance from individual’s home to original members of group Probit . % of original members within 10‐minute walk of individual’s home Probit . % of original members with same culture as individual Probit . Distance from individual’s home to original members of group . % of original members within 10‐minute walk of individual’s home Multinomial Logit . % of original members with same culture as individual . (1) . (2) . (3) . (4) . (5) . (6) . (A) Probit Specification  Dependent variable = 1 if client dropped out 0.038 −0.012** −0.218* (0.032) (0.006) (0.125) (B) Multinomial Logit, categorical dependent variable  Dependent variable = Dropped out with default Omitted Omitted omitted  Dependent variable = Dropped out without default −0.194 0.124** 1.712 (0.359) (0.059) (1.458)  Dependent variable = Remained in with default −0.935* 0.203** 2.397* (0.540) (0.081) (1.295)  Dependent variable = Remained in without default −0.373 0.144** 2.536** (0.312) (0.061) (1.127)  Observations 635 635 635 635 635 635  Observed probability of dropout 0.235 0.235 0.235  Log‐likelihood −234.98 −232.73 −234.59 −428.38 −426.10 −429.65  Pseudo‐Rsquared 0.321 0.327 0.322 0.304 0.308 0.302  Groups 42 42 42 42 42 42 Key Independent variable:Specification . Distance from individual’s home to original members of group Probit . % of original members within 10‐minute walk of individual’s home Probit . % of original members with same culture as individual Probit . Distance from individual’s home to original members of group . % of original members within 10‐minute walk of individual’s home Multinomial Logit . % of original members with same culture as individual . (1) . (2) . (3) . (4) . (5) . (6) . (A) Probit Specification  Dependent variable = 1 if client dropped out 0.038 −0.012** −0.218* (0.032) (0.006) (0.125) (B) Multinomial Logit, categorical dependent variable  Dependent variable = Dropped out with default Omitted Omitted omitted  Dependent variable = Dropped out without default −0.194 0.124** 1.712 (0.359) (0.059) (1.458)  Dependent variable = Remained in with default −0.935* 0.203** 2.397* (0.540) (0.081) (1.295)  Dependent variable = Remained in without default −0.373 0.144** 2.536** (0.312) (0.061) (1.127)  Observations 635 635 635 635 635 635  Observed probability of dropout 0.235 0.235 0.235  Log‐likelihood −234.98 −232.73 −234.59 −428.38 −426.10 −429.65  Pseudo‐Rsquared 0.321 0.327 0.322 0.304 0.308 0.302  Groups 42 42 42 42 42 42 *** 99% significance; ** 95% significance; * 90% significance. Marginal effects of probit reported. Standard errors corrected for clustering at the group level. Control variables not shown for distance to FINCA (town Centre), town dummy, neighbourhood dummies, age, education, marital status, siblings, children, no. in household, year and tenure of group. Open in new tab Table 7 Group Outcomes 
Default, Savings and Dropout
OLS . Average % default in group . Total savings deposits . Mandatory savings deposits . Voluntary savings deposits . % return on savings . % Dropout from programme . (1) . (2) . (3) . (4) . (5) . (6) . Average distance of original members to current, uninvited members 0.156** 0.162 2.688 −2.526 0.000 0.055* (0.066) (3.110) (2.075) (2.170) (0.003) (0.030) n = 245 n = 245 n = 245 n = 245 n = 245 n = 245 Average % of original members within 10‐minute walk of current, uninvited members −1.290*** 57.738** 18.094 39.644** 0.040** −0.441* (0.426) (21.905) (11.620) (15.559) (0.019) (0.265) n = 245 n = 245 n = 245 n = 245 n = 245 n = 245 Average probability that original member is of same culture as current, uninvited member −0.348 −25.835 −3.722 −22.113 −0.017 −0.348* (0.562) (31.488) (13.747) (21.476) (0.028) (0.198) n = 245 n = 245 n = 245 n = 245 n = 245 n = 245 . Average % default in group . Total savings deposits . Mandatory savings deposits . Voluntary savings deposits . % return on savings . % Dropout from programme . (1) . (2) . (3) . (4) . (5) . (6) . Average distance of original members to current, uninvited members 0.156** 0.162 2.688 −2.526 0.000 0.055* (0.066) (3.110) (2.075) (2.170) (0.003) (0.030) n = 245 n = 245 n = 245 n = 245 n = 245 n = 245 Average % of original members within 10‐minute walk of current, uninvited members −1.290*** 57.738** 18.094 39.644** 0.040** −0.441* (0.426) (21.905) (11.620) (15.559) (0.019) (0.265) n = 245 n = 245 n = 245 n = 245 n = 245 n = 245 Average probability that original member is of same culture as current, uninvited member −0.348 −25.835 −3.722 −22.113 −0.017 −0.348* (0.562) (31.488) (13.747) (21.476) (0.028) (0.198) n = 245 n = 245 n = 245 n = 245 n = 245 n = 245 *** 99% significance; ** 95% significance; * 90% significance Each cell is a separate specification. Standard errors corrected for clustering at the group level in all specifications. Groups weighted evenly. Specifications include the following control variables (See AppendixTable 2 for results on control variables): Average distance to FINCA/town centre (for geographic proximity), % who live within 10‐minutes of town centre (for geographic proximity), % indigenous (for cultural similarity), % Western (for cultural similarity), town dummy, average age, average education, average no. in household, average siblings, average no. of children, year and tenure of group Loan size estimated using approved loan amount, which is savings balance at end of prior cycle. Open in new tab Table 7 Group Outcomes 
Default, Savings and Dropout
OLS . Average % default in group . Total savings deposits . Mandatory savings deposits . Voluntary savings deposits . % return on savings . % Dropout from programme . (1) . (2) . (3) . (4) . (5) . (6) . Average distance of original members to current, uninvited members 0.156** 0.162 2.688 −2.526 0.000 0.055* (0.066) (3.110) (2.075) (2.170) (0.003) (0.030) n = 245 n = 245 n = 245 n = 245 n = 245 n = 245 Average % of original members within 10‐minute walk of current, uninvited members −1.290*** 57.738** 18.094 39.644** 0.040** −0.441* (0.426) (21.905) (11.620) (15.559) (0.019) (0.265) n = 245 n = 245 n = 245 n = 245 n = 245 n = 245 Average probability that original member is of same culture as current, uninvited member −0.348 −25.835 −3.722 −22.113 −0.017 −0.348* (0.562) (31.488) (13.747) (21.476) (0.028) (0.198) n = 245 n = 245 n = 245 n = 245 n = 245 n = 245 . Average % default in group . Total savings deposits . Mandatory savings deposits . Voluntary savings deposits . % return on savings . % Dropout from programme . (1) . (2) . (3) . (4) . (5) . (6) . Average distance of original members to current, uninvited members 0.156** 0.162 2.688 −2.526 0.000 0.055* (0.066) (3.110) (2.075) (2.170) (0.003) (0.030) n = 245 n = 245 n = 245 n = 245 n = 245 n = 245 Average % of original members within 10‐minute walk of current, uninvited members −1.290*** 57.738** 18.094 39.644** 0.040** −0.441* (0.426) (21.905) (11.620) (15.559) (0.019) (0.265) n = 245 n = 245 n = 245 n = 245 n = 245 n = 245 Average probability that original member is of same culture as current, uninvited member −0.348 −25.835 −3.722 −22.113 −0.017 −0.348* (0.562) (31.488) (13.747) (21.476) (0.028) (0.198) n = 245 n = 245 n = 245 n = 245 n = 245 n = 245 *** 99% significance; ** 95% significance; * 90% significance Each cell is a separate specification. Standard errors corrected for clustering at the group level in all specifications. Groups weighted evenly. Specifications include the following control variables (See AppendixTable 2 for results on control variables): Average distance to FINCA/town centre (for geographic proximity), % who live within 10‐minutes of town centre (for geographic proximity), % indigenous (for cultural similarity), % Western (for cultural similarity), town dummy, average age, average education, average no. in household, average siblings, average no. of children, year and tenure of group Loan size estimated using approved loan amount, which is savings balance at end of prior cycle. Open in new tab First, individuals with higher levels of social connection perhaps are not being punished after default as much as those with lower levels of social connection. This is consistent with the Rai and Sjöström (2004) model, in which individuals have information that the lender (FINCA) does not. As Rai and Sjöström (2004) discusses, the lending institution (FINCA) provides the framework to facilitate a risk‐sharing arrangement, hence uses the information that peers are able to gather (but FINCA is not) regarding each other’s ability to repay loans. Another possible story is that those with higher social connections receive alternate, but less severe, punishments, or are perhaps unpunishable. The test employed here cannot distinguish between lower dropout due to the higher cost of punishing someone you know or due to successful identification of individuals with true negative shocks. However, qualitative data support the monitoring explanation for the empirical finding in Columns 4–6. In a second survey (discussed in more detail below in Section 5), I asked current members about other members still in the programme who had default. In each instance, I asked an open‐ended question as to why that person was allowed to stay. I recorded the free‐form response of each member, and then categorised the answers. AppendixTable 3 shows these results. In 38 out of 44 instances of individuals having default but remaining, at least one of the other current members reported a negative shock that the individual experienced or reported evidence that the person did undergo some process of explanation to her peers in order to be allowed to remain. Conversations with FINCA Peru management also support the story that individuals with negative and observable shocks are forgiven if the shock is verified by someone else in the group. Hence, these results support the hypothesis that group liability, through the social connections of the members, provides incentives to members to monitor each other’s ability to repay loans. 5. Monitoring and Enforcement Activities 5.1. Data Collected In 2002, after the initial data collection reported above had been completed, I collected further data from FINCA‐Peru in Ayacucho on the monitoring and enforcement activities of clients. In a private interview, I asked each current client from 28 lending groups about each of the individuals that left their group or defaulted on their loan in the prior two loan cycles (the prior 5–8 months). Specifically, the questions included the following: (a) Do you remember the individual? (b) Did the person leave with default? (c) If so, why did the person go into default, and how did you acquire this information? (d) Did you know the person before joining the group? (e) Is the person an extended family member? (f) Is person a close friend? (g) Does the person live near your home? (h) Have you ever visited the business of the other person? In addition, I asked them to compare their current relationship now that the person had left the programme to their relationship with that person when they were in the programme. Specifically, I asked: (a) Do you buy or sell goods from the person more, the same, or less frequently? (b) Has your friendship become stronger, the same or weaker? (c) Has your trust in this person become stronger, the same, or weaker? Table 8 shows the summary tabulations on these questions. The primary goal in doing this was to observe whether default is correlated with the destruction of social relationships. If it is, this supports the idea that social relationships help to enforce group lending contracts through the threat of reputation loss (or other informal punishment paths). The empirical analysis on these data contains two tests. First, I examine monitoring activities by looking at how much accurate information current members have about those who recently left or had default. Second, I examine punishment by looking at whether relationships deteriorated after default. 5.2. Monitoring: Knowledge and Awareness of Causes of Default In a developing country setting, lenders often cannot observe outcomes of borrowers and hence are not able to assess ability to repay loans. Peer lending transfers responsibility of this task onto the peers; they should have access to better information (presumably through stronger social networks) and thus will be able to assess who can and cannot repay (Banerjee et al., 1994). If peers are monitoring each other, then they should have accurate information about each other’s outcomes and reasons for default. I observe this directly, by first asking individuals why their peers left the programme and/or went into default.32 I then examine the accuracy of these answers and whether it improves with stronger social connections. Table 9 presents these results. Next, I ask each respondent whether the individual who left had default when they left and then create a dependent variable for the accuracy of their information. This variable equals to one if the respondent’s information was correct. Columns 1 and 2 in Table 9 report these results. Using a probit specification, I then examine whether those with stronger social connections are more likely to get this question right. Those of similar cultural background are more likely to answer correctly this question (significant at 90% or 95%, depending on whether cultural similarity is scored as a binary variable if similar, or as the absolute difference between the two culture scores). Living closer to each other is not a predictor of having accurate information once I control for knowing the person before hand. Having visited the business of the borrower and knowing the individual beforehand are strong predictors of correctly knowing whether the individual had default. This supports a monitoring theory as a key mechanism through which group lending works. Table 8 Survey of Current Members about Dropouts/Defaults
Summary Tabulations . All . With Default . Without Default . Freq . % . Freq . % . Freq . % . (a)  Observations   Number of lending groups 28   Number of current members 459   Number of dropouts/defaults (all dropouts/ defaults from prior two loan cycles) 575 119 20.7 456 79.3   Number of dropouts 550 117 21.3 433 78.7   Number of pairwise relationships between current members and dropouts/defaulters 9,337 1900 20.3 7439 79.7   Number of defaulters who remained in the programme 44  Basic information on relationships   Number of pairwise relationships 9,337 100.0 1900 20.3 7439 79.7   Instances of recognising name of dropout (hence interview continued) 4,073 43.6 778 19.1 3295 80.9   Lived within 10‐minute walk of the dropout 587 6.3 86 14.7 501 85.3   Family member 121 1.3 14 11.6 107 88.4   Knew the dropout before joining the lending group 431 4.6 59 13.7 372 86.3   Has visited the business of the dropout 107 1.1 7 6.5 100 93.5   Culture score within 1 point 1,344 14.4 225 16.7 1119 83.3  Accuracy of information on default   Current member thought that the dropout left with default 475 5.1 305 64.2 170 35.8   Current member thought that the dropout left without default (& remembered individual) 3,598 88.3 473 13.1 3125 86.9 . All . With Default . Without Default . Freq . % . Freq . % . Freq . % . (a)  Observations   Number of lending groups 28   Number of current members 459   Number of dropouts/defaults (all dropouts/ defaults from prior two loan cycles) 575 119 20.7 456 79.3   Number of dropouts 550 117 21.3 433 78.7   Number of pairwise relationships between current members and dropouts/defaulters 9,337 1900 20.3 7439 79.7   Number of defaulters who remained in the programme 44  Basic information on relationships   Number of pairwise relationships 9,337 100.0 1900 20.3 7439 79.7   Instances of recognising name of dropout (hence interview continued) 4,073 43.6 778 19.1 3295 80.9   Lived within 10‐minute walk of the dropout 587 6.3 86 14.7 501 85.3   Family member 121 1.3 14 11.6 107 88.4   Knew the dropout before joining the lending group 431 4.6 59 13.7 372 86.3   Has visited the business of the dropout 107 1.1 7 6.5 100 93.5   Culture score within 1 point 1,344 14.4 225 16.7 1119 83.3  Accuracy of information on default   Current member thought that the dropout left with default 475 5.1 305 64.2 170 35.8   Current member thought that the dropout left without default (& remembered individual) 3,598 88.3 473 13.1 3125 86.9 (b) .  Change in relationship: Current members reporting about dropouts . Worse . Same . Better . . Obs . Freq . % . Freq . % . Freq . % .   Friendship, if respondent reported that dropout left without default 8,862 245 2.8 8,598 97.0 19 0.2   Friendship, if respondent reported that dropout left with default 475 57 12.0 418 88.0 ‐ 0.0   Trust, if respondent reported that dropout left without default 8,862 87 1.0 8,765 98.9 10 0.1   Trust, if respondent reported that dropout left with default 475 28 5.9 447 94.1 ‐ 0.0   Buying/selling, if respondent reported that dropout left without default 8,862 65 0.7 8,791 99.2 6 0.1   Buying/selling, if respondent reported that dropout left with default 475 10 2.1 464 97.7 1 0.2   Speaking outside of meeting, if respondent reported that dropout left without default 8,862 554 6.3 8,287 93.5 21 0.2   Speaking outside of meeting, if respondent reported that dropout left with default 475 88 18.5 386 81.3 1 0.2 (b) .  Change in relationship: Current members reporting about dropouts . Worse . Same . Better . . Obs . Freq . % . Freq . % . Freq . % .   Friendship, if respondent reported that dropout left without default 8,862 245 2.8 8,598 97.0 19 0.2   Friendship, if respondent reported that dropout left with default 475 57 12.0 418 88.0 ‐ 0.0   Trust, if respondent reported that dropout left without default 8,862 87 1.0 8,765 98.9 10 0.1   Trust, if respondent reported that dropout left with default 475 28 5.9 447 94.1 ‐ 0.0   Buying/selling, if respondent reported that dropout left without default 8,862 65 0.7 8,791 99.2 6 0.1   Buying/selling, if respondent reported that dropout left with default 475 10 2.1 464 97.7 1 0.2   Speaking outside of meeting, if respondent reported that dropout left without default 8,862 554 6.3 8,287 93.5 21 0.2   Speaking outside of meeting, if respondent reported that dropout left with default 475 88 18.5 386 81.3 1 0.2 Open in new tab Table 8 Survey of Current Members about Dropouts/Defaults
Summary Tabulations . All . With Default . Without Default . Freq . % . Freq . % . Freq . % . (a)  Observations   Number of lending groups 28   Number of current members 459   Number of dropouts/defaults (all dropouts/ defaults from prior two loan cycles) 575 119 20.7 456 79.3   Number of dropouts 550 117 21.3 433 78.7   Number of pairwise relationships between current members and dropouts/defaulters 9,337 1900 20.3 7439 79.7   Number of defaulters who remained in the programme 44  Basic information on relationships   Number of pairwise relationships 9,337 100.0 1900 20.3 7439 79.7   Instances of recognising name of dropout (hence interview continued) 4,073 43.6 778 19.1 3295 80.9   Lived within 10‐minute walk of the dropout 587 6.3 86 14.7 501 85.3   Family member 121 1.3 14 11.6 107 88.4   Knew the dropout before joining the lending group 431 4.6 59 13.7 372 86.3   Has visited the business of the dropout 107 1.1 7 6.5 100 93.5   Culture score within 1 point 1,344 14.4 225 16.7 1119 83.3  Accuracy of information on default   Current member thought that the dropout left with default 475 5.1 305 64.2 170 35.8   Current member thought that the dropout left without default (& remembered individual) 3,598 88.3 473 13.1 3125 86.9 . All . With Default . Without Default . Freq . % . Freq . % . Freq . % . (a)  Observations   Number of lending groups 28   Number of current members 459   Number of dropouts/defaults (all dropouts/ defaults from prior two loan cycles) 575 119 20.7 456 79.3   Number of dropouts 550 117 21.3 433 78.7   Number of pairwise relationships between current members and dropouts/defaulters 9,337 1900 20.3 7439 79.7   Number of defaulters who remained in the programme 44  Basic information on relationships   Number of pairwise relationships 9,337 100.0 1900 20.3 7439 79.7   Instances of recognising name of dropout (hence interview continued) 4,073 43.6 778 19.1 3295 80.9   Lived within 10‐minute walk of the dropout 587 6.3 86 14.7 501 85.3   Family member 121 1.3 14 11.6 107 88.4   Knew the dropout before joining the lending group 431 4.6 59 13.7 372 86.3   Has visited the business of the dropout 107 1.1 7 6.5 100 93.5   Culture score within 1 point 1,344 14.4 225 16.7 1119 83.3  Accuracy of information on default   Current member thought that the dropout left with default 475 5.1 305 64.2 170 35.8   Current member thought that the dropout left without default (& remembered individual) 3,598 88.3 473 13.1 3125 86.9 (b) .  Change in relationship: Current members reporting about dropouts . Worse . Same . Better . . Obs . Freq . % . Freq . % . Freq . % .   Friendship, if respondent reported that dropout left without default 8,862 245 2.8 8,598 97.0 19 0.2   Friendship, if respondent reported that dropout left with default 475 57 12.0 418 88.0 ‐ 0.0   Trust, if respondent reported that dropout left without default 8,862 87 1.0 8,765 98.9 10 0.1   Trust, if respondent reported that dropout left with default 475 28 5.9 447 94.1 ‐ 0.0   Buying/selling, if respondent reported that dropout left without default 8,862 65 0.7 8,791 99.2 6 0.1   Buying/selling, if respondent reported that dropout left with default 475 10 2.1 464 97.7 1 0.2   Speaking outside of meeting, if respondent reported that dropout left without default 8,862 554 6.3 8,287 93.5 21 0.2   Speaking outside of meeting, if respondent reported that dropout left with default 475 88 18.5 386 81.3 1 0.2 (b) .  Change in relationship: Current members reporting about dropouts . Worse . Same . Better . . Obs . Freq . % . Freq . % . Freq . % .   Friendship, if respondent reported that dropout left without default 8,862 245 2.8 8,598 97.0 19 0.2   Friendship, if respondent reported that dropout left with default 475 57 12.0 418 88.0 ‐ 0.0   Trust, if respondent reported that dropout left without default 8,862 87 1.0 8,765 98.9 10 0.1   Trust, if respondent reported that dropout left with default 475 28 5.9 447 94.1 ‐ 0.0   Buying/selling, if respondent reported that dropout left without default 8,862 65 0.7 8,791 99.2 6 0.1   Buying/selling, if respondent reported that dropout left with default 475 10 2.1 464 97.7 1 0.2   Speaking outside of meeting, if respondent reported that dropout left without default 8,862 554 6.3 8,287 93.5 21 0.2   Speaking outside of meeting, if respondent reported that dropout left with default 475 88 18.5 386 81.3 1 0.2 Open in new tab Table 9 Loan Monitoring, Direct Evidence
 Results from Survey of Current Members about Recent Defaulters and Dropouts
Probit Binary dependent variables: . Respondent correctly reported whether dropout had default . Respondent reported knowing why dropout had default . (1) . (2) . (3) . (4) . Culture score within 1 point of dropout/defaulter 0.049* 0.028 (0.027) (0.045) Absolute difference between culture scores −0.014** −0.018** (0.006) (0.009) Lives within 10 minutes of other dropout/defaulter 0.008 0.007 −0.022 −0.024 (0.031) (0.031) (0.025) (0.024) Respondent is Indigenous (relative to ‘mixed’) −0.034*** −0.033*** −0.028** −0.027** (0.012) (0.012) (0.014) (0.014) Respondent is Western (relative to ‘mixed’) −0.105*** −0.100*** −0.052*** −0.049*** (0.017) (0.017) (0.016) (0.016) Dropout/defaulter is Indigenous 0.009 0.016 −0.003 −0.008 (0.044) (0.044) (0.050) (0.046) Dropout/defaulter is Western 0.078 0.098 0.045 0.105 (0.059) (0.061) (0.082) (0.096) Extended family member 0.087 0.086 0.002 0.001 (0.095) (0.095) (0.076) (0.073) Knew the dropout/defaulter before being a member 0.491*** 0.491*** 0.173** 0.169** (0.028) (0.028) (0.072) (0.072) Has visited the business of the dropout/defaulter 0.594*** 0.594*** 0.260** 0.267** (0.029) (0.029) (0.133) (0.131) Mean of Binary Dependent Variable 0.3674 0.3674 0.0757 0.0757 Number of Observations 9337 9337 1900 1900 Pseudo R‐Squared 0.0563 0.0566 0.0332 0.0383 Binary dependent variables: . Respondent correctly reported whether dropout had default . Respondent reported knowing why dropout had default . (1) . (2) . (3) . (4) . Culture score within 1 point of dropout/defaulter 0.049* 0.028 (0.027) (0.045) Absolute difference between culture scores −0.014** −0.018** (0.006) (0.009) Lives within 10 minutes of other dropout/defaulter 0.008 0.007 −0.022 −0.024 (0.031) (0.031) (0.025) (0.024) Respondent is Indigenous (relative to ‘mixed’) −0.034*** −0.033*** −0.028** −0.027** (0.012) (0.012) (0.014) (0.014) Respondent is Western (relative to ‘mixed’) −0.105*** −0.100*** −0.052*** −0.049*** (0.017) (0.017) (0.016) (0.016) Dropout/defaulter is Indigenous 0.009 0.016 −0.003 −0.008 (0.044) (0.044) (0.050) (0.046) Dropout/defaulter is Western 0.078 0.098 0.045 0.105 (0.059) (0.061) (0.082) (0.096) Extended family member 0.087 0.086 0.002 0.001 (0.095) (0.095) (0.076) (0.073) Knew the dropout/defaulter before being a member 0.491*** 0.491*** 0.173** 0.169** (0.028) (0.028) (0.072) (0.072) Has visited the business of the dropout/defaulter 0.594*** 0.594*** 0.260** 0.267** (0.029) (0.029) (0.133) (0.131) Mean of Binary Dependent Variable 0.3674 0.3674 0.0757 0.0757 Number of Observations 9337 9337 1900 1900 Pseudo R‐Squared 0.0563 0.0566 0.0332 0.0383 In all specifications, standard errors corrected for clustering across observations regarding the same dropout/default individual. Similar’ culture score defined as within 1 point of each other after scoring each on a scale of 0 to 8, western to indigenous. Indicator variable included to capture any missing culture data. Marginal effects reported for coefficients in probit model. * Significant at 10%; ** Significant at 5%; *** Significant at 1%. Open in new tab Table 9 Loan Monitoring, Direct Evidence
 Results from Survey of Current Members about Recent Defaulters and Dropouts
Probit Binary dependent variables: . Respondent correctly reported whether dropout had default . Respondent reported knowing why dropout had default . (1) . (2) . (3) . (4) . Culture score within 1 point of dropout/defaulter 0.049* 0.028 (0.027) (0.045) Absolute difference between culture scores −0.014** −0.018** (0.006) (0.009) Lives within 10 minutes of other dropout/defaulter 0.008 0.007 −0.022 −0.024 (0.031) (0.031) (0.025) (0.024) Respondent is Indigenous (relative to ‘mixed’) −0.034*** −0.033*** −0.028** −0.027** (0.012) (0.012) (0.014) (0.014) Respondent is Western (relative to ‘mixed’) −0.105*** −0.100*** −0.052*** −0.049*** (0.017) (0.017) (0.016) (0.016) Dropout/defaulter is Indigenous 0.009 0.016 −0.003 −0.008 (0.044) (0.044) (0.050) (0.046) Dropout/defaulter is Western 0.078 0.098 0.045 0.105 (0.059) (0.061) (0.082) (0.096) Extended family member 0.087 0.086 0.002 0.001 (0.095) (0.095) (0.076) (0.073) Knew the dropout/defaulter before being a member 0.491*** 0.491*** 0.173** 0.169** (0.028) (0.028) (0.072) (0.072) Has visited the business of the dropout/defaulter 0.594*** 0.594*** 0.260** 0.267** (0.029) (0.029) (0.133) (0.131) Mean of Binary Dependent Variable 0.3674 0.3674 0.0757 0.0757 Number of Observations 9337 9337 1900 1900 Pseudo R‐Squared 0.0563 0.0566 0.0332 0.0383 Binary dependent variables: . Respondent correctly reported whether dropout had default . Respondent reported knowing why dropout had default . (1) . (2) . (3) . (4) . Culture score within 1 point of dropout/defaulter 0.049* 0.028 (0.027) (0.045) Absolute difference between culture scores −0.014** −0.018** (0.006) (0.009) Lives within 10 minutes of other dropout/defaulter 0.008 0.007 −0.022 −0.024 (0.031) (0.031) (0.025) (0.024) Respondent is Indigenous (relative to ‘mixed’) −0.034*** −0.033*** −0.028** −0.027** (0.012) (0.012) (0.014) (0.014) Respondent is Western (relative to ‘mixed’) −0.105*** −0.100*** −0.052*** −0.049*** (0.017) (0.017) (0.016) (0.016) Dropout/defaulter is Indigenous 0.009 0.016 −0.003 −0.008 (0.044) (0.044) (0.050) (0.046) Dropout/defaulter is Western 0.078 0.098 0.045 0.105 (0.059) (0.061) (0.082) (0.096) Extended family member 0.087 0.086 0.002 0.001 (0.095) (0.095) (0.076) (0.073) Knew the dropout/defaulter before being a member 0.491*** 0.491*** 0.173** 0.169** (0.028) (0.028) (0.072) (0.072) Has visited the business of the dropout/defaulter 0.594*** 0.594*** 0.260** 0.267** (0.029) (0.029) (0.133) (0.131) Mean of Binary Dependent Variable 0.3674 0.3674 0.0757 0.0757 Number of Observations 9337 9337 1900 1900 Pseudo R‐Squared 0.0563 0.0566 0.0332 0.0383 In all specifications, standard errors corrected for clustering across observations regarding the same dropout/default individual. Similar’ culture score defined as within 1 point of each other after scoring each on a scale of 0 to 8, western to indigenous. Indicator variable included to capture any missing culture data. Marginal effects reported for coefficients in probit model. * Significant at 10%; ** Significant at 5%; *** Significant at 1%. Open in new tab Lastly, I ask respondents why an individual did not repay their loan. Restricting the sample to those who did have some default, I find that cultural similarity does correlate with being more likely to know the cause of default (see Table 9, Columns 3 and 4). Moreover, knowing someone before joining is also a positive and stronger (in magnitude) predictor of having this information. Geographic proximity, on the other hand, does not correlate with such knowledge. Again, simple monitoring activities, such as having visited the business of the borrower and knowing the individual beforehand, are strong predictors of knowing why someone had default. 5.3. Punishment: Changes in Relationships with Those Who Dropout of the Programme Peer pressure to repay the loan is often cited as a benefit of group lending (Besley and Coate, 1995; Ghatak and Guinnane, 1999). If this is in fact a mechanism through which group lending generates repayment, then after default one should observe some destruction of social relationships. This is testable using the survey we conducted of current members regarding those who dropped out of the programme. For each individual who left, I asked the current member what has happened to their relationship: Did it remain the same, improve, or worsen? For each question (business transactions, trust and friendship), there are three possible outcomes: worsen (−1), stay the same (0) or improve (1). Tabulations of these responses are shown in Panel B of Table 8. Whereas 2.8% of relationships deteriorated when there was no default, 12.0% deteriorated when there was default. Regarding the reported trust, the difference is 1.0% versus 5.9%, and regarding buying and selling goods from each other, the difference is 0.7% versus 2.1%.33 In addition, when the dropout has default, the respondent is far more likely to report having spoken to that person outside of the bank meeting (e.g., at their business or home), 18.5% versus 6.3% for those without default. This is direct evidence of the monitoring and enforcement activity, since it is those with default that are visited by current members in order to observe their ability to repay and convince them to repay. Regarding improvements in the relationships, those with default are far less likely to experience improvements in their relationships. In fact, no instances of improvement in friendship or trust exist after an individual leaves in default, whereas 0.3% of individuals report an improvement in trust or friendship with an individual after they left without default. Regardless, the small frequency of reported improvements suggests that any short term gain in social networks among current members tends to diminish as individuals leave the programme. 6. Conclusion In response to abysmal repayment rates and unsustainable projects (Adams et al., 1984; Kahlily and Meyer, 1993; Yaron, 1994), the past few decades have seen dramatic changes in the design of credit projects. Four mechanism design changes stand out: (1) the use of group liability to reduce screening, monitoring and delivery costs, (2) the promise of repeat lending as a repayment incentive, (3) the use of regular and more frequent payments, and (4) the offer, or sometimes requirement, of savings. Despite these significant changes, there has been little empirical research conducted to help organisations understand the effect of these innovations (Banerjee, 2002). In particular, the decision of whether to impose joint liability on borrowers is a central choice that many organisations face, yet few have studied empirically. This research finds evidence to support one hypothesis behind group lending: that monitoring and enforcement activities do improve group lending outcomes, and that social connections, broadly defined, facilitate the monitoring and enforcement of joint liability loan contracts. Social connections might have this effect simply through lowering the cost of gathering information about each other (i.e., a monitoring story), or through a social capital story in which more connected individuals trust each other more and value each other’s relationships more. Note that this social capital story encompasses both actions taken to protect one’s relationships, and also actions taken merely out of altruism towards those similar to you. I find that both cultural similarity and geographic concentration lead to improved group lending outcomes (specifically, higher repayment rates savings rates, and returns on savings). There is also suggestive evidence that social connections help groups distinguish between true negative shocks and mere reneging, and that those who have negative shocks are forgiven and thus allowed to continue borrowing. Furthermore, I find direct evidence of effective monitoring, such as knowledge and awareness of each other’s default status and causes, as well as direct evidence of punishment, such as deterioration of relationships. The monitoring activities specifically occur through the same cultural channels found to predict repayment and savings. This further establishes the causal link between cultural similarity and repayment rates and savings. These findings show that peer lending programmes can be more effective if groups are more concentrated geographically and similar culturally. However, the conclusion does not support creating entirely homogenous groups, either geographically or culturally, since extreme situations are not observed in these data. Complete homogeneity might result in collusive activities or may make punishment more difficult. Furthermore, the findings should not be construed as an endorsement of group lending over individual lending, since the sample consists entirely of group borrowers, and those who opt for group lending may be influenced differently by peer pressure. In fact, in work directly comparing group to individual liability, Gine and Karlan (2006) conduct a randomised control trial in which some pre‐existing groups (hence screened by peers with group liability) are converted into individual liability and others remain under group liability. This experiment allows for a clean test of whether monitoring and enforcement (ex post activities) by the peers leads to higher or lower repayment than if conducted by the bank alone. No difference in repayment is found between group liability and individual liability. Thus, although here I find that within group lending programmes, groups with higher social capital perform better than groups with lower social capital, one should not extrapolate from that to conclude that group lending works better than individual lending. Although this article examines the link between informal social connections and repayment of loans, it speaks to a larger issue of how nonmarket institutions and forces can help overcome market failures. These findings support a growing literature on the importance of informal networks for development. Further research to understand how these networks can best be harnessed or better yet, developed, is critical. Data Appendix A.1. Survey Data Collection Process The primary survey data were collected from January to May, 2000 by a team of 10 local surveyors, and in 2002 by a team of 2 surveyors. Three surveys were completed in 2000: an individual survey conducted publicly at the weekly meeting, a private individual survey, and a former member survey, and one private individual survey was conducted in 2002. Table A1 Correlations between Geographic and Cultural Concentrations and Direct Social Capital Measures Tobit . Homes known of members when joined . Number of members with whom client has bought or sold goods . Instances of direct borrowing or lending between members . (1) . (2) . (3) . Average distance of original members of group −0.005*** −0.014*** −0.008** (0.001) (0.004) (0.003) % of original members within 10‐minute walk 1.544* 1.656 2.414 (0.878) (2.037) (3.326) % of original members with same culture 1.857** −1.091 2.186 (0.729) (2.221) (2.059) No. of observations censored at zero 227 300 538 Observations 948 948 946 . Homes known of members when joined . Number of members with whom client has bought or sold goods . Instances of direct borrowing or lending between members . (1) . (2) . (3) . Average distance of original members of group −0.005*** −0.014*** −0.008** (0.001) (0.004) (0.003) % of original members within 10‐minute walk 1.544* 1.656 2.414 (0.878) (2.037) (3.326) % of original members with same culture 1.857** −1.091 2.186 (0.729) (2.221) (2.059) No. of observations censored at zero 227 300 538 Observations 948 948 946 *** 99% significance; ** 95% significance; * 90% significance Each column represents a separate tobit specification with the social interaction measure as the dependent variable. Standard errors corrected for clustering at the group level. Includes controls for neighbourhood, distance to FINCA, and culture score. Open in new tab Table A1 Correlations between Geographic and Cultural Concentrations and Direct Social Capital Measures Tobit . Homes known of members when joined . Number of members with whom client has bought or sold goods . Instances of direct borrowing or lending between members . (1) . (2) . (3) . Average distance of original members of group −0.005*** −0.014*** −0.008** (0.001) (0.004) (0.003) % of original members within 10‐minute walk 1.544* 1.656 2.414 (0.878) (2.037) (3.326) % of original members with same culture 1.857** −1.091 2.186 (0.729) (2.221) (2.059) No. of observations censored at zero 227 300 538 Observations 948 948 946 . Homes known of members when joined . Number of members with whom client has bought or sold goods . Instances of direct borrowing or lending between members . (1) . (2) . (3) . Average distance of original members of group −0.005*** −0.014*** −0.008** (0.001) (0.004) (0.003) % of original members within 10‐minute walk 1.544* 1.656 2.414 (0.878) (2.037) (3.326) % of original members with same culture 1.857** −1.091 2.186 (0.729) (2.221) (2.059) No. of observations censored at zero 227 300 538 Observations 948 948 946 *** 99% significance; ** 95% significance; * 90% significance Each column represents a separate tobit specification with the social interaction measure as the dependent variable. Standard errors corrected for clustering at the group level. Includes controls for neighbourhood, distance to FINCA, and culture score. Open in new tab Table A2 Control Variables Results from Default, Savings, and Dropout Tables
Tobit, OLS, and Probit . Default . Total Savings . Dropout . Typical Results . Probit . Typical Results . OLS . Typical Results . Probit . (1) . (2) . (3) . (4) . (5) . (6) . Indigenous mixed 0.031 neg −6.940 pos 0.071 Western pos* 0.015 pos 2.750 pos 0.009 Distance to town centre mixed 0.007 pos** 0.014* pos 0.010 Ayacucho neg* −0.337* pos 24.992 neg*** 0.490*** No. of children = 0 pos 0.031 neg −4.740 pos 0.068 No. of children pos 0.007 neg −0.529 neg −0.011 Age neg −0.006 pos 0.374 neg −0.004 Age‐squared pos 0.000 neg −0.001 pos 0.000 Spouse pos** 0.057** neg −4.766 pos*** 0.092*** Finished high school neg −0.022 pos 6.724 neg −0.019 No. of siblings neg 0.000 pos* 2.129 neg −0.003 No. of women in household mixed 0.006 neg −2.650 neg** −0.034* No. of men in household mixed −0.005 pos 0.555 neg** −0.043** . Default . Total Savings . Dropout . Typical Results . Probit . Typical Results . OLS . Typical Results . Probit . (1) . (2) . (3) . (4) . (5) . (6) . Indigenous mixed 0.031 neg −6.940 pos 0.071 Western pos* 0.015 pos 2.750 pos 0.009 Distance to town centre mixed 0.007 pos** 0.014* pos 0.010 Ayacucho neg* −0.337* pos 24.992 neg*** 0.490*** No. of children = 0 pos 0.031 neg −4.740 pos 0.068 No. of children pos 0.007 neg −0.529 neg −0.011 Age neg −0.006 pos 0.374 neg −0.004 Age‐squared pos 0.000 neg −0.001 pos 0.000 Spouse pos** 0.057** neg −4.766 pos*** 0.092*** Finished high school neg −0.022 pos 6.724 neg −0.019 No. of siblings neg 0.000 pos* 2.129 neg −0.003 No. of women in household mixed 0.006 neg −2.650 neg** −0.034* No. of men in household mixed −0.005 pos 0.555 neg** −0.043** *** 99% significance; ** 95% significance; * 90% significance ‘Typical Results’ summarises the typical result across the various permutations of specifications, which depend on which measure of social capital is included and, in the case of default, whether a tobit, probit or OLS, is employed. The representative examples in columns 2, 4, and 6 use the second geographic dispersion measure, % who live within a 5‐minute walk. Column 2 corresponds to Table 4, Column 3, Row 2. Column 4 corresponds to Table 5, Column 1, Row 2. Column 6 corresponds to Table 6, Column 5. Open in new tab Table A2 Control Variables Results from Default, Savings, and Dropout Tables
