The Economic Journal

Armstrong,, Mark;Wright,, Julian

doi: 10.1111/j.1468-0297.2009.02276.xpmid: N/A

Abstract We analyse charges levied by mobile telephone networks to deliver calls. We integrate two literatures: one analysing calls from the fixed network, where predicted unregulated termination charges are too high, and one analysing calls from rival mobile networks, where predicted charges are too low. In practice, however, networks adopt uniform charges for terminating both kinds of traffic, as do regulators. We show how incorporating wholesale arbitrage and demand‐side substitution helps to reconcile theory with practice. In our framework, the unregulated charge is uniform and typically lies between the efficient and monopoly benchmarks. There remains a rationale for regulation, albeit reduced. There is an important set of markets in which monopoly prices emerge even when competition is intense. That is to say, while industry profit is not excessive overall, there is an inefficient balance of prices: too high for some services, too low for others. Familiar examples involve consumer ‘lock‐in’ of various kinds, including markets with switching costs.1 In these markets the typical pattern of prices involves ‘bargains‐then‐ripoffs’, so that firms attract new consumers with generous deals up‐front and consumers pay high prices once locked in. If competition is vigorous, the profits from locked‐in consumers are transferred to new consumers, and the lifetime profitability of a consumer is approximately zero. In a sense, a consumer’s ‘future self’ is exploited by her ‘present self’. A related, but distinct, set of markets exhibit what might be termed ‘competitive bottlenecks’. Here, firms compete to attract one group of consumers. For natural technological or geographical reasons, each consumer in this group wishes to deal with just one firm. A second group of consumers wishes to interact with the first group and, because each consumer in the first group deals exclusively with one firm, that firm can charge the second group high prices (or pay low input prices) for access to its captive customers. If competition is vigorous, the monopoly profits generated by the first group are passed back to these consumers in the form of subsidised service. Here, the first group may be said to exploit the second. Examples of this kind of market include: newspaper advertising (where most readers tend to read a single newspaper due to time constraints and so a newspaper can charge high fees to advertisers for access to its captive readers) and supermarkets (where a consumer tends to visit just one shop over the relevant time horizon and so suppliers are prepared to be paid low input prices for access to these captive shoppers).2 However, perhaps the leading example of a competitive bottleneck is call termination on mobile telephone networks. Call termination refers to the service whereby a network completes – or ‘terminates’– a call made to one of its subscribers by a caller on another network. We focus on mobile, as opposed to fixed, networks since the former are currently much more competitive than the latter in most countries, and the issue of regulation there is more controversial. Mobile networks compete for subscribers but in the absence of regulation they may charge other networks excessively to talk to their subscribers. Concerns about mobile call termination being a bottleneck, and the associated high charges for calling mobile subscribers, have led to regulation of termination charges around the world. For example, caps on mobile call termination charges are applied throughout the European Union. In other countries, notably the United States, mobile termination charges are indirectly regulated (at low levels) through reciprocity requirements with the fixed‐line networks. There are two broad types of call termination on mobile telephone networks: termination of calls made by callers on the fixed‐line telephone network (fixed‐to‐mobile, or FTM, termination) and termination of calls made from other mobile networks (termed mobile‐to‐mobile, or MTM, termination).3 In the literature to date, FTM and MTM termination have been largely treated separately, with contrasting market failures highlighted. One aim of this article is to integrate these two strands. Broadly speaking, FTM call termination viewed in isolation is likely to involve unilateral monopoly pricing if left unchecked. The vast majority of mobile subscribers join just one mobile network and so callers on the fixed telephone network must route calls through a subscriber’s chosen network. No matter how competitive the market for mobile subscribers may be, a mobile network holds a monopoly over – and can set high charges for – delivering calls to its subscribers. That is to say, as was shown by Armstrong (2002, section 3.1) and Wright (2002), FTM call termination gives rise to a competitive bottleneck and regulators need to worry about excessive termination charges. The study of MTM termination (often termed ‘two‐way interconnection’ in the literature) has led to quite a different focus: whether mobile networks can use a negotiated termination charge to relax retail competition for subscribers.4 They can do this by setting MTM termination charges below cost. This causes off‐net calls (i.e., calls from one mobile network to another mobile network) to be cheaper than on‐net calls (calls from one mobile network to the same network), so that consumers prefer to join the smaller network. This in turn reduces each network’s incentive to attract subscribers. Thus, when networks can coordinate on the choice of a MTM termination charge, they may have an incentive to choose