Tobit, OLS, and Probit . Default . Total Savings . Dropout . Typical Results . Probit . Typical Results . OLS . Typical Results . Probit . (1) . (2) . (3) . (4) . (5) . (6) . Indigenous mixed 0.031 neg −6.940 pos 0.071 Western pos* 0.015 pos 2.750 pos 0.009 Distance to town centre mixed 0.007 pos** 0.014* pos 0.010 Ayacucho neg* −0.337* pos 24.992 neg*** 0.490*** No. of children = 0 pos 0.031 neg −4.740 pos 0.068 No. of children pos 0.007 neg −0.529 neg −0.011 Age neg −0.006 pos 0.374 neg −0.004 Age‐squared pos 0.000 neg −0.001 pos 0.000 Spouse pos** 0.057** neg −4.766 pos*** 0.092*** Finished high school neg −0.022 pos 6.724 neg −0.019 No. of siblings neg 0.000 pos* 2.129 neg −0.003 No. of women in household mixed 0.006 neg −2.650 neg** −0.034* No. of men in household mixed −0.005 pos 0.555 neg** −0.043** . Default . Total Savings . Dropout . Typical Results . Probit . Typical Results . OLS . Typical Results . Probit . (1) . (2) . (3) . (4) . (5) . (6) . Indigenous mixed 0.031 neg −6.940 pos 0.071 Western pos* 0.015 pos 2.750 pos 0.009 Distance to town centre mixed 0.007 pos** 0.014* pos 0.010 Ayacucho neg* −0.337* pos 24.992 neg*** 0.490*** No. of children = 0 pos 0.031 neg −4.740 pos 0.068 No. of children pos 0.007 neg −0.529 neg −0.011 Age neg −0.006 pos 0.374 neg −0.004 Age‐squared pos 0.000 neg −0.001 pos 0.000 Spouse pos** 0.057** neg −4.766 pos*** 0.092*** Finished high school neg −0.022 pos 6.724 neg −0.019 No. of siblings neg 0.000 pos* 2.129 neg −0.003 No. of women in household mixed 0.006 neg −2.650 neg** −0.034* No. of men in household mixed −0.005 pos 0.555 neg** −0.043** *** 99% significance; ** 95% significance; * 90% significance ‘Typical Results’ summarises the typical result across the various permutations of specifications, which depend on which measure of social capital is included and, in the case of default, whether a tobit, probit or OLS, is employed. The representative examples in columns 2, 4, and 6 use the second geographic dispersion measure, % who live within a 5‐minute walk. Column 2 corresponds to Table 4, Column 3, Row 2. Column 4 corresponds to Table 5, Column 1, Row 2. Column 6 corresponds to Table 6, Column 5. Open in new tab Table A3 Qualitative Responses on Monitoring of Default Why did X not repay her loan? . Why was X allowed to remain in the group even after she had default? . number . % . . Evidence of monitoring? . Number . % . Do not know 260 54.6 Do not know no 174 52.3 Business was not going well 51 10.7 Bank needed members no 44 13.2 Health 50 10.5 Talked to the members yes 25 7.5 Family problems 46 9.7 Talked to director yes 18 5.4 Travel 31 6.5 Had family problems (sickness, accident) yes 16 4.8 Robbery 13 2.7 She was responsible/ punctual (in paying) yes 14 4.2 She lent it to someone else 8 1.7 She lent it to someone else yes 11 3.3 Legal problems 4 0.8 She got sick yes 7 2.1 Death in family 3 0.6 She wanted to stay no 4 1.2 Did not want to pay 3 0.6 She was traveling yes 4 1.2 Studies 3 0.6 Trust yes 4 1.2 Had other debt 1 0.2 Said they would improve/ be more responsible yes 3 0.9 She was a con artist 1 0.2 Talked to credit officer yes 3 0.9 Stopped working 1 0.2 Robbery yes 2 0.6 Work 1 0.2 Business was not going well yes 1 0.3 476 100.0 Son left for schooling yes 1 0.3 Had an accident yes 1 0.3 Car broke down yes 1 0.3 333 100.0 Why did X not repay her loan? . Why was X allowed to remain in the group even after she had default? . number . % . . Evidence of monitoring? . Number . % . Do not know 260 54.6 Do not know no 174 52.3 Business was not going well 51 10.7 Bank needed members no 44 13.2 Health 50 10.5 Talked to the members yes 25 7.5 Family problems 46 9.7 Talked to director yes 18 5.4 Travel 31 6.5 Had family problems (sickness, accident) yes 16 4.8 Robbery 13 2.7 She was responsible/ punctual (in paying) yes 14 4.2 She lent it to someone else 8 1.7 She lent it to someone else yes 11 3.3 Legal problems 4 0.8 She got sick yes 7 2.1 Death in family 3 0.6 She wanted to stay no 4 1.2 Did not want to pay 3 0.6 She was traveling yes 4 1.2 Studies 3 0.6 Trust yes 4 1.2 Had other debt 1 0.2 Said they would improve/ be more responsible yes 3 0.9 She was a con artist 1 0.2 Talked to credit officer yes 3 0.9 Stopped working 1 0.2 Robbery yes 2 0.6 Work 1 0.2 Business was not going well yes 1 0.3 476 100.0 Son left for schooling yes 1 0.3 Had an accident yes 1 0.3 Car broke down yes 1 0.3 333 100.0 Data come from 2001 survey of current members. Each member was asked privately about the default of all members who had default in the prior two loans. Left side represents 206 different individuals. Of those 206 individuals, 100 had at least one person report why she did not repay her loan. Right side represents 44 different individuals. Of those 44 individuals, 38 had at least one person report a ‘monitoring’ explanation for her remaining in the programme. Open in new tab Table A3 Qualitative Responses on Monitoring of Default Why did X not repay her loan? . Why was X allowed to remain in the group even after she had default? . number . % . . Evidence of monitoring? . Number . % . Do not know 260 54.6 Do not know no 174 52.3 Business was not going well 51 10.7 Bank needed members no 44 13.2 Health 50 10.5 Talked to the members yes 25 7.5 Family problems 46 9.7 Talked to director yes 18 5.4 Travel 31 6.5 Had family problems (sickness, accident) yes 16 4.8 Robbery 13 2.7 She was responsible/ punctual (in paying) yes 14 4.2 She lent it to someone else 8 1.7 She lent it to someone else yes 11 3.3 Legal problems 4 0.8 She got sick yes 7 2.1 Death in family 3 0.6 She wanted to stay no 4 1.2 Did not want to pay 3 0.6 She was traveling yes 4 1.2 Studies 3 0.6 Trust yes 4 1.2 Had other debt 1 0.2 Said they would improve/ be more responsible yes 3 0.9 She was a con artist 1 0.2 Talked to credit officer yes 3 0.9 Stopped working 1 0.2 Robbery yes 2 0.6 Work 1 0.2 Business was not going well yes 1 0.3 476 100.0 Son left for schooling yes 1 0.3 Had an accident yes 1 0.3 Car broke down yes 1 0.3 333 100.0 Why did X not repay her loan? . Why was X allowed to remain in the group even after she had default? . number . % . . Evidence of monitoring? . Number . % . Do not know 260 54.6 Do not know no 174 52.3 Business was not going well 51 10.7 Bank needed members no 44 13.2 Health 50 10.5 Talked to the members yes 25 7.5 Family problems 46 9.7 Talked to director yes 18 5.4 Travel 31 6.5 Had family problems (sickness, accident) yes 16 4.8 Robbery 13 2.7 She was responsible/ punctual (in paying) yes 14 4.2 She lent it to someone else 8 1.7 She lent it to someone else yes 11 3.3 Legal problems 4 0.8 She got sick yes 7 2.1 Death in family 3 0.6 She wanted to stay no 4 1.2 Did not want to pay 3 0.6 She was traveling yes 4 1.2 Studies 3 0.6 Trust yes 4 1.2 Had other debt 1 0.2 Said they would improve/ be more responsible yes 3 0.9 She was a con artist 1 0.2 Talked to credit officer yes 3 0.9 Stopped working 1 0.2 Robbery yes 2 0.6 Work 1 0.2 Business was not going well yes 1 0.3 476 100.0 Son left for schooling yes 1 0.3 Had an accident yes 1 0.3 Car broke down yes 1 0.3 333 100.0 Data come from 2001 survey of current members. Each member was asked privately about the default of all members who had default in the prior two loans. Left side represents 206 different individuals. Of those 206 individuals, 100 had at least one person report why she did not repay her loan. Right side represents 44 different individuals. Of those 44 individuals, 38 had at least one person report a ‘monitoring’ explanation for her remaining in the programme. Open in new tab The public individual survey included questions for which the answers were public information, such as how many homes of the others someone knows, how someone joined the group, and from how many others each person has bought or sold a product or service. These questions were done publicly for three reasons. First, individuals are more likely to speak truthfully for fear of others seeing them be untruthful. Second, other individuals were able to help out with certain answers, such as when respondents had a difficult time understanding the questions. Third, this procedure was significantly faster because each question did not need to be repeated for each and every person. I conducted these surveys with the assistance of one or two employees in order to communicate with the Quechua‐speaking respondents. The private individual survey was conducted privately by one of the 10 surveyors. These questions were more personal and included certain subjective questions for other related research. The former‐member survey sought to gather basic demographic information, such as location of home, cultural characteristics, religious affiliation, and social connections with members of the group. When possible, this information was gathered from current members but otherwise was conducted in the home or business of the former member. A.2. Formulation of Cultural Measures For each individual a simple cultural index was calculated which equally weights four physical characteristics: hair, dress, language, and hat. For each category, the individual receives a zero, one, or two, zero being the most Western and two being the most indigenous. A borrower wearing her hair in braided pigtails receives two points, in a long and flowing style (i.e., probably recently in pigtails or easily put in pigtails) receives one point, and in a short or curled‐styled receives zero points. A Spanish only speaker receives zero points, a bilingual speaker receives one point, and a Quechua‐only speaker receives two points. A woman wearing an indigenous hat receives two points, while a woman with no hat receives no points. Last, a woman wearing a pollera, an indigenous‐style skirt, receives two points, a woman wearing Western‐style clothing receives zero points, and those in the middle receive one point. In total, each person receives between zero and eight points. Individuals with a score of zero or one are categorised as Western, and individuals with a score of five or more are categorised as indigenous. The results reported in this article are robust to various formulations and combinations of these cultural measures. A.3. Relevance of Measures of Social Connections The cultural and geographic concentration indices are correlated with several direct measures of social connections. First, more indigenous individuals tend to sit together at group meetings. This is also true, but to a lesser extent, of the Western individuals. Similarly, individuals tend to sit next to those who live closer to them. Empirically I test this by comparing the mean probability that the person in the next seat is of the same culture to the mean probability that a randomly chosen person from the group is of the same culture. Table 2 shows this comparison in the Seating Arrangements section. For uninvited individuals, the probability rises from 23% to 26% (significant to 95%). For invited individuals the probability rises from 24% to 26% (significant to 95%). Similarly, the same comparison holds with respect to distance between members. Both uninvited and invited members live one minute and two minutes, respectively, closer to the person seated next to them (insignificant for uninvited, significant at 95% for invited).34 Second, participants reported several direct measures of social and business interactions, and these responses were correlated with both cultural and geographic dispersion. Five questions were asked: (1) how many homes they knew of others in the group, (2) from how many others they have purchased a good or service, (3) to how many others they have sold a good or service, (4) from how many others they have borrowed directly, and (5) to how many others they have lent directly. These questions are not good measures for the primary analysis, since the current information is endogenous and the questions asked in recall are both suspect and mostly invariate (few people say they knew anyone when they joined). These questions do, however, provide evidence supporting the social connection measures used in the heart of this article. Geographic dispersion and cultural similarity are correlated with these direct measures of social connection, as shown in Data AppendixTable 1. The first question, how many homes they knew personally, is correlated significantly with both geographic proximity at 99% and cultural similarity at 95% (column 1). The second and third questions (combined) are correlated significantly with geographic proximity at 99% (column 2), but not with cultural similarity. The fourth and fifth questions (combined) on direct borrowing and lending also are correlated significantly with geographic proximity at 95% but not with cultural similarity. Footnotes 1 " This definition is similar to Adler and Kwon’s (2000) internal social capital. 2 " See Zeller (1998), Wydick (1999) and Ahlin and Townsend (2007) (this Feature). 3 " Furthermore, as shown in the Data Appendix, I find that the geographic and cultural concentration indices are correlated with several direct measures of social interaction, such as whether individuals have bought or sold from each other, know each others’ homes, borrow directly from each other, and sit next to each other in group meetings. Karlan (2005) finds that both the cultural and the geographical concentration indices are correlated with cognitive social capital measures, as measured by behaviour in a trust game and a public goods game. See Krishna and Shrader (1999) and Uphoff and Wijayaratna (2000) for a discussion of social capital measures. 4 " See Gine and Karlan (2006) for an experiment in the Philippines in which group liability lending groups were randomly assigned to be converted to individual liability groups, or to remain as‐is. No change in default was observed one year after the conversion. 5 " Hence, the interest received by an individual is equal to her pro‐rata share of the net interest earned by the group. The net interest is equal to the sum of all interest earned by the group on the internal loans minus the sum of all defaults. 6 " See Besley et al. (1993) for a discussion of rotating savings and credit associations, or ROSCAs. 7 " This occurs when there is no current opening in one’s own bank. So an individual may be referred by a client of FINCA but placed into a group without the referring member. 8 " In the entire sample, I observed only three instances of individuals coming in groups of three, and no groups larger than that. 9 " Clearly, this could be causally related in either direction. 10 " Some groups do this more carefully than others. Anecdotally, groups with higher overall repayment rates were more likely to follow through with such proactive monitoring activities. 11 " The fungibility of money potentially makes this particular monitoring action no better than observing that they are working. 12 " In the extreme, family members have been shown consistently to overcome information asymmetry problems, for example, in the used car market. See Pollack (1985). 13 " A third concern involves the formation of small groups within the larger group and then collusion among the members of the smaller group. Suppose a bank has many small, well‐connected groups. Suppose a small group decides to collude whereby one member does not repay while the others report that indeed she has no capacity to repay due to some calamity. In an individual setting with imperfect monitoring, this individual might repay. However, in this setting, the promise of false monitoring by her immediate peers in fact guarantees that she is not monitored. The entire small group could not go into default because then there would be no ‘good’ client to report back to the group. Naturally, if the entire bank divides into mini‐groups with each mini‐group using this strategy, this could lead to the unraveling of the group as a whole. I found no anecdotal evidence to support this possibility at FINCA‐Peru. See Genicot and Ray (2003) for a theoretical discussion of such dynamics. 14 " Ayacucho is a town in the Andes with a population of 150,000. The Shining Path, the communist‐oriented faction from the 1980s civil war, was started in Ayacucho. 15 " The Data Appendix discusses the formulation of these measures and provides evidence supporting the relevance of these as social capital measures. 16 " See http://www.karlan.net for copies of the survey instruments. 17 " Again, due to the endogeneity of the social connections for invited members, the group‐level measure is best calculated by averaging the uninvited person’s connection to the original members of the group, rather than by measuring the connections of current group members to each other. 18 " To verify that the observed variation was consistent with a random process, I conducted a Monte Carlo simulation in which 500 sets of 42 similarly sized groups were formed randomly from the entire sample. I then verified that the actual mean and standard deviation for both the geographic and cultural measures fell within the middle 95% of the distribution of each statistic in the Monte Carlo simulation. Furthermore, the size of the group is not correlated with the measures of geographic and cultural concentration, so the measure does not appear to be a construct of, for example, endogenous group size or missing data. 19 " This is a test of the exogeneity of the social capital measures, whereas the Monte Carlo simulation referred to in the previous footnote verifies that the small sample of each group was not sufficiently large as to remove any variation across groups. Such variation is necessary to identify generate an interesting enough range of observed values for the exogenous variable. 20 " Since this measure does not incorporate distance between neighbourhoods, I do not use it for the primary analysis. 21 " Without data on population by neighbourhood, I use the total sample of all banks to generate general population estimates. The area was then broken into a grid with 43 different neighbourhoods. 22 " Similar to geographic dispersion, the measure of cultural dispersion is where si is the share of the bank with culture score i, and xi is the share of the general population with culture score i. Similarly, E(CD), given random selection, is given by 23 " Note that uninvited individuals are more likely than invited individuals to live near those in other groups. This indicates that uninvited individuals are more likely than invited individuals to come from the centre of town but does not imply any bias in the group formation process itself. 24 " See Morduch (1999a, b). 25 " It does not, however, capture direct travel costs. 26 " Potential loan amount is equal to the client’s last loan amount plus their accumulated savings. 27 " Control variables also include distance to FINCA (town centre), Ayacucho versus Huanta dummy, age, age‐squared, marital status, siblings, children, and number of persons in household. 28 " Results remain similar when all three measures included in each specification. 29 " The results are insignificant but negative, with higher cultural similarity predicting lower savings. When geographic concentration is omitted from the specification, the coefficient for cultural similarity falls to zero when predicting total savings. 30 " Although, most who leave do so without default. See Panel A of Table 8. 31 " See Rai and Sjöström (2001) for a theoretical discussion of how cross‐reporting can efficiently induce repayment. 32 " AppendixTable 3 shows the qualitative responses received to this question. 33 " An alternative explanation for the deterioration of buying/selling is that individuals who leave after default are more likely to close their business. 34 " The distance between invited members could be less than that for uninvited for one of two reasons. First, individuals tend to invite other household members or neighbours to the bank (more so than they do by culture). Second, for logistical reasons, individuals will walk to the meetings with their neighbours or household members. Then, if walking into the meeting in a group, it would be awkward to then separate and sit apart from each other. If an immediate neighbour or household member is in the bank, then one of them most likely invited the other. References Adams , D. , Graham , D. and Von Pischke , J. ( 1984 ). Undermining Rural Development with Cheap Credit , Boulder, CO: Westview Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Adler , P. and Kwon , S. ( 2000 ). ‘Social capital: the good, the bad, and the ugly’, in ( E. Lesser, ed.), Knowledge and Social Capital: Foundations and Applications , pp. 89 – 115 , Boston: Butterworth‐Heineman . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Ahlin , C. and Townsend , R. ( 2007 ). ‘Using repayment data to test across models of joint liability lending’ , Economic Journal , vol. 117, pp. F11 – F51 . OpenURL Placeholder Text WorldCat Alesina , A. , Baqir , R. and Hoxby , C. ( 2004 ). ‘Political jurisdictions in heterogeneous communities’ , Journal of Political Economy , vol. 112 ( 2 ), pp. 348 – 96 . Google Scholar Crossref Search ADS WorldCat Alesina , A. and La Ferrara , E. ( 2000 ). ‘The determinants of trust’ , NBER Working Paper 7621. Armendariz de Aghion , B. ( 1999 ). ‘On the design of a credit agreement with peer monitoring’ , Journal of Development Economics , vol. 60 , pp. 79 – 104 . Google Scholar Crossref Search ADS WorldCat Armendariz de Aghion , B. and Gollier , C. ( 2000 ). ‘Peer group formation in an adverse selection model’ , Economic Journal , vol. 110 ( 465 ), pp. 632 – 43 . Google Scholar Crossref Search ADS WorldCat Banerjee , A. ( 2002 ). ‘The uses of game theory: against a purely positive interpretation of game theoretic results’ , MIT Working Paper. Banerjee , A. , Besley , T. and Guinnane , T. ( 1994 ). ‘Thy neighbor’s keeper: the design of a credit cooperative with theory and a test’ , Quarterly Journal of Economics , vol. 109 ( 2 ), pp. 491 – 515 . Google Scholar Crossref Search ADS WorldCat Besley , T. and Coate , S. ( 1995 ). ‘Group lending, repayment incentives and social collateral’ , Journal of Development Economics , vol. 46 ( 1 ), pp. 1 – 18 . Google Scholar Crossref Search ADS WorldCat Besley , T. , Coate , S. and Loury , G. ( 1993 ). ‘The economics of rotating savings and credit associations’ , American Economic Review , vol. 83 , pp. 782 – 810 . OpenURL Placeholder Text WorldCat Busch , M. and Reinhardt , E. ( 1999 ). ‘Industrial location and protection: the political and economic geography of U.S. nontariff barriers’ , American Journal of Political Science , vol. 43 ( 3 ), pp. 1028 – 50 . Google Scholar Crossref Search ADS WorldCat Cassar , A. , Crowley , L. and Wydick , B. ( 2007 ). ‘The effect of social capital on group loan repayment: evidence from artefactual field experiments’ , Economic Journal , vol. 117, pp. F85 – F106 . OpenURL Placeholder Text WorldCat Collier , P. ( 1998 ). ‘Social capital and poverty ’, World Bank Social Capital Initiative Working Paper No. 4. Cull , R. , Demirgüç‐Kunt , A. and Morduch , J. ( 2007 ). ‘Financial performance and outreach: a global analysis of leading microbanks’ , Economic Journal , vol. 117, pp. F107 – F133 . OpenURL Placeholder Text WorldCat De Soto , H. ( 2000 ). The Mystery of Capital: Why Capitalism Triumphs in the West and Fails Everywhere Else , New York, NY: Basic Books . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Diamond , D.W. ( 1984 ). ‘Financial intermediation and delegated monitoring’ , Review of Economic Studies , vol. 51 ( 3 ), pp. 393 – 414 . Google Scholar Crossref Search ADS WorldCat Ellison , G. and Glaeser , E. ( 1997 ). ‘Geographic concentration in U.S. manufacturing industries: a dartboard approach’ , Journal of Political Economy , vol. 105 ( 5 ), pp. 889 – 927 . Google Scholar Crossref Search ADS WorldCat Genicot , G. and Ray , D. ( 2003 ). ‘Endogenous group formation in risk‐sharing arrangements’ , Review of Economic Studies , vol. 70 ( 1 ), pp. 87 – 113 . Google Scholar Crossref Search ADS WorldCat Ghatak , M. , ( 1999 ). ‘Group lending, local information and peer selection’ , Journal of Development Economics , vol. 60 , pp. 27 – 50 . Google Scholar Crossref Search ADS WorldCat Ghatak , M. , ( 2000 ). ‘Screening by the company you keep: joint liability lending and the peer selection effect’ , Economic Journal , vol. 110 , pp. 601 – 31 . Google Scholar Crossref Search ADS WorldCat Ghatak , M. and Guinnane , T. ( 1999 ). ‘The economics of lending with joint liability: a review of theory and practice’ , Journal of Development Economics , vol. 60 , (October), pp. 195 – 228 . Google Scholar Crossref Search ADS WorldCat Gine , X. , Jakiela , P., Karlan , D. and Morduch , J. ( 2006 ). ‘Microfinance games’ , Yale University, Working Paper, No. 936 Gine , X. and Karlan , D. ( 2006 ). ‘Group versus individual liability: a field experiment in the Philippines’ , Yale University working paper. Glaeser , E. , Laibson , D., Scheinkman , J. and Soutter , C. ( 2000 ). ‘Measuring Trust’ , Quartery Journal of Economics , vol. 115 , (August), pp. 811 – 46 . Google Scholar Crossref Search ADS WorldCat Guiso , L. , Sapienza , P. and Zingales , L. ( 2004 ). ‘The role of social capital in financial development’ , American Economic Review , vol. 94 ( 3 ), (June), pp. 526 – 56 . Google Scholar Crossref Search ADS WorldCat Kahily , M.A. and Meyer , R. ( 1993 ). ‘The political economy of rural loan recovery: evidence from Bangladesh’ , Savings and Development , vol. 17 ( 1 ), pp. 23 – 38 . OpenURL Placeholder Text WorldCat Karlan , D. ( 2005 ). ‘Using experimental economics to measure social capital and predict real financial decisions’ , American Economic Review , vol. 95 ( 5 ), pp. 1688 – 99 . Google Scholar Crossref Search ADS WorldCat Krishna , A. and Shrader , E. ( 1999 ). ‘Social capital assessment tool’ , unpublished manuscript, prepared for the World Bank Conference on Social Capital and Poverty Reduction. La Ferrara , E. ( 2003 ). ‘Kin groups and reciprocity: a model of credit transactions in Ghana’ , American Economic Review , vol. 93 ( 5 ), (December), pp. 1730 – 51 . Google Scholar Crossref Search ADS WorldCat Morduch , J. ( 1999a ). ‘The role of subsidies in microfinance: evidence from the Grameen Bank’ , Journal of Development Economics , vol. 60 , pp. 229 – 48 . Google Scholar Crossref Search ADS WorldCat Morduch , J. ( 1999b ). ‘The microfinance promise’ , Journal of Economic Literature , vol. 37 ( 4 ), (December), pp. 1563 – 2190 . Google Scholar Crossref Search ADS WorldCat Pollack , R.A. , ( 1985 ). ‘A transaction cost approach to families and households’ , Journal of Economic Literature , vol. 23 ( 2 ), (June), pp. 581 – 608 . OpenURL Placeholder Text WorldCat Putnam , R. (with R. Leonardi and R. Y. Nanetti), ( 1993 ). Making Democracy Work , Princeton, NJ: Princeton University Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Rai , A. and Sjöström , T. ( 2004 ). ‘Is Grameen lending efficient? Repayment incentives and insurance in village economies’ , Review of Economic Studies (forthcoming). OpenURL Placeholder Text WorldCat Sadoulet , L. and Carpenter , S. ( 1999 ). ‘Risk‐matching in credit groups: evidence from Guatemala’ , ECARE Working Paper. Sharma , M. and Zeller , M., ( 1997 ). ‘Repayment performance in group‐based credit programs in Bangladesh: an empirical analysis’ , World Development , vol. 25 ( 10 ), (October), pp. 1731 – 42 . Google Scholar Crossref Search ADS WorldCat Stiglitz , J. ( 1990 ). ‘Peer monitoring and credit markets’ , World Bank Economic Review vol. 4 ( 3 ), pp. 351 – 66 . Google Scholar Crossref Search ADS WorldCat Uphoff , N. and Wijayaratna , C.M. ( 2000 ). ‘Demonstrated benefits from social capital: the productivity of farmer organizations in Gal Oya, Sri Lanka’ , World Development , vol. 28 , pp. 11 ‐ ??? . Google Scholar Crossref Search ADS WorldCat Van Bastelaer , T. ( 1999 ). ‘Does social capital facilitate the poor’s access to credit?’ , World Bank Social Capital Initiative Working Paper. Varian , H. ( 1990 ). ‘Monitoring agents with other agents’ , Journal of Institutional and Theoretical Economics , vol. 146 ( 1 ), pp. 153 – 74 . OpenURL Placeholder Text WorldCat Varian , H. ( 2001 ). ‘Economic scene’ , New York Times . November 21, section C, page 2, column 1. OpenURL Placeholder Text WorldCat Wydick , B. ( 1999 ). ‘Can social cohesion be harnessed to repair market failures? Evidence from Group Lending in Guatemala’ , Economic Journal , vol. 109 ( 457 ), pp. 2015 – 75 . Google Scholar Crossref Search ADS WorldCat Yaron , J. ( 1994 ). ‘What makes rural financial markets successful?’ , World Bank Research Observer , vol. 9 ( 1 ), pp. 49 – 70 . Google Scholar Crossref Search ADS WorldCat Zeller , M. ( 1998 ). ‘Determinants of repayment performance in credit groups: the role of program design, intragroup risk pooling, and social cohesion’ , Economic Development and Cultural Change , vol. 46 ( 3 ), pp. 599 – 620 . Google Scholar Crossref Search ADS WorldCat Author notes " I thank my advisors Abhijit Banerjee, Dora Costa, Esther Duflo and Sendhil Mullainathan for their guidance throughout this research. I thank eight anonymous referees, four editors and Niels Hermes, Robert Lensink. I also thank Alexander Aganine, Beatriz Armendariz de Aghion, Robin Burgess, Ben Jones, Leigh Linden, Norman Loayza, Cade Massey, Jonathan Morduch, Ashok Rai, Laura Schechter, Richard Thaler, Ashley Timmer, Chris Udry, Justin Wolfers and Eric Zitzewitz and participants at the 2006 Groningen conference on microfinance, 2001 NEUDC, 2001 LACEA PEG/NIP and 2004 ASSA, and at seminars at Harvard/M.I.T., Princeton, Yale, Maryland, Texas, Williams, Johns Hopkins, UC‐Irvine, Berkeley, UCLA, University of Washington, Michigan, Georgetown, Miami and the Univeristy of Natal, South Africa. Last, but not least, thanks to Iris Lanao, Aquiles Lanao and Morena Lanao from FINCA and my field team, particularly Alcides Medina, Fatima Oriundo and Jeny Yucra. The research reported herein was supported by the SSRC, the Russell Sage Foundation, the M.I.T. George Shultz Fund and the Center for Retirement Research at Boston College pursuant to a grant from the Social Security Administration. All views and errors are mine. © The Author(s). Journal compilation © Royal Economic Society 2007

Greenaway,, David;Kneller,, Richard

doi: 10.1111/j.1468-0297.2007.02018.xpmid: N/A

Abstract A rapidly expanding literature on firm heterogeneity and firm level globalisation strategies has developed over the last decade. There are new insights on why some firms export and others do not, why some firms fail to survive in export markets and some choose to produce overseas rather than export. This article provides a synthesis and evaluation of this literature. It reviews both new theories of firms in an open economy context and the extensive microeconometric evidence base, which has now developed. It highlights the implications of this evidence base for policy and includes an assessment of how the research agenda may evolve. Interest in a range of aspects of firm and plant level adjustment to trade liberalisation and falling trade costs has exploded in recent years, and a new literature is leading to significant re‐thinking of key drivers of the globalisation process: cross‐border trade and cross‐border investment. Like the last revolution in thinking in international trade (sometimes called ‘new trade theory’) which incorporated imperfect competition as a response to empirical observation of intra‐industry trade, this new literature was also triggered by empirical observation, particularly the work of Bernard and Jensen (1995). That paper drew attention to the fact that exporting and non‐exporting firms co‐existed in the same industry but were marked by clear defining characteristics.1 The development of the literature since then into a progressive research programme has been fuelled by two complementary developments. First, major theoretical breakthroughs associated with Melitz (2003), Helpman et al. (2004) and Bernard, Eaton et al. (2003) among others have resulted in new ways of thinking about firm heterogeneity and participation in international markets. Second, the growing availability of micro level datasets has facilitated detailed analysis of firm level adjustment in a large number of countries. One dimension which has received particularly close attention is the relationship between firm level productivity, entry to and survival in export markets. Following Bernard and Jensen (1995) there is now an extensive body of empirical analyses on a large number of industrialised, transitional and developing countries. This addresses not only the characteristics of firms which enter export markets, but also those markers likely to be associated with survival. In addition, recent analysts have turned their attention to the issue of why firms choose to export rather than engage in direct production overseas. For both, the interaction of sunk costs and productivity heterogeneity is key. At the most basic level what this literature adds to our understanding of export behaviour is clear: a combination of sunk costs and heterogeneity in the underlying characteristics of firms explains why not all firms export.2 We have moved from the new trade theory world of representative firms, where all firms export, to one in which firms are heterogeneous and some export, some do not. But the literature goes beyond this, for example to the recognition of potential complementarity between exporting and foreign direct investment (FDI), which challenges the traditional view of multinationals as different from other firms, with exporting and FDI being substitute strategies. Helpman et al. (2004) and others build on the Brainard (1987, 1993) model, which stresses trade‐offs between proximity and concentration, but differ in that the export or FDI choice is predetermined by firm productivity. This provides a basis for understanding globalisation in a broader context and therefore in understanding how changes to the costs of exporting or foreign direct investment change production patterns within industries and across countries. Within this literature, the direction of causation between productivity and internationalisation has been controversial. It has become something of a stylised fact that ex‐ante productivity determines the choice of whether or not to export. In other words, firms have to become more productive before they export and causality runs from productivity to exports. Causality in the opposite direction is less clear. One can think of plausible reasons why a presence in export markets might raise productivity after entry, for instance exposure to best practise technology and learning, but the empirical evidence is mixed. More generally, when studying the determinants of entry and exit from markets, most researchers include measures of international trade in the industry and at the firm level, with the notion that firm death is less likely when the firm is an exporter or in an industry in which exposure to imports is low. Entry and exit then lead to aggregate productivity changes as market shares change. These are important issues from a policy perspective. Export promotion policies of one form or another are pervasive the world over, as a glance at a random sample of World Trade Organisation (WTO) Trade Policy Reviews would confirm. These can take many (transparent and opaque) forms and are often general rather than targeted. The point to note at this stage however is that if not all firms have the appropriate attributes to export, some may simply self select into export subsidies. So the literature is sharpening this policy debate. In this article we provide a critical review of this new literature. Because it is growing so fast, we limit ourselves to firm heterogeneity, exporting and FDI. We begin our appraisal with a review of new theories of the firm and international trade. In Section 2 we then focus on productivity, entry and survival, taking in evidence on exchange rates, agglomeration and changes in the policy environment. Section 3 moves on to exporting and FDI. In addition to evaluating these as alternative strategies we also examine links between the decision to establish production facilities overseas and exporting. In Section 4 we discuss the emerging research agenda including for example new thinking on the boundaries of the firm, outsourcing and offshoring, associated with Antras (2003) and Antras and Helpman (2004). We also look more closely at the policy context in this Section. Section 5 concludes. 1. New Theories of the Firm and International Trade Although the standard workhorse Heckscher‐Ohlin model of international trade has profit maximising firms in the background, operating under constant returns to scale, their boundaries are not well defined and they have no deterministic role in determining the pattern or commodity composition of trade. Economic activity takes place in sectors and international competitiveness is fashioned by relative factor endowments between potential trading partners. ‘New trade theory’ associated with Krugman (1979) and others builds on Dixit‐Stiglitz monopolistic competition and explicitly has firms. However in that framework all firms export, because each produces a unique variety that consumers, who have ‘love of variety’ preference functions, want. In this setting any trade costs just absorb a proportion of a firm’s foreign revenue but do not stop it from exporting. Although new trade theory gave us new insights into the determinants of trade, a world where all firms export is manifestly at odds with what we observe in the real world, where some export and others in the same industry do not. The reason why this happens in the models of Krugman (1979) and others is that firms do not face fixed costs of exporting. The business community would take it as axiomatic that entering export markets incurs sunk costs: market research has to be done; option appraisals completed; existing products have to be modified; new distribution networks set up and so on. Clerides et al. (1998) were one of the first to model this explicitly in a discrete choice framework. In their model, more productive firms with lower marginal costs earn higher gross profits from producing, but not all firms export. Only those with sufficiently high profits to cover the sunk costs do so. This intuitively appealing result leads to the conclusion that self‐selection is fundamental – sunk costs and firm heterogeneity interact and the most productive firms self‐select into export markets.3 Its corollary is that firms have to raise productivity before they enter. So it follows that there is a direct connection between productivity and exporting (but if policymakers want to exploit that, they should target support at potential rather than actual exporters). But this may not be the end of the story. Clerides et al. (1998) also raise the possibility of ‘learning by exporting’. In other words, once a firm has entered export markets, productivity growth may receive a further boost. They model this as an upward shift in the (stochastic) process that determines firms’ productivity and it can be rationalised in various ways. For example, actual involvement in export markets could sharpen incentives to innovate by raising returns to innovation, a possibility modelled by Holmes and Schmitz (2001). A second possibility is that export markets are more competitive than domestic markets, forcing firms to reduce X‐inefficiency. Here, learning results in business process re‐engineering for example. The point is that if learning by exporting occurs, firm productivity may grow after entry as well as before. If this were the case, it provides a plausible mechanism underpinning export‐led growth, though it also complicates the calculation that faces policy makers. Ultimately it is an empirical issue to which we turn in Section 2. Everything we have said so far refers to intra‐firm productivity. At the macro‐level we often associate productivity growth with inter‐sectoral reallocation, classically the shift of resources from agriculture to manufacturing. Can we say anything in the current context about inter‐firm reallocation and industry productivity growth? The pioneering paper here is Melitz (2003), which is set out schematically in Figure 1 from Falvey et al.(2005). He builds a dynamic industry model with heterogeneous firms operating in (Dixit‐Stiglitz) monopolistically competitive industries. Firms incur a fixed cost to export. However, each has to make a productivity draw from an exogenous distribution which determines whether they produce and export, and an endogenously determined productivity threshold determines who does and does not export.4 The interaction of these raises industry productivity. First, there is a rationalisation effect. Exporting increases expected profit, which induces entry, pushes up the productivity threshold for survival and drives out the least efficient firms in a Schumpterian wave of ‘creative destruction’. Clearly this raises average industry productivity. Second, exporting allows the most productive firms to expand and causes less productive firms to contract. The productivity distribution that results is set out in Figure 2. This reallocation effect again acts to raise average industry productivity. This model, despite its microeconomic structure, helps us understand the correlation between exports and growth widely observed at the macro level. Fig. 1. Open in new tabDownload slide Productivity Uncertainty and Firm Entry/Exit Fig. 1. Open in new tabDownload slide Productivity Uncertainty and Firm Entry/Exit Fig. 2. Open in new tabDownload slide Productivity Heterogeneity and Industry Reallocation Fig. 2. Open in new tabDownload slide Productivity Heterogeneity and Industry Reallocation Melitz (2003) is an important model linking heterogeneous firms and industry productivity, with exporting being a key factor. It is not the only model to point to causal links between exporting and industry productivity. This is also a key output of Bernard, Eaton, et al. (2003). Their industrial organisation structure is different but they still derive rationalisation and reallocation effects, however, the former is driven by import competition and the latter from exporters penetrating more markets. Jean (2002) also identifies import driven and export driven contributors to industry productivity growth, in a two‐country setting with differences in relative efficiencies across countries. The core Melitz (2003) model is now being developed in various ways. Helpman et al. (2004) extend it to consider the decision to set up an overseas affiliate. As in Melitz (2003) increased globalisation is likely to lead to firm exit, where the probability is decreasing in whether the firm is an exporter or multinational firm. We return to this in Section 3. A number of recent papers extend Melitz to consider asymmetries between countries. Melitz and Ottaviano (2003) examine differences in the extent of competition between countries (proxied by differences in size) on equilibrium outcomes following trade liberalisation. They find that because competition is ‘tougher’ in the large country, product choice is greater, average productivity higher, but firm survival lower, because new entrants have a higher probability of failure. Trade liberalisation increases competition in both countries thereby raising aggregate productivity but these effects are felt disproportionately in the big country (because it attracts a disproportionate number of firms). In Falvey et al. (2004) countries differ in the efficiency with which they use frontier technology. One interesting finding is that self‐selection is stronger for industries in which the degree of substitution across products is higher. Therefore the probability of firm closure may be negatively correlated with the level of intra‐industry trade. They also find the higher the average efficiency of the country the more likely firms are to survive in the export market, but the less likely they are to survive in the more efficient country, which leads us to expect that trade structure is important. The pattern of trade is determined by the physical size of countries and size of the efficiency gap. For a given efficiency difference, as the size falls, domestic production of the differentiated product falls. By contrast, for a given size difference, as the efficiency gap rises, domestic production of the differentiated product rises. The effect of falling trade costs is to raise the minimum productivity needed to survive‐it raises the self‐selection cut‐off point. This effect is strongest in the more efficient country. The approach of Bernard et al. (2007) is to combine heterogeneous firms with Helpman and Krugman (1985) assumptions of imperfect competition and scale economies, and Heckscher‐Ohlin differences in factor endowments. The model generates predictions about reallocations of resources across industries by firms. Finally, Bernard, Redding and Schott (2003) develop a model to explain an alternative form of exit to death‐industry switching. Productivity levels are again shown to be important, albeit in the context of a closed economy. Here product switching depends on the fixed costs associated with production of different products and heterogeneity in productivity. More productive firms endogenously choose to produce products with higher sunk costs. Although that paper does not identify a role for international competition in firm choices, an effect from increased openness to trade is possible to envisage. Firms alter their output mix towards industries in which they have a comparative advantage and therefore avoid competition from countries in industries where they do not. For OECD countries this is more likely towards the use of technologies with higher costs, where this decision is dependent on firm productivity. As we can see from this brief review of this theoretical literature,5 modelling exporting activity at the firm level throws up a range of possible channels through which exporting might be causally linked to firm and industry productivity. We now turn to the econometric analysis of these issues. 2. Evidence on Productivity, Export Market Entry and Survival As we have seen, theory points to differing performance characteristics of exporters and non‐exporters. But do these differences result from the decision to export or do only ‘good’ firms become exporters? This question of causality between exports and productivity, sparked in part by the ongoing debate over the relationship between openness and growth at the aggregate level6 has, by some margin, received most attention within the micro literature on exports. Thus, we first consider determinants of export market entry and exit as well as evidence on potential feedback from export market participation into firm performance. To provide some structure we begin with evidence relating to participation in export markets more generally. According to Melitz (2003) and others, participation decisions are determined completely by a combination of sunk‐costs and firm productivity. Although in empirical counterparts to this, the set of firm characteristics has been extended to include factors such as size, age, human capital, capital‐intensity, ownership and so on, these predictions are supported by the evidence. While there are differences in the exact methodology employed (the choice over logit or probit models and attempts to correct for bias from inclusion of lagged export status of the firm) results are for the most part robust, a point made forcefully in Wagner (2007). Some if not all firm level variables are strongly correlated with export market entry. It follows that episodes of entry and exit should be predicted by periods of change in these characteristics (which we discuss below). Of the explanatory variables, that relating to persistence (proxied by lagged export status) almost always explains most of the variation in the data. Exporting next period is strongly correlated with exporting this period, even when other determinants of persistence have been controlled for. Its coefficient is usually interpreted as evidence of sunk‐costs. While the exact magnitude varies across studies, past participation increases the probability that a firm will continue to export by between 36% in the US (Bernard and Jensen, 2004a) and 90% in Italy (Bugamelli and Infante, 2002). Entry is therefore likely to be determined by changes in sunk‐costs. As Das et al. (2001) show these are most relevant for those firms who export little, the ‘fringe players in export markets’ (Tybout, 2003). But what are these changes that produce waves of entry and exit? The three contributors most often discussed are exchange rates, policy innovation and agglomeration effects. 2.1. Exchange Rates Macroeconomic evidence on the effect on trade of exchange rate levels and volatility suggests effects that are either significant but small in magnitude, or insignificant (Pozo, 1992; Chowdhury, 1993; Parley and Wei, 1993).7 This implies that exchange rate movements play little or no role as a sunk cost. The micro evidence suggests however that these results are a product of aggregation and exchange rates are important. In the presence of sunk‐costs the export responsiveness of exchange rate changes is likely to be higher amongst current exporters compared to non‐exporters. That is, changes in exchange rates are more likely to lead to changes in the intensive rather than extensive margin. Bernard and Jensen (2004b) for example, study the export response of US manufacturing plants to dollar depreciation in the 1980s, and report that 87% of the expansion was from increased export intensity and 13% from entry of new firms. A similarly strong correlation is reported by Bugamelli and Infante (2002) and Bernard and Jensen (2004a). Whilst useful for future comparative work, this approach does not provide a complete explanation of micro responses for three reasons. First, Das et al. (2004) find significant cross‐industry variation in the effects of exchange rate movements. Simulating a 20% devaluation for three Colombian industries they report that the magnitude of industry response depends on previous export exposure, homogeneity of expected profit flows between firms and their proximity to the export market entry threshold. Ten years after devaluation the industry level effect varies between 14 and 107% (although unfortunately they do not break this into that generated by new entrants and that from existing exporters). Second, devaluation can also lead to substantial exit. According to Blalock and Roy (2007) the 2 to 1 devaluation of the Indonesian rupiah against the US dollar between 1996 and 1998 did not lead to an aggregate export boom. Deeper analysis showed that although there was an expansion of export activity by established exporters and new entry by non‐exporters, new activity was offset by cessation of exporting by previous exporters. Bernard and Jensen (2004b) also find evidence of exit for the US. Blalock and Roy (2007) offer an explanation: firms that ceased exporting were no more likely to report liquidity constraints, or infrastructure problems, compared to firms that continued to export and were no less productive; they were however less likely to be foreign and less likely to have made R&D or training investments. These same variables predicted which firms would become new exporters. An alternative explanation can be found in Maloney and Azevado (1995), where in a model in which firms export to diversify revenue streams fitted to Mexican data, exchange rate volatility and the co‐movement of domestic and foreign demand shocks can lead to counter‐intuitive movements in export volumes following changes in exchange rates. Finally, as we also note below, all of the detailed micro level analysis of exchange rate movements has been of episodes during which the domestic currency depreciated. It is not known whether the effect of appreciation is symmetric. 2.2. Policy Innovation Export decisions are likely to be influenced by the environment in which the firm operates, where policy changes may impact on both intensive and extensive margins. For example, were policy to lead to within firm improvement in productivity perhaps because of increased competition or reduced costs of intermediate imports, it may be more likely that non‐exporters enter export markets, but also easier for current exporters to increase export sales to existing or new markets. Unfortunately however we have little evidence on what aspects of policy are important for export volumes. In fact the evidence is concentrated in just five studies across two types of policy, trade liberalisation and export promotion, the results for which are summarised in Table 1.8 Table 1 Evidence on Policy Intervention and Firm Export Responses Authors . Sample . Policy intervention . Outcome . Alvarez (2004) Chile, 1990–96 Trade shows No effect on export market success Trade missions No effect on export market success Exporter committees Positive effect on export market success Baldwin and Gu (2004) Canada, 1984–96 Canadian‐US commodity tariff rates 4.5% reduction in Canadian tariffs increased the probability of exporting by 24% and export intensity by 46% percent Bernard and Jensen (2004a) US, 1984–92 State expenditures on export promotion Insignificant effect on export market participation Görg et al. (2007) Ireland, 1983–98 Capital grants, training grants, rent subsidies, employment grants, feasibility study grants, technology acquisition grants, loan guarantees, research and development grants In a matched sample large grants lead to additional exports. No evidence of additional entry. Withdrawal of grants does not lead to exit. Authors . Sample . Policy intervention . Outcome . Alvarez (2004) Chile, 1990–96 Trade shows No effect on export market success Trade missions No effect on export market success Exporter committees Positive effect on export market success Baldwin and Gu (2004) Canada, 1984–96 Canadian‐US commodity tariff rates 4.5% reduction in Canadian tariffs increased the probability of exporting by 24% and export intensity by 46% percent Bernard and Jensen (2004a) US, 1984–92 State expenditures on export promotion Insignificant effect on export market participation Görg et al. (2007) Ireland, 1983–98 Capital grants, training grants, rent subsidies, employment grants, feasibility study grants, technology acquisition grants, loan guarantees, research and development grants In a matched sample large grants lead to additional exports. No evidence of additional entry. Withdrawal of grants does not lead to exit. Open in new tab Table 1 Evidence on Policy Intervention and Firm Export Responses Authors . Sample . Policy intervention . Outcome . Alvarez (2004) Chile, 1990–96 Trade shows No effect on export market success Trade missions No effect on export market success Exporter committees Positive effect on export market success Baldwin and Gu (2004) Canada, 1984–96 Canadian‐US commodity tariff rates 4.5% reduction in Canadian tariffs increased the probability of exporting by 24% and export intensity by 46% percent Bernard and Jensen (2004a) US, 1984–92 State expenditures on export promotion Insignificant effect on export market participation Görg et al. (2007) Ireland, 1983–98 Capital grants, training grants, rent subsidies, employment grants, feasibility study grants, technology acquisition grants, loan guarantees, research and development grants In a matched sample large grants lead to additional exports. No evidence of additional entry. Withdrawal of grants does not lead to exit. Authors . Sample . Policy intervention . Outcome . Alvarez (2004) Chile, 1990–96 Trade shows No effect on export market success Trade missions No effect on export market success Exporter committees Positive effect on export market success Baldwin and Gu (2004) Canada, 1984–96 Canadian‐US commodity tariff rates 4.5% reduction in Canadian tariffs increased the probability of exporting by 24% and export intensity by 46% percent Bernard and Jensen (2004a) US, 1984–92 State expenditures on export promotion Insignificant effect on export market participation Görg et al. (2007) Ireland, 1983–98 Capital grants, training grants, rent subsidies, employment grants, feasibility study grants, technology acquisition grants, loan guarantees, research and development grants In a matched sample large grants lead to additional exports. No evidence of additional entry. Withdrawal of grants does not lead to exit. Open in new tab Evidence on trade liberalisation suggests an effect on both intensive and extensive margins.9Blalock and Gertler (2004) find that liberalisation in Indonesia between 1990 to 1996 doubled the number of exporters, while in their study of the effects of NAFTA on Canadian firms, Baldwin and Gu (2003) report increases in both the number of exporters (the share of plants that export increased from 37 to 53% between 1984 and 1990) and export intensity (in 48% of exporters). Using more sophisticated econometric techniques, they find the effect of policy on the export entry decision to be substantial. The 4.5% reduction in Canadian‐US tariffs that occurred increased the probability of exporting by 63%. Export promotion is pervasive, and most governments intervene in one way or another, ranging from providing infrastructure support to offering direct export subsidies. Empirical evidence is again mixed, although this may be a result of both the question asked and level of detail available. Both Bernard and Jensen (2004a) and Alvarez (2004) find an insignificant effect from export promotion schemes, the former for exporters versus non‐exporters; the latter for permanent versus sporadic exporters. Alvarez (2004) does however find differences in detail. Trade missions and trade shows do not increase the probability that a firm will become a permanent exporter, whereas market studies and arranged meetings with clients, authorities and experts do, even when controlling for other firm and industry determinants. Finally, it is worth noting the evidence of self‐selection when evaluating export promotion schemes, a problem thus far not dealt with. Alvarez (2004) finds that established exporters are much more likely to have used public instruments for export promotion than sporadic exporters. More detailed information on the payment of grants to firms is available for Ireland, as discussed by Görg et al. (2007). Using matching to control for selection problems, the authors find only limited success from intervention; large grants can induce existing exporters to expand overseas sales further but fail to encourage additional entry from those that did not previously export. 2.3. Agglomeration Compared to the scrutiny of productivity spillovers, where some 40 studies were evaluated in Görg and Greenaway (2004), the literature on export spillovers is limited. It also concentrates on spillovers from the presence of other multinational firms within the same industry or region. As can be seen from Table 2 only Aitken et al. (1997), Clerides et al. (1998),,Bernard and Jensen (2004a) and Greenaway and Kneller (2003) consider spillovers from other exporters and only Greenaway and Kneller (2003), Sjoholm (2003) and Kneller and Pisu (2007) allow for spillovers from outside the region or industry. Table 2 Evidence on Agglomeration and Firm Export Responses Agglomeration Authors . Sample . Measure of agglomeration . Export Participation* . Export Share . Aitken et al. (1997) Mexico, 1986–89 Foreign MNE share of exports by state & industry + State industry share of national exports − Barrios et al. (2003) Spain, 1990–98 Foreign MNE share of exports by industry 0 0 Foreign MNE share of R&D by industry 0 + Bernard and Jensen (2004a) US, 1984–92 No. of exporters in region 0 No. of exporters in industry − No. of exporters in region & industry 0 Clerides et al. (1998) Colombia, Mexico and Morocco Exporters per industry or region + Greenaway and Kneller (2003) UK, 1989–2002 No. of exporters in industry (SIC‐3) & region + New exporters in industry & region + Greenaway et al. (2004) UK, 1992–96 Foreign MNE share of employment by industry + + Foreign MNE share of exports by industry + + Kneller and Pisu (2005) UK, 1988–98 Horizontal‐industry‐region domestic sales + + Horizontal‐industry‐region export sales + 0 Horizontal industry domestic sales 0 0 Horizontal industry exports 0 + Forward vertical linkages + 0 Backward vertical linkages 0 + Kokko et al. (1997) Uruguay, 1990 Foreign firms created post 1973 + Ruane and Sutherland (2007) Ireland, 1991–98 Foreign MNE share of employment by industry + + Foreign MNE share of exports by industry − − Sjoholm (2003) Indonesia, 1980–91 Foreign MNE share of output by region 0 Swenson (2005)† China, 1997–2003 No. of multinational firms in city + No. of multinational firms in city and industry + Exports by multinational in a city + Exports by multinationals in a city and industry − Relative transaction density in a city + Authors . Sample . Measure of agglomeration . Export Participation* . Export Share . Aitken et al. (1997) Mexico, 1986–89 Foreign MNE share of exports by state & industry + State industry share of national exports − Barrios et al. (2003) Spain, 1990–98 Foreign MNE share of exports by industry 0 0 Foreign MNE share of R&D by industry 0 + Bernard and Jensen (2004a) US, 1984–92 No. of exporters in region 0 No. of exporters in industry − No. of exporters in region & industry 0 Clerides et al. (1998) Colombia, Mexico and Morocco Exporters per industry or region + Greenaway and Kneller (2003) UK, 1989–2002 No. of exporters in industry (SIC‐3) & region + New exporters in industry & region + Greenaway et al. (2004) UK, 1992–96 Foreign MNE share of employment by industry + + Foreign MNE share of exports by industry + + Kneller and Pisu (2005) UK, 1988–98 Horizontal‐industry‐region domestic sales + + Horizontal‐industry‐region export sales + 0 Horizontal industry domestic sales 0 0 Horizontal industry exports 0 + Forward vertical linkages + 0 Backward vertical linkages 0 + Kokko et al. (1997) Uruguay, 1990 Foreign firms created post 1973 + Ruane and Sutherland (2007) Ireland, 1991–98 Foreign MNE share of employment by industry + + Foreign MNE share of exports by industry − − Sjoholm (2003) Indonesia, 1980–91 Foreign MNE share of output by region 0 Swenson (2005)† China, 1997–2003 No. of multinational firms in city + No. of multinational firms in city and industry + Exports by multinational in a city + Exports by multinationals in a city and industry − Relative transaction density in a city + Notes.* + the effect is positive and significant, − the effect is negative and significant, 0 the effect is insignificant and/or changes sign and/or significance through the paper. †These regressions relate to the 2‐stage Probit regressions reported in Table 3 and excluding natural resource intensive sectors. Table 3 Evidence on Export Market Entry Effects and Firms Authors . Sample . Methodology . Pre‐entry difference . Post‐entry difference . Self‐Selection versus Learning Aw et al. (2000) Korea, 1983–93 and Taiwan (China), 1981–91 New Exporters vs. non‐exporters 5+% TFP Taiwan
? TFP Korea 6+%Δ TFP Taiwan
? Δ TFP Korea Baldwin and Gu (2003) Canada, 1974–96 New Exporters vs. non‐exporters 3%ΔLP, 0%ΔTFP 6%ΔLP, 2%ΔTFP Bernard and Jensen (1999) US, 1984–92 New Exporters vs. non‐exporters 6% TFP, 7–8% LP 3%ΔTFP, 3%ΔLP–short run
1%ΔTFP, 1–2%ΔLP–medium run 
1%ΔTFP, 1–2%ΔLP–long run Bernard and Jensen, (2004b) US, 1983–92 New Exporters vs. non‐exporters 3% TFP 6% TFP, 2%ΔTFP Bernard and Wagner (1997) Germany, 1978–92 New Exporters vs. non‐exporters 5% LP, 0%ΔLP 5%ΔLP Castellani (2002) Italy, 1989–94 Exporters vs. non‐exporters + TFP, 0 ΔTFP Damijan et al. (2006) Slovenia, 1994–2002 Exporters vs. non‐exporters 0% TFP 0% TFP t0
0% TFP when export to non‐OECD countries t1
11+% TFP when export to OECD countries t1 Delgado et al. (2002) Spain, 1991–96 New Exporters vs. non‐exporters
Stochastic dominance + TFP 0 ΔTFP Greenaway and Yu (2004) UK chemicals industry, 1990–2000 Dynamic panel 10% increase in exports = 1% TFP, 6% LP Hahn (2004) Korea, 1990–98 New Exporters vs. non‐exporters 4% TFP
 7% TFP Hansson and Lundin (2004) Sweden, 1990–99 New Exporters vs. non‐exporters 0%ΔTFP, 0%ΔLP 0%ΔTFP, 5%ΔLP Isgut (2001) Colombia, 1981–91 New Exporters vs. non‐exporters 20% LP, 4%ΔLP 5%ΔLP1 Kraay (1999) China, 1988–92 Dynamic panel 1s.d. increase in exports = 2% TFP, 13% LP Liu et al. (1999) Taiwan, 1989–93 New Exporters vs. non‐exporters 0%ΔLP, 6%ΔTFP 7%ΔLP, 0%ΔTFP Self‐Selection with Endogenous Productivity Change
Post‐entry effects Arnold and Hussinger (2005a) Germany, 1992–00 Matched D‐i‐D + ΔTFP
non‐matched sample 0%ΔTFP
matched sample Baldwin and Gu (2003) Canada, 1974–96 GMM 3.4% LP, 0% TFP
non‐matched sample 5.5%LP, 1.7%TFP
non‐matched sample
11%LP, 1%TFP
GMM results Bigsten et al. (2000) 4 African countries 1992–95 Dynamic system + ΔTechnical efficiency Blalock and Gertler (2004) Indonesian firms, 1990–96 1.Fixed effects
2. IV–OP & LP
3. timing 3. 0%ΔTFP 1. 5% TFP
2. 2–5% TFP
3. 4%ΔTFP Clerides et al. (1998) Colombia 1981–91, Mexico, 1986–90 and Morocco 1984–91 GMM Colombia + LP 
Mexico 0 LP
Morocco + LP Colombia +LP 
Mexico 0 LP
Morocco + LP
 De Loecker (2004) Slovenia, 1994–2000 Matched D‐i‐D 22%TFP t0 Girma et al. 2003) UK, 1988–98 Matched D‐i‐D 0%ΔTFP, 0%ΔLP 
in matched sample
1%ΔTFP, 0%ΔLP 
in unmatched sample ΔTFP:2%ΔLP:2%
in matched sample
ΔTFP:2%ΔLP:1%
in unmatched sample Greenaway and Kneller (2003) UK, 1989–2002 Matched D‐i‐D 0%ΔTFP, 0%ΔLP 
in matched sample ΔTFP:3%ΔLP:5.5%
Effect stronger when interacted with export share Greenaway, Gullstrand and Knellar(2005) Sweden, 1980–97 Matched D‐i‐D 0%ΔLP 
0%ΔTFP 0%ΔLP
0%ΔTFP Van Biesebroeck (2005) 9 African countries, 1992–96 GMM 35%TFP Wagner (2002) Germany, 1978–89 matching 0% LP 0%ΔLP Self‐Selection with Endogenous Productivity Change 
 Pre‐entry effects Alvarez and López (2005) Chile, 1990–96 Matched D‐i‐D + ΔINV, + ΔSKILL
+ TFP, + LP 
non‐matched results 0%ΔTFP, ?%ΔLP
matched sample López (2004) Chile, 1990–96 New Exporters vs. non‐exporters + ΔINV, 0%ΔDOMSALE
+ ΔTFP Authors . Sample . Methodology . Pre‐entry difference . Post‐entry difference . Self‐Selection versus Learning Aw et al. (2000) Korea, 1983–93 and Taiwan (China), 1981–91 New Exporters vs. non‐exporters 5+% TFP Taiwan
? TFP Korea 6+%Δ TFP Taiwan
? Δ TFP Korea Baldwin and Gu (2003) Canada, 1974–96 New Exporters vs. non‐exporters 3%ΔLP, 0%ΔTFP 6%ΔLP, 2%ΔTFP Bernard and Jensen (1999) US, 1984–92 New Exporters vs. non‐exporters 6% TFP, 7–8% LP 3%ΔTFP, 3%ΔLP–short run
1%ΔTFP, 1–2%ΔLP–medium run 
1%ΔTFP, 1–2%ΔLP–long run Bernard and Jensen, (2004b) US, 1983–92 New Exporters vs. non‐exporters 3% TFP 6% TFP, 2%ΔTFP Bernard and Wagner (1997) Germany, 1978–92 New Exporters vs. non‐exporters 5% LP, 0%ΔLP 5%ΔLP Castellani (2002) Italy, 1989–94 Exporters vs. non‐exporters + TFP, 0 ΔTFP Damijan et al. (2006) Slovenia, 1994–2002 Exporters vs. non‐exporters 0% TFP 0% TFP t0
0% TFP when export to non‐OECD countries t1
11+% TFP when export to OECD countries t1 Delgado et al. (2002) Spain, 1991–96 New Exporters vs. non‐exporters
Stochastic dominance + TFP 0 ΔTFP Greenaway and Yu (2004) UK chemicals industry, 1990–2000 Dynamic panel 10% increase in exports = 1% TFP, 6% LP Hahn (2004) Korea, 1990–98 New Exporters vs. non‐exporters 4% TFP
 7% TFP Hansson and Lundin (2004) Sweden, 1990–99 New Exporters vs. non‐exporters 0%ΔTFP, 0%ΔLP 0%ΔTFP, 5%ΔLP Isgut (2001) Colombia, 1981–91 New Exporters vs. non‐exporters 20% LP, 4%ΔLP 5%ΔLP1 Kraay (1999) China, 1988–92 Dynamic panel 1s.d. increase in exports = 2% TFP, 13% LP Liu et al. (1999) Taiwan, 1989–93 New Exporters vs. non‐exporters 0%ΔLP, 6%ΔTFP 7%ΔLP, 0%ΔTFP Self‐Selection with Endogenous Productivity Change
Post‐entry effects Arnold and Hussinger (2005a) Germany, 1992–00 Matched D‐i‐D + ΔTFP
non‐matched sample 0%ΔTFP
matched sample Baldwin and Gu (2003) Canada, 1974–96 GMM 3.4% LP, 0% TFP
non‐matched sample 5.5%LP, 1.7%TFP
non‐matched sample
11%LP, 1%TFP
GMM results Bigsten et al. (2000) 4 African countries 1992–95 Dynamic system + ΔTechnical efficiency Blalock and Gertler (2004) Indonesian firms, 1990–96 1.Fixed effects
2. IV–OP & LP
3. timing 3. 0%ΔTFP 1. 5% TFP
2. 2–5% TFP
3. 4%ΔTFP Clerides et al. (1998) Colombia 1981–91, Mexico, 1986–90 and Morocco 1984–91 GMM Colombia + LP 
Mexico 0 LP
Morocco + LP Colombia +LP 
Mexico 0 LP
Morocco + LP
 De Loecker (2004) Slovenia, 1994–2000 Matched D‐i‐D 22%TFP t0 Girma et al. 2003) UK, 1988–98 Matched D‐i‐D 0%ΔTFP, 0%ΔLP 
in matched sample
1%ΔTFP, 0%ΔLP 
in unmatched sample ΔTFP:2%ΔLP:2%
in matched sample
ΔTFP:2%ΔLP:1%
in unmatched sample Greenaway and Kneller (2003) UK, 1989–2002 Matched D‐i‐D 0%ΔTFP, 0%ΔLP 
in matched sample ΔTFP:3%ΔLP:5.5%
Effect stronger when interacted with export share Greenaway, Gullstrand and Knellar(2005) Sweden, 1980–97 Matched D‐i‐D 0%ΔLP 
0%ΔTFP 0%ΔLP
0%ΔTFP Van Biesebroeck (2005) 9 African countries, 1992–96 GMM 35%TFP Wagner (2002) Germany, 1978–89 matching 0% LP 0%ΔLP Self‐Selection with Endogenous Productivity Change 
 Pre‐entry effects Alvarez and López (2005) Chile, 1990–96 Matched D‐i‐D + ΔINV, + ΔSKILL
+ TFP, + LP 
non‐matched results 0%ΔTFP, ?%ΔLP
matched sample López (2004) Chile, 1990–96 New Exporters vs. non‐exporters + ΔINV, 0%ΔDOMSALE
+ ΔTFP Notes: Where possible the results refer to a comparison of new exporters versus non‐exporters. TFP = total factor productivity, LP = labour productivity, Δ = growth + the difference relative to the control group is positive and significant, − the difference relative to the control group is negative and significant, 0 the difference relative to the control group is insignificant, ? the difference relative to the control group changes sign and/or significance through the paper. These results refer to firms that survive in export markets, as reported in Table 10 and for value added per worker. Castellani (2002) compares exporters versus non‐exporters. Open in new tab Table 3 Evidence on Export Market Entry Effects and Firms Authors . Sample . Methodology . Pre‐entry difference . Post‐entry difference . Self‐Selection versus Learning Aw et al. (2000) Korea, 1983–93 and Taiwan (China), 1981–91 New Exporters vs. non‐exporters 5+% TFP Taiwan
? TFP Korea 6+%Δ TFP Taiwan
? Δ TFP Korea Baldwin and Gu (2003) Canada, 1974–96 New Exporters vs. non‐exporters 3%ΔLP, 0%ΔTFP 6%ΔLP, 2%ΔTFP Bernard and Jensen (1999) US, 1984–92 New Exporters vs. non‐exporters 6% TFP, 7–8% LP 3%ΔTFP, 3%ΔLP–short run
1%ΔTFP, 1–2%ΔLP–medium run 
1%ΔTFP, 1–2%ΔLP–long run Bernard and Jensen, (2004b) US, 1983–92 New Exporters vs. non‐exporters 3% TFP 6% TFP, 2%ΔTFP Bernard and Wagner (1997) Germany, 1978–92 New Exporters vs. non‐exporters 5% LP, 0%ΔLP 5%ΔLP Castellani (2002) Italy, 1989–94 Exporters vs. non‐exporters + TFP, 0 ΔTFP Damijan et al. (2006) Slovenia, 1994–2002 Exporters vs. non‐exporters 0% TFP 0% TFP t0
0% TFP when export to non‐OECD countries t1
11+% TFP when export to OECD countries t1 Delgado et al. (2002) Spain, 1991–96 New Exporters vs. non‐exporters
Stochastic dominance + TFP 0 ΔTFP Greenaway and Yu (2004) UK chemicals industry, 1990–2000 Dynamic panel 10% increase in exports = 1% TFP, 6% LP Hahn (2004) Korea, 1990–98 New Exporters vs. non‐exporters 4% TFP
 7% TFP Hansson and Lundin (2004) Sweden, 1990–99 New Exporters vs. non‐exporters 0%ΔTFP, 0%ΔLP 0%ΔTFP, 5%ΔLP Isgut (2001) Colombia, 1981–91 New Exporters vs. non‐exporters 20% LP, 4%ΔLP 5%ΔLP1 Kraay (1999) China, 1988–92 Dynamic panel 1s.d. increase in exports = 2% TFP, 13% LP Liu et al. (1999) Taiwan, 1989–93 New Exporters vs. non‐exporters 0%ΔLP, 6%ΔTFP 7%ΔLP, 0%ΔTFP Self‐Selection with Endogenous Productivity Change
Post‐entry effects Arnold and Hussinger (2005a) Germany, 1992–00 Matched D‐i‐D + ΔTFP
non‐matched sample 0%ΔTFP
matched sample Baldwin and Gu (2003) Canada, 1974–96 GMM 3.4% LP, 0% TFP
non‐matched sample 5.5%LP, 1.7%TFP
non‐matched sample
11%LP, 1%TFP
GMM results Bigsten et al. (2000) 4 African countries 1992–95 Dynamic system + ΔTechnical efficiency Blalock and Gertler (2004) Indonesian firms, 1990–96 1.Fixed effects
2. IV–OP & LP
3. timing 3. 0%ΔTFP 1. 5% TFP
2. 2–5% TFP
3. 4%ΔTFP Clerides et al. (1998) Colombia 1981–91, Mexico, 1986–90 and Morocco 1984–91 GMM Colombia + LP 
Mexico 0 LP
Morocco + LP Colombia +LP 
Mexico 0 LP
Morocco + LP
 De Loecker (2004) Slovenia, 1994–2000 Matched D‐i‐D 22%TFP t0 Girma et al. 2003) UK, 1988–98 Matched D‐i‐D 0%ΔTFP, 0%ΔLP 
in matched sample
1%ΔTFP, 0%ΔLP 
in unmatched sample ΔTFP:2%ΔLP:2%
in matched sample
ΔTFP:2%ΔLP:1%
in unmatched sample Greenaway and Kneller (2003) UK, 1989–2002 Matched D‐i‐D 0%ΔTFP, 0%ΔLP 
in matched sample ΔTFP:3%ΔLP:5.5%
Effect stronger when interacted with export share Greenaway, Gullstrand and Knellar(2005) Sweden, 1980–97 Matched D‐i‐D 0%ΔLP 
0%ΔTFP 0%ΔLP
0%ΔTFP Van Biesebroeck (2005) 9 African countries, 1992–96 GMM 35%TFP Wagner (2002) Germany, 1978–89 matching 0% LP 0%ΔLP Self‐Selection with Endogenous Productivity Change 
 Pre‐entry effects Alvarez and López (2005) Chile, 1990–96 Matched D‐i‐D + ΔINV, + ΔSKILL
+ TFP, + LP 
non‐matched results 0%ΔTFP, ?%ΔLP
matched sample López (2004) Chile, 1990–96 New Exporters vs. non‐exporters + ΔINV, 0%ΔDOMSALE
+ ΔTFP Authors . Sample . Methodology . Pre‐entry difference . Post‐entry difference . Self‐Selection versus Learning Aw et al. (2000) Korea, 1983–93 and Taiwan (China), 1981–91 New Exporters vs. non‐exporters 5+% TFP Taiwan
? TFP Korea 6+%Δ TFP Taiwan
? Δ TFP Korea Baldwin and Gu (2003) Canada, 1974–96 New Exporters vs. non‐exporters 3%ΔLP, 0%ΔTFP 6%ΔLP, 2%ΔTFP Bernard and Jensen (1999) US, 1984–92 New Exporters vs. non‐exporters 6% TFP, 7–8% LP 3%ΔTFP, 3%ΔLP–short run
1%ΔTFP, 1–2%ΔLP–medium run 
1%ΔTFP, 1–2%ΔLP–long run Bernard and Jensen, (2004b) US, 1983–92 New Exporters vs. non‐exporters 3% TFP 6% TFP, 2%ΔTFP Bernard and Wagner (1997) Germany, 1978–92 New Exporters vs. non‐exporters 5% LP, 0%ΔLP 5%ΔLP Castellani (2002) Italy, 1989–94 Exporters vs. non‐exporters + TFP, 0 ΔTFP Damijan et al. (2006) Slovenia, 1994–2002 Exporters vs. non‐exporters 0% TFP 0% TFP t0
0% TFP when export to non‐OECD countries t1
11+% TFP when export to OECD countries t1 Delgado et al. (2002) Spain, 1991–96 New Exporters vs. non‐exporters
Stochastic dominance + TFP 0 ΔTFP Greenaway and Yu (2004) UK chemicals industry, 1990–2000 Dynamic panel 10% increase in exports = 1% TFP, 6% LP Hahn (2004) Korea, 1990–98 New Exporters vs. non‐exporters 4% TFP
 7% TFP Hansson and Lundin (2004) Sweden, 1990–99 New Exporters vs. non‐exporters 0%ΔTFP, 0%ΔLP 0%ΔTFP, 5%ΔLP Isgut (2001) Colombia, 1981–91 New Exporters vs. non‐exporters 20% LP, 4%ΔLP 5%ΔLP1 Kraay (1999) China, 1988–92 Dynamic panel 1s.d. increase in exports = 2% TFP, 13% LP Liu et al. (1999) Taiwan, 1989–93 New Exporters vs. non‐exporters 0%ΔLP, 6%ΔTFP 7%ΔLP, 0%ΔTFP Self‐Selection with Endogenous Productivity Change
Post‐entry effects Arnold and Hussinger (2005a) Germany, 1992–00 Matched D‐i‐D + ΔTFP
non‐matched sample 0%ΔTFP
matched sample Baldwin and Gu (2003) Canada, 1974–96 GMM 3.4% LP, 0% TFP
non‐matched sample 5.5%LP, 1.7%TFP
non‐matched sample
11%LP, 1%TFP
GMM results Bigsten et al. (2000) 4 African countries 1992–95 Dynamic system + ΔTechnical efficiency Blalock and Gertler (2004) Indonesian firms, 1990–96 1.Fixed effects
2. IV–OP & LP
3. timing 3. 0%ΔTFP 1. 5% TFP
2. 2–5% TFP
3. 4%ΔTFP Clerides et al. (1998) Colombia 1981–91, Mexico, 1986–90 and Morocco 1984–91 GMM Colombia + LP 
Mexico 0 LP
Morocco + LP Colombia +LP 
Mexico 0 LP
Morocco + LP
 De Loecker (2004) Slovenia, 1994–2000 Matched D‐i‐D 22%TFP t0 Girma et al. 2003) UK, 1988–98 Matched D‐i‐D 0%ΔTFP, 0%ΔLP 
in matched sample
1%ΔTFP, 0%ΔLP 
in unmatched sample ΔTFP:2%ΔLP:2%
in matched sample
ΔTFP:2%ΔLP:1%
in unmatched sample Greenaway and Kneller (2003) UK, 1989–2002 Matched D‐i‐D 0%ΔTFP, 0%ΔLP 
in matched sample ΔTFP:3%ΔLP:5.5%
Effect stronger when interacted with export share Greenaway, Gullstrand and Knellar(2005) Sweden, 1980–97 Matched D‐i‐D 0%ΔLP 
0%ΔTFP 0%ΔLP
0%ΔTFP Van Biesebroeck (2005) 9 African countries, 1992–96 GMM 35%TFP Wagner (2002) Germany, 1978–89 matching 0% LP 0%ΔLP Self‐Selection with Endogenous Productivity Change 
 Pre‐entry effects Alvarez and López (2005) Chile, 1990–96 Matched D‐i‐D + ΔINV, + ΔSKILL
+ TFP, + LP 
non‐matched results 0%ΔTFP, ?%ΔLP
matched sample López (2004) Chile, 1990–96 New Exporters vs. non‐exporters + ΔINV, 0%ΔDOMSALE
+ ΔTFP Notes: Where possible the results refer to a comparison of new exporters versus non‐exporters. TFP = total factor productivity, LP = labour productivity, Δ = growth + the difference relative to the control group is positive and significant, − the difference relative to the control group is negative and significant, 0 the difference relative to the control group is insignificant, ? the difference relative to the control group changes sign and/or significance through the paper. These results refer to firms that survive in export markets, as reported in Table 10 and for value added per worker. Castellani (2002) compares exporters versus non‐exporters. Open in new tab Open in new tab Table 2 Evidence on Agglomeration and Firm Export Responses Agglomeration Authors . Sample . Measure of agglomeration . Export Participation* . Export Share . Aitken et al. (1997) Mexico, 1986–89 Foreign MNE share of exports by state & industry + State industry share of national exports − Barrios et al. (2003) Spain, 1990–98 Foreign MNE share of exports by industry 0 0 Foreign MNE share of R&D by industry 0 + Bernard and Jensen (2004a) US, 1984–92 No. of exporters in region 0 No. of exporters in industry − No. of exporters in region & industry 0 Clerides et al. (1998) Colombia, Mexico and Morocco Exporters per industry or region + Greenaway and Kneller (2003) UK, 1989–2002 No. of exporters in industry (SIC‐3) & region + New exporters in industry & region + Greenaway et al. (2004) UK, 1992–96 Foreign MNE share of employment by industry + + Foreign MNE share of exports by industry + + Kneller and Pisu (2005) UK, 1988–98 Horizontal‐industry‐region domestic sales + + Horizontal‐industry‐region export sales + 0 Horizontal industry domestic sales 0 0 Horizontal industry exports 0 + Forward vertical linkages + 0 Backward vertical linkages 0 + Kokko et al. (1997) Uruguay, 1990 Foreign firms created post 1973 + Ruane and Sutherland (2007) Ireland, 1991–98 Foreign MNE share of employment by industry + + Foreign MNE share of exports by industry − − Sjoholm (2003) Indonesia, 1980–91 Foreign MNE share of output by region 0 Swenson (2005)† China, 1997–2003 No. of multinational firms in city + No. of multinational firms in city and industry + Exports by multinational in a city + Exports by multinationals in a city and industry − Relative transaction density in a city + Authors . Sample . Measure of agglomeration . Export Participation* . Export Share . Aitken et al. (1997) Mexico, 1986–89 Foreign MNE share of exports by state & industry + State industry share of national exports − Barrios et al. (2003) Spain, 1990–98 Foreign MNE share of exports by industry 0 0 Foreign MNE share of R&D by industry 0 + Bernard and Jensen (2004a) US, 1984–92 No. of exporters in region 0 No. of exporters in industry − No. of exporters in region & industry 0 Clerides et al. (1998) Colombia, Mexico and Morocco Exporters per industry or region + Greenaway and Kneller (2003) UK, 1989–2002 No. of exporters in industry (SIC‐3) & region + New exporters in industry & region + Greenaway et al. (2004) UK, 1992–96 Foreign MNE share of employment by industry + + Foreign MNE share of exports by industry + + Kneller and Pisu (2005) UK, 1988–98 Horizontal‐industry‐region domestic sales + + Horizontal‐industry‐region export sales + 0 Horizontal industry domestic sales 0 0 Horizontal industry exports 0 + Forward vertical linkages + 0 Backward vertical linkages 0 + Kokko et al. (1997) Uruguay, 1990 Foreign firms created post 1973 + Ruane and Sutherland (2007) Ireland, 1991–98 Foreign MNE share of employment by industry + + Foreign MNE share of exports by industry − − Sjoholm (2003) Indonesia, 1980–91 Foreign MNE share of output by region 0 Swenson (2005)† China, 1997–2003 No. of multinational firms in city + No. of multinational firms in city and industry + Exports by multinational in a city + Exports by multinationals in a city and industry − Relative transaction density in a city + Notes.* + the effect is positive and significant, − the effect is negative and significant, 0 the effect is insignificant and/or changes sign and/or significance through the paper. †These regressions relate to the 2‐stage Probit regressions reported in Table 3 and excluding natural resource intensive sectors. Table 3 Evidence on Export Market Entry Effects and Firms Authors . Sample . Methodology . Pre‐entry difference . Post‐entry difference . Self‐Selection versus Learning Aw et al. (2000) Korea, 1983–93 and Taiwan (China), 1981–91 New Exporters vs. non‐exporters 5+% TFP Taiwan
? TFP Korea 6+%Δ TFP Taiwan