too low a charge. This low termination charge acts – somewhat counter‐intuitively – to harm mobile subscribers, who prefer the intense competition which accompanies a higher charge. This analysis gives rise to a puzzle. Even though the theory predicts that networks will want to set MTM termination charges that are too low, to the best of our knowledge no regulator has taken such a concern seriously. Rather, as with FTM termination, policy has acted to prevent firms choosing high MTM termination charges. Section 2 presents a benchmark model in which both FTM and MTM calls are present but where the two termination charges are determined separately. This integrates the existing literature and captures the opposing incentives networks face in their determination of the two charges. In section 3 we show this puzzle about MTM termination can be reconciled by taking into account the practical constraint of wholesale arbitrage. By this we mean that a mobile network cannot maintain a high FTM termination charge together with a low MTM termination charge, since the fixed network could then ‘transit’ its calls via another mobile network and so end up paying the lower MTM rate (plus a small transit charge). As a result, a network is forced to set an (approximately) uniform termination charge for FTM and MTM traffic. We establish that this requirement can, by itself, often make sense of the regulatory concern that unregulated charges are too high rather than too low. If the networks collectively set a uniform termination charge, their preferred charge will reflect the relative importance of FTM and MTM termination. If the former is more important, firms will want to set a uniform termination charge which is too high rather than too low (Section 3.1). If instead networks choose their uniform termination charges unilaterally (as perhaps is more plausible), then the temptation to extract termination profits always dominates the incentive to set a low termination charge in order to relax network competition. The result is that unregulated termination charges will be set above the efficient level.5 Nevertheless, a network’s incentive to set monopoly termination charges is mitigated. Setting the uniform termination charge too high strengthens network effects, making the firms tougher rivals. To avoid this effect, networks will keep their termination charges below the monopoly level (unless market expansion possibilities are very significant). Additional effects arise when we take into account that fixed‐line callers are often also mobile subscribers, so that the two groups are not disjoint as assumed in the standard competitive bottleneck story. Such callers can make MTM calls instead of FTM calls if the former are cheaper. In section 3 we find this demand‐side substitution between FTM and MTM further weakens the competitive bottleneck of call termination and brings the equilibrium charge closer to the efficient level. As a result, the welfare gains from regulation are smaller than the previous literature may have indicated, although the predicted termination charge without regulation remains excessive. Another important impact of demand‐side substitution is to reduce the need to regulate the retail markup on FTM calls, since market power here is constrained by the competing MTM service. Finally, Section 4 briefly discusses the robustness of our results to vertical integration between fixed and mobile networks, with Section 5 providing some concluding remarks. Before turning to the analysis, though, we review some relevant aspects of the UK mobile sector and its regulation. This will be used to motivate our theoretical framework and to discuss the policy implications of our results throughout the article. 1. Mobile Call Termination in the UK Since their inception, regulatory price caps on mobile termination in the UK have faced considerable resistance from mobile networks, as well as support from fixed‐line networks. They have also resulted in a large amount of paperwork. Since 1997, there have been several thousand pages of publicly released government reports and submissions on mobile call termination in the UK. Using materials from some of these reports, in this Section we give a brief overview of the most relevant aspects of the UK mobile market. The regulation of mobile termination dates back to at least 1998 when Oftel (then the UK telecommunications regulator) tried formally to control the FTM termination charges set by the two largest mobile networks Cellnet (the precursor to the current O2) and Vodafone. The proposed regulation was challenged by these mobile networks, leading to a competition enquiry – see MMC (1999). This enquiry did not investigate MTM termination charges, nor did it investigate FTM termination charges levied by the two networks, Orange and T‐Mobile, which had entered only recently at that point. The enquiry concluded that the two established networks’ FTM termination charges were too high in relation to cost and, based on its recommendations, Oftel regulated their FTM termination charges with a price cap. The imminent expiry of this cap led to a 2002 Competition Commission enquiry – see Competition Commission (2003). The Commission upheld Oftel’s proposed regulation, which covered all four mobile networks and both FTM and MTM (voice) termination charges. (SMS termination was not covered in the enquiry, nor by subsequent regulation to date in the UK.) Shortly before the 2002 enquiry, a fifth network, H3G, entered the market, although this immature network was excluded from that investigation. Subsequent reviews by Ofcom (the current UK telecommunications regulator) in 2004 and 2007 extended these regulations. As of 2007, all five networks are subject to price caps for call termination, with reductions in these caps applying from 2007 through to 2011, at which point they will be reviewed again (see Ofcom (2007b)). Table 1 gives an indication of the history of average mobile termination charges in recent years (for all UK networks). Table 1 
Average Mobile Termination Charges (in pence per minute)* . 2001 . 2002 . 2003 . 2004 . 2005 . all UK networks: 11.1 10.7 9.9 7.9 5.9 . 2001 . 2002 . 2003 . 2004 . 2005 . all UK networks: 11.1 10.7 9.9 7.9 5.9 *See Ofcom (2006a, Fig. 3.38). Open in new tab Table 1 
Average Mobile Termination Charges (in pence per minute)* . 2001 . 2002 . 2003 . 2004 . 2005 . all UK networks: 11.1 10.7 9.9 7.9 5.9 . 2001 . 2002 . 2003 . 2004 . 2005 . all UK networks: 11.1 10.7 9.9 7.9 5.9 *See Ofcom (2006a, Fig. 3.38). Open in new tab Thus, termination charges approximately halved over five years, in large part due to tightened regulation. However, the networks do not all set the same average level of charges, as shown in Table 2 which reports the termination charges set by the five networks in 2006. In 2006, the newest entrant, H3G, faced softer regulation than the established networks and took advantage of this to set termination charges which were substantially higher than its rivals. Table 2 
The Impact of Asymmetric Regulation on Termination Charges (March 2006)* . Daytime . Evening . Weekends . O2 6.4 6.3 3.1 Orange 7.6 5.4 4.3 T‐Mobile 8.1 4.0 4.0 Vodafone 8.5 3.4 2.8 H3G 15.6 10.8 2.5 . Daytime . Evening . Weekends . O2 6.4 6.3 3.1 Orange 7.6 5.4 4.3 T‐Mobile 8.1 4.0 4.0 Vodafone 8.5 3.4 2.8 H3G 15.6 10.8 2.5 *Taken from Ofcom (2006b, Fig. 1). Open in new tab Table 2 
The Impact of Asymmetric Regulation on Termination Charges (March 2006)* . Daytime . Evening . Weekends . O2 6.4 6.3 3.1 Orange 7.6 5.4 4.3 T‐Mobile 8.1 4.0 4.0 Vodafone 8.5 3.4 2.8 H3G 15.6 10.8 2.5 . Daytime . Evening . Weekends . O2 6.4 6.3 3.1 Orange 7.6 5.4 4.3 T‐Mobile 8.1 4.0 4.0 Vodafone 8.5 3.4 2.8 H3G 15.6 10.8 2.5 *Taken from Ofcom (2006b, Fig. 1). Open in new tab Fig. 1. Open in new tabDownload slide Call Termination on Mobile Networks Fig. 1. Open in new tabDownload slide Call Termination on Mobile Networks The mobile industry in the UK currently consists of four mature and roughly equal sized mobile networks, each with around 15 million subscribers, and a recent entrant (H3G) with fewer subscribers (as shown in Table 3). Table 3 
Subscriber Numbers in 2001 and 2005* . Vodafone . O2 . Orange . T‐Mobile . H3G . Active subscriber numbers in 2001 (m) 11.0 11.1 12.4 10.3 n/a Active subscriber numbers in 2005 (m) 14.8 17.0 14.9 15.3 3.5 . Vodafone . O2 . Orange . T‐Mobile . H3G . Active subscriber numbers in 2001 (m) 11.0 11.1 12.4 10.3 n/a Active subscriber numbers in 2005 (m) 14.8 17.0 14.9 15.3 3.5 *See Ofcom (2006a, Fig. 3.40). The current total number of subscribers exceeds the official UK population. This can be explained by the fact that some people have two or more phone subscriptions; for example, one for business and one for personal use; or because they continue to hold an old subscription which they no longer actively use. This is consistent with evidence from Ofcom (2006a, Fig. 3.46) that 10% of UK households do not have access to a mobile phone, similar to the proportion without a fixed‐line phone. Open in new tab Table 3 
Subscriber Numbers in 2001 and 2005* . Vodafone . O2 . Orange . T‐Mobile . H3G . Active subscriber numbers in 2001 (m) 11.0 11.1 12.4 10.3 n/a Active subscriber numbers in 2005 (m) 14.8 17.0 14.9 15.3 3.5 . Vodafone . O2 . Orange . T‐Mobile . H3G . Active subscriber numbers in 2001 (m) 11.0 11.1 12.4 10.3 n/a Active subscriber numbers in 2005 (m) 14.8 17.0 14.9 15.3 3.5 *See Ofcom (2006a, Fig. 3.40). The current total number of subscribers exceeds the official UK population. This can be explained by the fact that some people have two or more phone subscriptions; for example, one for business and one for personal use; or because they continue to hold an old subscription which they no longer actively use. This is consistent with evidence from Ofcom (2006a, Fig. 3.46) that 10% of UK households do not have access to a mobile phone, similar to the proportion without a fixed‐line phone. Open in new tab This compares to a total of around only 8 million subscribers at the time of initial regulation in 1998 – see Competition Commission (2003, Fig. 6.1). Total annual retail revenue for mobile networks in 2007 was about £13 billion and mobile call termination generated annual revenue of around £2.5 billion (Ofcom, 2007b, p. 7). As mobile penetration has grown, the importance of MTM calls in generating termination revenue has risen. Whereas at the time of the Competition Commission’s enquiry in 2002, nearly three‐quarters of mobile termination revenue was from FTM traffic, now it is only about one‐third of the total.6 We show in Section 3 that the volume of FTM traffic relative to MTM traffic plays an important role in determining the equilibrium termination charge. Importantly, and in contrast to the situation in many countries, all mobile networks in the UK are separately owned from the significant fixed networks. (Until 2001, however, Cellnet was owned by the principal fixed network, British Telecom.) This means that concerns arising in other jurisdictions about vertical price squeezes and foreclosure are not likely to be an issue in the UK. In Section 4 we discuss briefly how our analysis applies when one of the mobile networks is integrated with the fixed‐line network. The regulated termination charges in Table 1 were calculated using two kinds of markup over estimates of marginal