? Δ TFP Korea Baldwin and Gu (2003) Canada, 1974–96 New Exporters vs. non‐exporters 3%ΔLP, 0%ΔTFP 6%ΔLP, 2%ΔTFP Bernard and Jensen (1999) US, 1984–92 New Exporters vs. non‐exporters 6% TFP, 7–8% LP 3%ΔTFP, 3%ΔLP–short run
1%ΔTFP, 1–2%ΔLP–medium run 
1%ΔTFP, 1–2%ΔLP–long run Bernard and Jensen, (2004b) US, 1983–92 New Exporters vs. non‐exporters 3% TFP 6% TFP, 2%ΔTFP Bernard and Wagner (1997) Germany, 1978–92 New Exporters vs. non‐exporters 5% LP, 0%ΔLP 5%ΔLP Castellani (2002) Italy, 1989–94 Exporters vs. non‐exporters + TFP, 0 ΔTFP Damijan et al. (2006) Slovenia, 1994–2002 Exporters vs. non‐exporters 0% TFP 0% TFP t0
0% TFP when export to non‐OECD countries t1
11+% TFP when export to OECD countries t1 Delgado et al. (2002) Spain, 1991–96 New Exporters vs. non‐exporters
Stochastic dominance + TFP 0 ΔTFP Greenaway and Yu (2004) UK chemicals industry, 1990–2000 Dynamic panel 10% increase in exports = 1% TFP, 6% LP Hahn (2004) Korea, 1990–98 New Exporters vs. non‐exporters 4% TFP
 7% TFP Hansson and Lundin (2004) Sweden, 1990–99 New Exporters vs. non‐exporters 0%ΔTFP, 0%ΔLP 0%ΔTFP, 5%ΔLP Isgut (2001) Colombia, 1981–91 New Exporters vs. non‐exporters 20% LP, 4%ΔLP 5%ΔLP1 Kraay (1999) China, 1988–92 Dynamic panel 1s.d. increase in exports = 2% TFP, 13% LP Liu et al. (1999) Taiwan, 1989–93 New Exporters vs. non‐exporters 0%ΔLP, 6%ΔTFP 7%ΔLP, 0%ΔTFP Self‐Selection with Endogenous Productivity Change
Post‐entry effects Arnold and Hussinger (2005a) Germany, 1992–00 Matched D‐i‐D + ΔTFP
non‐matched sample 0%ΔTFP
matched sample Baldwin and Gu (2003) Canada, 1974–96 GMM 3.4% LP, 0% TFP
non‐matched sample 5.5%LP, 1.7%TFP
non‐matched sample
11%LP, 1%TFP
GMM results Bigsten et al. (2000) 4 African countries 1992–95 Dynamic system + ΔTechnical efficiency Blalock and Gertler (2004) Indonesian firms, 1990–96 1.Fixed effects
2. IV–OP & LP
3. timing 3. 0%ΔTFP 1. 5% TFP
2. 2–5% TFP
3. 4%ΔTFP Clerides et al. (1998) Colombia 1981–91, Mexico, 1986–90 and Morocco 1984–91 GMM Colombia + LP 
Mexico 0 LP
Morocco + LP Colombia +LP 
Mexico 0 LP
Morocco + LP
 De Loecker (2004) Slovenia, 1994–2000 Matched D‐i‐D 22%TFP t0 Girma et al. 2003) UK, 1988–98 Matched D‐i‐D 0%ΔTFP, 0%ΔLP 
in matched sample
1%ΔTFP, 0%ΔLP 
in unmatched sample ΔTFP:2%ΔLP:2%
in matched sample
ΔTFP:2%ΔLP:1%
in unmatched sample Greenaway and Kneller (2003) UK, 1989–2002 Matched D‐i‐D 0%ΔTFP, 0%ΔLP 
in matched sample ΔTFP:3%ΔLP:5.5%
Effect stronger when interacted with export share Greenaway, Gullstrand and Knellar(2005) Sweden, 1980–97 Matched D‐i‐D 0%ΔLP 
0%ΔTFP 0%ΔLP
0%ΔTFP Van Biesebroeck (2005) 9 African countries, 1992–96 GMM 35%TFP Wagner (2002) Germany, 1978–89 matching 0% LP 0%ΔLP Self‐Selection with Endogenous Productivity Change 
 Pre‐entry effects Alvarez and López (2005) Chile, 1990–96 Matched D‐i‐D + ΔINV, + ΔSKILL
+ TFP, + LP 
non‐matched results 0%ΔTFP, ?%ΔLP
matched sample López (2004) Chile, 1990–96 New Exporters vs. non‐exporters + ΔINV, 0%ΔDOMSALE
+ ΔTFP Authors . Sample . Methodology . Pre‐entry difference . Post‐entry difference . Self‐Selection versus Learning Aw et al. (2000) Korea, 1983–93 and Taiwan (China), 1981–91 New Exporters vs. non‐exporters 5+% TFP Taiwan
? TFP Korea 6+%Δ TFP Taiwan
? Δ TFP Korea Baldwin and Gu (2003) Canada, 1974–96 New Exporters vs. non‐exporters 3%ΔLP, 0%ΔTFP 6%ΔLP, 2%ΔTFP Bernard and Jensen (1999) US, 1984–92 New Exporters vs. non‐exporters 6% TFP, 7–8% LP 3%ΔTFP, 3%ΔLP–short run
1%ΔTFP, 1–2%ΔLP–medium run 
1%ΔTFP, 1–2%ΔLP–long run Bernard and Jensen, (2004b) US, 1983–92 New Exporters vs. non‐exporters 3% TFP 6% TFP, 2%ΔTFP Bernard and Wagner (1997) Germany, 1978–92 New Exporters vs. non‐exporters 5% LP, 0%ΔLP 5%ΔLP Castellani (2002) Italy, 1989–94 Exporters vs. non‐exporters + TFP, 0 ΔTFP Damijan et al. (2006) Slovenia, 1994–2002 Exporters vs. non‐exporters 0% TFP 0% TFP t0
0% TFP when export to non‐OECD countries t1
11+% TFP when export to OECD countries t1 Delgado et al. (2002) Spain, 1991–96 New Exporters vs. non‐exporters
Stochastic dominance + TFP 0 ΔTFP Greenaway and Yu (2004) UK chemicals industry, 1990–2000 Dynamic panel 10% increase in exports = 1% TFP, 6% LP Hahn (2004) Korea, 1990–98 New Exporters vs. non‐exporters 4% TFP
 7% TFP Hansson and Lundin (2004) Sweden, 1990–99 New Exporters vs. non‐exporters 0%ΔTFP, 0%ΔLP 0%ΔTFP, 5%ΔLP Isgut (2001) Colombia, 1981–91 New Exporters vs. non‐exporters 20% LP, 4%ΔLP 5%ΔLP1 Kraay (1999) China, 1988–92 Dynamic panel 1s.d. increase in exports = 2% TFP, 13% LP Liu et al. (1999) Taiwan, 1989–93 New Exporters vs. non‐exporters 0%ΔLP, 6%ΔTFP 7%ΔLP, 0%ΔTFP Self‐Selection with Endogenous Productivity Change
Post‐entry effects Arnold and Hussinger (2005a) Germany, 1992–00 Matched D‐i‐D + ΔTFP
non‐matched sample 0%ΔTFP
matched sample Baldwin and Gu (2003) Canada, 1974–96 GMM 3.4% LP, 0% TFP
non‐matched sample 5.5%LP, 1.7%TFP
non‐matched sample
11%LP, 1%TFP
GMM results Bigsten et al. (2000) 4 African countries 1992–95 Dynamic system + ΔTechnical efficiency Blalock and Gertler (2004) Indonesian firms, 1990–96 1.Fixed effects
2. IV–OP & LP
3. timing 3. 0%ΔTFP 1. 5% TFP
2. 2–5% TFP
3. 4%ΔTFP Clerides et al. (1998) Colombia 1981–91, Mexico, 1986–90 and Morocco 1984–91 GMM Colombia + LP 
Mexico 0 LP
Morocco + LP Colombia +LP 
Mexico 0 LP
Morocco + LP
 De Loecker (2004) Slovenia, 1994–2000 Matched D‐i‐D 22%TFP t0 Girma et al. 2003) UK, 1988–98 Matched D‐i‐D 0%ΔTFP, 0%ΔLP 
in matched sample
1%ΔTFP, 0%ΔLP 
in unmatched sample ΔTFP:2%ΔLP:2%
in matched sample
ΔTFP:2%ΔLP:1%
in unmatched sample Greenaway and Kneller (2003) UK, 1989–2002 Matched D‐i‐D 0%ΔTFP, 0%ΔLP 
in matched sample ΔTFP:3%ΔLP:5.5%
Effect stronger when interacted with export share Greenaway, Gullstrand and Knellar(2005) Sweden, 1980–97 Matched D‐i‐D 0%ΔLP 
0%ΔTFP 0%ΔLP
0%ΔTFP Van Biesebroeck (2005) 9 African countries, 1992–96 GMM 35%TFP Wagner (2002) Germany, 1978–89 matching 0% LP 0%ΔLP Self‐Selection with Endogenous Productivity Change 
 Pre‐entry effects Alvarez and López (2005) Chile, 1990–96 Matched D‐i‐D + ΔINV, + ΔSKILL
+ TFP, + LP 
non‐matched results 0%ΔTFP, ?%ΔLP
matched sample López (2004) Chile, 1990–96 New Exporters vs. non‐exporters + ΔINV, 0%ΔDOMSALE
+ ΔTFP Notes: Where possible the results refer to a comparison of new exporters versus non‐exporters. TFP = total factor productivity, LP = labour productivity, Δ = growth + the difference relative to the control group is positive and significant, − the difference relative to the control group is negative and significant, 0 the difference relative to the control group is insignificant, ? the difference relative to the control group changes sign and/or significance through the paper. These results refer to firms that survive in export markets, as reported in Table 10 and for value added per worker. Castellani (2002) compares exporters versus non‐exporters. Open in new tab Table 3 Evidence on Export Market Entry Effects and Firms Authors . Sample . Methodology . Pre‐entry difference . Post‐entry difference . Self‐Selection versus Learning Aw et al. (2000) Korea, 1983–93 and Taiwan (China), 1981–91 New Exporters vs. non‐exporters 5+% TFP Taiwan
? TFP Korea 6+%Δ TFP Taiwan
? Δ TFP Korea Baldwin and Gu (2003) Canada, 1974–96 New Exporters vs. non‐exporters 3%ΔLP, 0%ΔTFP 6%ΔLP, 2%ΔTFP Bernard and Jensen (1999) US, 1984–92 New Exporters vs. non‐exporters 6% TFP, 7–8% LP 3%ΔTFP, 3%ΔLP–short run
1%ΔTFP, 1–2%ΔLP–medium run 
1%ΔTFP, 1–2%ΔLP–long run Bernard and Jensen, (2004b) US, 1983–92 New Exporters vs. non‐exporters 3% TFP 6% TFP, 2%ΔTFP Bernard and Wagner (1997) Germany, 1978–92 New Exporters vs. non‐exporters 5% LP, 0%ΔLP 5%ΔLP Castellani (2002) Italy, 1989–94 Exporters vs. non‐exporters + TFP, 0 ΔTFP Damijan et al. (2006) Slovenia, 1994–2002 Exporters vs. non‐exporters 0% TFP 0% TFP t0
0% TFP when export to non‐OECD countries t1
11+% TFP when export to OECD countries t1 Delgado et al. (2002) Spain, 1991–96 New Exporters vs. non‐exporters
Stochastic dominance + TFP 0 ΔTFP Greenaway and Yu (2004) UK chemicals industry, 1990–2000 Dynamic panel 10% increase in exports = 1% TFP, 6% LP Hahn (2004) Korea, 1990–98 New Exporters vs. non‐exporters 4% TFP
 7% TFP Hansson and Lundin (2004) Sweden, 1990–99 New Exporters vs. non‐exporters 0%ΔTFP, 0%ΔLP 0%ΔTFP, 5%ΔLP Isgut (2001) Colombia, 1981–91 New Exporters vs. non‐exporters 20% LP, 4%ΔLP 5%ΔLP1 Kraay (1999) China, 1988–92 Dynamic panel 1s.d. increase in exports = 2% TFP, 13% LP Liu et al. (1999) Taiwan, 1989–93 New Exporters vs. non‐exporters 0%ΔLP, 6%ΔTFP 7%ΔLP, 0%ΔTFP Self‐Selection with Endogenous Productivity Change
Post‐entry effects Arnold and Hussinger (2005a) Germany, 1992–00 Matched D‐i‐D + ΔTFP
non‐matched sample 0%ΔTFP
matched sample Baldwin and Gu (2003) Canada, 1974–96 GMM 3.4% LP, 0% TFP
non‐matched sample 5.5%LP, 1.7%TFP
non‐matched sample
11%LP, 1%TFP
GMM results Bigsten et al. (2000) 4 African countries 1992–95 Dynamic system + ΔTechnical efficiency Blalock and Gertler (2004) Indonesian firms, 1990–96 1.Fixed effects
2. IV–OP & LP
3. timing 3. 0%ΔTFP 1. 5% TFP
2. 2–5% TFP
3. 4%ΔTFP Clerides et al. (1998) Colombia 1981–91, Mexico, 1986–90 and Morocco 1984–91 GMM Colombia + LP 
Mexico 0 LP
Morocco + LP Colombia +LP 
Mexico 0 LP
Morocco + LP
 De Loecker (2004) Slovenia, 1994–2000 Matched D‐i‐D 22%TFP t0 Girma et al. 2003) UK, 1988–98 Matched D‐i‐D 0%ΔTFP, 0%ΔLP 
in matched sample
1%ΔTFP, 0%ΔLP 
in unmatched sample ΔTFP:2%ΔLP:2%
in matched sample
ΔTFP:2%ΔLP:1%
in unmatched sample Greenaway and Kneller (2003) UK, 1989–2002 Matched D‐i‐D 0%ΔTFP, 0%ΔLP 
in matched sample ΔTFP:3%ΔLP:5.5%
Effect stronger when interacted with export share Greenaway, Gullstrand and Knellar(2005) Sweden, 1980–97 Matched D‐i‐D 0%ΔLP 
0%ΔTFP 0%ΔLP
0%ΔTFP Van Biesebroeck (2005) 9 African countries, 1992–96 GMM 35%TFP Wagner (2002) Germany, 1978–89 matching 0% LP 0%ΔLP Self‐Selection with Endogenous Productivity Change 
 Pre‐entry effects Alvarez and López (2005) Chile, 1990–96 Matched D‐i‐D + ΔINV, + ΔSKILL
+ TFP, + LP 
non‐matched results 0%ΔTFP, ?%ΔLP
matched sample López (2004) Chile, 1990–96 New Exporters vs. non‐exporters + ΔINV, 0%ΔDOMSALE
+ ΔTFP Authors . Sample . Methodology . Pre‐entry difference . Post‐entry difference . Self‐Selection versus Learning Aw et al. (2000) Korea, 1983–93 and Taiwan (China), 1981–91 New Exporters vs. non‐exporters 5+% TFP Taiwan
? TFP Korea 6+%Δ TFP Taiwan
? Δ TFP Korea Baldwin and Gu (2003) Canada, 1974–96 New Exporters vs. non‐exporters 3%ΔLP, 0%ΔTFP 6%ΔLP, 2%ΔTFP Bernard and Jensen (1999) US, 1984–92 New Exporters vs. non‐exporters 6% TFP, 7–8% LP 3%ΔTFP, 3%ΔLP–short run
1%ΔTFP, 1–2%ΔLP–medium run 
1%ΔTFP, 1–2%ΔLP–long run Bernard and Jensen, (2004b) US, 1983–92 New Exporters vs. non‐exporters 3% TFP 6% TFP, 2%ΔTFP Bernard and Wagner (1997) Germany, 1978–92 New Exporters vs. non‐exporters 5% LP, 0%ΔLP 5%ΔLP Castellani (2002) Italy, 1989–94 Exporters vs. non‐exporters + TFP, 0 ΔTFP Damijan et al. (2006) Slovenia, 1994–2002 Exporters vs. non‐exporters 0% TFP 0% TFP t0
0% TFP when export to non‐OECD countries t1
11+% TFP when export to OECD countries t1 Delgado et al. (2002) Spain, 1991–96 New Exporters vs. non‐exporters
Stochastic dominance + TFP 0 ΔTFP Greenaway and Yu (2004) UK chemicals industry, 1990–2000 Dynamic panel 10% increase in exports = 1% TFP, 6% LP Hahn (2004) Korea, 1990–98 New Exporters vs. non‐exporters 4% TFP
 7% TFP Hansson and Lundin (2004) Sweden, 1990–99 New Exporters vs. non‐exporters 0%ΔTFP, 0%ΔLP 0%ΔTFP, 5%ΔLP Isgut (2001) Colombia, 1981–91 New Exporters vs. non‐exporters 20% LP, 4%ΔLP 5%ΔLP1 Kraay (1999) China, 1988–92 Dynamic panel 1s.d. increase in exports = 2% TFP, 13% LP Liu et al. (1999) Taiwan, 1989–93 New Exporters vs. non‐exporters 0%ΔLP, 6%ΔTFP 7%ΔLP, 0%ΔTFP Self‐Selection with Endogenous Productivity Change
Post‐entry effects Arnold and Hussinger (2005a) Germany, 1992–00 Matched D‐i‐D + ΔTFP
non‐matched sample 0%ΔTFP
matched sample Baldwin and Gu (2003) Canada, 1974–96 GMM 3.4% LP, 0% TFP
non‐matched sample 5.5%LP, 1.7%TFP
non‐matched sample
11%LP, 1%TFP
GMM results Bigsten et al. (2000) 4 African countries 1992–95 Dynamic system + ΔTechnical efficiency Blalock and Gertler (2004) Indonesian firms, 1990–96 1.Fixed effects
2. IV–OP & LP
3. timing 3. 0%ΔTFP 1. 5% TFP
2. 2–5% TFP
3. 4%ΔTFP Clerides et al. (1998) Colombia 1981–91, Mexico, 1986–90 and Morocco 1984–91 GMM Colombia + LP 
Mexico 0 LP
Morocco + LP Colombia +LP 
Mexico 0 LP
Morocco + LP
 De Loecker (2004) Slovenia, 1994–2000 Matched D‐i‐D 22%TFP t0 Girma et al. 2003) UK, 1988–98 Matched D‐i‐D 0%ΔTFP, 0%ΔLP 
in matched sample
1%ΔTFP, 0%ΔLP 
in unmatched sample ΔTFP:2%ΔLP:2%
in matched sample
ΔTFP:2%ΔLP:1%
in unmatched sample Greenaway and Kneller (2003) UK, 1989–2002 Matched D‐i‐D 0%ΔTFP, 0%ΔLP 
in matched sample ΔTFP:3%ΔLP:5.5%
Effect stronger when interacted with export share Greenaway, Gullstrand and Knellar(2005) Sweden, 1980–97 Matched D‐i‐D 0%ΔLP 
0%ΔTFP 0%ΔLP
0%ΔTFP Van Biesebroeck (2005) 9 African countries, 1992–96 GMM 35%TFP Wagner (2002) Germany, 1978–89 matching 0% LP 0%ΔLP Self‐Selection with Endogenous Productivity Change 
 Pre‐entry effects Alvarez and López (2005) Chile, 1990–96 Matched D‐i‐D + ΔINV, + ΔSKILL
+ TFP, + LP 
non‐matched results 0%ΔTFP, ?%ΔLP
matched sample López (2004) Chile, 1990–96 New Exporters vs. non‐exporters + ΔINV, 0%ΔDOMSALE
+ ΔTFP Notes: Where possible the results refer to a comparison of new exporters versus non‐exporters. TFP = total factor productivity, LP = labour productivity, Δ = growth + the difference relative to the control group is positive and significant, − the difference relative to the control group is negative and significant, 0 the difference relative to the control group is insignificant, ? the difference relative to the control group changes sign and/or significance through the paper. These results refer to firms that survive in export markets, as reported in Table 10 and for value added per worker. Castellani (2002) compares exporters versus non‐exporters. Open in new tab Open in new tab In line with evidence of spillovers more generally, results are somewhat mixed. Some studies identify strong positive spillover effects (Aitken et al., 1997; Kokko et al., 1997; Greenaway et al., 2004; Greenaway and Kneller, 2003) others have either found none and in some cases negative impacts (Bernard and Jensen, 2004a; Sjoholm, 2003; Barrios et al., 2003; Ruane and Sutherland, 2005). Kneller and Pisu (2007) and Swenson (2005) find mixed evidence, depending on the channel considered. Beyond country specific differences there is no obvious pattern to these inconsistencies. This is best seen from a comparison of Greenaway et al. (2004), Barrios et al. (2003) and Ruane and Sutherland (2005) which all focus on European countries, measure foreign presence in the same way, and use a similar methodology. Greenaway et al. (2004) measure foreign presence in the UK as the sum of industry employment or output and, in an attempt to separate competition from information effects, add exports from foreign multinationals as a proportion of total exports in the industry. They find both the likelihood of exporting and export share are increasing in the industry‐level foreign presence index, even controlling for firm and industry level characteristics. They report less clear results for the index measuring export activities of foreign firms, this being positive and weakly significant for the export decision and positive and insignificant in the decision of how much to export. By contrast, Barrios et al. (2003) for Spain find no evidence of an effect on the export decision from MNEs or the export share. Ruane and Sutherland (2005) also use a Heckman selection model to account for interdependence between export participation and export share decisions, but with contrasting results. They find positive effects from foreign presence of multinationals and negative effects from their export share on both export and export share decisions, with a suggestion the latter is due to US multinationals. They attribute this to the use of Ireland as an export platform to the EU. They argue export spillovers are unlikely where the country is an export platform because competition with domestic firms in local markets is limited. The use of spillovers from other exporters does not appear to improve this. Aitken et al. (1997) and Bernard and Jensen (2004a) find no effect from such measures, whereas Greenaway and Kneller (2003) do. While positive and insignificant effects are relatively easy to explain in this context, negative effects are more puzzling. Ruane and Sutherland (2005) explain theirs by Ireland being an export platform, thus multinationals have less contact with indigenous firms. It is not clear however why this makes Irish firms less likely to export. Perhaps more plausible is the congestion argument of Swenson (2005): competition with multinationals raises prices in product markets forcing domestic firms up their average cost curves for example; or, perhaps higher costs result from congestion of local infrastructure. 2.4. Consequences of Export Market Entry Entry can have a number of different impacts on the firm and aggregate economy. Some have provoked less discussion than others. For example there is widespread evidence of an aggregate productivity effect through resource reallocation (Bernard and Jensen, 2004a ; Hansson and Lundin, 2004; Falvey et al., 2004). The area given greatest attention however, is direction of causality between exporting and within‐firm changes in productivity. We focus on that, although other important effects might relate to survival probability of exporters (Bernard and Wagner, 1997; Bernard and Jensen, 1999). At the simplest level this literature can be seen as a test between self‐selection and learning, and indeed this was explicit in the earliest studies. The umbrella label ‘learning’ in fact contains three separate channels. First, interaction with foreign competitors and customers provides information about process and product reducing costs and raising quality, which can be interpreted as learning by exporting. Second exporting allows firms to increase scale.10 Finally increased competition in foreign markets forces firms to be more efficient and stimulates innovation. However this fails to recognise how the hypothesis under test has evolved, to one of a bi‐causal relationship. Self‐selection is important, but leads also to endogenous changes in productivity either as a result of learning by exporting or learning to export. In the earliest literature the hypothesis under test was clearly one of self‐selection versus learning. The arguments in favour of the former are most powerfully put by Bernard and Jensen (1999, 2004b). In their study of US plants they found productivity growth of exporters was not significantly different from non‐exporters, independent of whether productivity was measured as labour productivity or TFP. This implies that the productivity distribution of firms in any given industry does not widen continuously over time, or put differently the growth effects from learning are not permanent. They also provided evidence that out of the pool of non‐exporters, new exporters were already among the best and differed significantly from the average non‐exporter. Whilst there is some country specific sensitivity in the magnitude of any difference in performance, a reasonable summary would be that the results of Bernard and Jensen (1999) for the US are replicated for most other countries (see Table 3).11 Export market entry is associated with significant changes in performance around the point at which export sales begin. This argument for self‐selection is therefore based on a comparison between established exporters and non‐exporters and a difference in the performance of new export firms around the point of entry which is not permanent. Future entrants have many of the right characteristics that make them likely to export and faster productivity growth than non‐exporters when they do. But, after a short period they become indistinguishable from other exporters. The strong conclusions reached by Bernard and Jensen (1999) in favour of self‐selection led quickly to an adaptation of the hypothesis being tested to one of self‐selection versus a bi‐causal relationship. Recognising that new exporters appeared to already have many of the right characteristics to become exporters one can test whether the surge in productivity associated with entry was explained by the decision to become an exporter, or whether the productivity surge led to the export decision. As a consequence of the change in focus, methodology also evolved, with attempts to control for self‐selection using either instrumental variable or matching techniques (alone or in combination with difference in differences). As argued in Van Biesebroeck (2005) not controlling for self‐selection will overstate evidence of learning for new exporters in the data. Instrumental variable approaches have usually been estimated using GMM; see for example Van Biesebroeck (2005); Baldwin and Gu (2003). Whilst they have the advantage of being relatively easy to estimate one faces the perennial question of instrument validity. By contrast, matching attempts to reduce heterogeneity between new and non‐exporters by using observable firm characteristics. It has the disadvantage of removing observations from the data set and requiring specific assumptions about non‐observable factors such as managerial ability. Establishing causality is probably the most challenging issue facing researchers in this area. Our view is that matching offers the sounder foundation, but we leave arguments to which of these methodologies should be preferred to Blundell and Costa Dias (2000) and focus instead on results from each. The impact of applying these alternative techniques has been largely to confirm self‐selection is more important than learning. For example, comparisons of new exporters and non‐exporters without controlling for selection in Germany (Bernard and Wagner, 1997) and the UK (Girma, Greenaway and Kneller, 2004) shows significant pre‐entry differences in performance, whereas differences are not evident with methods controlling for selection. Yet whilst evidence of post‐entry productivity changes are reported for the UK (Girma et al., 2005b) they are not for Germany (Wagner, 2002). Indeed whilst both GMM and matching advance on simply comparing new exporters with all non‐export firms, they do not guarantee post‐entry productivity changes will be observed. As Table 3 shows, more studies report evidence for learning than fail to find such effects, although it is perhaps worth noting these tend to be studies that use matching. So what explains this divergence? Two issues have been explored, heterogeneity and timing. Some have argued that learning is likely to be specific to some firms, such as those that are young (Delgado et al. 2002; Fernandes and Isgut, 2005), or highly exposed to export markets (Kraay, 1999; Castellani, 2002; Girma, Görg and Strobl, 2004; Damijan et al., 2007). Others have found post‐entry changes depend on existing industry characteristics, productivity changes are lower in industries in which current exposure to foreign firms (through arms length trade and FDI) is high (Greenaway and Kneller, 2003). While it is difficult to conclude against such effects, heterogeneity should not be allowed to become an easy excuse for inconsistencies across studies. To establish heterogeneity will require evidence that the same mechanisms (such as age or foreign market exposure) are important across countries. The learning by exporting hypothesis attributes part of the change in productivity to the endogenous decision to start exporting. More recently López (2004) and Alvarez and López (2005) have questioned the timing issue, arguing that productivity changes occur after the decision to start exporting, that is they may pre‐date the point at which export sales begin.12 Firms invest in new technologies leading to pre‐entry changes in productivity: they learn to export rather than learn by exporting. This takes the view that learning effects are neither inevitable nor automatic but require investments in domestic technology (Keller, 2004). While this might be seen by some as an unfair shift of the goalposts, it is consistent with a test of exogenous versus endogenous changes in productivity associated with exporting. It has also existed as an idea within the case study literature for some time (see the review by Pack, 2000) and a number of studies report anecdotal evidence (López 2004; Alvarez and López, 2005; Van Biesebroeck, 2005; and Blalock and Gertler, 2004). Empirical testing of this using micro data sets becomes more difficult owing to the unobservable nature of the time at which the decision to start to export is made, and the likelihood that preparation time varies across firms. As López (2004) notes however, without information on timing of the decision, the time path of an endogenous change in productivity is likely to look similar to that of an exogenous change and it becomes harder to conclude that observed productivity changes are orthogonal to the export entry decision. Using an econometric approach Aw et al. (2006) study the evolution of productivity and R&D for exporters in Taiwanese electronics. They find that those that do not invest in R&D have lower productivity growth than those that just export, which in turn is lower than those firms that invest in both.13 They argue these findings are consistent with an interpretation that R&D investments are necessary for firms to benefit from their exposure to international markets. López (2004) develops the same idea for domestic sales and investment. He finds investment and productivity rises in the pre‐entry period but domestic sales are flat and argues this is consistent with investment in technology for sales to foreign but not domestic markets. Endogenous pre‐entry changes in productivity offer an interesting possibility for future research, though current analysis raises questions. First, a simple growth accounting approach suggests that if investment rises and output remains flat, productivity should fall. Simultaneous increases in investment and productivity would therefore seem an unlikely combination, unless of course there are reductions in other inputs. Here more detailed data on equipment and R&D investment would help. Second, how are we to interpret evidence of post‐entry changes in productivity? The most obvious explanation is overlap between the benefits to new technology with the point at which sales start, perhaps due to lags in their effects due to learning. An alternative might be a difference between firms that are passive and active in their export decision. Discussions with those involved in export promotion in the UK suggest both occur frequently. For those firms that are passive, no pre‐entry investments are made and productivity changes are likely to occur with the start of export sales. Ultimately perhaps issues surrounding timing of the decision and investment in new plant, equipment or personnel are difficult to answer with available data, which offers insufficient detail. While case studies offer one solution, perhaps a more interesting approach is that used by Baldwin and Gu (2004) who combine micro data with questionnaires about export behaviour. They find evidence consistent with changes in scale, increased efficiency through competition and learning. Canadian exporters used more foreign technologies, were more likely to have R&D collaboration with foreign firms and improved the flow of information about foreign technologies to Canadian firms. That also led to increased innovation and investments in absorptive capacity. 2.5 Determinants and Consequences of Exit As with export market entry, the literature on exit splits into determinants and consequences. A reasonable expectation would be that exit should be symmetric to entry. To some extent this is so. Exit from export markets is correlated with similar firm level variables as entry: it is less likely the larger, more productive and more human capital intensive the firm, and the lower the ratio of exports to domestic sales; see for example Greenaway and Kneller (2003) and Blalock and Roy (2005). Industry determinants have been less well researched. For example, research that focuses on the effect of exchange rate changes considers periods of domestic currency depreciation, when exports are likely to expand (Bernard and Jensen, 2004b, Das et al., 2004; Blalock and Roy, 2005). Thus far no one has considered whether the effect of appreciation is symmetric, although evidence of substantial export market exit in the presence of a depreciation of the Indonesia rupiah by Blalock and Roy (2005) suggests it is not. The set of industry variables is extended by Greenaway and Kneller (2003) to include import penetration and intra‐industry trade, as well as industry sunk costs. Conditional on firm level variables they find exit is more likely in industries with low sunk‐costs, (because re‐entry is easier) and those with high levels of intra‐industry trade. No role for import penetration was found which is consistent with Melitz (2003), where self‐selection is driven not by an increase in imports but the pull of export markets. The literature on consequences of exit is somewhat larger. As with entry, self‐selection appears to be important. Export quitters tend to have lower productivity compared to firms that continue (Aw et al., 2000; Baldwin and Gu, 2003; Girma et al., 2003) and no significant difference from, or in some cases, lower productivity (growth) than non‐exporters (Bernard and Jensen, 1999; Hansson and Lundin, 2004; Hahn, 2004). Firms seem to self‐select out of export markets just as they do into them. One caveat might be made from an often overlooked feature of the data, the comparison of new exporters with entrants: evidence presented across studies comparing entrants and quitters suggests the latter have higher productivity. As with entry the effect of exit on productivity produces mixed results. Of those not conditioning for self‐selection Hansson and Lundin (2004) and Hahn, (2004) find no obvious post‐exit productivity changes, whereas Girma et al. (2003) and Blalock and Gertler (2004) report similar results conditioning on self‐selection. By contrast, for the US Bernard and Jensen (1999, 2004b) report post‐exit changes, not controlling for self‐selection. On balance, it would seem that self‐selection is important, weaker firms are likely to exit, but unlike entry there is little impact on productivity of this choice. 3. Exporting and Foreign Direct Investment 3.1. Exports versus FDI At the simplest level, exports and FDI are substitute channels for firms globalising.14 The conditions for foreign production become more favourable relative to exporting as the size of the foreign market increases and costs of exporting increase; and less favourable as costs of setting up foreign production grow. This is the proximity‐concentration trade‐off explained by Brainard (1993). The contribution of Helpman et al. (2004) to this is analogous to Melitz (2003) contribution to the basic model of trade with representative firms. Adding heterogeneity allows this choice to differ across firms within the same industry and thus determines which firms export and which become multinational. The interesting properties of the model in this regard are generated through the assumptions of different costs (largely fixed) associated with serving domestic and foreign markets (through FDI or exports), along with heterogeneity in productivity across firms. As we have seen sunk‐costs of exporting are typically thought to include fixed costs of research into product compliance, distribution networks, advertising and so on. Goods exported are also subject to transportation costs. The fixed costs of FDI are the duplication of costs in establishing domestic production facilities. They are assumed to be greater than those of exporting, FDI eliminates variable transport costs, but involves higher fixed costs. Heterogeneous productivity then ensures self‐selection. Only the most productive firms become multinationals; firms whose productivity falls in an intermediate range export and the least productive only sell domestically. Helpman et al. (2004) assume the decision to establish foreign production facilities is based purely on considerations of market access. All FDI is horizontally motivated. Head and Ries (2003) demonstrate that when there are factor price and market size differentials, firms invest abroad for vertical motives also: the ordering of the productivity distribution between multinationals and non‐multinationals can even be reversed. If the foreign country is small and offers some cost advantage, for a certain range of the parameter of the model, the least productive firms locate abroad whereas more productive ones produce at home. In this case, low productivity enterprises have a greater incentive to pay the FDI sunk costs because they use more intensively the factor whose overseas price is low. Empirical tests of the heterogeneous firm model have generally followed one of two lines. First, testing within industries for substitution between exports and FDI related to productivity differences. Second, testing the cross‐industry/country predictions – the volume of exports relative to FDI we might expect. Whilst there is a large literature comparing productivity levels of multinationals against non‐multinationals and exporters against non‐exporters, there are only a small number of studies that compare exporters and multinationals. In part this is because it is a relatively new question, in part because for many countries information on which domestic firms export and which are multinational is not available. As can be seen from Table 4 two basic approaches to this question are evident. The first follows Head and Ries (2003) in comparing mean values (in some cases conditional on other firm and industry characteristics), see for example Castellani and Zanfei (2007) and Kimura and Kioyata (2004). The second follows Girma et al. (2005a) in using Kolmogrov‐Smirnov tests of stochastic dominance, see Girma, Görg and Strobl (2004), Arnold and Hussinger (2005b) and Wagner (2005). This approach compares the cumulative distribution of productivity for different types of firms and not just the mean. Despite the difference in methodology, the prediction with regard to exports versus FDI would appear to have strong support, Head and Ries (2003) being the exception), while ironically that between exporters and non‐exporters less so. Whilst explaining differences across a small number of studies is never easy, several report a bias towards large firms, and therefore a bias against finding significant productivity differences, and there is a suggestion that this is most severe in Head and Ries (2003), who use information on publicly listed firms. Table 4 Evidence on Relative Productivity of Exporters and Multinationals Authors . Sample . Methodology . Exporters vs. non‐exporters . MNEs vs. exporters . Arnold and Hussinger (2005b) Germany, 1996–2002 K‐S tests of stochastic dominance + + Castellani and Zanfei (2007) Italy, 1994–96 OLS 0 + Girma, Görg and Strobl (2004) Ireland, 2000 K‐S tests of stochastic dominance 0 + Girma et al. (2005a) UK, 1990–95 K‐S tests of stochastic dominance + + Head and Ries (2003)2 Japan, 1989 OLS 0 0 Kimura and Kiyota (2004) Japan, 1994–2000 OLS + + Wagner (2005) Germany,1995 K‐S tests of stochastic dominance + + Authors . Sample . Methodology . Exporters vs. non‐exporters . MNEs vs. exporters . Arnold and Hussinger (2005b) Germany, 1996–2002 K‐S tests of stochastic dominance + + Castellani and Zanfei (2007) Italy, 1994–96 OLS 0 + Girma, Görg and Strobl (2004) Ireland, 2000 K‐S tests of stochastic dominance 0 + Girma et al. (2005a) UK, 1990–95 K‐S tests of stochastic dominance + + Head and Ries (2003)2 Japan, 1989 OLS 0 0 Kimura and Kiyota (2004) Japan, 1994–2000 OLS + + Wagner (2005) Germany,1995 K‐S tests of stochastic dominance + + Notes: + the effect is positive and significant, − the effect is negative and significant, 0 the effect is insignificant and/or changes sign and/or significance through the paper. Head and Ries do find predictions in support of the model for size characteristics. Open in new tab Table 4 Evidence on Relative Productivity of Exporters and Multinationals Authors . Sample . Methodology . Exporters vs. non‐exporters . MNEs vs. exporters . Arnold and Hussinger (2005b) Germany, 1996–2002 K‐S tests of stochastic dominance + + Castellani and Zanfei (2007) Italy, 1994–96 OLS 0 + Girma, Görg and Strobl (2004) Ireland, 2000 K‐S tests of stochastic dominance 0 + Girma et al. (2005a) UK, 1990–95 K‐S tests of stochastic dominance + + Head and Ries (2003)2 Japan, 1989 OLS 0 0 Kimura and Kiyota (2004) Japan, 1994–2000 OLS + + Wagner (2005) Germany,1995 K‐S tests of stochastic dominance + + Authors . Sample . Methodology . Exporters vs. non‐exporters . MNEs vs. exporters . Arnold and Hussinger (2005b) Germany, 1996–2002 K‐S tests of stochastic dominance + + Castellani and Zanfei (2007) Italy, 1994–96 OLS 0 + Girma, Görg and Strobl (2004) Ireland, 2000 K‐S tests of stochastic dominance 0 + Girma et al. (2005a) UK, 1990–95 K‐S tests of stochastic dominance + + Head and Ries (2003)2 Japan, 1989 OLS 0 0 Kimura and Kiyota (2004) Japan, 1994–2000 OLS + + Wagner (2005) Germany,1995 K‐S tests of stochastic dominance + + Notes: + the effect is positive and significant, − the effect is negative and significant, 0 the effect is insignificant and/or changes sign and/or significance through the paper. Head and Ries do find predictions in support of the model for size characteristics. Open in new tab The second strand of the literature concerns itself with proximity‐concentration predictions, the relative level of exports to FDI. Helpman et al. (2004) predict FDI will be more common relative to exports, the greater the dispersion of productivity levels within an industry. The data requirements of such a test are demanding however, particularly with regard to foreign sales by domestic multinationals and measures of dispersion within an industry. They use US data and regress the ratio of exports to FDI (measured by sales of overseas affiliates) on traditional proximity‐concentration variables, unit costs of trade and plant fixed costs, as well as a new variable, within industry dispersion. They consistently find that dispersion has the expected effect on relative sales: industries in which firm size is highly dispersed are associated with relatively more FDI than exports. 