termination costs. The first markup is designed to tax fixed‐line callers to subsidise mobile network use in order to stimulate mobile network expansion. As is explained further in Section 2.3, this is consistent with the positive network externalities generated by additional subscribers. The second markup reflects an intended contribution to a mobile network’s fixed and common costs. In Section 2.4, we argue that including fixed and common costs in this context may be a flawed policy unless there is significant scope for expansion in the mobile market. The price caps for FTM and MTM termination have been set equal to each other, as were the actual FTM and MTM termination charges set by networks. This contrasts with our initial analysis in Section 2, where in the absence of regulation we argue that networks may wish to set lower charges for terminating MTM calls (although often the welfare‐maximising charges are similar for the two kinds of calls.) It is worth noting there has been no regulatory constraint that prevents networks setting MTM termination charges which are lower than FTM termination charges. In Table 4 we give an idea of average per‐minute retail prices for on‐net and off‐net MTM calls, as well as FTM calls and text messages. The decline in off‐net MTM and FTM call charges evident here is no doubt partly due to the fall in termination charges documented in Table 1. However, the decline in off‐net MTM call charges has been particularly dramatic and this is likely to reflect that a growing number of calling plans include some free off‐net MTM calls. Despite the narrowing of the differentials between off‐net and on‐net calls, though, the difference remains striking. Due in part to this price differential, Table 5 shows the volumes of off‐net and on‐net calls to be unbalanced. With equal off‐net and on‐net charges and four roughly symmetric networks, one might expect that off‐net traffic would be approximately three times greater than on‐net traffic, rather than having a lower volume than on‐net traffic. Table 5 
Shares of Minutes of Types of Mobile Calls* . Off‐net MTM . On‐net MTM . mobile to fixed . % in 2001 14.9 31.0 54.1 % in 2005 25.8 34.8 39.4 . Off‐net MTM . On‐net MTM . mobile to fixed . % in 2001 14.9 31.0 54.1 % in 2005 25.8 34.8 39.4 *Data from Ofcom (2006a, Fig. 3.50). There is a typo in the original document in which the proportion of on‐net and off‐net MTM calls was reversed. Open in new tab Table 5 
Shares of Minutes of Types of Mobile Calls* . Off‐net MTM . On‐net MTM . mobile to fixed . % in 2001 14.9 31.0 54.1 % in 2005 25.8 34.8 39.4 . Off‐net MTM . On‐net MTM . mobile to fixed . % in 2001 14.9 31.0 54.1 % in 2005 25.8 34.8 39.4 *Data from Ofcom (2006a, Fig. 3.50). There is a typo in the original document in which the proportion of on‐net and off‐net MTM calls was reversed. Open in new tab Table 4 
Average Call Charges, Pence Per Minute/Message* . Off‐net MTM calls . On‐net MTM calls . FTM calls . Text messages . 2001 26.2 5.9 14.4 8.1 2005 11.3 4.2 11.5 6.3 . Off‐net MTM calls . On‐net MTM calls . FTM calls . Text messages . 2001 26.2 5.9 14.4 8.1 2005 11.3 4.2 11.5 6.3 *The price of FTM calls is taken from Ofcom (2006a, Fig. 3.22). It is a complicated and somewhat arbitrary task to give precise estimates for the prices of the various types of calls and messages originating on mobile networks. This is because mobile networks each offer a wide variety of tariffs, with different monthly rentals (where applicable) corresponding to different volumes of inclusive call minutes and text messages. Other than those for FTM calls, the numbers in Table 4 are taken from Ofcom (2006a, Fig. 3.39), although the method of calculation is not clear from that document. Open in new tab Table 4 
Average Call Charges, Pence Per Minute/Message* . Off‐net MTM calls . On‐net MTM calls . FTM calls . Text messages . 2001 26.2 5.9 14.4 8.1 2005 11.3 4.2 11.5 6.3 . Off‐net MTM calls . On‐net MTM calls . FTM calls . Text messages . 2001 26.2 5.9 14.4 8.1 2005 11.3 4.2 11.5 6.3 *The price of FTM calls is taken from Ofcom (2006a, Fig. 3.22). It is a complicated and somewhat arbitrary task to give precise estimates for the prices of the various types of calls and messages originating on mobile networks. This is because mobile networks each offer a wide variety of tariffs, with different monthly rentals (where applicable) corresponding to different volumes of inclusive call minutes and text messages. Other than those for FTM calls, the numbers in Table 4 are taken from Ofcom (2006a, Fig. 3.39), although the method of calculation is not clear from that document. Open in new tab In addition to the price differential documented in Table 4, there are at least two other reasons why call volumes are biased towards on‐net calls. First, ‘closed user groups’, i.e., groups of subscribers who predominantly make calls within their own group, may be present. Often, such groups have their network subscription decision made centrally (e.g., by their employer’s procurement office) and to a single network. To the extent these groups are widespread, this will boost the share of on‐net calls in the market. Second, there may be some substitution between MTM and FTM calls. A mobile subscriber, when she is in the home or office, has a choice between calling another mobile subscriber by means of either her fixed line or mobile phone. In many cases, she will just use the cheaper alternative. With the charges in Table 4, this implies she will often want to make an on‐net MTM call if the recipient is on the same mobile network, although probably not for off‐net MTM calls. This will amplify the bias towards on‐net call volumes, and we discuss this issue in Section 3.3. Since 2003, Ofcom has determined that the mobile retail market is effectively competitive.7 In contrast, UK regulators have consistently ruled that each mobile network has a monopoly of call termination on its own network, for the reason that there is no practical way for people to call someone on the go without calling the person’s mobile phone and having the call terminated by the mobile network to which that person has subscribed. This is not to suggest that there are absolutely no substitution possibilities. One objective of this article is to explore the extent to which this conclusion remains robust to substitution possibilities, both at the wholesale level on the supply side and at the consumer level on the demand side.8 2. Separate FTM and MTM Termination Charges The principal purpose of the benchmark model in this Section is to contrast the pricing of MTM termination with the pricing of FTM termination when each charge is chosen separately. That is, in this model we assume that firms can freely set distinct termination charges for FTM and MTM traffic. Provided FTM retail call charges are regulated to be equal to cost, we see that welfare is maximised when both the FTM and MTM termination charges are equal to cost. Without regulation, we see that mobile networks in this model will wish to set an excessive FTM termination charge (in fact, the monopoly charge), while they wish to set too low a MTM termination charge. Readers less interested in the details of the modelling can skip to Section 2.4 for a discussion of the findings. To model MTM calls, a standard framework of two‐way interconnection between symmetric networks is adopted, based on Laffont et al. (1998b) and Gans and King (2001) in which two symmetric networks labelled i = 1,2 offer mobile telephone services. The networks are assumed to negotiate an industry‐wide MTM termination charge, denoted a. Mobile subscribers are assumed to be identical in terms of their demand for calls to other mobile subscribers. With this simplification, if subscriber j faces a per‐minute charge p for calling subscriber k, j will choose to make (an average of) q(p) minutes of calls to k. Thus, each subscriber is equally likely to wish to call any other subscriber. Let v(p) be the consumer surplus associated with the demand function q(p), so that v′(p) ≡ −q(p). Added to this framework is a model of FTM termination described in Armstrong (2002, section 3.1) and Wright (2002). There is a fixed‐line network, from which a demand for FTM calls is generated. As in the UK, we assume this fixed sector is separately owned from the mobile sector. Each mobile network unilaterally chooses a termination charge for completing FTM calls and network i ’s FTM termination charge is denoted Ai. (Where possible, we use upper case notation for calls from the fixed network and lower case notation for calls from mobile networks.) We assume that in the first stage firms negotiate an industry‐wide MTM termination charge, a, and subsequently set their FTM termination charges, Ai, together with their retail tariffs to mobile customers. If the retail price for FTM calls to mobile network i is Pi per minute, suppose that there are Q(Pi) FTM minutes of calls to each subscriber on network i.9 We assume the fixed network can set different call charges to different mobile networks to reflect the networks’ different FTM termination charges.10 In general, the price Pi will be an increasing function of the FTM termination charges, Ai, which we write Pi = P(Ai). For instance, it may be that (1) where C is the fixed network’s marginal cost of originating a call. In this case, the FTM call charge is equal to the fixed network’s total cost of making such calls. Such pricing could arise as a result of the regulation of the fixed network or competition between fixed networks.11 (In Section 3.3 we argue that substitution between fixed and mobile calls might induce the FTM charges in (1).) Let V(·) be the consumer surplus function associated with the demand function Q(·), so that V′(P) ≡ −Q(P). Define (2) to be a mobile network’s profit, per subscriber, from providing termination services for the fixed network when its FTM termination charge is A. Each mobile firm is assumed to incur a marginal cost cO of originating a call and a marginal cost cT of terminating a call, so the actual marginal cost of a MTM call is cO + cT. In addition, there is a fixed cost f of serving each mobile subscriber, which includes the subscriber’s handset, billing costs, and so on. For now, assume that FTM and MTM calls are independent markets, and that the call charge in one market does not affect the demand in the other market. Figure 1 depicts our stylised model of the mobile industry. Denote firm i’s on‐net MTM call charge by pi and its off‐net MTM call charge by . In addition, the firm charges a fixed (rental) charge ri for subscribing. If firm i’s market share is si, its subscribers make a fraction si of their mobile calls on‐net and the remaining 1−si calls off‐net. Then a subscriber’s utility if she joins that network is (3) We assume a Hotelling specification for subscriber network choice, and the market share of network i given the pair of utilities {u1,u2} available from the networks is (4) Here, t is a parameter which represents the degree of product differentiation in the market for mobile subscribers.12 Note that in this benchmark model there is an exogenously fixed number of mobile subscribers, which is normalised to 1. (Appendix A analyses a more complicated but more realistic model where the number of mobile subscribers increases if networks offer better deals, and this analysis is summarised in Section 2.3.) Given a choice a for the industry MTM termination charge, suppose that network i chooses its own charges to be (). Then network i’s profit is (5) This consists of the retail profit from supplying service to its subscribers, the profit from providing termination for the rival mobile network, and the profit from providing termination for the fixed network. 