3.2. Exports by MNEs Whilst in a single product world exports and FDI are substitutes, even if this choice is determined exogenously by productivity levels, in practice multinationals also export. Indeed many report that foreign multinationals contribute disproportionately to exports compared to employment or output shares (Baldwin and Gu, 2003; Kneller and Pisu, 2004). To some extent this should be expected, a well‐established result is the superior performance of foreign owned firms with respect to employment, wages and productivity, all of which are important determinants of exports. Should the export decision of multinational firms be modelled as identical to that of domestic firms however? What little evidence there is suggests not. Kneller and Pisu (2004) find that even controlling for characteristics, foreign firms are more likely to export than indigenous ones, and export more intensively. So what explains export decisions of multinationals? Modelling has developed along two lines: export platform FDI and complementarity, broadly distinguished by the number of product lines the firm is assumed to produce.15 Export platform FDI is typically defined as the establishment of foreign production facilities and allocation of part or all of the output to serve a third country. It therefore refers to exports of a single product line, where these are not to the home country. Complementarity refers instead to multi‐product firms, to multiple stages of production and to export and FDI flows from the home to foreign countries: exports and FDI become positively correlated if there are horizontal or vertical complementarities across product lines. Theories of export platform FDI have developed by adding more countries and stages of production to traditional theories of FDI and in more recent developments in cross‐firm heterogeneity, FDI becomes complex. Vertical FDI occurs when the stages of production are located in more than one country; and horizontal when the same stage is located in more than one country. Vertical FDI is factor seeking; horizontal, market seeking. When there are more than two countries and more than two stages of production, multinationals are likely to undertake more complex FDI choices which involve intra‐firm trade and export platform FDI. The effect of adding more countries is to allow for the possibility of a horizontal motive for export platform FDI, adding more stages allows for a vertical motive. Motta and Norman (1996), motivated by the observation that much FDI is between countries in regional trading blocks, consider three identical countries and a single stage of production. Costs of production do not differ between countries but costs of trading do (because two either enter a free trade agreement or raise external barriers against the third). If we start from an equilibrium where each firm exports to the other two countries from its home base, raising external barriers or creating a free trade area encourages the outside firm to set up production facilities inside the free trade area and export to the other country in the bloc. Where the outside country chooses to locate production in and export from is left undetermined. Again, because of identical costs neither of the inside countries choose export platform FDI as a strategy. The conditions under which export platform FDI is likely have been analysed by Ekholm et al. (2003) where there are two identical countries in the North (A and B) one in the South, and multiple stages of production. Each firm produces intermediates and a final good. Firms must provide headquarter services from their home northern country but can choose where to produce intermediates as well as assembling the final product. Two of the countries, one northern (A) and one southern are members of a free trade area. The drivers of the model include assumptions about the size of the (marginal) cost advantage of southern firms and trading costs between different sets of countries. The free trade area between A and the Southern country means it is always optimal for the northern country to locate production in the South and export home (owing to the cost advantage from doing so). Therefore, unlike Motta and Norman (1996), when there are no vertical motives for FDI, the country inside the free trade area always has a motive to undertake export platform FDI. For the other northern country (B) the model predicts three outcomes. First, no FDI: firm B produces at home and exports to the free trade area; second, export‐platform FDI: firm B produces the good to be sold at home domestically, whereas the final product sold in the other northern country is produced in the South and exported; third, vertical FDI (hybrid MNE): firm B locates all production in the South and exports to both markets in the North. The last is hybrid because toward the home country, the firm undertakes vertical FDI whereas, toward the other Northern country, it undertakes a pure form of export platform FDI. Which strategy is adopted depends on the size of the (marginal) cost advantage to Southern firms, and trade costs. As the cost advantage of Southern firms increases we move from the first equilibrium to the second and when the cost advantage of locating in the South becomes large enough all production moves there. Similarly as trade costs between the Southern and two Northern countries fall, the Northern firm outside the FTA finds it competitive to move from exporting to the FTA, to export platform FDI, to locating all production in the Southern country. This has similarities to Motta and Norman (1996). The predictions of these models are driven primarily on cross‐country differences in costs. Grossman et al. (2003), developing the complex FDI model of Yeaple (2003), show that firm characteristics may also be important. If firms in the same industry are heterogeneous in productivity they may make different choices, even though costs of exporting and FDI are the same. They assume three countries (two North and one South); firms must provide headquarter services, produce intermediates and assemble the final product. Their analysis allows for the coexistence in the same sector of a rich array of profitable FDI strategies. In brief, the general lesson is that least productive firms will not undertake FDI. More productive firms choose complex strategies that involve a mix of FDI and exports. In most situations these can be classified as neither purely horizontal nor purely vertical, and involve the export of intermediates and/or final products. Models of export platform FDI simplify the analysis to a single product firm (albeit with multiple stages of production). An alternative set of models consistent with the idea that multinationals may also export comes from the literature on complementarity (Head and Ries, 2004). Again there are horizontal and vertical elements to this. In a multi‐product firm, exports and FDI become positively correlated if there are horizontal or vertical complementarities across product lines. For example, in the case of horizontal complementarities increased demand for the good supplied by foreign production may lead to increased demand for all goods produced by that firm, some of which may be supplied through arms‐length trade. For vertical complementarities the establishment of a plant in a foreign country to produce or assemble final goods will displace the exports of this product, but at the same time increase exports of intermediates from the home country. Net complementarity may arise if the displaced export of the final good is more than compensated by increased exports of intermediates. Empirical evidence on the export decision of multinationals has concentrated largely on direction of correlation, whether positive or negative, rather than explanation. In all cases, at the firm level, this relationship has been found to be positive, for example Lipsey and Weiss (1984) for the US, Swedenborg (1985) for Sweden, and Lipsey et al. (2000) and Kiyota and Urata (2005) for Japan. Attempts at understanding the explanation for any correlation are limited to Head and Ries (2003), Kiyota and Urata (2005) and Girma et al. (2005a). The first two test for the effect of vertical FDI on exports using export demand equations for the firm (both for Japan) and find similar results. Head and Ries (2001) find complementarity between exports and FDI for the most vertically integrated firms and substitution can be found for the least integrated, whereas Kiyota and Utata (2005) find that intra‐firm exports grow faster than total exports‐with increased FDI some of the inter‐firm exports shift to intra‐firm exports. By contrast Girma et al. (2005b) test for export platform FDI for the UK. They find foreign multinationals tend to acquire domestic firms that export – they cherry‐pick the best firms. However there are differences in the post‐acquisition export trajectories of acquired firms according to whether they is inside or outside the EU. For firms outside, export intensity rises, whereas it falls for firms inside. This appears consistent with export platform motives as discussed by Motta and Norman (1996). 4. Future Research Issues and Policy Dimensions 4.1. Future Research Issues A review of the Tables associated with this evaluation and references appended confirm how rapidly the literature has grown. It has also generated genuinely new insights, particularly with regard to the determinants of exporting. However, it is also a progressive research agenda in the sense that there is both unfinished business and new research questions being raised. As we have seen, some aspects of the export decision have received more attention than others. For example, while much is known about the characteristics of exporters and non‐exporters and what happens when a firm enters export markets, relatively little empirical work has been conducted around the question of choices that firms make between exports and FDI. To a degree this is data driven, given the demanding requirements of the underlying models. Since little may change with respect to data availability, or at least change only slowly, this suggests that future empirical work is likely to continue along current lines, with some spread to questions where the data constraints are not so severe. Tests of export‐FDI models are also likely to remain specific to more data rich countries such as the US, Japan and Sweden. A new strand of empirical analysis does appear to be emerging from the predictions of the heterogeneous firm models that provides some insight about the export‐FDI choice of firms however. That is the dynamic consequences of changes in the costs of exports and FDI. Perhaps the earliest example of this is by Pavcnik (2002), who studies the within firm and between firm productivity effects of trade liberalisation in Chile. Although the evidence base points unambiguously to the crucial role of sunk costs, little research has as yet focused on what these are, and how agglomeration, exchange rates and policy changes affect them. Whilst many researchers go through the motions of commenting on (for example) changes in product design, setting up distribution channels and so on as possible sources, that is generally as far as it goes. Sharper insights are needed if we really are to understand firm heterogeneity. This will rely on merging datasets and/or firm and industry specific survey based enquiry. A recent example of the former, which investigates the role of access to credit is Greenaway, Guariglia and Kneller (2005). A fourth issue, which again depends on merging datasets is the role, if any, of the origin and destination of trade/FDI. As we saw in Section 1 (extensions of the Melitz model to incorporate country asymmetries) and Section 3 (North‐South FDI models) origin and destination are likely to affect outcomes. Moreover, they may be key to understanding some of the empirical findings reported in Section 2. For example, it may be that potential learning from exporting is fashioned by the markets into which one exports. Finally, a new strand of research is being pioneered by Antras (2003) and Antras and Helpman (2004) exploring the implications of heterogeneity for the boundaries of the firm and strategies for outsourcing and insourcing of activities. This is a potentially rich vein of research, yielding new insights into globalisation and industrial organisation. Empirically however research here will be even more challenging given the need for disaggregated data on trade in intermediates, mapped on to firm specific information. 4.2. Policy Dimensions Intervention to promote exports is very widespread – every WTO Trade Policy Review16 contains a chapter on ‘Measures Directly Affecting Exports’ and there are always measures to report. These range from intervention to improve market intelligence (public support for trade missions), to sector specific fiscal intervention (tax concessions or duty drawbacks), to export processing zones (free zones). Such a widespread commitment to a specific policy agenda is unusual and the commitment to export promotion has historically been driven by a presumption that export growth and output growth are positively correlated. Although theoretical models linking openness and economic growth are not unequivocal, a large empirical literature points to a positive correlation, even if the direction of causality is controversial. Be that as it may, the key point is that intervention is motivated by macroeconometric evidence. Does the microeconometric evidence we have reviewed reinforce or undermine a case for active promotion? López (2005) asks this question and concludes that it reinforces the macro evidence. He argues that even if self‐selection is the key driver of export market entry, it may nevertheless be ‘conscious self selection’, especially in developing countries. What he means is that firms consciously improve their productivity with the international market in mind, rather than the best firms just starting to export. Policy intervention could than stimulate more conscious self‐selection and deliver a productivity boost. Clearly if learning by exporting does occur, productivity gains are boosted further. Moreover, if there are spillovers, perhaps because non‐exporting firms learn to export from other (domestic or multinational) exporting firms, the case is strengthened. This is a plausible argument, though it could only underpin a case for general rather than targeted intervention. López (2005) himself stresses the importance of reducing (overseas) barriers to exports, which clearly aligns with other arguments for trade liberalisation. To this should be added internal barriers to export, chief among which is domestic import protection, since as the incidence of protection literature shows, import tariffs are taxes on exporting. If sunk costs are important, one can think of intervention to improve aspects of infrastructure as relevant – improving information flows, promoting clustering and so on. If policy makers wanted evidence to support intervention targeted at specific sectors or firms, that would require much more information than we have access to at present. For example, are entry costs higher for small firms? is access to credit a barrier? and so on. In the absence of more robust evidence, targeted intervention to support exporting firms is subject to the same risks as identifying so‐called infant industries and the record on that front is not a good one. 5. Conclusions This article has synthesised and evaluated a new literature linking firms, trade and cross‐border investment. Its starting point was a well‐known feature of the real world, firms that export and others that do not co‐exist in the same industries. Until recently, this was not well explained by core trade models. This has changed with the development of heterogeneous firm models. These explain how firms that export are more productive and this, together with the reallocation of output which occurs as less productive firms contract or go out of business, points to a direct link between exporting and productivity. The framework has been extended to allow for the fact that some firms choose to produce overseas rather than export. The empirical literature has grown fast and as we have seen extends across a large number of industrialised, transitional and developing countries. Moreover this literature points to a number of regularities: exporting firms do tend to be larger and more productive than non‐exporters; sunk costs appear to be important; multinational firms tend to be more productive than domestic firms. Other evidence is less conclusive however, such as that relating to learning by exporting. We have learned a lot in a remarkably short space of time, but as we saw in the last section, a rich research agenda has been thrown‐up and this is a literature that will continue to grow. Footnotes 1 " In so doing this paper fits into a broader literature on the within‐industry heterogeneity of firms such as Olley and Pakes (1996), Roberts and Tybout (1996) and Aw et al. (1997). 2 " Earlier and related insights into the role of sunk costs in sluggish adjustment of trade responses to exchange rate fluctuations are attributable to Baldwin (1988) and Baldwin and Krugman (1989). 3 " In a muliti‐country setting, between firm productivity differences can generate intra‐industry trade in these models. 4 " Ederington and McCalman (2004) develop a model of firm heterogeneity with the opposite outcome. Heterogeneity is a consequence of the decision of some firms to start to export. 5 " A more comprehensive review of the theoretical literature can be found in Helpman (2005). 6 " See for example Rodriguez and Rodrik (2000) and Greenaway et al. (2002) and see López (2005) for an evaluation of micro and macro evidence. 7 " This contrasts with the large estimated currency union effects of Rose and Stanley (2005). 8 " We concentrate on evidence of trade liberalisation on export volumes at the firm level. There is a larger literature, see for example Pavcnik (2002), Roberts and Tybout (1996) or Tybout (2003) for references, that discusses the productivity impacts of such changes and Head and Ries (1999) and Roberts and Tybout (1991) for the effect on firm size. Given the link between exports, firm size and productivity these might be seen as indirect evidence of the export effect of policy changes. 9 " The Table does not include the results from Blalock and Gertler (2004) because of a lack of formal econometric evidence in the paper. 10 " Evidence from Tybout and Westbrook (1995) suggests that this may be an unimportant source of efficiency change. 11 " The evidence for Sweden (Hansson and Lundin, 2004; Greenaway, Gullstrand and Kneller (2005) and Slovenia (Damijan et al., 2007) are exceptions. 12 " Alvarez and López (2005) label pre‐entry effects as ‘learning to export’ compared to ‘learning by exporting’ for post‐entry effects. The common element between these is the effect of the decision to export on the firms productivity. 13 " A number of papers have found that exporters have higher levels of R&D but do not establish the direction of causality, see for example Bleaney and Wakelin (2002) and Roper and Love (2002) for the UK, Bernard and Jensen (1995) for the US, Aw et al. (2006) for Taiwan and Baldwin and Gu (2004) for Canada. 14 " We concentrate here on the evidence at the level of the firm. The issue of complementarity and substitution between exports and FDI has been studied at many other levels of aggregation, a summary of the evidence for which can be found in the Head and Ries (2004). 15 " Helpman (2005) takes a somewhat broader view of this question adding a discussion of the role of incomplete contracts for firms internationalisation and offshoring decisions. 16 " The WTO’s Trade Policy Review Mechanism ensures that the trade policies of Members are audited on a regular basis. For the ‘big three’ (US, EU and Japan) this means every two years; for the smallest Members, it takes place every seven years. References Aitken , B , Hanson , G. and Harrison , A ( 1997 ). ‘Spillovers, foreign investment and export behavior’ , Journal of International Economics , vol. 43 , pp. 103 – 32 . Google Scholar Crossref Search ADS WorldCat Alvarez , R. , ( 2004 ). ‘Sources of export success in small and medium‐sized enterprises: the impact of public programs’ , International Business Review , vol. 13 , pp. 383 – 400 . Google Scholar Crossref Search ADS WorldCat Alvarez , R. and López , R.A. ( 2005 ). ‘Exporting and firm performance: evidence from Chilean plants’ , Canadian Journal of Economics , vol. 38 , pp. 1384 – 400 . Google Scholar Crossref Search ADS WorldCat Antras , P. ( 2003 ). ‘Firms, contracts and trade structure’ , Quarterly Journal of Economics , vol. 118 , pp. 1375 – 418 . Google Scholar Crossref Search ADS WorldCat Antras , P. and Helpman , E. ( 2004 ). ‘Global sourcing’ , Journal of Political Economy , vol. 112 , pp. 552 – 80 . Google Scholar Crossref Search ADS WorldCat Arnold , J. and Hussinger , K. ( 2005a ). ‘Export behavior and firm productivity in German manufacturing: a firm level analysis’ , Review of World Economics/Weltwirtschaftliches Archiv , vol. 141 , pp. 219 – 43 . Google Scholar Crossref Search ADS WorldCat Arnold , J. and Hussinger , K. ( 2005b ). ‘Exports versus FDI in German manufacturing: firm performance and participation in international markets’ , mimeo, World Bank. Aw , B.Y. , Chen , X. and Roberts , M.J., ( 1997 ). ‘Plant level evidence on productivity differentials, turnover and exports in Taiwanese manufacturing’ , mimeo, Pennsylvania State University . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Aw , B.Y. , Chung , S. and Roberts , M.J., ( 2000 ). ‘Productivity and turnover in the export market: micro‐level evidence from the Republic of Korea and Taiwan (China)’ , World Bank Economic Review , vol. 14 , pp. 65 – 90 . Google Scholar Crossref Search ADS WorldCat Aw , B.Y. , Roberts , M.J., and Winston , T. ( 2007 ). ‘The complementary role of exports and R&D investments as sources of productivity growth’ , The World Economy , vol. 30 , pp. 83 – 104 . Google Scholar Crossref Search ADS WorldCat Baldwin , R. ( 1988 ). ‘Hysteresis in import prices: the Beachhead effect’ , American Economic Review , vol. 78 , pp. 773 – 85 . OpenURL Placeholder Text WorldCat Baldwin , R. and Krugman , P.R. ( 1989 ) ‘Persistent trade effects of large exchange rate shocks’ , Quarterly Journal of Economics , vol. 419 , pp. 635 – 54 . Google Scholar Crossref Search ADS WorldCat Baldwin , J.R. and Gu , W. ( 2003 ). ‘Export market participation and productivity performance in Canadian manufacturing’ , Canadian Journal of Economics , vol. 36 , pp. 634 – 57 . Google Scholar Crossref Search ADS WorldCat Baldwin , J.R. and Gu , W. ( 2004 ). ‘Trade liberalisation: export‐market participation, productivity growth and innovation’ , Oxford Review of Economic Policy , vol. 20 , pp. 372 – 92 . Google Scholar Crossref Search ADS WorldCat Barrios S. , Görg , H. and Strobl , E. ( 2003 ). ‘Explaining firms’ export behaviour: R&D, spillovers and the destination market’ , Oxford Bulletin of Economics and Statistics , vol. 65 , pp. 475 – 96 . Google Scholar Crossref Search ADS WorldCat Bernard , A. , Eaton , J., Jensen , J.B. and Kortum , S. ( 2003 ). ‘Plants and productivity in international trade’ , American Economic Review , vol. 93 , pp. 1268 – 90 . Google Scholar Crossref Search ADS WorldCat Bernard , A. and Jensen , J.B. ( 1995 ). ‘Exporters, jobs and wages in US manufacturing: 1976–1987’ , Brookings Papers on Economic Activity , Microeconomics , pp. 67 – 119 . OpenURL Placeholder Text WorldCat Bernard , A. and Jensen , J.B. ( 1999 ). ‘Exceptional exporters performance: cause, effect or both?’ , Journal of International Economics , vol. 47 , pp. 1 – 25 . Google Scholar Crossref Search ADS WorldCat Bernard , A. and Jensen , J.B. ( 2004a ). ‘Why some firms export’ , Review of Economics and Statistics , vol. 86 , pp. 561 – 9 . Google Scholar Crossref Search ADS WorldCat Bernard , A. and Jensen , J.B. ( 2004b ) ‘Entry, expansion and intensity in the U.S. export boom, 1987–1992’ , Review of International Economics , vol. 12 , pp. 662 – 75 . Google Scholar Crossref Search ADS WorldCat Bernard , A.B. , Redding , S. and Schott , P. ( 2003 ) ‘Product choice and product switching’ , NBER Working Paper 9789. Bernard , A.B. , Redding , S. and Schott , P. ( 2007 ) ‘Comparative advantage and heterogeneous firms’ , Review of Economic Studies , vol. 74 , pp. 31 – 66 . Google Scholar Crossref Search ADS WorldCat Bernard , A. and Wagner , J. ( 1997 ). ‘Exports and success in German manufacturing’ , Review of World Economics/Weltwirtschaftliches Archiv , vol. 133 , pp. 134 – 57 . Google Scholar Crossref Search ADS WorldCat Bigsten , A. , Collier , P., Decron , S., Fafchamps , M., Gauthier , B., Gunning , J.W., Habarurema , J., Oduro , A., Oostendrop , R., Pattilito , C., Soderbom , M., Teal , F. and Zeufack , A. ( 2000 ). ‘Exports and firm efficiency in African manufacturing’ , Centre for the Study of African Economies , WPS/2000–16, Oxford University. Blalock , G. and Gertler , P.J. ( 2004 ). ‘Learning from exporting revisited in less developed setting’ , Journal of Development Economics , vol. 75 , pp. 397 – 416 . Google Scholar Crossref Search ADS WorldCat Blalock , G. and Roy , S. ( 2007 ). ‘A firm level examination of the exports puzzle: why East Asian exports didn’t increase after 1997–1998 financial crisis?’ , The World Economy , vol. 30 , pp. 39 – 59 . Google Scholar Crossref Search ADS WorldCat Bleaney , M.F. and Wakelin , K. ( 2002 ). ‘Efficiency, innovation and exports’ , Oxford Bulletin of Economics and Statistics , vol. 64 , pp. 3 – 15 . Google Scholar Crossref Search ADS WorldCat Blundell , R. and Costa Dias , M ( 2000 ). ‘Evaluation methods for non‐experimental data’ , Fiscal Studies , vol. 21 , pp. 427 – 68 . Google Scholar Crossref Search ADS WorldCat Brainard , S.L. ( 1987 ). ‘A simple theory of multinational corporations and trade with a trade‐off between proximity and concentration’ , American Economic Review , vol. 84 , pp. 447 – 65 . OpenURL Placeholder Text WorldCat Brainard , S.L. ( 1993 ). ‘A simple theory of multinational corporations and trade with a trade‐off between proximity and concentration’ , NBER Working Paper No. 4269. Bugamelli , M. and Infante L. ( 2002 ). ‘Sunk costs to exports’ , Bank of Italy Research Papers. Castellani , D. ( 2002 ). ‘Export behaviour and productivity growth: evidence from Italian manufacturing firms’ , Weltwirtschaftliches Archiv , vol. 138 , pp. 605 – 28 . Google Scholar Crossref Search ADS WorldCat Castellani , D. and Zanfei , A. ( 2004 ). ‘Internationalisation, innovation and productivity: how do firms differ in Italy?’ , The World Economy, (forthcoming). Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Chowdhury , A.R. ( 1993 ). ‘Does exchange rate volatility depress trade flows? Evidence from error correction models’ , Review of Economics and Statistics , vol. 75 , pp. 700 – 6 . Google Scholar Crossref Search ADS WorldCat Clerides , S. , Lach , S. and Tybout , J. ( 1998 ). ‘Is learning by exporting important? Micro‐dynamic evidence from Columbia, Mexico and Morocco’ , Quarterly Journal of Economics , vol. 113 , pp. 903 – 48 . Google Scholar Crossref Search ADS WorldCat Damijan , J. , Polanec S. and Prašnikar J. ( 2007 ). ‘Self‐selection, export market heterogeneity and productivity improvements: firm level evidence from Slovenia’ , The World Economy , vol. 30 , pp. 135 – 55 . Google Scholar Crossref Search ADS WorldCat Das , S. , Roberts , M.J. and Tybout , J. ( 2001 ). ‘Micro‐foundations of export dynamics’ , mimeo, Pennsylvania State University. Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Das , S. , Roberts , M.J. and Tybout , J. ( 2004 ). ‘Market entry costs, producer heterogeneity, and export dynamics’ , NBER Working Paper 8629. De Loecker , J. ( 2004 ). ‘Do exports generate higher productivity? Evidence from Slovenia’ , Katholieke Universiteit Leuven , LICOS Discussion Paper 151/2004. Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Delgado , M. , Fariñas , J. and Ruano , S. ( 2002 ). ‘Firm productivity and export markets: a non‐parametric approach’ , Journal of International Economics , vol. 57 , pp. 397 – 422 . Google Scholar Crossref Search ADS WorldCat Ederington , J. and McCalman , P. ( 2004 ). ‘Endogenous firm heterogeneity and the dynamics of trade liberalisation’ , University of Kentucky Working Paper . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Ekholm , K. , Forslid , R. and Markusen , J.R. ( 2003 ). ‘Export‐platform foreign direct investment’ , NBER Working Paper No. 9517. Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Falvey , R. , Greenaway , D, Gullstrand , J. and Yu , Z. ( 2004 ). ‘Exports, restructuring and industry productivity growth’ , GEP Research Paper 04/40, Leverhulme Centre for Research on Globalisation and Economic Policy , University of Nottingham. Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Falvey , R. , Greenaway , D and Yu , Z. ( 2004 ). ‘Efficiency differentials and intra‐industry trade’ , GEP Research Paper 04/05, Leverhulme Centre for Research on Globalisation and Economic Policy , University of Nottingham. Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Fernandes , A.M. and Isgut , A. ( 2005 ). ‘Learning‐by‐doing, learning‐by‐exporting, and productivity: evidence from Colombia’ , World Bank Policy Research Working Paper No. 3544. Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Girma , S. , Görg , H. and Strobl , E. ( 2004 ). ‘Exports, international investment, and plant performance: evidence from a non‐parametric test’ , Economics Letters , vol. 83 , pp. 317 – 24 . Google Scholar Crossref Search ADS WorldCat Girma , S. , Greenaway , D. and Kneller , R. ( 2003 ) ‘Export market exit and performance dynamics: a causality analysis of matched firms’ , Economics Letters , vol. 80 , pp. 181 – 7 . Google Scholar Crossref Search ADS WorldCat Girma , S. , Greenaway , D. and Kneller , R. ( 2004 ). ‘Does exporting increase productivity? A microeconometric analysis of matched firms’ , Review of International Economics , vol. 12 , pp. 855 – 66 . Google Scholar Crossref Search ADS WorldCat Girma , S. , Kneller , R. and Pisu , M. ( 2005a ). ‘Exports versus FDI: an empirical test’ , Weltwirtschaftliches Archiv , vol. 12 , pp. 855 – 66 . OpenURL Placeholder Text WorldCat Girma , S. , Kneller , R. and Pisu , M. ( 2005b ). ‘Trade creation, destruction and replacement in regional trade agreements: micro level evidence for the UK’ , Review of International Economics, (forthcoming). Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Görg , H. and Greenaway , D. ( 2004 ). ‘Much ado about nothing? Do domestic firms really benefit from foreign direct investment?’ , World Bank Research Observer , vol. 19 , pp. 171 – 97 . Google Scholar Crossref Search ADS WorldCat Görg , H. , Henry , M. and Strobl , E. ( 2007 ). ‘Grant support and exporting activity: Evidence from Irish manufacturing’ , Review of Economics and Statistics, (forthcoming). Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Greenaway , D. , Guariglia , A. and Kneller , R. ( 2005 ). ‘Do financial factors affect exporting decisions?’ , GEP Research Paper 05/28, Leverhulme Centre for Research on Globalisation and Economic Policy , University of Nottingham. Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Greenaway , D. , Gullstrand , J., and Kneller , R. ( 2005 ). ‘Exporting may not always boost firm level productivity’ , Weltwirtschaftliches Archiv , vol. 141 , pp. 561 – 82 . Google Scholar Crossref Search ADS WorldCat Greenaway , D. , and Kneller , R. ( 2003 ). ‘Exporting, productivity and agglomeration: a matched difference in difference analysis of matched firms’ . GEP Research Paper 03/45, Leverhulme Centre for Research on Globalisation and Economic Policy , University of Nottingham. Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Greenaway , D. , Morgan , W., and Wright , P. ( 2002 ). ‘Trade liberalisation and growth in developing countries’ , Journal of Development Economics , vol. 67 , pp. 229 – 44 . Google Scholar Crossref Search ADS WorldCat Greenaway , D. , Sousa , N. and Wakelin , K. ( 2004 ). ‘Do domestic firms learn to export from multinationals?’ , European Journal of Political Economy , vol. 20 , pp. 1027 – 44 . Google Scholar Crossref Search ADS WorldCat Greenaway , D. , and Yu , Z. ( 2004 ). ‘Firm level interactions between exporting and productivity: industry‐specific evidence’ , Weltwirtschaftliches Archiv , vol. 140 , pp. 376 – 92 . Google Scholar Crossref Search ADS WorldCat Grossman , G.M. , Helpman E., and Adam Szeidl ( 2003 ). ‘Optimal integration strategies for the multinational firm’ , NBER Working Paper 10189. Hahn C.H. ( 2004 ). ‘Exporting and performance of plants: evidence from Korean manufacturing’ , NBER Working Paper 10208. Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Hansson , P. and Lundin N. ( 2004 ) ‘Exports as indicator on or a promoter of successful Swedish manufacturing firms in the 1990s’ , Weltwirtschaftliches Archiv , vol. 140 , pp. 415 – 45 . Google Scholar Crossref Search ADS WorldCat Head , K. and Ries , J. ( 1999 ). ‘Rationalization effects of tariff reductions’ , Journal of International Economics , vol. 47 , pp. 295 – 320 . Google Scholar Crossref Search ADS WorldCat Head K. and Ries , J ( 2001 ). ‘Overseas investment and firm exports’ , Review of International Economics , vol. 9 , pp. 108 – 22 . Google Scholar Crossref Search ADS WorldCat Head , K. and Ries , J. ( 2003 ). ‘Heterogeneity and the foreign direct investment versus exports decision of Japanese manufacturers’ , Journal of the Japanese and International Economies , vol. 17 , pp. 448 – 67 . Google Scholar Crossref Search ADS WorldCat Head , K. and Ries J. ( 2004 ). ‘Exporting and FDI as alternative strategies’ , Oxford Review of Economic Policy , vol. 20 , pp. 409 – 23 . Google Scholar Crossref Search ADS WorldCat Helpman , E. ( 2005 ). ‘Trade, FDI and the organisation of production’ , mimeo, Harvard University. Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Helpman , E. and Krugman , P. ( 1985 ). Market Structure and Foreign Trade , Cambridge, MA: MIT Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Helpman , E. , Melitz , M. and Yeaple , S. ( 2004 ). ‘Export versus FDI’ , American Economic Review , vol. 94 , pp. 300 – 16 . Google Scholar Crossref Search ADS WorldCat Holmes , T.J. and Schmitz , J.A. ( 2001 ). ‘A gain from trade: from unproductive to productive entrepreneurship’ , Journal of Monetary Economics , vol. 47 , pp. 417 – 46 . Google Scholar Crossref Search ADS WorldCat Isgut , A. ( 2001 ). ‘What’s different about exporters? Evidence from Colombian manufacturing’ , Journal of Development Studies , vol. 37 , pp. 57 – 82 . Google Scholar Crossref Search ADS WorldCat Jean , S. ( 2002 ). International trade and firms’ heterogeneity under monopolistic competition’ , Open Economics Review , vol. 13 , pp. 291 – 311 . Google Scholar Crossref Search ADS WorldCat Keller , W. ( 2004 ) ‘International technology diffusion’ , Journal of Economic Literature , vol. 42 , pp. 752 – 82 . Google Scholar Crossref Search ADS WorldCat Kimura , F. and Kiyota , K. ( 2004 ). ‘Exports, FDI and productivity of firm: cause and effect’ , Faculty of Business Administration , Yokohama National University, Working Paper 216. Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Kneller , R. and Pisu , M. ( 2004 ). ‘Export‐oriented FDI in the UK’ , Oxford Review of Economic Policy , vol. 20 , pp. 424 – 39 . Google Scholar Crossref Search ADS WorldCat Kneller , R. and Pisu , M. ( 2007 ). ‘Industrial linkages and export spillovers from FDI’ , The World Economy, vol. 30 , pp. 105 – 34 . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Kokko , A. , Tansini , R. and Zejan , M., ( 1997 ). ‘Trade regimes and spillover effects of FDI’ , mimeo, Stockholm School of Economics. Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Kiyota , K. and Urata , S. ( 2005 ). ‘The role of multinational firms in international trade: the case of Japan’ , RIETI Discussion Paper Series 05‐E‐012, Yokohama National University. Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Kraay , A. ( 1999 ). ‘Exports and economic performance: evidence from a panel of Chinese enterprises’ , Revue d’Economie du Developpement , vol. 2 , pp. 183 – 207 . Google Scholar Crossref Search ADS WorldCat Krugman , P. ( 1979 ). ‘A model of balance of payments crises’ , Journal of Money, Credit and Banking , vol. 11 , pp. 311 – 25 . Google Scholar Crossref Search ADS WorldCat Lipsey , R.E. , Ramstetter , E. and Blomström , M. ( 2000 ). ‘Outward FDI and parent exports and employment: Japan, the United States, and Sweden’ , Global Economy Quarterly , vol. 1 , pp. 285 – 302 . OpenURL Placeholder Text WorldCat Lipsey , R.E. , and Weiss , M.Y. ( 1984 ). ‘Foreign production and exports of individual firms’ , Review of Economics and Statistics , vol. 66 , pp. 304 – 8 . Google Scholar Crossref Search ADS WorldCat Liu , J‐T. , Tsou , M‐W, and Hammitt , J.K. ( 1999 ). ‘Export activity and productivity: evidence from the Taiwan electronics industry’ , Weltwirtschaftliches Archiv , vol. 135 , pp. 675 – 91 . Google Scholar Crossref Search ADS WorldCat López , R.A. ( 2004 ). ‘Self‐selection into the export markets: a conscious decision?’ , mimeo, Department of Economics, Indiana University. Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC López , R.A. ( 2005 ). ‘Trade and growth: reconciling the macroeconomic and microeconomic evidence’ , Journal of Economic Surveys vol. 19 , pp. 623 – 48 . Google Scholar Crossref Search ADS WorldCat Melitz , M. ( 2003 ). ‘The impact of trade on intra‐industry reallocations and aggregate industry productivity’ , Econometrica , vol. 71 , pp. 1695 – 725 . Google Scholar Crossref Search ADS WorldCat Melitz , M. and Ottaviano , G. ( 2003 ). ‘Market size, trade and productivity’ , mimeo, Harvard University. Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Motta , M. and Norman , G. ( 1996 ). ‘Does economic integration cause foreign direct investment?’ , International Economic Review , vol. 37 (4) , pp. 757 – 83 . Google Scholar Crossref Search ADS WorldCat Olley , S. and Pakes , A. ( 1996 ). ‘The dynamics of productivity in the telecommunications equipment industry’ , Econometrica , vol. 42 , pp. 217 – 42 . OpenURL Placeholder Text WorldCat Pack , H. ( 2000 ). ‘Modes of technology transfer at the firm level’ , mimeo, University of Pennsylvania. Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Parley , D. and Wei , S. ( 1993 ). ‘Insignificant and inconsequential hypothesis: the case of US, bilateral trade’ , Review of Economics and Statistics , vol. 95 , pp. 606 – 15 . Google Scholar Crossref Search ADS WorldCat Pavcnik N. ( 2002 ). ‘Trade liberalization, exit and productivity improvements: evidence from Chilean plants’ , Review of Economic Studies , vol. 69 , pp. 245 – 76 . Google Scholar Crossref Search ADS WorldCat Pozo , S. ( 1992 ). ‘Are flexible exchange rates really more volatile? Evidence from the early 1990s’ , Applied Economics , vol. 3 , pp. 87 – 105 . OpenURL Placeholder Text WorldCat Roberts , M.J. and Tybout , J. ( 1991 ). ‘Size rationalization and trade exposure in developing countries’, in ( R. Baldwin ed.), Empirical Studies of Commercial Policy , Chicago: Chicago Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Roberts , M.J. and Tybout , J. ( 1996 ). Industrial Evolution in Developing Countries , Oxford: Oxford University Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Rodriguez , F. , and Rodrik , D., ( 2000 ). ‘Trade policy and economic growth: a skeptic’s guide to the cross‐national evidence’, in ( B. Bernanke and K. S. Rogoff, eds.), Macroeconomics Annual , MIT Press for NBER , Cambridge, MA. Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Roper , S. and Love , J.H. ( 2002 ). ‘Innovation and export performance: evidence from the UK and German manufacturing plants’ , Research Policy , vol. 31 , pp. 1087 – 102 . Google Scholar Crossref Search ADS WorldCat Rose , A. and Stanley , T. ( 2005 ). ‘A meta‐analysis of the effect of common currencies on international trade’ , Journal of Economic Surveys , vol. 19 , pp. 347 . Google Scholar Crossref Search ADS WorldCat Ruane , F. and Sutherland J. ( 2007 ). ‘Foreign direct investment and export spillovers: how do export platforms fare?’ , European Journal of Political Economy, (forthcoming). Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Sjoholm , F. ( 2003 ). ‘Which Indonesian firms export? The importance of foreign networks’ , Papers in Regional Science , vol. 82 , pp. 333 – 50 . Google Scholar Crossref Search ADS WorldCat Swedenborg , B. ( 1985 ). ‘Sweden’, in ( J. Dunning, ed.), Multinational Enterprises, Economic Structure, and International Competitiveness , Chichester: Wiley/IRM Series on Multinationals . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Swenson , D.L. ( 2005 ). ‘Multinationals and the creation of Chinese trade linkages’ , mimeo , University of California, Davis. Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Tybout , J. ( 2003 ). ‘Plant and firm level evidence on new trade theories’, in ( E. Kwan Choi and J. Harrigan, eds.), Handbook of International Economics , pp. 388 – 415 , Oxford: Blackwell . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Tybout , J. and Westbrook , M.D. ( 1995 ). ‘Trade liberalization and dimensions of efficiency change in Mexican manufacturing industries’ , Journal of International Economics , vol. 31 , pp. 53 – 78 . Google Scholar Crossref Search ADS WorldCat Van Biesebroeck , J. ( 2005 ). ‘Exporting raises productivity in Sub‐Saharan manufacturing plants’ , Journal of International Economics , vol. 67 , pp. 373 – 91 . Google Scholar Crossref Search ADS WorldCat Wagner , J. ( 2002 ). ‘The causal effects of exports on firm size and labor productivity: first evidence from a matching approach’ , Economics Letters , vol. 77 , pp. 287 – 92 . Google Scholar Crossref Search ADS WorldCat Wagner , J. ( 2005 ). ‘Exports, foreign direct investment and productivity: evidence from German firm level data’ , mimeo , University of Lueneberg. Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Wagner , J. ( 2007 ). ‘Exports and productivity: a survey of the evidence from firm level data’ , The World Economy , vol. 30 , pp. 60 – 82 . Google Scholar Crossref Search ADS WorldCat Yeaple , S.R. ( 2003 ). ‘The complex integration strategies of multinationals and cross country dependencies in the structure of foreign direct investment’ , Journal of International Economics , vol. 60 , pp. 293 – 314 . Google Scholar Crossref Search ADS WorldCat Author notes " The authors acknowledge helpful comments on an earlier draft from three anonymous referees, Roberto Alvarez, Daniel Bernhofen, Ricardo Lopez, Jim Markusen, Horst Raff, participants at the Singapore Economic Review Annual Conference 2005, the Otago Trade Workshop 2006 and at a SUFE‐Orebro Conference in Shanghai in 2005. Financial support for The Leverhulme Trust under Programme Grant F114/BF is also gratefully acknowledged. © The Author(s). Journal compilation © Royal Economic Society 2007

Glaeser, Edward, L.