2.1. Fixed‐to‐mobile Call Termination First, consider network i ’s incentive to choose its FTM termination charge, Ai. Expression (5) shows that each network’s unregulated FTM termination charge will be chosen to maximise its profits from FTM call termination, F(·). This is a dominant strategy for each network, regardless of choices for retail tariffs and the MTM termination charge. By setting Ai to maximise F, each firm will be able to subsidise subscribers to the maximum extent, thereby increasing market share without having to lower profit per subscriber. We denote the resulting unregulated termination charge by AM, to indicate it is the termination charge that would be chosen by a monopoly mobile network. Welfare (as measured by the sum of consumer surplus and profit) generated by FTM calls when the termination charge is A is (6) which simplifies to As one would expect, this is maximised by setting the FTM call charge equal to the cost of such calls: where AW denotes the welfare‐maximising FTM termination charge. This ensures fixed‐line callers face a FTM price equal to the true marginal cost of their calls, C + cT. When the FTM call charge is equal to the fixed network’s cost, so that (1) holds, welfare is maximised with a FTM termination charge equal to cost so that AW = cT. On the other hand, if P(A) > C + A then welfare is maximised by setting AW < cT to counteract the markup present in the FTM retail charge. 2.2. Mobile‐to‐mobile Call Termination Suppose that the mobile networks each set the FTM termination charge A, which could be AM or any other (perhaps regulated) level of A. (The choice of a will turn out not to depend on A in this benchmark model.) First, we derive equilibrium call charges given A and a. In a symmetric equilibrium, each network will serve half the market. Therefore, from (3) and (4), firm i ’s market share is unchanged if it modifies its charges in such a way that (7) is unchanged. From (5) and (7), in a symmetric situation the network’s profit in terms of ) is It follows that pi is chosen to maximise v(p) + (p − cO − cT)q(p) and is chosen to maximise v(p) + (p − cO − a)q(p). Therefore, in equilibrium each network will set the on‐net call charge p and off‐net call charge given by (8) so that call charges are equal to the respective marginal costs of making calls.13 Having found the equilibrium call charges, we complete the analysis of retail tariff decisions by considering the choice of rental charge, r. Analogously to (2), write (9) for the profit from MTM termination when the MTM termination charge is a. From (5) and (8), network i ’s profit is (10) Expression (4) implies that firm i ’s market share si satisfies where and . Solving this explicitly in terms of si implies that (11) Finally, substituting (11) into (10), maximising with respect to ri and setting ri = rj = r shows that the equilibrium rental charge is (12) From (3) and (12), subscriber utility is equal to (13) Clearly, subscriber utility increases with both the termination charges A and a (at least for A up to the monopoly level AM). This observation implies that firms and the regulator can use relatively high termination charges as a means to expand the number of mobile subscribers, as discussed in the next section. The final variable to determine is the industry choice of a. Substituting (12) into (10) shows that industry profit in the mobile sector is (14) In particular, the FTM termination charge A has no impact on equilibrium profits in the mobile sector. By contrast, mobile networks do care about the MTM termination charge. Without regulation, the industry will choose a to maximise (14), which clearly entails a < cT. Thus, as in Gans and King (2001, Proposition 2), mobile networks prefer a MTM termination charge set below cost. In fact, if demand is concave (q′′≤0), networks prefer the lowest feasible MTM termination charge. If negative termination charges are not feasible, this implies that networks prefer a so‐called ‘bill‐and‐keep’ arrangement, whereby the MTM termination charge is equal to zero. Since the efficient MTM call charge, for both on‐net and off‐net calls, is equal to the cost cO + cT, expression (8) implies that the efficient MTM termination charge is a = cT, just as with FTM termination. In particular, optimal regulatory policy treats the two termination charges symmetrically. We deduce that unregulated firms in this model will choose a MTM termination charge which is too low relative to the efficient level, in contrast to their incentives concerning FTM termination. 