doi: 10.1111/j.1468-0297.2007.02031_4.xpmid: N/A

Review IV Daron Acemoglu and James Robinson’s magisterial book Economic Origins of Dictatorship and Democracy forcefully makes the case for using economic theory to understand democratisation. The book fuses two different strands and styles of research. Their use of formal models to understand democratisation has its antecedents in the positive political economy work of Barry Weingast, John Roemer, Herschel Grossman and others. Their ambitious objectives and liberal use of historical examples makes them the heirs of Barrington Moore, Samuel Huntington and Theda Skocpol. The combination of these approaches will hopefully ensure that their work will appeal not only to economists but also political scientists and in so doing will help to ensure that future work on democratisation will be grounded on the firmer soil of formal models. The most important thing about this book is its combination of theory and great historical breadth. Its models are inspired by the important historical episodes and help make sense of much of political history. The book makes strong and often convincing claims that it will be hard for anyone working in this area to ignore. For people who have been thinking about these topics for decades, the book is a challenge that needs facing. For younger researchers just beginning their work on democratisation, the book will be the natural starting part. Both groups of researchers will need to accept the language of formal theory to make sense of the contributions of Acemoglu and Robinson. The Argument of the Book Their book addresses two separate questions. First, the authors ask why political institutions matter at all. Why do people actually bother to change the political rules instead of just taking resources? Second, the authors ask what factors will determine the move from dictatorship to democracy. Why has democracy had such an easier time in England than in Argentina? Their answer to the first question is that institutions are commitment devices. Without changing the rules of the game, promises are hollow. Kings may promise things under duress, but they will quickly renege on their promises unless institutions have been put in place that keep them to their word. We need only remember how quickly John I sought (and received) Papal dispensation for reneging on the Magna Carta to see the obvious wisdom in this view. Acemoglu and Robinson favour the story of Richard II acceding to Wat Tyler’s demand and then completely disowning his promises but history is replete with similar examples. Because of this commitment problem, revolutionary actors seek institutional change, not just material goods. Even when the aims of revolutionaries are purely material, only through institutional change can their material aims be achieved in the long run. There are other less materialist approaches to regime change. Some have argued about the primacy of ideology, which is essentially explaining the demand for democracy with a taste for democracy. Unless those tastes are endogenised in a meaningful way, such theories will always seem unsatisfying and ad hoc. Acemoglu and Robinson pay much attention to elites changing institutions in response to popular pressure. The Reform Acts of the nineteenth century are particularly piquant examples of this phenomenon. They give less time to the revolutionaries who succeed and then end up writing the rules themselves. The Weimar constitution is better seen as the product of socialists, like Ebert, writing a revolution to entrench themselves after the Kaiser’s downfall than elites responding to external pressure. But this distinction is minor and not always that clear. Does it matter if the current elite rewrites the constitution at the point of a gun or if the gun is used to make them step aside so that the constitution is written by someone else? In either case, change is being forced by what Acemoglu and Robinson call de facto political power. And in either case, institutional change is being sought to ensure that future rents will flow more freely towards the group that currently has that de facto power. A more serious question is how effective institutional change can be as a commitment device. In some cases, new constitutions can be readily undermined by the old order. Between 1865 and 1876, American Northerners forced institutional change on the states of the Confederacy. Between 1876 and 1900, Southerners completely undermined much of the spirit of this change (albeit without reintroducing slavery), all the while sticking to the letter of the law (sort of). Even Hitler’s subversion of Weimar more or less followed the rules. Coups or counter‐revolutions that ignore the new democratic rules entirely are even more common. Acemoglu and Robinson put forward a flow diagram where de facto political power creates political institutions that then persist and ultimately create de jure political power. There is an alternative view that accepts the first part of their causal chain – political institutions certainly reflect de facto power – but that is sceptical about the transforming power of those institutions. The view that institutions are irrelevant is untenable. If they did not matter at all, then people would not spend so much time fighting over them. But there is at least a viable view that institutions primarily reflect de facto power, and only to a modest degree determine that power. The justly famous work by Acemoglu and Robinson, together with Simon Johnson (Acemoglu, Johnson and Robinson, 2001), on the lasting imprint of colonial origins can be interpreted through either view. According to the either view, colonies with small European minorities bent on exploiting native and slave populations developed institutions that reflected the unequal power in these societies. Colonies with larger populations of permanent European settlers had more democratic institutions reflecting the innate equality of these societies. According to the institutionalist view, these early institutions then determined modern institutions and modern incomes. The underlying power view suggests that the early differences in education and wealth in the population have persisted to this day and institutions continue to reflect these underlying inequalities of human and physical capital. The fact that one can disagree with the Acemoglu and Robinson’s confidence in institutions does not in any sense undermine the importance of this book. Their work is the starting point for debate. They have put forward a clear view and even more importantly set down the rules of engagement that combine formal models, regressions and history. Future work that debates these issues will only underline the importance of their work. Democracy Across the World Acemoglu and Robinson answer their second question – what determines the level of democracy – by pointing to incentives and resources. In a simple two‐class model of elites and citizens, the four broadly relevant criteria are the resources of citizens to force change, the resources of the elites to repress, and the incentives facing both groups to want or oppose democracy. They also offer a more complex three‐group model that throws in middle classes for good measure but again focuses on incentives and political resources. The focus on incentives and resources is surely the right starting point for any analysis of democracy and revolution. They are surely right that both forces have certainly mattered. The English elites who accepted electoral change in the nineteenth century did so in part because they thought that this change would not destroy their world. The rise of nineteenth century democracy certainly owes something to the ability of urbanised and, I would argue, more educated, masses to push change. While this broad framework is surely the right starting point for debate, there is plenty of room within that framework for disagreement and I will try to highlight the possible fissures that could be the starting point for future debate. At the broadest level, one can wonder whether incentives or political resources are more important. Acemoglu and Robinson give credit to both factors, but it is hard not to think that they are particularly attracted to the importance of incentives. This is unsurprising. They are economists after all, and as a profession, we have been attracted to the importance of incentives for centuries. In the context of democratisation, their focus on incentives and particularly on the incentives facing elites, leads them to focus on inequality. When highly unequal societies democratise, classic median voter results suggest that large‐scale redistribution will ensue. Elites therefore fight harder in more unequal societies to defend non‐democratic regimes. This presents one explanation for why highly unequal Argentina had such a rocky road to democracy relative to the more equal Anglo‐Saxon countries. The explanation is neat and elegant and surely has some elements of truth. Of course, even if incentives – not resources – are paramount, then we must ask about what inequality does to the incentives facing the disenfranchised. As Acemoglu and Robinson note, the poor stand to gain more as society becomes more unequal and this would make them more likely to fight for change. Acemoglu and Robinson show that inequality has a non‐monotonic impact on democratisation where the incentive effects work primarily to increase the incentives for the disenfranchised when inequality is low and to increase the incentives for the incumbents when inequality is high. This finding is clever, but the non‐monotonicity is far from general. It is an artifact of the timing of the model where elites move first and then masses respond. In a more general setting, inequality impacts both groups incentives and can increase, decrease or not impact democratisation at all. More generally, I do not think that any change that increases the transfers to the poor associated with the move to democracy can be said unequivocally to reduce the likelihood that the democracy will succeed or fail. Those transfers increase the incentives for one group and decrease them for the other. The net impact depends on which group has a more elastic response to this change in incentives. If your vision of regime change is Victorian England then it does seem reasonable to see the great parliamentarians as the primary actors and to think that these were highly rational men whose decisions were shaped regularly by costs and benefits. Their opponents, on the other hand, may well be seen as being somewhat less driven by clear economic considerations. However, in other settings – Imperial Austro‐Hungary for example – the old order seems guided by its own, often inexplicable motives, while the opposition seems far more rational. I am sceptical of the value of incentive‐based theories because I doubt that reasonably general models will yield tight predictions. If there were forces that clearly increased the net gains from Democratising, perhaps by ensuring that the gains to the masses were quite large while the losses to the elites were few, then economics would surely predict that those changes in incentives would predict democracy. The process of industrialisation with its need for large‐scale investment may represent such a shift. The gains from Democracy are not gains from a large franchise, but rather the gains from abandoning arbitrary autocratic rule in favour of a regime that can protect property. In this case, gains may be much higher than losses but without net gains to society as a whole, I have trouble seeing how incentives offer clear predictions about when we see democracy. On the other hand, I have much more sympathy for the view that political resources – the ability to force change and protect current regimes – drive democracy and dictatorship. According to this view, inequality matters not because of the incentives it provides to elites and masses but because economic inequality is closely related to political inequality. If the masses have few resources – either financial or educational – then we should not expect them to be able to overturn the better financed and educated elites. The important aspect about inequality, according to this view, is not that it makes the elites want to preserve a non‐democratic status quo more avidly but that inequality enables the elites to protect themselves better, because they are facing a weaker opposition. I think it is easy to argue for the primacy of resources over incentives. The great wave of European revolutions ‐ Russia, Germany, Austro‐Hungary – in the early twentieth century did not reflect changing incentives. The elite had not lost its desire to rule. These revolutions occurred because the elite had lost its ability to repress in the chaos of World War I. The wave of franchise reform certainly makes more sense as reflecting an increase in the ability of agitators to impose costs than as an increase in the incentives facing those agitators. Certainly, it is easy to argue that the important factor in Democratic uprisings is not the incentives to democratise, but rather the ability to effectively organise and push change. Similar arguments could be made about the defence of Democracy against coups. A democracy always faces threats from either an external or internal coup. The defence of democracy requires the masses to marshal significant political resources to defend their rights. It is at least arguable that the US has been better able to defend its democracy than its Latin American neighbours because political resources, particularly education, have been spread more widely across its population. It is at least arguable that the difference between France in 1789, when democracy led quickly to dictatorship and France in 1871, when democracy stuck, is an increasingly broad base of political skills that could fight back challenges of military adventurers like Napoleon and Boulanger. None of this contradicts Acemoglu and Robinson and all of it should be seen as a debate over how much attention should be given to the different forces that they highlight. Indeed, they are quite explicit in noting that ‘before the nineteenth century, the disenfranchised segments of society were scattered in rural areas; therefore, we may think of the threat of revolution as less severe because it was very difficult for them to organise’ (p. 68). However, despite this acknowledgment, there is much less in the book about the causes of de facto power than about the sources of incentives for democratisation. After reading this book, I have come to see my own work in this area with Andrei Shleifer and Giacomo Ponzetto (Glaeser et al., 2006) as exploring what we perceive to be the understudied topic of the sources of de facto power. We take the somewhat reductionist view that education is the prime determinant of that power. Masses are effective when they are educated and ineffective otherwise. Education contributes partially by making groups savvier in their activities and also by increasing the ability to organise. Much of human capital comes in social training and more education increasing civic and social organisation along almost every dimension. The available evidence even suggests that the educated make better soldiers. The spread of education then determines the success of democracy against dictatorship. Dictatorships provide sharp incentives for an elite to fight for the regime; democracies provide weak incentives for the masses. Only when the masses are adept both at fighting and overcoming free rider problems can democracy be defended against dictatorship. Our emphasis on education can be seen as part of a worldview that emphasises innate economic conditions rather than institutions, but it also can be seen as a logical extension of Acemoglu and Robinson that focuses particularly on the sources of political resources and de facto power. The book is important and insightful. I do not agree with all of its particulars and I have many petty disagreements with them over particular political episodes. But the work is still enormously significant and should order the debate on these topics for years to come. It is a great contribution to the field. References Acemoglu , Daron , Johnson , Simon and Robinson , James ( 2001 ). ‘The colonial origins of comparative development: an empirical investigation’ , American Economic Review , vol. 91 , pp. 1369 – 401 . Google Scholar Crossref Search ADS WorldCat Glaeser , Edward , Ponzetto , Giacomo and Shleifer , Andre ( 2006 ). ‘Why does democracy need education?’ , NBER Working Paper, no.12128. © The Author(s). Journal compilation © Royal Economic Society 2007

Cassar,, Alessandra;Crowley,, Luke;Wydick,, Bruce

doi: 10.1111/j.1468-0297.2007.02016.xpmid: N/A

Abstract An important question to microfinance is the relevance of existing social capital in target communities to the performance of group lending. This research presents evidence from field experiments in South Africa and Armenia, in which subjects participate in trust and microfinance games. We present evidence that personal trust between group members and social homogeneity are more important to group loan repayment than general societal trust or acquaintanceship between members. We also find some evidence of reciprocity: those who have been helped by other group members in the past are more likely to contribute in the future. During the past decade, exploring the role that social capital plays in economic behaviour has emerged as one of the most fascinating and fertile areas of economic research. Although precise definitions of social capital are notoriously difficult to pin down, one of the early pioneers of the concept, Coleman (1988), defines social capital as ‘social structure that facilitates certain actions of actors within the structure’. In his definition, Coleman specifically highlights the roles of mutual obligation, expectations and trustworthiness, social norms, social sanctions, and the transmission of information. Important studies in both developed and developing countries have analysed the impact of social capital in economic relationships. Putnam’s celebrated work, Making Democracy Work: Civic Traditions in Modern Italy (1993) and Bowling Alone: America’s Declining Social Capital (1995), brought attention to the role that social capital plays in the development of the modern state. Udry’s ground‐breaking (1994) work in Nigeria illustrated how the social capital existing in traditional societies may allow for more efficient credit contracts than in developed economies with weaker social capital. This research uses experimental methods to estimate the importance of social capital to the success of group lending, a commonly used tool to deliver credit to the poor in developing countries. Although group lending has been shown to be correlated with higher portfolio quality in microfinance institutions (see, for example, Cull et al. (2007) in this Feature), empirical work that has tried to isolate the influence of social capital on group loan repayment has faced a number of challenges. First, social capital and its various components are notoriously hard to measure. Moreover, groups often self‐select over different components of social capital, thus making it endogenous to actual loan repayment. While some recent work, such as the articles in this Feature by Ahlin and Townsend (2007) and by Karlan, (2007) has made important inroads in ameliorating these difficulties, our research investigates the influence of social capital using a different approach. We examine the effect of different components of relational social capital on group loan repayment by carrying out microfinance experiments on pools of subjects that reflect the characteristics of actual microloan recipients in Nyanga, South Africa and Berd, Armenia. In short, data from our experiments indicate that relational social capital in the form of personal trust between individuals and social homogeneity within groups has a positive effect on borrowing group performance. In contrast, we find that social capital as measured by simple acquaintanceship with other individuals or an individual’s general trust in society via responses to the standard General Social Survey questions has little effect on group performance. 1. Introduction Economists have developed numerous theories that seek to explain the high repayment rates frequently associated with group lending in developing countries. These theories can be roughly divided into three categories: (1) those that view the relational aspects of social capital as central to the performance of group lending; (2) those that view the informational aspects of social capital as central to the performance of group lending; and (3) those that view the merits of group lending (relative to individual lending) solely through its innate properties as a joint‐liability contract, where social capital plays little or no role. The distinction is important. If the first two groups of theories hold, the existing level of social capital in the form of strong personal relationships or local information may be critical to group lending’s success. If the third group of theories holds, then group lending may succeed whether or not it is implemented among borrowers with high levels of existing social capital. Our experiments primarily seek to ascertain the influence of relational social capital on group performance, and the type of relational social capital that is most critical to it. However, we believe our results also may have implications for informational social capital since, in practice, borrowers use private information to self‐select over aspects of relational social capital that may be conducive to the performance of their borrowing group. In borrowing groups with high levels of relational capital, strong social ties generate trust that other group members will contribute their share toward repaying group loans, thus making it worthwhile for each individual to repay. Moreover, because group members are jointly liable for repayment of the loan of each group member, they have an incentive to pressure fellow members who fail to maximise the probability that their own share of the group loan will be repaid. Ostensibly, the stronger the ties between group members, the greater the potential exists for social sanctions, and thus the more likely these sanctions are to lie off the equilibrium path, implying higher group loan repayment rates. The best known paper in this category is that of Besley and Coate (1995), who argue that without the potential for social sanctions, group lending may offer little if any advantage over individual lending. However, given that sanctions are sufficiently strong, group lending in their model is able to curtail the moral hazard associated with loan repayment. Social sanctions, combined with peer monitoring also play a role in papers such as Stiglitz (1990), Banerjee et al. (1994) and Armendáriz de Aghion (1999), though in focusing on peer monitoring, social sanctions are typically assumed to be exogenous. In the model of Wydick (2001), sanctions in the form of group expulsion are endogenous in that they represent a credible threat that comprises part of a perfect Bayesian equilibrium punishment strategy. Given a sufficiently low level of peer monitoring between borrowers, it is rational for group members to replace a defaulting member with a new member, even when there is no informational evidence of risky borrower behaviour. In a high‐information environment, expulsions and group replacements are only carried out if there is observable evidence of risky behaviour. The threat of social sanctions over and above group expulsion, however, only adds to the incentive to undertake safe investments. While the threat of social sanctions can clearly discipline borrowers in many of these papers, it is often unclear if simple group expulsion, and the resulting loss of low‐interest credit, is able to act as a strong substitute for social sanctions. Papers in the second category focus on the heightened informational flows that exist in high social‐capital areas, and their impact on group loan repayment. Foremost among these are papers by Van Tassel (1999) and Ghatak (1999) who both demonstrate that the borrower self‐selection process used in most group lending schemes improves repayment rates through mitigating adverse selection in credit markets. If borrowers have clear information over the riskiness of one another’s projects, they sort themselves into homogeneous groups through an assortative matching process. Van Tassel’s model in particular shows how a lender can offer a set of individual and group loan contracts such that only high‐ability borrowers will accept the group loan contract in equilibrium. The intuition is similar to the way insurance companies offer separate car insurance contracts to single and married drivers: insurance companies know that married drivers tend to be safer, and that it would be irrational to get married simply to pay less for car insurance. In Ghatak’s model, risky borrowers internalise their externality on the group through being yoked with other risky borrowers. Safe borrowers are drawn back into the credit pool as the equilibrium interest rate is reduced, thus increasing repayment rates. In both models, existing social capital is important only in that it facilitates informational flow between borrowers; social sanctions are unnecessary to their results. A third view of group lending downplays the influence of existing social capital in the performance of group lending altogether. The advantages of group lending over individual lending rest on neither the potential for social sanctions nor informational flows between members. Instead, the potential advantage of group lending arises simply from the terms of a joint liability contract. The best example of this view is Armendáriz de Aghion and Gollier (2000). They show that, in a pool of ‘safe’ and ‘risky’ borrowers, if the higher return realised by a risky borrower in the good state of nature is (uniquely) sufficient to cover for a defaulting group member, then the group lending contract can reduce the equilibrium interest rate and induce higher repayment rates relative to individual lending. The interesting point about their result is that unlike the models of Van Tassel and Ghatak, it does not rely on borrowers having an informational advantage over the lender. Their model is, however, sensitive to changes in assumptions about borrower returns. Some empirical work has sought to discriminate between these three classes of group lending theories but results have been mixed. Wenner (1995) provides some evidence that active screening and social pressure among members of 25 Costa Rican credit groups improved group performance. Zeller (1998) finds credit group performance positively related to social cohesion within groups. Wydick (1999) finds that while peer monitoring appears to have some positive effect on group loan repayment, strong social ties within groups appear to make it more difficult to pressure fellow members to repay loans. More recent research on larger microfinance data sets has yielded fascinating, but somewhat contradictory, results on this question. Gómez and Santor (2003) use a statistical matching model to compare default rates of 1,389 individual and group borrowers in a Canadian lending institution. Based on observable variables, they find both selection and incentive effects to be operational in explaining lower default rates for group loans relative to individual loans. Moreover, incentive effects appear to be strengthened when ‘low trust’ groups are removed from the sample, leaving groups within which there existed a higher degree of trust before applying for the loan. Their results, however, depend on the assumption that unobservable characteristics are uncorrelated with borrowing group formation. If borrowing groups admit members based on characteristics unobservable to the researcher, the results could overstate group‐borrowing effects. Nevertheless, their findings present evidence in favour of the positive effects of informational and relational social capital on group loan repayment. The conclusions of Ahlin and Townsend (2007) (this Feature) contrast somewhat with those of Gómez and Santor. Ahlin and Townsend’s (2007) logit estimation results support the group self‐selection models in the wealthier central region near Bangkok, and the models emphasising the importance of social sanctions in the poorer, northeastern Thailand. Yet the fact that they find strong social ties within borrowing groups to be negatively correlated with group repayment causes them to challenge the idea that group lending works through its ability to harness all types of existing social capital. They argue that aspects of social capital that facilitate social penalties for non‐repayment of group loans can be helpful to group lending, while social capital that inhibits social penalties can be harmful. The particular angle that we take with our research is most similar to that of Abbink et al. (2006), Ginéet al. (2005) and Karlan (2005), who use experimental methods to analyse group lending repayment. We use the taxonomy developed by Harrison and List (2004) to categorise our own work within this body of experimental research. Abbink et al. (2006) carry out a conventional lab experiment in which students in the social sciences at the University of Erfurt participate in a microfinance game. Student subjects were formed into 31 borrowing groups of varying sizes; groups were rewarded with subsequent ‘loans’ upon repayment of the previous loan. The game involves a stochastic element: each student‐borrower faces a 1/6 probability of a negative shock, forcing her to depend on fellow members to repay the amount due on the group loan. The researchers are able to draw interesting conclusions about the effect of group size, gender and social ties on loan repayment. To isolate the effect of social ties, they used two separate recruitment techniques. Some groups were formed of students registering individually for the experiment, minimising the degree of social ties between members. Other participants registered together in groups; in these groups social ties were stronger. Some of their results are intriguing. Self‐selected groups contributed mightily in the first round, but cooperation tended to fizzle among these groups in later rounds, while the cooperation of the randomly chosen groups started lower, but became more stable than the self‐selected groups as the rounds progressed. Their results show that social ties within groups induce higher, but less stable, group loan repayment and that the performance of borrowing groups with initially weak social ties may grow with experience together in group loan repayment. Ginéet al. (2005) carry out a framed field experiment in which subjects in central Lima received a ‘loan’ of 100 points. A framed field experiment differs from an artefactual field experiment in that the experimenter attempts to replicate, or ‘frame’ the experiment in the context of the actual task under study (in this case, group loan repayment). Subjects in their study had to invest their loan in either a safe project (yielding a certain return of 200 points) or risky project (yielding a return of 600 points with probability 1/2 and zero otherwise). In their experiment the researchers introduce multiple rounds contingent on project success, joint‐liability, complete information on one’s partner’s project choice and outcome, communication between partners and election of partners. The varying permutations allow the authors to identify the importance of dynamic incentives, insurance, monitoring, free‐riding, and group formation, respectively. Taken together, Ginéet al. (2005) find evidence that group lending may actually induce moral hazard (through risk‐taking and free‐riding) rather than reduce it, though group self‐selection counteracts some of these problems. Karlan’s (2005) research employs an artefactual field experiment, which he then links to observational data. An artefactual experiment differs from a conventional lab experiment in that it uses a non‐standard subject pool that is pertinent to the issue being studied: members of 41 female borrowing groups in a Peruvian microfinance program in Karlan’s research replace the usual student subjects. He then tracks the behaviour of these same subjects over the course of one year after they received real microfinance loans. Initially, experimenters had each of the subjects play the trust game in which either 0, 1, 2, or 3 coins are passed from player A to a partner, player B. If at least one coin was passed, the experimenters matched the contribution to player B, who could pass back as many coins as desired to player A. Karlan (2005) finds that some characteristics related to cultural homogeneity such as both partners being indigenous, living nearby, and attending the same church are correlated with player A originally passing more coins, though social cohesion has a much weaker affect on the number of coins passed from B back to A. Over the course of the following year as borrowers repay their loans, the propensity for a borrower to pass coins in the role of player A is actually correlated with a lower level of savings and a higher rate of group expulsion/dropout. Karlan accounts for this result by noting that a higher propensity for a player A to pass coins may reflect a higher propensity to gamble rather than a higher propensity to trust. Additionally he finds that positive responses by borrowers to General Social Survey questions intended to measure social capital are negatively correlated with default and group dropout. Taken together, his results indicate moderate support for importance of existing social capital between members to group lending but, specifically, the importance of innate trustworthiness, as opposed to trustworthiness driven by the fear of social sanctions. Our research consists of both artefactual and framed field experiment components, in that we employ the trust game used by Karlan, (2005) and the microfinance game of Abbink et al. (2006) respectively. While Karlan’s (2005) work was carried out among indigenous peoples of Western Hemisphere, we choose two very different locations: Nyanga, South Africa for a smaller pilot study and Berd, Armenia for our main study. As Ahlin and Townsend (2007) show in this Feature, the relative effects of different joint‐liability mechanisms may display considerable variation between clients and geographic regions. Thus we see it as advantageous to look for similarities and differences in the relationship between existing social ties and group loan repayment between substantially different subject pools and geographical areas. We favour the microfinance game developed by Abbink et al. (2006) because it effectively captures the idea that group lending is heavily dependent on dynamic incentives. Individuals have an incentive to repay group loans if they believe a critical mass of other members will do the same in order to receive future group loans. The belief that other members will contribute in the current round is partially a function of the social capital that exists within the borrowing group, which may be a product of a borrowing group member’s (a) generalised trust in her society as a whole, (b) level of acquaintanceship with fellow group members, (c) specific trust toward group members, or (d) trust that emerges from early rounds of positive experience with other members in group loan repayment. We use virtually the same experimental methodology for our smaller study in South Africa as we do in Armenia, though our questions to ascertain the level of social cohesion within microfinance game groups obviously needed to be distinct between sites (e.g. there are no clans in Armenia, and no post‐Perestroika generation in South Africa). We use results from our trust games to obtain measures of trust and trustworthiness for our microfinance games. We also include measures of existing levels of trust and social capital between the subjects in our 36 microfinance game groups such as age, intensity of relationship, years members have known one another, whether a subject would be willing to lend another subject money, and distance between their homes. If our different measures of relational social capital within our exogenously formed borrowing groups are significantly associated with superior borrowing group performance in our experiments, then we would interpret this as evidence that these aspects of relational social capital may matter to real‐world group loan repayment. To the extent that these measures of relational social capital are insignificantly related to borrowing group performance, we would take this as evidence that variability in borrowing group performance may be due to other factors which we do not account for in our experiments, such as peer monitoring or contractual variations. We include results from both individual group member and group repayment decisions. Our results indicate first that specific trust between a borrower and other individual group members appears to be relatively more important than trust in society as a whole for group loan repayment. This holds true for our subjects in both South Africa and Armenia. We find that group loan repayment appears to be more heavily associated with affirmative answers to questions such as, ‘Would you lend (person x) 1000 drams?’ than questions from the General Social Survey intended to measure broadly existing trust in society. Second, we find moderate evidence that social homogeneity in borrowing groups may be helpful. Having a larger number of one’s own clan as members in the group spurred individual contributions in South Africa, while having a high number and homogeneous makeup of long‐term local residents facilitated group repayment in Armenia. Third, we find mere acquaintanceship between members to be unrelated to group performance. Since social sanctions are generally ineffectual without at least weak social ties between individuals, our study suggests that potential social sanctions may not be the most important component of relational social capital to influence group loan repayment; interpersonal trust appears to be more important. We also find that when group repayment begins to break down from random shocks or non‐contribution, individuals withhold their own contributions, apparently to avoid getting burned by contributing to a losing cause. But our results also reveal evidence of reciprocity: when a member experiencing a negative shock is helped by others to repay the group loan, the benefiting member is more likely to contribute in the subsequent round. The remainder of our article is organised as follows: Section 2 provides details of our experimental methodology in Nyanga and Berd. Section 3 presents and discusses results from our experimental data. Section 4 concludes with a summary of how our results compare with those of the existing empirical literature. 2. Experimental Design 2.1. Locations We conducted a smaller pilot experiment at the SHAWCO1 Senior Centre in Nyanga, Cape Town, South Africa, pop. 24,003. Nyanga is a poor town, made up of nearly all black residents, and where annual per capita income is approximately 30,553 rand (US$4,460) (Republic of South Africa 2003 National Census). The HIV prevalence rate in Nyanga is one of the highest in the area. The subjects were identified by the neighbourhood representatives of the local municipality and experienced SHAWCO staff as women who fit the profile of the typical microcredit borrower in the region: eighteen years of age and older, either employed or available for work,2 and willing to participate in the experiment that took place from June 10th to July 10th, 2004. From the pool of potential subjects, a systematic sampling took place whereby a subject fitting the profile from every fifth eligible household was selected to participate. We conducted our second, larger experiment at the Artig Business Company (ABC) in Berd, Tavush Marz, Armenia (pop. 8,700), with per capita income 1,830,000 drams (US$3,900), roughly comparable to Nyanga. The subjects were identified by the ABC using the same criteria established above, with the experimental period lasting from March 19th to April 6th, 2005. In both experiments, women who had a previous professional relation with either the SHAWCO in Nyanga or the ABC in Armenia or who had ever been part of a joint‐liability borrowing group were excluded from the subject pool. In Nyanga, 87 women completed the general survey, 62 of them participated in the trust game experiment and 60 participated in the microfinance experiment.3 In Berd, 160 women completed the general survey and participated in the trust game experiment, and 156 of them participated in the microfinance experiment. 2.2. Survey In the Nyanga experiment, the subjects filled out a 38‐question survey, which took approximately 15 to 20 minutes to complete. The survey contained demographic questions such as age, length of residency, spoken languages, clan name as well as questions related to the various affiliations of the subject, her level of participation in groups and associations (e.g. political organisations, churches, ROSCAS etc.). In Berd, the subjects filled out a 26‐question survey that also required about 15 to 20 minutes to complete.4 In addition to questions related to demographic characteristics and the subject’s involvement in society, the Berd questionnaire included three attitudinal questions from the General Social Survey (GSS) that relate to trust, also used in Karlan (2005).5 The subjects were guaranteed a minimum of 1,500 drams upon completion of the survey and the two follow‐up activities with final payment depending on the outcomes of the games. (We were careful not to mention the words ‘game’ or ‘play’ in favour of the more neutral terms ‘activity’ and ‘decision making’.) After completing the surveys, the subjects participated in the trust game and microfinance game experiments. In Berd we alternated the order in which the experiments were played to account for the possible dependence of one game’s results from the results of the game previously played. 2.3. Trust Game As in the original trust game design of Berg et al. (1995), pairs of individuals are randomly matched and assigned the role of either ‘sender’ or ‘receiver’. In the Berd experiment, our largest subject pool, we ran two kinds of treatments: a treatment with equal initial endowments (senders and receivers starting with 1,000 drams), as well as a treatment with unequal endowments (senders starting with 1,000 drams and receivers starting with 400 drams). In Nyanga we ran only the unequal starting endowment treatment with senders starting with 25 rand and receivers starting with 10 rand. We used the treatment with unequal endowments because it more closely represents an actual microfinance situation in which both initial assets and returns are seldom equal between members, as well as to explore fairness issues in the trust game. The trust game consists of two stages. In the first stage, the sender has to choose how much of the initial endowment to send to the receiver (the ratio of the amount sent to the initial endowment is considered a measure of trust). The amount sent is then multiplied by three by the experimenter and passed to the receiver. In the second stage, the receiver then has the opportunity to return some of the received amount back to the sender (the ratio of the amount returned to the amount received is considered a measure of trustworthiness). In Berd, approximately two weeks after completing the general survey, twelve groups of 10 to 18 subjects, were formed and allocated to the different games, depending on whether they were chosen to play the trust game before or after participating in the microfinance game. In addition, as we explain below, we did control for whether the subjects began their professional lives before (or during) perestroika or post‐perestroika. The reading of the instructions occurred in front of the entire group. During the actual playing of the game, the pairings remained anonymous. In Nyanga, the trust game experiments were played between pairs of individuals from opposite sides of town with no previous level of social connection. Approximately one week after completing the general survey, six groups of fourteen to eighteen subjects were randomly formed and, over the course of four days, were called to the SHAWCO Senior Center. As in Glaeser et al. (2000), the subjects saw with whom they were matched but we ensured that they had never met one another before the game to control for the effect of personal social connectedness on trust behaviour. Individuals who arrived together or talked with each other were not paired together; otherwise, they were paired so as to maximise the physical distance between their households and, therefore, to minimise the chances that they had a personal relationship (corroborated by an exit interview). The pairings were not made public before the reading of the instructions.6 The subjects were then divided into two groups, senders and receivers. One pair at a time, they proceeded into a different room where a second experimenter ran the trust game experiment and administered the exit questionnaire. Summary results from the trust games are given in the appendix in Table A1. 