2.3. Mobile Market Expansion One limitation of the benchmark model just outlined is that the number of mobile subscribers was constant and did not change in response to the deals on offer. However, historically at least, the market has expanded dramatically (see Table 3). While in a recent survey only 10% of households in the UK market were without access to a mobile phone (Ofcom, 2006a, p. 157), there may nonetheless remain some marginal subscribers, and inducing these subscribers to have a mobile phone could still generate externalities to others.14 Incorporating market expansion effects into the model has important implications for the analysis of FTM and MTM termination charges. It will also help to explain why mobile firms resisted regulated reductions in their FTM termination charges, and will play a key role in our analysis of the coordinated choice of a uniform termination charge in section 3.1. In this section we modify the benchmark model so that the number of subscribers served by network i is ni = φ(ui,uj) rather than (4), where the total number of mobile subscribers, denoted N ≡ n1 + n2, is no longer constant. Subscriber utility at firm i is modified from (3) to be (15) Here, v0 is a subscriber’s utility from other mobile services, most notably from calls to the fixed and international network. (This parameter plays no role when the market size is constant and is ignored elsewhere in the paper.) Firm i ’s profit is modified from (5) to be It is immediate that each firm will set its FTM termination charge Ai to maximise profits from termination F(·), so A = AM in equilibrium. The potential for market expansion has no impact on a mobile network’s incentive to set a high FTM termination charge, since even with a fixed market size the termination charge was set at its profit‐maximising level. Likewise, when the MTM termination charge is a, equilibrium call charges still reflect calling costs, so that (8) holds. Therefore, firm i’s profit is (16) Incorporating elastic subscriber participation into the benchmark model introduces market‐level network effects (as more subscribers join, it becomes more attractive for others to get a mobile phone). In order to make further progress, it is convenient to specify subscriber demand φ explicitly. To that end, suppose that (4) is modified in the following way: if the two networks offer utilities u1 and u2, then firm i attracts (17) subscribers. (This model of consumer demand, which is chosen for its relative analytical tractability, is sometimes known as the ‘Hotelling model with hinterlands’.) Here, λ ≥ 0 represents the magnitude of market expansion possibilities. In order to ensure market expansion is non‐explosive, λ cannot be too large and the following condition is assumed: (18) Using this model, in Appendix A we establish that the equilibrium market size N increases with F(A), reflecting the fact that high FTM termination profits will feed through into attractive tariffs for mobile subscribers, which will in turn induce more people to subscribe to mobile networks. This implies mobile profits increase as the FTM termination charge is increased above cost. The efficient FTM termination charge is also now above cost, since greater subscription will benefit all users (fixed and mobile) since they have more people to call. Nevertheless, the efficient FTM termination charge is still below the unregulated level, AM. Starting from this profit‐maximising charge, a small reduction in the charge has a second‐order impact on the profit from call termination, and hence only a second‐order impact on the number of mobile subscribers. However, it has a first‐order impact on the retail price of FTM calls. Therefore, welfare rises with a reduction in the FTM charge from the unregulated level.15 In sum, when there is potential for market expansion there remains a rationale for regulatory control of the FTM termination charge, albeit reduced. What about the MTM termination charge when market expansion is possible? Appendix A shows that firms still jointly prefer to choose a MTM termination charge which is below cost, just as in the benchmark model. However, the efficient MTM termination charge is, like the FTM termination charge, now above cost. The reason is that mobile subscriber surplus is increased by a high MTM termination charge (as in the benchmark model). In sum, if in this extended model networks can feasibly set different termination charges for FTM and MTM traffic, then without regulation networks would choose too high an FTM charge and too low a MTM charge. The socially efficient charges, though, are now both above the cost of supplying termination (and are typically set at different levels, if feasible). 2.4. Discussion 2.4.1. FTM termination: Despite competition between mobile networks in the retail market, the equilibrium FTM termination charge is equal to the monopoly charge. The result does not depend on the competitiveness of the market for subscribers (it does not depend on product differentiation t and it would not depend on the number of firms if our benchmark model was extended to allow more firms). Thus, it is perfectly possible that one side of the mobile market (the retail market for mobile subscribers) is highly competitive, yet the other side (FTM call termination) is essentially a series of monopolies. By contrast, the socially efficient FTM termination charge is lower. If market expansion possibilities are not significant, the efficient charge is approximately equal to the cost of providing termination; if expansion is possible then the efficient charge will be somewhat above cost in order to stimulate market expansion but it is always below the unregulated level.16 Similarly, in our benchmark setting even a small (or new) mobile network has the ability and incentive to set high FTM termination charges. An illustration of what happens if some firms are regulated and some are not was presented in Table 2 above, where, on weekdays at least, H3G took advantage of its then less regulated position to set termination charges which were roughly double those levied by its more regulated rivals. In the benchmark model, competition between networks to attract subscribers meant higher FTM termination charges resulted in lower prices for mobile subscribers. Each subscriber brings a profit F(A) and competition acts to pass these profits onto the subscribers themselves. Correspondingly, if regulation squeezes