2.4. Microfinance Game The microfinance experiment follows Abbink et al. (2006), with some minor modifications. A group of six individuals receive a loan of 30 rand (3,000 drams in Berd), for which all group members are jointly liable for repayment. The loan enables each member of the group to invest 5 rand (500 drams) in an individual risky project. All projects are of the same type and the probability of success is 5/6. In the event of a successful project, the investor receives a project payoff of 12 rand (1200 drams). If the project fails, however, the subject receives zero. After the outcomes of the projects are realised, the group loan plus interest must be repaid. We assumed a group loan interest rate of 20%, so that the group is liable to repay a total amount of 36 rand (3600 drams). The individuals whose project failed cannot contribute to group loan repayment, so the group debt is split among those individuals whose projects succeed and decide to contribute. Information on the individual project’s success or failure is private so that no other member of the group can ascertain whether a group partner’s defaults are due to project failure or strategic decision making. In this environment, loan repayment may ensue in the absence of contract enforceability. Since the debt is evenly divided among those individuals who are both able and willing to contribute, the fewer the number of contributors, the higher their burden. Since contributions can only be financed from the current project payoffs, full repayment is only possible if at least half of the group members (three subjects) decide to contribute. At the end of the round, the players are informed about the number of contributors, but not their identities, and their resulting payoff (one’s project payoff minus own share of repayment). If the group fulfils its repayment obligation, the game continues into a further round, which proceeds in the same way with the same group members. If more than half the group members default, regardless of whether the default is strategic or due to project failure, then the group cannot repay the full amount, and no further rounds are played. We like this feature of the game because it replicates the dynamic incentives utilised by most microfinance institutions, which make follow‐up loans conditional on the full repayment of previous loans. One aspect of the Abbink et al. (2006) experiment has been questioned by some researchers – by Armendáriz de Aghion and Morduch (2005) for example – namely that the results of the experiment are more difficult to interpret because participants are told that it will consist of a finite number of rounds (ten), leading to the traditional unravelling problem in which non‐contribution in all rounds is a subgame‐perfect equilibrium. We consequently modify their experiment slightly by creating, after the sixth round, a 1/6 probability that a group continues for another round. To minimise contamination from subjects taking into account an impending end‐game, we utilise data from rounds one to six in our analysis. (Our fundamental results are unchanged by excluding the small amount of data from later rounds.) To isolate the effect of social capital on group performance most efficiently, the microfinance game groups were formed so as to maximise the number of group members who shared the same clan name in Nyanga. To carry out tests for social homogeneity within groups, some groups were ‘stacked’ with individuals who shared the same clan name, which were made public during pre‐game introductions; otherwise, they were randomly formed. The microfinance experiments were played about one week after the trust game experiments. In the Berd experiment, one‐third of the microfinance groups were formed by those who began their working lives before or during perestroika, one‐third by individuals who began their working lives post‐perestroika, and one‐third was mixed (we used a cut‐off age of 36 to identify this). The experiments were played either one week before the trust games or one week after, depending on the subject pool. Subjects knew who belonged to their microfinance group in order to test for the effect of heterogeneity on repayment. 3. Empirical Results Our estimates use two separate units of observation. First, we look at the repayment behaviour of individuals in the microfinance games as a function of (a) negative shocks to themselves and the other five group members; (b) contributions by other members; (c) measures of acquaintanceship and personal trust level between the given individual and other members in the group; (d) measures of generalised trust by the given individual in the society and culture around them; (e) results from the trust game; and (f) social/cultural group homogeneity between the individual member and other group members. (Means and standard deviations of our independent variables are provided in Table 1. Summaries of group longevity and contribution rates in Berd and Nyanga are in Table 2.) Ideally for our type of unbalanced panel data, one would like to employ fixed or random effects estimation. However, the time‐invariant nature of most of the important variables in our study precludes fixed‐effect estimation. We carried out a Hausman test for the feasibility of a random effects estimator (which can be used on time‐invariant data) vis‐à‐vis fixed effects on our four time‐variant variables but the Hausman test rejected the null hypothesis of exogeneity of these right‐hand‐side variables at the 1% level. Instead, we run OLS on the average contribution of each individual for every round in which the borrower was able to contribute to group repayment (experienced no shock), the only weakness with this approach being possible path dependence in the denominator of the dependent variable. We report this estimation for our pilot study in Nyanga and our larger study in Berd in Table 3. To act as a check on the estimation for our principal study in Berd, we employ a logit estimation on our pooled, unbalanced panel data in Table 4, being aware that such estimation does overweight the frequency of individuals in the sample from groups with longer duration. We therefore employ as an additional check on the estimation, which is also included in Table 4, a logit estimation on individual rounds in Berd, for which there is no doubt of pure exogeneity or sample bias, but where estimation is performed on a sequence of smaller samples. To respect space constraints, we include rounds 2 to 5, before most groups had ceased repayment; the estimation on later rounds yield little additional insight, but are available upon request. We find remarkable consistency from our three types of estimation on individual repayment. Next, we examine the repayment behaviour at the borrowing group level using means and aggregates of many of these variables for each group of six borrowers. We present the results in Table 5, where we first show estimation for the 26 microfinance game groups in Berd. The dependent variable is the number of rounds of borrowing group survival (see Table 2), upon which we carry out OLS estimation. In this Table we also present some estimation where we combine the Berd‐Nyanga data set of 36 groups. For both Nyanga and Berd we create measures of group heterogeneity along significant social divisions in the respective societies. In Nyanga we focus on clan membership. But in Berd, due to the tremendous changes in post‐Soviet society, the greatest social division there is not ethnic but generational. Those who began work after Perestroika possess an economic and social outlook that is unusually distinct from that of the older generation. Along these lines we created statistically comparable indices of heterogeneity for our 36 borrowing groups, 10 in Nyanga and 26 in Berd. While the differences in heterogeneity are obviously quite distinct (ethnic vs. generational), we think it worthwhile to examine some aggregated measure of heterogeneity across our two study areas. We also examine heterogeneity in groups between ‘insiders’ and ‘outsiders’, creating an index that is more heterogeneous (homogeneous) if the group contains a larger (smaller) variation of long‐term residents and newcomers as measured by the standard deviation in the number of years of members’ local residency. We summarise our results below and juxtapose them to those of related empirical literature when appropriate comparisons can be made: Table 1 Summary Statistics Group repayment variables: . Berd, Armenia , σ . Nyanga, South Africa , σ . Individual repayment variables: (only used for Berd) . Berd, Armenia , σ . No. observations 26 10 No. observations 498 Number of rounds reached by group 4.192 3.000 Subject contributes in round 0.731 (1.650) (1.563) (0.444) Mean per period shocks received by group 1.119 1.358 Shock to subject period before 0.143 (0.678) (0.416) (0.350) Mean number of others acquaintances in group 1.489 0.733 Shocks to others in group period before 0.713 (1.033) (0.424) (0.800) Mean number others would loan to in group 1.590 0.450 Contrib. by others in group period before (dram) 3000.00 (0.849) (0.385) (416.94) Mean distance, km between members’ homes 26.727 10.851 Num. of acquaintances in group 1.382 (5.428) (1.705) (1.318) Mean fraction of life lived in area 17.867 23.730 Num. of group members subject would loan to 1.677 (13.114) (3.740) (1.753) Heterogeneity‐fraction life lived in area 0.243 0.750 Mean distance to others’ homes 26.668 (0.129) (0.339) (8.028) % members work after Perestroika/Same clan 0.587 0.400 Fraction of life lived in area 0.1684 (0.3589) (0.378) (0.266) Heterogeneity in peer group/ clan, (given by std.dev.) 0.298 0.477 Fraction of others in peer group 0.752 (0.243) (0.402) (0.250) Sender trust 0.428 0.334 Sender trust (only senders) 0.431 (0.158) (0.103) (0.246) Receiver trustworthiness 0.437 0.360 Receiver trustworthiness (only receivers) 0.441 (0.140) (0.192) (0.228) GSS1: Trust question 0.660 GSS1: Trust question 0.637 (0.200) (0.481) GSS2: Fairness question 0.679 GSS2: Fairness question 0.677 (0.194) (0.468) GSS3: Helpfulness quest. 0.346 GSS3: Helpfulness question 0.357 (0.210) (0.480) Group repayment variables: . Berd, Armenia , σ . Nyanga, South Africa , σ . Individual repayment variables: (only used for Berd) . Berd, Armenia , σ . No. observations 26 10 No. observations 498 Number of rounds reached by group 4.192 3.000 Subject contributes in round 0.731 (1.650) (1.563) (0.444) Mean per period shocks received by group 1.119 1.358 Shock to subject period before 0.143 (0.678) (0.416) (0.350) Mean number of others acquaintances in group 1.489 0.733 Shocks to others in group period before 0.713 (1.033) (0.424) (0.800) Mean number others would loan to in group 1.590 0.450 Contrib. by others in group period before (dram) 3000.00 (0.849) (0.385) (416.94) Mean distance, km between members’ homes 26.727 10.851 Num. of acquaintances in group 1.382 (5.428) (1.705) (1.318) Mean fraction of life lived in area 17.867 23.730 Num. of group members subject would loan to 1.677 (13.114) (3.740) (1.753) Heterogeneity‐fraction life lived in area 0.243 0.750 Mean distance to others’ homes 26.668 (0.129) (0.339) (8.028) % members work after Perestroika/Same clan 0.587 0.400 Fraction of life lived in area 0.1684 (0.3589) (0.378) (0.266) Heterogeneity in peer group/ clan, (given by std.dev.) 0.298 0.477 Fraction of others in peer group 0.752 (0.243) (0.402) (0.250) Sender trust 0.428 0.334 Sender trust (only senders) 0.431 (0.158) (0.103) (0.246) Receiver trustworthiness 0.437 0.360 Receiver trustworthiness (only receivers) 0.441 (0.140) (0.192) (0.228) GSS1: Trust question 0.660 GSS1: Trust question 0.637 (0.200) (0.481) GSS2: Fairness question 0.679 GSS2: Fairness question 0.677 (0.194) (0.468) GSS3: Helpfulness quest. 0.346 GSS3: Helpfulness question 0.357 (0.210) (0.480) Open in new tab Table 1 Summary Statistics Group repayment variables: . Berd, Armenia , σ . Nyanga, South Africa , σ . Individual repayment variables: (only used for Berd) . Berd, Armenia , σ . No. observations 26 10 No. observations 498 Number of rounds reached by group 4.192 3.000 Subject contributes in round 0.731 (1.650) (1.563) (0.444) Mean per period shocks received by group 1.119 1.358 Shock to subject period before 0.143 (0.678) (0.416) (0.350) Mean number of others acquaintances in group 1.489 0.733 Shocks to others in group period before 0.713 (1.033) (0.424) (0.800) Mean number others would loan to in group 1.590 0.450 Contrib. by others in group period before (dram) 3000.00 (0.849) (0.385) (416.94) Mean distance, km between members’ homes 26.727 10.851 Num. of acquaintances in group 1.382 (5.428) (1.705) (1.318) Mean fraction of life lived in area 17.867 23.730 Num. of group members subject would loan to 1.677 (13.114) (3.740) (1.753) Heterogeneity‐fraction life lived in area 0.243 0.750 Mean distance to others’ homes 26.668 (0.129) (0.339) (8.028) % members work after Perestroika/Same clan 0.587 0.400 Fraction of life lived in area 0.1684 (0.3589) (0.378) (0.266) Heterogeneity in peer group/ clan, (given by std.dev.) 0.298 0.477 Fraction of others in peer group 0.752 (0.243) (0.402) (0.250) Sender trust 0.428 0.334 Sender trust (only senders) 0.431 (0.158) (0.103) (0.246) Receiver trustworthiness 0.437 0.360 Receiver trustworthiness (only receivers) 0.441 (0.140) (0.192) (0.228) GSS1: Trust question 0.660 GSS1: Trust question 0.637 (0.200) (0.481) GSS2: Fairness question 0.679 GSS2: Fairness question 0.677 (0.194) (0.468) GSS3: Helpfulness quest. 0.346 GSS3: Helpfulness question 0.357 (0.210) (0.480) Group repayment variables: . Berd, Armenia , σ . Nyanga, South Africa , σ . Individual repayment variables: (only used for Berd) . Berd, Armenia , σ . No. observations 26 10 No. observations 498 Number of rounds reached by group 4.192 3.000 Subject contributes in round 0.731 (1.650) (1.563) (0.444) Mean per period shocks received by group 1.119 1.358 Shock to subject period before 0.143 (0.678) (0.416) (0.350) Mean number of others acquaintances in group 1.489 0.733 Shocks to others in group period before 0.713 (1.033) (0.424) (0.800) Mean number others would loan to in group 1.590 0.450 Contrib. by others in group period before (dram) 3000.00 (0.849) (0.385) (416.94) Mean distance, km between members’ homes 26.727 10.851 Num. of acquaintances in group 1.382 (5.428) (1.705) (1.318) Mean fraction of life lived in area 17.867 23.730 Num. of group members subject would loan to 1.677 (13.114) (3.740) (1.753) Heterogeneity‐fraction life lived in area 0.243 0.750 Mean distance to others’ homes 26.668 (0.129) (0.339) (8.028) % members work after Perestroika/Same clan 0.587 0.400 Fraction of life lived in area 0.1684 (0.3589) (0.378) (0.266) Heterogeneity in peer group/ clan, (given by std.dev.) 0.298 0.477 Fraction of others in peer group 0.752 (0.243) (0.402) (0.250) Sender trust 0.428 0.334 Sender trust (only senders) 0.431 (0.158) (0.103) (0.246) Receiver trustworthiness 0.437 0.360 Receiver trustworthiness (only receivers) 0.441 (0.140) (0.192) (0.228) GSS1: Trust question 0.660 GSS1: Trust question 0.637 (0.200) (0.481) GSS2: Fairness question 0.679 GSS2: Fairness question 0.677 (0.194) (0.468) GSS3: Helpfulness quest. 0.346 GSS3: Helpfulness question 0.357 (0.210) (0.480) Open in new tab Table 2 Frequencies of Failures and Contribution Decisions (number of groups, number of actual contributors) Round Number . Berd, Armenia . Nyanga, South Africa . % Failures . % Contributions . % Failures . % Contributions . 1 21.79 71.79 15.0 68.33 (26, 112) (10, 41) 2 16.0 62.67 25.93 48.15 (25, 94) (9, 26) 3 15.87 62.70 13.33 56.67 (21, 79) (5, 17) 4 14.58 57.29 33.33 50.0 (16, 55) (3, 9) 5 9.72 63.89 16.67 50.0 (12, 46) (2, 6) 6 11.11 70.37 33.33 66.67 (9, 38) (1, 4) Round Number . Berd, Armenia . Nyanga, South Africa . % Failures . % Contributions . % Failures . % Contributions . 1 21.79 71.79 15.0 68.33 (26, 112) (10, 41) 2 16.0 62.67 25.93 48.15 (25, 94) (9, 26) 3 15.87 62.70 13.33 56.67 (21, 79) (5, 17) 4 14.58 57.29 33.33 50.0 (16, 55) (3, 9) 5 9.72 63.89 16.67 50.0 (12, 46) (2, 6) 6 11.11 70.37 33.33 66.67 (9, 38) (1, 4) Open in new tab Table 2 Frequencies of Failures and Contribution Decisions (number of groups, number of actual contributors) Round Number . Berd, Armenia . Nyanga, South Africa . % Failures . % Contributions . % Failures . % Contributions . 1 21.79 71.79 15.0 68.33 (26, 112) (10, 41) 2 16.0 62.67 25.93 48.15 (25, 94) (9, 26) 3 15.87 62.70 13.33 56.67 (21, 79) (5, 17) 4 14.58 57.29 33.33 50.0 (16, 55) (3, 9) 5 9.72 63.89 16.67 50.0 (12, 46) (2, 6) 6 11.11 70.37 33.33 66.67 (9, 38) (1, 4) Round Number . Berd, Armenia . Nyanga, South Africa . % Failures . % Contributions . % Failures . % Contributions . 1 21.79 71.79 15.0 68.33 (26, 112) (10, 41) 2 16.0 62.67 25.93 48.15 (25, 94) (9, 26) 3 15.87 62.70 13.33 56.67 (21, 79) (5, 17) 4 14.58 57.29 33.33 50.0 (16, 55) (3, 9) 5 9.72 63.89 16.67 50.0 (12, 46) (2, 6) 6 11.11 70.37 33.33 66.67 (9, 38) (1, 4) Open in new tab 3.1. Shocks To no surprise, the results show that negative random shocks impact both individual and group repayment. In our Berd estimation we found individual repayment to be higher when a member had received a shock the period before (implying other members had covered for her). We interpret this as evidence of reciprocity among our subjects: as an individual has been helped by other group members, she is more likely to contribute given the opportunity the next time. The effect is large and statistically significant at the 99% level. For every additional negative shock received, it increases the fraction of the time she contributes on average, depending on our specification. This finding supports a large theoretical literature on the prevalence and importance of reciprocity among similar populations pioneered by Scott (1976) and Fafchamps (1992). Shocks to other members have a negative effect on individual contributions, which is statistically significant in Nyanga. It appears that when players sense that the end of the game is impending due to a lack of contributions by other members, this causes the individual to want to avoid being the ‘sucker’ who contributes futilely in the last round. In the group estimation in Table 5, we see that an increase in the mean number of shocks per round by one reduces the number of rounds the borrowing group continues to receive loans by about 1.7. Table 3 Individual Repayment Decisions – Nyanga and Berd Dependent Variable: Fraction of Times Repaid Divided by Opportunities to Repay† Variable: . —OLS Estimation— . Indiv. Repayment:
Nyanga, South Africa . Indiv. Repayment:
Berd, Armenia . No. of observations: . 53 . 53 . 151 . 151 . Intercept 1.296*** 1.294*** 0.621*** 0.643*** (0.364) (0.363) (0.208) (0.228) Mean contribution from others −0.537* −0.496 −0.250 −0.272 (0.349) (0.350) (0.195) (0.203) Mean shocks received–self −0.111 −0.096 0.350*** 0.341*** (0.277) (0.280) (0.138) (0.141) Mean shocks received–others −0.215* −0.198 −0.038 −0.038 (0.138) (0.139) (0.045) (0.046) Knows others in group 0.058 0.026 −0.010 −0.009 (0.108) (0.109) (0.022) (0.022) Mean would loan to each indiv. in group 0.015 0.036 0.037*** 0.037*** (0.122) (0.123) (0.016) (0.016) Mean Distance to others’ homes 0.004 0.002 0.007*** 0.006** (0.018) (0.018) (0.003) (0.003) Fraction of life lived in area −0.077 −0.051 0.001 0.001 (0.070) (0.072) (0.001) (0.001) Fraction of others in clan/peer group 0.178*** 0.183*** 0.073 0.078 (0.078) (0.078) (0.114) (0.115) Fraction of pre‐Perestroika subjects −0.122 −0.141 (0.113) (0.117) Self pre‐Perestrioka dummy 0.082 0.086 (0.075) (0.077) Average sender trust from trust game −0.151 0.036 0.047 (0.286) (0.122) (0.127) Average receiver trustworthiness −0.293 0.081 0.086 (0.201) (0.123) (0.125) GSS1: Trust question −0.049 (0.062) GSS2: Fairness question 0.012 (0.061) GSS3: Helpfulness question 0.070 (0.057) R‐Squared 0.1882 0.2272 0.1309 0.1428 Adj R‐Squared 0.0439 0.0475 0.0559 0.0482 F‐Statistic 1.30 1.26 1.74 1.51 F‐Signif. 0.266 0.281 0.064 0.110 Variable: . —OLS Estimation— . Indiv. Repayment:
Nyanga, South Africa . Indiv. Repayment:
Berd, Armenia . No. of observations: . 53 . 53 . 151 . 151 . Intercept 1.296*** 1.294*** 0.621*** 0.643*** (0.364) (0.363) (0.208) (0.228) Mean contribution from others −0.537* −0.496 −0.250 −0.272 (0.349) (0.350) (0.195) (0.203) Mean shocks received–self −0.111 −0.096 0.350*** 0.341*** (0.277) (0.280) (0.138) (0.141) Mean shocks received–others −0.215* −0.198 −0.038 −0.038 (0.138) (0.139) (0.045) (0.046) Knows others in group 0.058 0.026 −0.010 −0.009 (0.108) (0.109) (0.022) (0.022) Mean would loan to each indiv. in group 0.015 0.036 0.037*** 0.037*** (0.122) (0.123) (0.016) (0.016) Mean Distance to others’ homes 0.004 0.002 0.007*** 0.006** (0.018) (0.018) (0.003) (0.003) Fraction of life lived in area −0.077 −0.051 0.001 0.001 (0.070) (0.072) (0.001) (0.001) Fraction of others in clan/peer group 0.178*** 0.183*** 0.073 0.078 (0.078) (0.078) (0.114) (0.115) Fraction of pre‐Perestroika subjects −0.122 −0.141 (0.113) (0.117) Self pre‐Perestrioka dummy 0.082 0.086 (0.075) (0.077) Average sender trust from trust game −0.151 0.036 0.047 (0.286) (0.122) (0.127) Average receiver trustworthiness −0.293 0.081 0.086 (0.201) (0.123) (0.125) GSS1: Trust question −0.049 (0.062) GSS2: Fairness question 0.012 (0.061) GSS3: Helpfulness question 0.070 (0.057) R‐Squared 0.1882 0.2272 0.1309 0.1428 Adj R‐Squared 0.0439 0.0475 0.0559 0.0482 F‐Statistic 1.30 1.26 1.74 1.51 F‐Signif. 0.266 0.281 0.064 0.110 ***95% Significance, **90% Significance, *85% Significance. Standard errors are given in parentheses. Open in new tab Table 3 Individual Repayment Decisions – Nyanga and Berd Dependent Variable: Fraction of Times Repaid Divided by Opportunities to Repay† Variable: . —OLS Estimation— . Indiv. Repayment:
Nyanga, South Africa . Indiv. Repayment:
Berd, Armenia . No. of observations: . 53 . 53 . 151 . 151 . Intercept 1.296*** 1.294*** 0.621*** 0.643*** (0.364) (0.363) (0.208) (0.228) Mean contribution from others −0.537* −0.496 −0.250 −0.272 (0.349) (0.350) (0.195) (0.203) Mean shocks received–self −0.111 −0.096 0.350*** 0.341*** (0.277) (0.280) (0.138) (0.141) Mean shocks received–others −0.215* −0.198 −0.038 −0.038 (0.138) (0.139) (0.045) (0.046) Knows others in group 0.058 0.026 −0.010 −0.009 (0.108) (0.109) (0.022) (0.022) Mean would loan to each indiv. in group 0.015 0.036 0.037*** 0.037*** (0.122) (0.123) (0.016) (0.016) Mean Distance to others’ homes 0.004 0.002 0.007*** 0.006** (0.018) (0.018) (0.003) (0.003) Fraction of life lived in area −0.077 −0.051 0.001 0.001 (0.070) (0.072) (0.001) (0.001) Fraction of others in clan/peer group 0.178*** 0.183*** 0.073 0.078 (0.078) (0.078) (0.114) (0.115) Fraction of pre‐Perestroika subjects −0.122 −0.141 (0.113) (0.117) Self pre‐Perestrioka dummy 0.082 0.086 (0.075) (0.077) Average sender trust from trust game −0.151 0.036 0.047 (0.286) (0.122) (0.127) Average receiver trustworthiness −0.293 0.081 0.086 (0.201) (0.123) (0.125) GSS1: Trust question −0.049 (0.062) GSS2: Fairness question 0.012 (0.061) GSS3: Helpfulness question 0.070 (0.057) R‐Squared 0.1882 0.2272 0.1309 0.1428 Adj R‐Squared 0.0439 0.0475 0.0559 0.0482 F‐Statistic 1.30 1.26 1.74 1.51 F‐Signif. 0.266 0.281 0.064 0.110 Variable: . —OLS Estimation— . Indiv. Repayment:
Nyanga, South Africa . Indiv. Repayment:
Berd, Armenia . No. of observations: . 53 . 53 . 151 . 151 . Intercept 1.296*** 1.294*** 0.621*** 0.643*** (0.364) (0.363) (0.208) (0.228) Mean contribution from others −0.537* −0.496 −0.250 −0.272 (0.349) (0.350) (0.195) (0.203) Mean shocks received–self −0.111 −0.096 0.350*** 0.341*** (0.277) (0.280) (0.138) (0.141) Mean shocks received–others −0.215* −0.198 −0.038 −0.038 (0.138) (0.139) (0.045) (0.046) Knows others in group 0.058 0.026 −0.010 −0.009 (0.108) (0.109) (0.022) (0.022) Mean would loan to each indiv. in group 0.015 0.036 0.037*** 0.037*** (0.122) (0.123) (0.016) (0.016) Mean Distance to others’ homes 0.004 0.002 0.007*** 0.006** (0.018) (0.018) (0.003) (0.003) Fraction of life lived in area −0.077 −0.051 0.001 0.001 (0.070) (0.072) (0.001) (0.001) Fraction of others in clan/peer group 0.178*** 0.183*** 0.073 0.078 (0.078) (0.078) (0.114) (0.115) Fraction of pre‐Perestroika subjects −0.122 −0.141 (0.113) (0.117) Self pre‐Perestrioka dummy 0.082 0.086 (0.075) (0.077) Average sender trust from trust game −0.151 0.036 0.047 (0.286) (0.122) (0.127) Average receiver trustworthiness −0.293 0.081 0.086 (0.201) (0.123) (0.125) GSS1: Trust question −0.049 (0.062) GSS2: Fairness question 0.012 (0.061) GSS3: Helpfulness question 0.070 (0.057) R‐Squared 0.1882 0.2272 0.1309 0.1428 Adj R‐Squared 0.0439 0.0475 0.0559 0.0482 F‐Statistic 1.30 1.26 1.74 1.51 F‐Signif. 0.266 0.281 0.064 0.110 ***95% Significance, **90% Significance, *85% Significance. Standard errors are given in parentheses. Open in new tab 3.2. Effect of Others’ Contributions Theory would posit that the contributions of other group members could have differential effects: Contributions by other members could generate peer effects or fairness effects that stimulate one’s own contribution, or it could provoke free‐rider problems. In Berd we find very modest evidence that the contributions of others in the prior period increase one’s own desire to contribute in the following period. While none of the coefficients is significant, the variable consistently carries a positive sign on repayment in every specification. Table 4 Individual Repayment Decisions in Microfinance Game–Berd Dependent Variable: 1 = Individual Contributes in Roundx Variable: . —Binary Logit on Pooled Panel Data— . Only Trustors (All Rounds) . Only Trustees (All Rounds) . All Rounds . Round 2 . Round 3 . Round 4 . Round 5 . Number of observations: 213 214 427 126 106 82 65 Intercept −2.981 −4.789 −3.335 −12.061* −2.630 −3.059 −5.935 (4.780) (4.360) (3.127) (8.163) (8.986) (8.685) (8.468) Subject contributed period before 1.936** 1.435 1.766*** 3.023* 2.077 2.956 2.744 (1.184) (1.064) (0.773) (2.027) (2.266) (2.165) (2.277) Shock to subject period before 1.452*** 0.770 1.131*** 1.096 1.412* 1.562* 3.871*** (0.545) (0.630) (0.401) (1.077) (0.993) (1.039) (1.583) Shocks to others in group period before −0.373* −0.179 −0.253* −0.043 −0.053 −0.375 −0.386 (0.247) 0.236 (0.165) (0.420) (0.541) (0.386) (0.668) Contributions by others in group – period before 0.0002 0.0007 0.0005 0.002 0.0004 0.0004 0.001 (0.001) (0.001) (0.0009) (0.002) (0.002) (0.002) (0.002) Num. of Acquaintances in group 0.159 −0.161 −0.054 −0.234 −0.231 0.179 −0.448* (0.163) (0.139) (0.099) (0.210) (0.238) (0.239) (0.306) Num. of group members subject would loan to 0.174* 0.280*** 0.241*** 0.508*** 0.447*** −0.080 0.353* (0.124) (0.120) (0.079) (0.191) (0.227) (0.162) (0.227) Mean distance to others’ homes 0.045* 0.054* 0.042*** 0.087*** 0.028 −0.005 0.100** (0.030) (0.021) (0.017) (0.036) (0.042) (0.037) (0.053) Fraction of life lived in area 0.005 0.008 0.008* 0.014* 0.005 −0.006 0.035** (0.008) (0.008) (0.005) (0.009) (0.010) (0.013) (0.021) Fraction of Others in peer group 0.057 0.347 0.180 0.631 −1.041 2.192** −1.218 (0.732) (0.742) (0.486) (0.919) (1.255) (1.187) (1.481) GSS1: Trust question −0.405 −0.175 −0.228 −0.174 0.023 −0.070 −1.474** (0.393) (0.402) (0.263) (0.556) (0.610) (0.633) (0.779) GSS2: Fairness question 0.614* 0.192 0.261 0.637 0.879* −1.013* 0.441 (0.419) (0.395) (0.260) (0.539) (0.598) (0.663) (0.805) GSS3: Helpfulness question 0.418 0.174 0.266 0.116 1.504** −0.092 −0.770 (0.386) (0.416) (0.267) (0.525) (0.827) (0.616) (0.726) Sender trust from trust game 0.137 (0.736) Receiver trustworthiness from trust game 1.523** (0.833) Likelihood ratio 29.4118 34.4995 53.4543 25.0006 34.9393 17.9668 18.3098 p‐value 0.0057 0.0010 <0.0001 0.0148 0.0005 0.1167 0.1066 Variable: . —Binary Logit on Pooled Panel Data— . Only Trustors (All Rounds) . Only Trustees (All Rounds) . All Rounds . Round 2 . Round 3 . Round 4 . Round 5 . Number of observations: 213 214 427 126 106 82 65 Intercept −2.981 −4.789 −3.335 −12.061* −2.630 −3.059 −5.935 (4.780) (4.360) (3.127) (8.163) (8.986) (8.685) (8.468) Subject contributed period before 1.936** 1.435 1.766*** 3.023* 2.077 2.956 2.744 (1.184) (1.064) (0.773) (2.027) (2.266) (2.165) (2.277) Shock to subject period before 1.452*** 0.770 1.131*** 1.096 1.412* 1.562* 3.871*** (0.545) (0.630) (0.401) (1.077) (0.993) (1.039) (1.583) Shocks to others in group period before −0.373* −0.179 −0.253* −0.043 −0.053 −0.375 −0.386 (0.247) 0.236 (0.165) (0.420) (0.541) (0.386) (0.668) Contributions by others in group – period before 0.0002 0.0007 0.0005 0.002 0.0004 0.0004 0.001 (0.001) (0.001) (0.0009) (0.002) (0.002) (0.002) (0.002) Num. of Acquaintances in group 0.159 −0.161 −0.054 −0.234 −0.231 0.179 −0.448* (0.163) (0.139) (0.099) (0.210) (0.238) (0.239) (0.306) Num. of group members subject would loan to 0.174* 0.280*** 0.241*** 0.508*** 0.447*** −0.080 0.353* (0.124) (0.120) (0.079) (0.191) (0.227) (0.162) (0.227) Mean distance to others’ homes 0.045* 0.054* 0.042*** 0.087*** 0.028 −0.005 0.100** (0.030) (0.021) (0.017) (0.036) (0.042) (0.037) (0.053) Fraction of life lived in area 0.005 0.008 0.008* 0.014* 0.005 −0.006 0.035** (0.008) (0.008) (0.005) (0.009) (0.010) (0.013) (0.021) Fraction of Others in peer group 0.057 0.347 0.180 0.631 −1.041 2.192** −1.218 (0.732) (0.742) (0.486) (0.919) (1.255) (1.187) (1.481) GSS1: Trust question −0.405 −0.175 −0.228 −0.174 0.023 −0.070 −1.474** (0.393) (0.402) (0.263) (0.556) (0.610) (0.633) (0.779) GSS2: Fairness question 0.614* 0.192 0.261 0.637 0.879* −1.013* 0.441 (0.419) (0.395) (0.260) (0.539) (0.598) (0.663) (0.805) GSS3: Helpfulness question 0.418 0.174 0.266 0.116 1.504** −0.092 −0.770 (0.386) (0.416) (0.267) (0.525) (0.827) (0.616) (0.726) Sender trust from trust game 0.137 (0.736) Receiver trustworthiness from trust game 1.523** (0.833) Likelihood ratio 29.4118 34.4995 53.4543 25.0006 34.9393 17.9668 18.3098 p‐value 0.0057 0.0010 <0.0001 0.0148 0.0005 0.1167 0.1066 ***95% Significance, **90% Significance, *85% Significance. Standard errors are given in parentheses. Open in new tab Table 4 Individual Repayment Decisions in Microfinance Game–Berd Dependent Variable: 1 = Individual Contributes in Roundx Variable: . —Binary Logit on Pooled Panel Data— . Only Trustors (All Rounds) . Only Trustees (All Rounds) . All Rounds . Round 2 . Round 3 . Round 4 . Round 5 . Number of observations: 213 214 427 126 106 82 65 Intercept −2.981 −4.789 −3.335 −12.061* −2.630 −3.059 −5.935 (4.780) (4.360) (3.127) (8.163) (8.986) (8.685) (8.468) Subject contributed period before 1.936** 1.435 1.766*** 3.023* 2.077 2.956 2.744 (1.184) (1.064) (0.773) (2.027) (2.266) (2.165) (2.277) Shock to subject period before 1.452*** 0.770 1.131*** 1.096 1.412* 1.562* 3.871*** (0.545) (0.630) (0.401) (1.077) (0.993) (1.039) (1.583) Shocks to others in group period before −0.373* −0.179 −0.253* −0.043 −0.053 −0.375 −0.386 (0.247) 0.236 (0.165) (0.420) (0.541) (0.386) (0.668) Contributions by others in group – period before 0.0002 0.0007 0.0005 0.002 0.0004 0.0004 0.001 (0.001) (0.001) (0.0009) (0.002) (0.002) (0.002) (0.002) Num. of Acquaintances in group 0.159 −0.161 −0.054 −0.234 −0.231 0.179 −0.448* (0.163) (0.139) (0.099) (0.210) (0.238) (0.239) (0.306) Num. of group members subject would loan to 0.174* 0.280*** 0.241*** 0.508*** 0.447*** −0.080 0.353* (0.124) (0.120) (0.079) (0.191) (0.227) (0.162) (0.227) Mean distance to others’ homes 0.045* 0.054* 0.042*** 0.087*** 0.028 −0.005 0.100** (0.030) (0.021) (0.017) (0.036) (0.042) (0.037) (0.053) Fraction of life lived in area 0.005 0.008 0.008* 0.014* 0.005 −0.006 0.035** (0.008) (0.008) (0.005) (0.009) (0.010) (0.013) (0.021) Fraction of Others in peer group 0.057 0.347 0.180 0.631 −1.041 2.192** −1.218 (0.732) (0.742) (0.486) (0.919) (1.255) (1.187) (1.481) GSS1: Trust question −0.405 −0.175 −0.228 −0.174 0.023 −0.070 −1.474** (0.393) (0.402) (0.263) (0.556) (0.610) (0.633) (0.779) GSS2: Fairness question 0.614* 0.192 0.261 0.637 0.879* −1.013* 0.441 (0.419) (0.395) (0.260) (0.539) (0.598) (0.663) (0.805) GSS3: Helpfulness question 0.418 0.174 0.266 0.116 1.504** −0.092 −0.770 (0.386) (0.416) (0.267) (0.525) (0.827) (0.616) (0.726) Sender trust from trust game 0.137 (0.736) Receiver trustworthiness from trust game 1.523** (0.833) Likelihood ratio 29.4118 34.4995 53.4543 25.0006 34.9393 17.9668 18.3098 p‐value 0.0057 0.0010 <0.0001 0.0148 0.0005 0.1167 0.1066 Variable: . —Binary Logit on Pooled Panel Data— . Only Trustors (All Rounds) . Only Trustees (All Rounds) . All Rounds . Round 2 . Round 3 . Round 4 . Round 5 . Number of observations: 213 214 427 126 106 82 65 Intercept −2.981 −4.789 −3.335 −12.061* −2.630 −3.059 −5.935 (4.780) (4.360) (3.127) (8.163) (8.986) (8.685) (8.468) Subject contributed period before 1.936** 1.435 1.766*** 3.023* 2.077 2.956 2.744 (1.184) (1.064) (0.773) (2.027) (2.266) (2.165) (2.277) Shock to subject period before 1.452*** 0.770 1.131*** 1.096 1.412* 1.562* 3.871*** (0.545) (0.630) (0.401) (1.077) (0.993) (1.039) (1.583) Shocks to others in group period before −0.373* −0.179 −0.253* −0.043 −0.053 −0.375 −0.386 (0.247) 0.236 (0.165) (0.420) (0.541) (0.386) (0.668) Contributions by others in group – period before 0.0002 0.0007 0.0005 0.002 0.0004 0.0004 0.001 (0.001) (0.001) (0.0009) (0.002) (0.002) (0.002) (0.002) Num. of Acquaintances in group 0.159 −0.161 −0.054 −0.234 −0.231 0.179 −0.448* (0.163) (0.139) (0.099) (0.210) (0.238) (0.239) (0.306) Num. of group members subject would loan to 0.174* 0.280*** 0.241*** 0.508*** 0.447*** −0.080 0.353* (0.124) (0.120) (0.079) (0.191) (0.227) (0.162) (0.227) Mean distance to others’ homes 0.045* 0.054* 0.042*** 0.087*** 0.028 −0.005 0.100** (0.030) (0.021) (0.017) (0.036) (0.042) (0.037) (0.053) Fraction of life lived in area 0.005 0.008 0.008* 0.014* 0.005 −0.006 0.035** (0.008) (0.008) (0.005) (0.009) (0.010) (0.013) (0.021) Fraction of Others in peer group 0.057 0.347 0.180 0.631 −1.041 2.192** −1.218 (0.732) (0.742) (0.486) (0.919) (1.255) (1.187) (1.481) GSS1: Trust question −0.405 −0.175 −0.228 −0.174 0.023 −0.070 −1.474** (0.393) (0.402) (0.263) (0.556) (0.610) (0.633) (0.779) GSS2: Fairness question 0.614* 0.192 0.261 0.637 0.879* −1.013* 0.441 (0.419) (0.395) (0.260) (0.539) (0.598) (0.663) (0.805) GSS3: Helpfulness question 0.418 0.174 0.266 0.116 1.504** −0.092 −0.770 (0.386) (0.416) (0.267) (0.525) (0.827) (0.616) (0.726) Sender trust from trust game 0.137 (0.736) Receiver trustworthiness from trust game 1.523** (0.833) Likelihood ratio 29.4118 34.4995 53.4543 25.0006 34.9393 17.9668 18.3098 p‐value 0.0057 0.0010 <0.0001 0.0148 0.0005 0.1167 0.1066 ***95% Significance, **90% Significance, *85% Significance. Standard errors are given in parentheses. Open in new tab 3.3. Personal Trust Our principal measure of personal trust is the question, ‘Would you lend (person x) 1000 drams (100 rand)?’ Answering yes to this question for increasing numbers of individuals in the group has a positive effect on both individual contributions and group longevity. The coefficient carries the (correct) positive sign in virtually all of the estimation, and is statistically significant at the 99% level in our most important estimation on the entire sample in Table 4. Abbink et al. (2006) find that self‐selected groups of friends, among whom a greater level of trust presumably exists, perform better than randomly formed groups, but only in initial rounds. Table 5 Group Repayment Decisions Dependent Variable: Number of Rounds Reach by Group in Microfinance Game, μ = 3.861 Variable: . —OLS Estimates— . Group Repayment: Berd, Armenia . Combined Estimation: Berd and Nyanga . Number of observations: . 25 . 24 . 24 . 35 . 35 . 34 . Intercept 10.023*** 9.132*** 9.887*** 4.390*** 7.024*** 7.638*** (2.041) (2.543) (2.720) (1.725) (1.989) (2.448) Mean per period shocks received by group −1.713*** −1.708*** −1.714*** −1.598*** −1.721*** −1.665*** (0.336) (0.387) (0.413) (0.372) (0.357) (0.400) Mean number of acquaintances in group 0.233 0.324 0.609* 0.057 0.134 0.013 (0.287) (0.331) (0.370) (0.308) (0.308) (0.346) Mean would loan to other indivs. in group 0.021 0.040 0.033 0.677*** 0.433 0.448 (0.318) (0.343) (0.389) (0.331) (0.334) (0.348) Mean distance between members’ homes −0.063 −0.053 −0.038 0.016 −0.018 −0.023 (0.048) (0.0540) (0.0515) (0.053) (0.053) (0.058) Mean fraction of life lived in area 0.075*** 0.090*** 0.067* (0.0332) (0.039) (0.041) Heterogeneity‐fraction life lived in area −0.116*** −0.122*** −0.097*** −0.043*** −0.051*** (0.030) (0.033) (0.035) (0.018) (0.022) % members work after Perestroika −1.656** −1.457* −1.630** (0.857) (9.324) (0.892) Heterogeneity in peer group/clan −0.473 0.423 0.775 −0.952 −1.193 (1.07) (1.437) (1.669) (1.090) (1.248) Sender trust from trust game −1.851 −2.515 0.924 (2.172) (2.211) (1.810) Receiver trustworthiness 2.021 2.600 −1.488 (2.167) (2.344) (1.648) GSS1: Trust question −2.890 (2.286) GSS2: Fairness question −0.971 (1.765) GSS3: Helpfulness question 2.036 (1.2016) Nyanga dummy 0.259 0.146 0.151 (1.136) (1.099) (1.175) R‐Squared 0.7126 0.7119 0.8015 0.4885 0.5713 0.5710 Adj R‐Squared 0.5774 0.5061 0.5670 0.4032 0.20161 0.4166 F‐Statistic 5.27 3.46 3.42 5.73 5.33 3.70 F‐Signif. 0.002 0.017 0.024 0.0008 0.0006 0.0046 Variable: . —OLS Estimates— . Group Repayment: Berd, Armenia . Combined Estimation: Berd and Nyanga . Number of observations: . 25 . 24 . 24 . 35 . 35 . 34 . Intercept 10.023*** 9.132*** 9.887*** 4.390*** 7.024*** 7.638*** (2.041) (2.543) (2.720) (1.725) (1.989) (2.448) Mean per period shocks received by group −1.713*** −1.708*** −1.714*** −1.598*** −1.721*** −1.665*** (0.336) (0.387) (0.413) (0.372) (0.357) (0.400) Mean number of acquaintances in group 0.233 0.324 0.609* 0.057 0.134 0.013 (0.287) (0.331) (0.370) (0.308) (0.308) (0.346) Mean would loan to other indivs. in group 0.021 0.040 0.033 0.677*** 0.433 0.448 (0.318) (0.343) (0.389) (0.331) (0.334) (0.348) Mean distance between members’ homes −0.063 −0.053 −0.038 0.016 −0.018 −0.023 (0.048) (0.0540) (0.0515) (0.053) (0.053) (0.058) Mean fraction of life lived in area 0.075*** 0.090*** 0.067* (0.0332) (0.039) (0.041) Heterogeneity‐fraction life lived in area −0.116*** −0.122*** −0.097*** −0.043*** −0.051*** (0.030) (0.033) (0.035) (0.018) (0.022) % members work after Perestroika −1.656** −1.457* −1.630** (0.857) (9.324) (0.892) Heterogeneity in peer group/clan −0.473 0.423 0.775 −0.952 −1.193 (1.07) (1.437) (1.669) (1.090) (1.248) Sender trust from trust game −1.851 −2.515 0.924 (2.172) (2.211) (1.810) Receiver trustworthiness 2.021 2.600 −1.488 (2.167) (2.344) (1.648) GSS1: Trust question −2.890 (2.286) GSS2: Fairness question −0.971 (1.765) GSS3: Helpfulness question 2.036 (1.2016) Nyanga dummy 0.259 0.146 0.151 (1.136) (1.099) (1.175) R‐Squared 0.7126 0.7119 0.8015 0.4885 0.5713 0.5710 Adj R‐Squared 0.5774 0.5061 0.5670 0.4032 0.20161 0.4166 F‐Statistic 5.27 3.46 3.42 5.73 5.33 3.70 F‐Signif. 0.002 0.017 0.024 0.0008 0.0006 0.0046 ***95% significance, **90% significance, *85% significance. Standard errors are given in parentheses. Open in new tab Table 5 Group Repayment Decisions Dependent Variable: Number of Rounds Reach by Group in Microfinance Game, μ = 3.861 Variable: . —OLS Estimates— . Group Repayment: Berd, Armenia . Combined Estimation: Berd and Nyanga . Number of observations: . 25 . 24 . 24 . 35 . 35 . 34 . Intercept 10.023*** 9.132*** 9.887*** 4.390*** 7.024*** 7.638*** (2.041) (2.543) (2.720) (1.725) (1.989) (2.448) Mean per period shocks received by group −1.713*** −1.708*** −1.714*** −1.598*** −1.721*** −1.665*** (0.336) (0.387) (0.413) (0.372) (0.357) (0.400) Mean number of acquaintances in group 0.233 0.324 0.609* 0.057 0.134 0.013 (0.287) (0.331) (0.370) (0.308) (0.308) (0.346) Mean would loan to other indivs. in group 0.021 0.040 0.033 0.677*** 0.433 0.448 (0.318) (0.343) (0.389) (0.331) (0.334) (0.348) Mean distance between members’ homes −0.063 −0.053 −0.038 0.016 −0.018 −0.023 (0.048) (0.0540) (0.0515) (0.053) (0.053) (0.058) Mean fraction of life lived in area 0.075*** 0.090*** 0.067* (0.0332) (0.039) (0.041) Heterogeneity‐fraction life lived in area −0.116*** −0.122*** −0.097*** −0.043*** −0.051*** (0.030) (0.033) (0.035) (0.018) (0.022) % members work after Perestroika −1.656** −1.457* −1.630** (0.857) (9.324) (0.892) Heterogeneity in peer group/clan −0.473 0.423 0.775 −0.952 −1.193 (1.07) (1.437) (1.669) (1.090) (1.248) Sender trust from trust game −1.851 −2.515 0.924 (2.172) (2.211) (1.810) Receiver trustworthiness 2.021 2.600 −1.488 (2.167) (2.344) (1.648) GSS1: Trust question −2.890 (2.286) GSS2: Fairness question −0.971 (1.765) GSS3: Helpfulness question 2.036 (1.2016) Nyanga dummy 0.259 0.146 0.151 (1.136) (1.099) (1.175) R‐Squared 0.7126 0.7119 0.8015 0.4885 0.5713 0.5710 Adj R‐Squared 0.5774 0.5061 0.5670 0.4032 0.20161 0.4166 F‐Statistic 5.27 3.46 3.42 5.73 5.33 3.70 F‐Signif. 0.002 0.017 0.024 0.0008 0.0006 0.0046 Variable: . —OLS Estimates— . Group Repayment: Berd, Armenia . Combined Estimation: Berd and Nyanga . Number of observations: . 25 . 24 . 24 . 35 . 35 . 34 . Intercept 10.023*** 9.132*** 9.887*** 4.390*** 7.024*** 7.638*** (2.041) (2.543) (2.720) (1.725) (1.989) (2.448) Mean per period shocks received by group −1.713*** −1.708*** −1.714*** −1.598*** −1.721*** −1.665*** (0.336) (0.387) (0.413) (0.372) (0.357) (0.400) Mean number of acquaintances in group 0.233 0.324 0.609* 0.057 0.134 0.013 (0.287) (0.331) (0.370) (0.308) (0.308) (0.346) Mean would loan to other indivs. in group 0.021 0.040 0.033 0.677*** 0.433 0.448 (0.318) (0.343) (0.389) (0.331) (0.334) (0.348) Mean distance between members’ homes −0.063 −0.053 −0.038 0.016 −0.018 −0.023 (0.048) (0.0540) (0.0515) (0.053) (0.053) (0.058) Mean fraction of life lived in area 0.075*** 0.090*** 0.067* (0.0332) (0.039) (0.041) Heterogeneity‐fraction life lived in area −0.116*** −0.122*** −0.097*** −0.043*** −0.051*** (0.030) (0.033) (0.035) (0.018) (0.022) % members work after Perestroika −1.656** −1.457* −1.630** (0.857) (9.324) (0.892) Heterogeneity in peer group/clan −0.473 0.423 0.775 −0.952 −1.193 (1.07) (1.437) (1.669) (1.090) (1.248) Sender trust from trust game −1.851 −2.515 0.924 (2.172) (2.211) (1.810) Receiver trustworthiness 2.021 2.600 −1.488 (2.167) (2.344) (1.648) GSS1: Trust question −2.890 (2.286) GSS2: Fairness question −0.971 (1.765) GSS3: Helpfulness question 2.036 (1.2016) Nyanga dummy 0.259 0.146 0.151 (1.136) (1.099) (1.175) R‐Squared 0.7126 0.7119 0.8015 0.4885 0.5713 0.5710 Adj R‐Squared 0.5774 0.5061 0.5670 0.4032 0.20161 0.4166 F‐Statistic 5.27 3.46 3.42 5.73 5.33 3.70 F‐Signif. 0.002 0.017 0.024 0.0008 0.0006 0.0046 ***95% significance, **90% significance, *85% significance. Standard errors are given in parentheses. Open in new tab We interpret our results on personal trust as some evidence for the importance of screening and self‐selection in borrowing groups; personal trust appears to play a far more important role than simple acquaintanceship. Mere acquaintanceship with other individuals in the group before the experiment (‘Do you know person x?’) is insignificant in virtually all of the estimation, and negatively significant in round 5 in Table 4. The implication is that group lending may not be successful when people simply know one another well; it is more likely to succeed where people can choose among a large number of trustworthy group members. Moreover, the data show distance between members’ homes to be (surprisingly) positively related to group performance. To the extent that someone needs to know another individual, or at least know of her and live somewhat close to her to impose some type of social sanction in response to suspected defections in the game, our results offer little support to Besley and Coate’s (1995) hypothesis that the potential for social sanctions is vital to group lending. Trust that others will contribute their share is far more significant in our study. 3.4. Generalised Trust While generalised trust in society is likely to be integral at a broader level, such as in the establishment of institutions and governance structures, positive answers to the General Social Survey questions proved to be insignificant as a determinant of behaviour in the microfinance game, and often carry an unexpected sign. This is consistent with the results of Bohnet and Frey (1999), who find that an accurate portrayal of cooperative behaviour is only revealed when ‘social distance’ diminishes and subjects interact with an identifiable person. Our finding contrasts somewhat in this respect with Karlan (2005) who finds that the GSS survey questions relative to societal trust were negatively associated with default among his sample of Peruvian microfinance borrowers. It is difficult to explain the insignificance of social capital reflected in the GSS questions, other than by noting that this fits a pattern in our empirical results. This pattern clearly points to the relative importance for group lending of personalised trust over generalised trust in society, and that answers to specific, contextual questions, such as ‘Would you lend (person x) 1000 drams (100 rand)?’ are a more powerful indicator of behaviour than generalised questions. Thus, if group members have an interest in being members of well‐functioning groups, then self‐selection should create endogenously formed groups with a high level of specific trust among members. Because self‐selection relies on specific trust and not generalised trust, our results would suggest that self‐selected groups should function better. This result appears consistent with what Ahlin and Townsend (2007) (this Feature) find on the importance of self‐selection and screening among their borrowers in central Thailand. 3.5. Effect of Trust Game Results We use our trust game to generate measures of both trust and trustworthiness that may be useful in understanding behaviour in our microfinance game. In short, consistent with Karlan (2005), in our experiment, we uncovered no evidence that trusting behaviour is at all positively related to greater rates of contribution to group loans. (He actually finds that it is negatively related, and interprets the result as possibly due to risk‐loving behaviour.) We find some evidence that trustworthiness is related to contributions, an effect that is fairly substantial in the Berd estimation: if a receiver returned all of the coins passed to him in the trust game, it increases his probability of contribution in the microfinance game by about 40 percentage points. Thus, a subject who was trustworthy as a receiver in the trust game tended to be a strong contributor in the microfinance game. Since players were anonymous in the trust game, the significance of the trustworthiness variable may reflect borrower quality or dependability, meaning that a community of dependable people may be likely to be a community of well‐performing borrowing groups. However, the coefficient on trustworthiness was insignificant in Nyanga and in the group estimation. 3.6. Social Homogeneity Many researchers and development practitioners have believed for some time that social cohesion has played a major role in credit group performance. Empirical evidence from actual field data has been mixed on the question with some such as Zeller (1998) finding positive effects of a variable counting the number of common characteristics among members. Karlan (2005) finds that ethnically homogeneous pairs are more trusting in the Peruvian experiments. Other results, such as Wydick (1999), find that the stronger the social ties between members, the less credible the threat of social sanctions becomes. Ahlin and Townsend (2007) (this Feature) also find evidence that existing social ties may hinder group loan repayment. The results from our field experiments lend measured empirical support to the idea that social homogeneity is a good thing for group loan repayment. In Nyanga, individual contributions are significantly associated at the 95% level with the number of members with the same clan name in the group as the observed individual, as seen in Table 3. We cannot control for clan type due to the large number of different clans in the Nyanga estimation, but since in Berd we have only two peer groups, we control for whether an individual was a member of the pre‐Perestroika generation and the fraction of pre‐Perestroika in the group, as we examine the effect of the fraction of members in the borrowing group who are in the same peer group as the individual. Though the sign of the coefficient points to homogeneity having a positive influence on contribution, it lacks statistical significance. Although two very different kinds of social heterogeneity characterise Nyanga and Berd, for the combined estimation in Table 5 we use a common diversity index of similar/dissimilar members (by clan name and pre and post Perestroika) and find the point estimate showing heterogeneity to have negative but statistically insignificant effects. Table 5 also shows heterogeneity in groups as measured by long‐term vs. short‐term residents. The coefficient on the standard deviation of number of years residing in the local area has the expected sign, and is significant at the 99% level in Berd and at the 95% level in the combined estimation. Taken in light of other research, our results support the idea that social and cultural group homogeneity is likely to exert a positive influence on loan repayment. 4. Conclusion Researchers face a puzzle in disentangling the diverse aspects of social capital and their influence on borrower behaviour in joint‐liability loan contracts. We view our experimental results as one piece to this puzzle. In contrast to other work, including the other articles in this Feature, we employ artefactual and framed field experiments that allow us to work at the problem from a particular angle through imposing a maximum degree of control and exogeneity on our estimation, while using subjects who closely represent the population of individuals that actually receives group loans in developing countries. We view this kind of experimental work relative to other techniques, such as estimation on field data, not as substitutes, but instead as complements in this puzzle‐solving process. Our goal for this research, and its contribution to this Feature, is that it be part of an effort that collectively triangulates on a better understanding of group lending. In this sense we cannot claim general inference from our findings, but instead that our results be viewed in light of the wide array of empirical methodologies focused on the same question. Taken in this context, it is interesting that our results support many of the traditional beliefs about group lending and work from other types of empirical studies. At the outset of this article we divided theories about group lending into three categories: those that emphasise the importance of relational social capital to group lending, those that emphasise the importance of informational social capital and those that emphasise the inherent contractual benefits of joint‐liability contracts. That we find socially heterogeneous groups consistently performing worse than socially homogeneous groups supports the notion that relational social capital matters to group lending. Social homogeneity appears to facilitate a confidence that other members will indeed repay, augmenting the belief that the group is likely to receive subsequent loans in the future and that those who do repay in early rounds will not get burned by non‐repayers. We also find evidence of reciprocity within groups as group members who have realised more shocks (and relied on others to pay for them in the past) are more likely to repay the next time they have the opportunity. Thus we find that social capital appears to grow with positive experiences from other members following through with repayment in the group. Additionally, we believe that our finding that personal trust between specific pairs of group members significantly affects performance in our microfinance games is significant. First, it implies that group lending is likely to be more successful when a borrower faces a pool of potential borrowing partners that contains a large number of people whom she personally trusts. Moreover, to the extent that borrowers have a choice within this pool, it supports the notion that informational social capital in the process of group self‐selection and screening is likely to matter in group lending. Although in our experiments borrowing groups are formed exogenously, if personal trust matters to group performance in practice, then borrowers will have an incentive to self‐select over this variable. In contrast, we find traditional measures of general, society‐wide social capital, such as reflected by the commonly used GSS questions, to be mostly insignificant. In many respects, this result is not surprising: repayment under dynamic incentives is individually rational when group members believe that the group as a whole will perform well enough to continue receiving future loans. Contributions (especially in early rounds) should depend on the confidence that borrowers have in the particular individuals within the borrowing group more than their confidence in society generally. That we find the strength of acquaintanceship between members and the distance between their homes to be insignificantly (and in some specifications negatively) related to group performance would seem to imply that the most important component of relational capital may be interpersonal trust between members rather than the underlying threat of social sanctions for non‐contribution. One caveat to our findings and those from similar research based on field data is that a high degree of social capital between group members is probably insufficient in and of itself to generate high repayment rates. For example, the relative importance of within‐group social capital in preventing defaults may well be weaker than the mere threat of group expulsion, the availability of alternative credit, or the intensity and quality of loan officer activity. As shown in Cull et al. (2007) (this Feature), other institutional factors matter greatly to borrower performance and the performance of microfinance institutions generally, such as investments in quality loan officers and other staff. There is probably no single factor that is alone responsible for the frequent success with group lending realised in such a wide variety of contexts in the developing world, but this research suggests that relational social capital between members appears to be one significant factor in this success. Appendix: Results of Trust Games Table A1 Sender’s Trust . —Berd— . Nyanga . (% Amount Sent to Receiver) . All games . Equal initial amounts . Unequal initial amounts . Unequal initial amounts . 0 – – – – (0–25%] 25.0 20.8 29.1 55.0 (25%–50%) 30.8 35.1 26.6 21.7 50% 29.5 28.6 30.4 – (50%–75%] – – – 20.0 (75%–100%) 7.1 10.4 3.8 3.3 100% 7.7 5.2 10.1 – Num. Obs. 156 77 79 60 Sender’s Trust . —Berd— . Nyanga . (% Amount Sent to Receiver) . All games . Equal initial amounts . Unequal initial amounts . Unequal initial amounts . 0 – – – – (0–25%] 25.0 20.8 29.1 55.0 (25%–50%) 30.8 35.1 26.6 21.7 50% 29.5 28.6 30.4 – (50%–75%] – – – 20.0 (75%–100%) 7.1 10.4 3.8 3.3 100% 7.7 5.2 10.1 – Num. Obs. 156 77 79 60 Open in new tab Table A1 Sender’s Trust . —Berd— . Nyanga . (% Amount Sent to Receiver) . All games . Equal initial amounts . Unequal initial amounts . Unequal initial amounts . 0 – – – – (0–25%] 25.0 20.8 29.1 55.0 (25%–50%) 30.8 35.1 26.6 21.7 50% 29.5 28.6 30.4 – (50%–75%] – – – 20.0 (75%–100%) 7.1 10.4 3.8 3.3 100% 7.7 5.2 10.1 – Num. Obs. 156 77 79 60 Sender’s Trust . —Berd— . Nyanga . (% Amount Sent to Receiver) . All games . Equal initial amounts . Unequal initial amounts . Unequal initial amounts . 0 – – – – (0–25%] 25.0 20.8 29.1 55.0 (25%–50%) 30.8 35.1 26.6 21.7 50% 29.5 28.6 30.4 – (50%–75%] – – – 20.0 (75%–100%) 7.1 10.4 3.8 3.3 100% 7.7 5.2 10.1 – Num. Obs. 156 77 79 60 Open in new tab Table A2 Receiver’s Trustworthiness . —Berd— . All games (% Amount sent back to sender) . (% Amount Sent to receiver) . 0 . (0–25%] . (25%–50%) . 50% . (50%–75%] . (75%–100%) . 100% . Num. Obs. . 0 – – – – – – – – (0–25%] – – 28.2 46.2 20.5 – 5.1 39 (25%–50%) – 29.2 39.6 10.4 8.3 8.3 4.2 48 50% – 26.1 39.1 – 17.4 13.0 4.4 46 (50%–75%] – – – – – – – – (75%–100%) – 63.6 18.2 – – 18.2 – 11 100% – 16.7 50.0 33.3 – – – 12 Num. Obs. 0 35 56 27 20 12 6 156 Receiver’s Trustworthiness . —Berd— . All games (% Amount sent back to sender) . (% Amount Sent to receiver) . 0 . (0–25%] . (25%–50%) . 50% . (50%–75%] . (75%–100%) . 100% . Num. Obs. . 0 – – – – – – – – (0–25%] – – 28.2 46.2 20.5 – 5.1 39 (25%–50%) – 29.2 39.6 10.4 8.3 8.3 4.2 48 50% – 26.1 39.1 – 17.4 13.0 4.4 46 (50%–75%] – – – – – – – – (75%–100%) – 63.6 18.2 – – 18.2 – 11 100% – 16.7 50.0 33.3 – – – 12 Num. Obs. 0 35 56 27 20 12 6 156 Open in new tab Table A2 Receiver’s Trustworthiness . —Berd— . All games (% Amount sent back to sender) . (% Amount Sent to receiver) . 0 . (0–25%] . (25%–50%) . 50% . (50%–75%] . (75%–100%) . 100% . Num. Obs. . 0 – – – – – – – – (0–25%] – – 28.2 46.2 20.5 – 5.1 39 (25%–50%) – 29.2 39.6 10.4 8.3 8.3 4.2 48 50% – 26.1 39.1 – 17.4 13.0 4.4 46 (50%–75%] – – – – – – – – (75%–100%) – 63.6 18.2 – – 18.2 – 11 100% – 16.7 50.0 33.3 – – – 12 Num. Obs. 0 35 56 27 20 12 6 156 Receiver’s Trustworthiness . —Berd— . All games (% Amount sent back to sender) . (% Amount Sent to receiver) . 0 . (0–25%] . (25%–50%) . 50% . (50%–75%] . (75%–100%) . 100% . Num. Obs. . 0 – – – – – – – – (0–25%] – – 28.2 46.2 20.5 – 5.1 39 (25%–50%) – 29.2 39.6 10.4 8.3 8.3 4.2 48 50% – 26.1 39.1 – 17.4 13.0 4.4 46 (50%–75%] – – – – – – – – (75%–100%) – 63.6 18.2 – – 18.2 – 11 100% – 16.7 50.0 33.3 – – – 12 Num. Obs. 0 35 56 27 20 12 6 156 Open in new tab Table A3 Receiver’s Trustworthiness . —Berd— . Equal initial amounts (% Amount sent back to sender) . (% Amount sent to receiver) . 0 . (0–25%] . (25%–50%) . 50% . (50%–75%] . (75%–100%) . 100% . No. Obs. . 0 – – – – – – – 0 (0–25%] – – 37.5 37.5 12.5 – 12.5 16 (25%–50%) – 37.0 37.0 11.1 – 7.4 7.4 27 50% – 9.1 54.6 – 18.2 18.2 – 22 (50%–75%] – – – – – – – 0 (75%–100%) – 75.0 25.0 – – – – 8 100% – 50.0 – 50.0 – – – 4 Num. Obs. 0 20 30 11 6 6 4 77 Receiver’s Trustworthiness . —Berd— . Equal initial amounts (% Amount sent back to sender) . (% Amount sent to receiver) . 0 . (0–25%] . (25%–50%) . 50% . (50%–75%] . (75%–100%) . 100% . No. Obs. . 0 – – – – – – – 0 (0–25%] – – 37.5 37.5 12.5 – 12.5 16 (25%–50%) – 37.0 37.0 11.1 – 7.4 7.4 27 50% – 9.1 54.6 – 18.2 18.2 – 22 (50%–75%] – – – – – – – 0 (75%–100%) – 75.0 25.0 – – – – 8 100% – 50.0 – 50.0 – – – 4 Num. Obs. 0 20 30 11 6 6 4 77 Open in new tab Table A3 Receiver’s Trustworthiness . —Berd— . Equal initial amounts (% Amount sent back to sender) . (% Amount sent to receiver) . 0 . (0–25%] . (25%–50%) . 50% . (50%–75%] . (75%–100%) . 100% . No. Obs. . 0 – – – – – – – 0 (0–25%] – – 37.5 37.5 12.5 – 12.5 16 (25%–50%) – 37.0 37.0 11.1 – 7.4 7.4 27 50% – 9.1 54.6 – 18.2 18.2 – 22 (50%–75%] – – – – – – – 0 (75%–100%) – 75.0 25.0 – – – – 8 100% – 50.0 – 50.0 – – – 4 Num. Obs. 0 20 30 11 6 6 4 77 Receiver’s Trustworthiness . —Berd— . Equal initial amounts (% Amount sent back to sender) . (% Amount sent to receiver) . 0 . (0–25%] . (25%–50%) . 50% . (50%–75%] . (75%–100%) . 100% . No. Obs. . 0 – – – – – – – 0 (0–25%] – – 37.5 37.5 12.5 – 12.5 16 (25%–50%) – 37.0 37.0 11.1 – 7.4 7.4 27 50% – 9.1 54.6 – 18.2 18.2 – 22 (50%–75%] – – – – – – – 0 (75%–100%) – 75.0 25.0 – – – – 8 100% – 50.0 – 50.0 – – – 4 Num. Obs. 0 20 30 11 6 6 4 77 Open in new tab Table A4 Receiver’s Trustworthiness . —Berd— . Unequal initial amounts (% Amount sent back to sender) . (% Amount sent to receiver) . 0 . (0–25%] . (25%–50%) . 50% . (50%–75%] . (75%–100%) . 100% . No. Obs. . 0 – – – – – – – 0 (0–25%] – – 21.7 52.2 26.1 – – 23 (25%–50%) – 19.1 42.9 9.5 19.1 9.5 – 21 50% – 41.7 25.0 – 16.7 8.3 8.3 24 (50%–75%] – – – – – – – 0 (75%–100%) – 33.3 – – – 66.7 – 3 100% – – 75.0 25.0 – – – 8 Num. Obs. 0 15 26 16 14 6 2 79 Receiver’s Trustworthiness . —Berd— . Unequal initial amounts (% Amount sent back to sender) . (% Amount sent to receiver) . 0 . (0–25%] . (25%–50%) . 50% . (50%–75%] . (75%–100%) . 100% . No. Obs. . 0 – – – – – – – 0 (0–25%] – – 21.7 52.2 26.1 – – 23 (25%–50%) – 19.1 42.9 9.5 19.1 9.5 – 21 50% – 41.7 25.0 – 16.7 8.3 8.3 24 (50%–75%] – – – – – – – 0 (75%–100%) – 33.3 – – – 66.7 – 3 100% – – 75.0 25.0 – – – 8 Num. Obs. 0 15 26 16 14 6 2 79 Open in new tab Table A4 Receiver’s Trustworthiness . —Berd— . Unequal initial amounts (% Amount sent back to sender) . (% Amount sent to receiver) . 0 . (0–25%] . (25%–50%) . 50% . (50%–75%] . (75%–100%) . 100% . No. Obs. . 0 – – – – – – – 0 (0–25%] – – 21.7 52.2 26.1 – – 23 (25%–50%) – 19.1 42.9 9.5 19.1 9.5 – 21 50% – 41.7 25.0 – 16.7 8.3 8.3 24 (50%–75%] – – – – – – – 0 (75%–100%) – 33.3 – – – 66.7 – 3 100% – – 75.0 25.0 – – – 8 Num. Obs. 0 15 26 16 14 6 2 79 Receiver’s Trustworthiness . —Berd— . Unequal initial amounts (% Amount sent back to sender) . (% Amount sent to receiver) . 0 . (0–25%] . (25%–50%) . 50% . (50%–75%] . (75%–100%) . 100% . No. Obs. . 0 – – – – – – – 0 (0–25%] – – 21.7 52.2 26.1 – – 23 (25%–50%) – 19.1 42.9 9.5 19.1 9.5 – 21 50% – 41.7 25.0 – 16.7 8.3 8.3 24 (50%–75%] – – – – – – – 0 (75%–100%) – 33.3 – – – 66.7 – 3 100% – – 75.0 25.0 – – – 8 Num. Obs. 0 15 26 16 14 6 2 79 Open in new tab Table A5 Receiver’s Trustworthiness . —Nyanga— . Unequal initial amounts All games (% Amount sent back to sender) . (% Amount sent to receiver) . 0 . (0—25%] . (25%–50%) . 50% . (50%–75%] . (75%–100%) . 100% . No. Obs. . 0 – – – – – – – – (0–25%] 12.9 19.35 61.29 – 6.45 – – 31 (25%–50%) – 38.46 15.38 30.77 – 15.38 – 13 50% – – – – – – – – (50%–75%] 33.33 33.33 16.67 – 16.67 – – 12 (75%–100%) – – – 100 – – – 2 100% – – – – – – – 0 Num. Obs. 8 15 23 6 4 2 0 58 Receiver’s Trustworthiness . —Nyanga— . Unequal initial amounts All games (% Amount sent back to sender) . (% Amount sent to receiver) . 0 . (0—25%] . (25%–50%) . 50% . (50%–75%] . (75%–100%) . 100% . No. Obs. . 0 – – – – – – – – (0–25%] 12.9 19.35 61.29 – 6.45 – – 31 (25%–50%) – 38.46 15.38 30.77 – 15.38 – 13 50% – – – – – – – – (50%–75%] 33.33 33.33 16.67 – 16.67 – – 12 (75%–100%) – – – 100 – – – 2 100% – – – – – – – 0 Num. Obs. 8 15 23 6 4 2 0 58 Open in new tab Table A5 Receiver’s Trustworthiness . —Nyanga— . Unequal initial amounts All games (% Amount sent back to sender) . (% Amount sent to receiver) . 0 . (0—25%] . (25%–50%) . 50% . (50%–75%] . (75%–100%) . 100% . No. Obs. . 0 – – – – – – – – (0–25%] 12.9 19.35 61.29 – 6.45 – – 31 (25%–50%) – 38.46 15.38 30.77 – 15.38 – 13 50% – – – – – – – – (50%–75%] 33.33 33.33 16.67 – 16.67 – – 12 (75%–100%) – – – 100 – – – 2 100% – – – – – – – 0 Num. Obs. 8 15 23 6 4 2 0 58 Receiver’s Trustworthiness . —Nyanga— . Unequal initial amounts All games (% Amount sent back to sender) . (% Amount sent to receiver) . 0 . (0—25%] . (25%–50%) . 50% . (50%–75%] . (75%–100%) . 100% . No. Obs. . 0 – – – – – – – – (0–25%] 12.9 19.35 61.29 – 6.45 – – 31 (25%–50%) – 38.46 15.38 30.77 – 15.38 – 13 50% – – – – – – – – (50%–75%] 33.33 33.33 16.67 – 16.67 – – 12 (75%–100%) – – – 100 – – – 2 100% – – – – – – – 0 Num. Obs. 8 15 23 6 4 2 0 58 Open in new tab Footnotes 1 " Student Health and Welfare Committee, a student‐run NGO sponsored by the University of Cape Town. 2 " The definition of ‘available for work’ considered whether the potential subject could participate in the Masizikhulise Project. 3 " Tests on self‐selection into the game in Nyanga revealed those who opted to participate in the microfinance game tended to be slightly poorer, were somewhat more religious and somewhat more politically inclined. 4 " Both the Berd and Nyanga surveys are available at http://www.usfca.edu/fac‐staff/acassar. 5 " The three GSS questions we used were the commonly administered trust question, ‘Generally speaking, would you say that most people can be trusted or that you can’t be to careful in dealing with people?’, the question on fairness, ‘Do you think most people would try to take advantage of you if they got the chance, or would they try to be fair?’, and the question on helpfulness, ‘Would you say that most of the time people try to be helpful, or that they are mostly just looking out for themselves?’. 6 " Instructions for all experiments are available at http://www.usfca.edu/fac‐staff/acassar. References Abbink , K. , Irlenbusch , B. and Renner , E. ( 2006 ). ‘Group size and social ties in microfinance institutions’ , Economic Inquiry, vol. 44 ( 4 ), pp. 614 – 28 . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Ahlin , C. and Townsend , R. ( 2007 ). ‘Using repayment data to test across models of joint liability lending’ , Economic Journal , vol. 117, pp. F11 – F51 . OpenURL Placeholder Text WorldCat Armendáriz de Aghion , B. ( 1999 ). ‘On the design of a credit agreement with peer monitoring’ , Journal of Development Economics , vol. 60 ( 1 ), pp. 79 – 104 . Google Scholar Crossref Search ADS WorldCat Armendáriz de Aghion , B. and Gollier , C. ( 2000 ). ‘Peer group formation in an adverse selection model’ , Economic Journal , vol. 110 , pp. 632 – 43 . Google Scholar Crossref Search ADS WorldCat Armendáriz de Aghion , B. and Morduch , J. ( 2005 ). The Economics of Microfinance , Cambridge MA: MIT Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Banerjee , A. , Besley , T. and Guinnane , T. ( 1994 ). ‘Thy neighbor’s keeper: design of a credit cooperative with theory and a test’ , Quarterly Journal of Economics , vol. 102 ( 9 ), pp. 491 – 515 . Google Scholar Crossref Search ADS WorldCat Berg , J. , Dickhaut , J. and McCabe , K. ( 1995 ). ‘Trust, reciprocity, and social history’ , Games and Economic Behavior , vol. 10 ( 1 ), pp. 122 – 42 . Google Scholar Crossref Search ADS WorldCat Besley , T. and Coate , S. ( 1995 ). ‘Group lending, repayment incentives and social collateral’ , Journal of Development Economics , vol. 46 ( 1 ), pp. 1 – 18 . Google Scholar Crossref Search ADS WorldCat Bohnet , I. and Frey , B. ( 1999 ). ‘Social distance and other‐regarding behavior in dictator games: comment’ , American Economic Review , vol. 89 ( 1 ), pp. 335 – 9 . Google Scholar Crossref Search ADS WorldCat Coleman , J. ( 1988 ). ‘Social capital in the creation of human capital’ , American Journal of Sociology , vol. 94 , pp. 95 – 120 . Google Scholar Crossref Search ADS WorldCat Cull , R. , Demirgüc‐Kunt , A. and Morduch , J. ( 2007 ). ‘Financial performance and outreach: a global analysis of leading micro banks’ , Economic Journal , vol. 117, pp. F107 – F133 . OpenURL Placeholder Text WorldCat Fafchamps , M. ( 1992 ). ‘Solitary networks in preindustrial societies: rational peasants with a moral economy’ , Economic Development and Cultural Change , vol. 40 , pp. 147 – 74 . Google Scholar Crossref Search ADS WorldCat Ghatak , M. ( 1999 ). ‘Group lending, local information and peer selection’ , Journal of Development Economics , vol. 60 ( 1 ), pp. 27 – 50 . Google Scholar Crossref Search ADS WorldCat Giné , X. , Jakiela , P., Karlan , D. and Morduch , J. ( 2005 ). ‘Microfinance games’ , Working Paper, Yale University and New York University. Glaeser , E. , Laibson , D., Scheinkman , J. and Soutter , C ( 2000 ). ‘Measuring trust’ , Quarterly Journal of Economics , vol. 115 , pp. 811 – 46 . Google Scholar Crossref Search ADS WorldCat Gómez , R. and Santor , E. ( 2003 ). ‘Do peer group members outperform individual borrowers? A test of peer group lending using Canadian microcredit data’ , Bank of Canada , Working Paper (October). OpenURL Placeholder Text WorldCat Harrison , G. and List , J. ( 2004 ). ‘Field experiments’ , Journal of Economic Literature , vol. 62 , pp. 1009 – 55 . Google Scholar Crossref Search ADS WorldCat Karlan , D. ( 2005 ). ‘Using experimental economics to measure social capital and predict financial decisions’ , American Economic Review , vol. 95 , pp. 1688 – 99 . Google Scholar Crossref Search ADS WorldCat Karlan , D. ( 2007 ). ‘Social connections and group banking’ , Economic Journal , vol. 117, pp. F52 – F84 . OpenURL Placeholder Text WorldCat Putnam , R. ( 1993 ). Making Democracy Work: Civic Traditions in Modern Italy , Princeton, NJ: Princeton University Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Scott , J. ( 1976 ). The Moral Economy of the Peasant , New Haven: Yale University Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Stiglitz , J. ( 1990 ). ‘Peer monitoring and credit markets’ , The World Bank Economic Review , vol. 4 ( 3 ), pp. 351 – 66 . Google Scholar Crossref Search ADS WorldCat Udry , C. ( 1994 ). ‘Risk and insurance in a rural credit market: an empirical investigation in northern Nigeria’ , Review of Economic Studies , vol. 61 ( 1 ), pp. 495 – 526 . Google Scholar Crossref Search ADS WorldCat Van Tassel , E. ( 1999 ). ‘Group lending under asymmetric information’ , Journal of Development Economics , vol. 60 ( 1 ), pp. 3 – 25 . Google Scholar Crossref Search ADS WorldCat Wenner , M. ( 1995 ). ‘Group credit: a means to improve information transfer and loan repayment performance’ , Journal of Development Studies , vol. 32 ( 1 ), pp. 263 – 81 . Google Scholar Crossref Search ADS WorldCat Wydick , B. ( 1999 ). ‘Can social cohesion be harnessed to mitigate market failures? Evidence from group lending in Guatemala’ , Economic Journal , vol. 109 , pp. 463 – 75 . Google Scholar Crossref Search ADS WorldCat Wydick , B. ( 2001 ). ‘Group lending under dynamic incentives as a borrower discipline device’ , Review of Development Economics , vol. 5 ( 3 ), pp. 406 – 20 . Google Scholar Crossref Search ADS WorldCat Zeller , M. ( 1998 ). ‘Determinants of repayment performance in credit groups: the role of program design, intragroup risk pooling, and social cohesion’ , Economic Development and Cultural Change , vol. 46 ( 3 ), pp. 599 – 620 . Google Scholar Crossref Search ADS WorldCat Author notes " The authors thank Chris Wendell for outstanding help with data collection, conference organisers Niels Hermes and Robert Lensink along with Chris Ahlin, Dan Friedman, Michael Jonas, Dean Karlan, Craig McIntosh, Stefan Klonner, Ashok Rai, Paul Ruud, Jana Vyrastekova and participants at the July 2005 Conference on Microfinance in Groningen, The Netherlands. Grant funding from the McCarthy Foundation is gratefully acknowledged. © The Author(s). Journal compilation © Royal Economic Society 2007

Spolaore,, Enrico

doi: 10.1111/j.1468-0297.2007.02031_3.xpmid: N/A

Review III Over the past few years, in a series of important articles,1 Daron Acemoglu and Jim Robinson have developed an influential theory about the creation and consolidation of democracy. In this book they summarise and extend their analytical results within a unified, coherent and elegant framework, and illustrate their insights with historical examples. Acemoglu and Robinson seek to reach beyond economists and formal political scientists who are already familiar with their articles. Their ambition is to convince a broader audience of political scientists, sociologists, historians, and other scholars that the ‘economic’ approach, based on stylised and simplified analytical models, brings valuable insights and a deeper understanding of complex political phenomena. The authors want to convince a potentially sceptical audience that heroic generalisations about democracy are good– a challenging task in an area of study in which ‘it seems that the ‘‘general propositions’’ are that there are no general propositions’ (p. 81). As a card‐carrying economist who shares their love for parsimonious and general models, I warmly endorse their methodological approach. In fact, I have been reading this book like a fan sitting on the sidelines and rooting for his favoured team, rejoicing when they score a spectacular goal, and suffering when they hit a post or send the ball out. Now I am ready to engage in what American football commentators call ‘Monday morning quarterbacking’ (‘il processo del lunedì’, for Italian readers). Their kick start is perfect: Occam’s razor. ‘Entities should not be multiplied beyond necessity’‘Be as simple as possible’. A more ‘nerdy’ version of Occam’s razor would be: add complications to your model only as long as the marginal benefit from doing that, in terms of explanatory power, is larger than the marginal cost. Acemoglu and Robinson know that the marginal cost of complexity is high, especially in a book addressed to a broad audience and with ambitions to provide a unified understanding of a vast and diverse set of events. Hence they strive to provide a simple, unified and flexible framework, which is maintained, with gradual extensions, through the whole book. The main questions, ideas and facts are introduced in Part I. Part II provides a user‐friendly introduction to basic modelling, while the central results about the creation and consolidation of democracy are presented in Part III. Part IV covers a rich set of extensions and applications. Part V contains the authors’ concluding remarks about areas for further research and the future of democracy. In the simplest version of their model, there are only two groups: the elites and the majority of the citizens. The citizens prefer democracy because it allows them to implement their favoured redistributive policies. For that same reason, the elites prefer to keep dictatorial power. In the absence of democracy, social unrest and the threat of revolution may induce the elites to make concessions. But concessions may not be credible in a dictatorship because current policies can be modified by the elites themselves in future periods. In contrast, democracy credibly transfers power to the citizens. When the costs of repression are high enough and promises of concessions are not credible, the elites may be forced to democratise as the only way to forestall costly disorder and revolution. Democratic institutions consolidate when the elites face low incentives to overthrow democracy in a coup and revert to dictatorship. This is a very crisp and insightful model, which clearly links preferences over economic policies (taxes) and political institutions (democracy) to fundamentals (material endowments). But, as a Monday‐morning quarterback, I dare say that the authors could have used Occam’s razor even more aggressively, in order to point out their different assumptions and insights to the reader. Their basic model collapses two logically distinct sets of assumptions. One set underpins the basic structure of the power game in which the elites choose current policies but can also choose state variables that affect the future distribution of power (institutional change). For example, at the beginning the authors could have presented a simpler, more general game with a group A that holds power dictatorially, and a (larger) group B that is excluded from power. Group A prefers policy a, and group B prefers policy b. By choosing more ‘democracy’, group A can increase group B’s probability of holding power in the future (for simplicity, the probability could go from 0 to 1, as in the book’s basic model). All results about credibility, repression etc. could perhaps have been illustrated within this simpler, more general framework. Once the general logical structure of the argument had been presented, the authors could have moved to their second set of assumptions – that is, political conflict as being about income redistribution across class lines. This approach might have allowed the authors to separate their general insight about democratisation as a commitment technology (democratisation ‘locks in’ temporary de facto power – a neat extension of an idea originally developed by North and Weingast (1989) in a different context) from the important but more specific application to redistributive conflict between the ‘rich’ and the ‘poor’. An additional question is whether the best way to model redistributive conflict within a society is by using a simplified Meltzer and Richard (1981) setting. Certainly, the approach has pedagogical advantages because of the popularity of this framework in political‐economy articles and textbooks. But the authors are well aware that the empirical record provides little support for the standard Meltzer‐Richard predictions. In particular, extensive empirical investigations have not found the implied positive relationship between inequality and redistribution in democracies, therefore casting a shadow over the mechanisms underpinning the book’s results about the relationship between inequality and democratisation.2 Acemoglu and Robinson do provide extensions that partially ‘insulate’ their model from some of those counterfactual implications. For example, in subsections about ‘alternative political identities’ (e.g., races or ethnic groups), they extend the basic framework to allow for conflict over income redistribution between two different groups which are not defined along class lines (e.g., each group includes some rich and some poor).3 Such extension is conceptually useful and realistic, but uncouples the relevant concept of inequality from standard inequality measures, making the basic theory difficult to test, at least with currently available data. As Acemoglu and Robinson recognise, a fundamental issue that complicates the empirical study of the relations among democracy, inequality and redistribution is that all three are endogenous in the long run, and affect each other in complex ways. Still, considering the well‐known creativity and industriousness of the authors, a loyal fan might have expected a more consistent effort to bring the model to the data, and clearly show the explanatory power of the economic approach to the sceptical audience mentioned above. However, the authors do present a broad review of stylised facts about democracy in chapter 3, and argue that these facts are loosely supportive of their main insights.4 Their fascinating discussion of case studies is limited to a handful of well‐chosen examples (mainly Britain, Argentina, Singapore and South Africa). An interesting question is the extent to which these examples are representative of a larger group of countries, and illustrate the main mechanisms in their framework, For instance, to what extent is Singapore representative of Taiwan or South Korea? And, in explaining Singapore’s stable non‐democracy, how paramount is the relatively low level of income inequality, rather than other characteristics (say, huge rates of economic growth) that may set Singapore apart from different countries with similar levels of inequality? When analysing non‐democratic regimes, the authors mainly focus on ‘right‐wing’ dictatorships, as they did in their original articles.5 In the simplest version of the framework, revolutions against the rich are ‘off the equilibrium path’, and even when they take place, they do not lead to ‘post‐revolutionary’ dictatorships. In this respect, this book is quite different from its illustrious inspiration, Social Origins of Dictatorship and Democracy by Barrington Moore (1966), which included a detailed analysis of communism. The lack of an explicit role for communist parties and regimes may miss a big chunk of the political mechanisms and conflicts at work during the twentieth century. For instance, a disproportionate number of dictatorships in the database provided by Beck et al. (2001) are classified as ‘left‐wing’. It is almost paradoxical that a book so closely inspired by a ‘Marxian’ focus on conflict over class lines should ignore socialist and communist regimes. The authors acknowledge this gap in the conclusions (p. 357) and suggest this as a promising area for future research. The relationship between economic structure and democracy is covered in Chapter 9, which contains an interesting discussion of how industrialisation and urbanisation may have made repression more costly and democracy less costly for the elites, hence spurring democratisation and redistribution.6 Perhaps the chapter could have included a more extensive discussion of alternative theories – for example, the hypothesis that industrialisation might have led to more education and less inequality as the optimal choice of capitalists who directly benefit from the accumulation of human capital by the masses (Galor and Moav, 2006). Future investigations may provide ‘horse races’ among competing theories to assess their relative empirical relevance. A key open question in the book is: Why are democratic institutions more persistent – and hence more credible – than mere ‘policy promises’? What is special about giving de jure power to the citizens through democratic institutions? In their formal models, the authors deal with this issue through an extreme application of Occam’s razor: by assumption. That is, they assume that the introduction of democratic institutions will change the distribution of power.7 This is an acceptable simplification at this level of abstraction, but it sidesteps the important issue of how institutional change (namely, democratisation) actually affects state variables that determine the distribution of power in the future. Through what mechanisms and under what conditions does de jure democratisation actually work? After all, history is full of ‘paper constitutions’ with little or no effects. For example, as the authors recall (pp. 140–2), after the Potemkin mutiny in 1905 and other major social unrest in Russia, Tsar Nicholas II granted freedom of speech and association, and established that no new laws could be introduced without the agreement of an elected body (the Duma). However, the Tsar repealed most of those reforms later on (with sad long‐term consequences for himself and his family). The authors argue that the ineffectiveness of these Tsarist reforms were due to the fact that Nicholas II did not go far enough in extending de jure power to his citizens. But how would more or less de jure power map into actual de facto power? How would the effectiveness of de jure democratisation depend on the willingness of people with de facto power (the military, the police, the bureaucracy) to obey elected officials and follow democratic procedures? Why do people with de facto power (say, the military) accept to obey civilian rule in some circumstances but not in others? Does a full‐fledged analysis of these concepts require a still‐to‐be‐developed economic theory of ‘legitimacy’? Acemoglu and Robinson informally discuss some of these issues in their book but an explicit formalisation is left for further research. Another topic that may deserve future attention is the study of the diffusion of democratic institutions across different countries and societies. Acemoglu and Robinson’s approach focuses on domestic conditions within individual countries, but it has often been noticed that democratisations tend to proceed in ‘waves’ (Huntington, 1991). In chapter 10 the authors consider the joint effects of globalisation forces as a possible explanation for the simultaneous adoption of democracy in several societies.8 Alternative channels could depend directly on the forces affecting the diffusion of democratic institutions themselves, as ‘institutional innovations’ that are imitated, with changes and adaptations, across different societies and cultures. In conclusion, I should point out, once again, that my remarks are ‘Monday morning quarterbacking’– and from a fan who, at the end of the day, is happy that his team has scored and won the game. Overall, this book is very successful at illustrating the intellectual benefits from approaching complex issues with simplified but insightful analytical models. And a major benefit from this approach is, indeed, that it stimulates so many questions! I strongly recommend this fascinating and ambitious book, and expect that it will pave the way for further advances in the growing field of political economy. Footnotes 1 " Most notably, Acemoglu and Robinson (2000, 2001). 2 " Specifically, in their setting the relationship can be U‐shaped, with democratisation occurring at intermediate levels of inequality, while low‐inequality societies and high‐inequality societies stay non‐democratic (pp. 189–93). 3 " An interesting issue for the theory of democratisation, which is not explored in the book, is the possibility that the main conflict between these political groups might not be over income distribution but over other policies (language, religion etc.). 4 " A systematic attempt to study the empirical relations between democracy and redistribution is in Boix (2003), who focuses on wealth distribution and factor mobility as key determinants of democratisation. 5 " For instance, in a footnote to their article on transitions, they wrote ‘Obviously, in practice, there are dictatorships that are against the interests of the richer segments of society, such as socialist dictatorships or some African regimes, but they fall outside the scope of our model’ (Acemoglu and Robinson, 2001) 6 " In particular, the redistributive costs of democracy may be larger for landowners than for the owners of mobile capital, as also stressed in Boix (2003). 7 " In a recent extension of the framework, Acemoglu and Robinson (2006) consider the possibility that increases in the citizens’de jure power be offset by additional investment in de facto power by the elites, therefore leaving the relative balance of power unchanged, or even more tilted towards the elites. However, this mechanism still assumes that, ceteris paribus, democratisation increases the power of the citizens. 8 " A different set of international mechanisms linking the process of democratisation across countries may emerge under the alternative hypothesis, briefly mentioned in the book, that democracy is introduced to increase governments’ ability to carry external wars (Ticchi and Vindigni, 2003). References Acemoglu , Daron and Robinson , James A. ( 2000 ). ‘Why did the West extend the franchise? Growth, inequality, and democracy in historical perspective’ , Quarterly Journal of Economics , vol. 105 , pp. 1167 – 99 . Google Scholar Crossref Search ADS WorldCat Acemoglu , Daron and Robinson , James A. ( 2001 ). ‘A theory of political transitions’ , American Economic Review , vol. 91 , pp. 938 – 63 . Google Scholar Crossref Search ADS WorldCat Acemoglu , Daron and Robinson , James A. ( 2006 ). ‘Persistence of power, elites and institutions’ , unpublished, Department of Economics, MIT and Department of Government , Harvard University. Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Beck , Thorsten , Clarke , George, Groff , Alberto, Keefer , Philip and Walsh , Patrick ( 2001 ). ‘New tools in comparative political economy: the database of political institutions’ , World Bank Economic Review , vol. 15 ( 1 ), pp. 165 – 76 . Google Scholar Crossref Search ADS WorldCat Boix , Carles ( 2003 ). Democracy and Redistribution , Cambridge and New York: Cambridge University Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Galor , Oded and Moav , Omer ( 2006 ). ‘Das Human Kapital: a theory of the demise of the class structure’ , Review of Economic Studies , vol. 73 , pp. 85 – 117 . Google Scholar Crossref Search ADS WorldCat Huntington , Samuel P. ( 1991 ). The Third Wave. Democratization in the Late Twentieth Century , Norman and London: University of Oklahoma Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Meltzer , Allan H. and Richard , Scott F. ( 1981 ). ‘A rational theory of the size of government’ , Journal of Political Economy , vol. 89 , pp. 914 – 27 . Google Scholar Crossref Search ADS WorldCat Moore , Barrington ( 1966 ). The Social Origins of Dictatorship and Democracy: Lord and Peasant in the Making of the Modern World , Boston: Beacon Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC North , Douglass C. and Weingast , Barry R. ( 1989 ). ‘Constitutions and commitment: the evolution of institutions governing public choice in seventeenth‐century England’ , Journal of Economic History , vol. 49 , pp. 803 – 32 . Google Scholar Crossref Search ADS WorldCat Ticchi , Davide and Vindigni , Andrea ( 2003 ). ‘On wars and political development. the role of international conflict in the democratization of the West’ , unpublished, Department of Economics, University of Urbino and Department of Politics , Princeton University. Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC © The Author(s). Journal compilation © Royal Economic Society 2007