out profits in FTM call termination, this will lead to price rises for mobile subscribers. This result is often termed the ‘waterbed’ effect. In the benchmark model with constant market size (Sections 2.1 and 2.2) there is a 100% waterbed effect, in that reduced profits from one source are completely clawed back from subscribers so that the overall profit impact is zero. Thus, without market expansion effects, firms should not object to regulatory intervention to bring each firm’s FTM termination charge down from the unregulated level to the socially efficient level. Nevertheless, in reality mobile networks do object to such a policy. Allowing for market expansion (Section 2.3), we found that the resulting waterbed effect is less than 100%. As such, mobile networks will collectively strictly prefer a higher FTM termination charge, and so have an incentive to lobby against proposed regulation to bring down FTM termination charges to efficient levels. A feature of a strong waterbed effect is that unregulated monopoly profits from FTM call termination are largely passed onto mobile subscribers in the form of low rental charges (e.g., free or subsidised handsets). To the extent this occurs, the market failure associated with FTM termination does not necessarily lead to excessive profits by mobile networks, but rather to a sub‐optimal balance of retail prices. Of course, this observation should not affect the welfare analysis: if high margins on FTM call termination lead to negative margins on services to mobile subscribers, there is allocative inefficiency regardless of whether overall profits in the mobile sector are excessive or not. Relatedly, mobile firms ‘advanced the argument that, because most people had a mobile phone, what they lost in high termination charges they gained in low access and outbound call charges’.17 However, even if all fixed‐line subscribers had a mobile phone, this argument is not correct: since high termination charges lead to allocative inefficiency, the total ‘size of the cake’ is shrunk, and the gain from handset subsidies is smaller than the losses caused by high FTM call charges.18 Nevertheless, as we discuss in Section 3, if FTM calls can substitute for MTM calls, the degree of overlap between the populations of fixed‐line and mobile subscribers does become relevant. Recall from Section 1 that UK policy has been to allow mobile networks to recover a portion of their fixed costs via a surcharge on FTM termination charges. While such a procedure is fairly standard in setting access prices for a regulated fixed‐line network, it is less clear that this policy should be applied to competitive mobile networks. If a strong waterbed effect operates, setting higher FTM termination charges will not provide networks with a significant contribution towards their fixed and common costs, and the policy will simply act as an (inefficient) transfer of surplus between fixed and mobile subscribers. In this environment, setting high termination charges in an attempt to allow for fixed and common cost recovery is likely to be a flawed policy. On the other hand, if market expansion possibilities are significant, the waterbed effect is weaker and above‐cost termination charges could, if necessary, be used to help cover the networks’ fixed costs (which are not part of our formal model), as well as being used to stimulate market expansion. An implicit assumption in the benchmark model is that fixed‐line callers alone determine the number and length of FTM calls. If in practice the volume of FTM calls is jointly determined by the caller and receiver, then high FTM termination charges could lead mobile networks to pay subscribers for receiving FTM calls (rather than merely reducing rental charges), so as to stimulate FTM call volumes. For instance, in 2006 H3G announced that it would pay its subscribers 5 pence per minute (ppm) for receiving calls, a marketing tactic which surely was motivated by its highly profitable termination charges seen in Table 2. This last point is also helpful in understanding the difference between the ‘caller‐pays’ regime as used in the UK, in which unregulated FTM termination charges and call charges are relatively high, and a ‘receiver‐pays’ regime used in the US, whereby the price of a FTM call is low but mobile subscribers often incur charges for receiving calls. In the US, reciprocity requirements imposed by regulators mean that FTM termination charges are set equal to those for mobile‐to‐fixed termination, which are quite likely below the cost of mobile termination.19 With little termination revenue, mobile networks may choose to charge mobile subscribers directly to recover their costs and/or to induce their subscribers to discourage incoming traffic. For instance, with a charge for receiving calls, a subscriber may keep her phone switched off to eliminate incoming calls, something which is often inefficient. The key point is not that the problem of high FTM termination charges is solved by imposing a receiver‐pays regime, but rather that a regulatory requirement to set low FTM termination charges may induce networks to charge subscribers for receiving calls. Low FTM termination charges will typically harm mobile subscribers and so reduce mobile penetration when subscriber numbers are elastic.20 2.4.2. MTM termination Why do firms wish to set a below‐cost MTM termination charge, with or without market expansion possibilities, if free to do so? Unless firms set a low termination charge, call charges (as in (8) above) will be such that it is more expensive to call off‐net than on‐net. In such a situation, subscribers will, all else equal, prefer to join a larger network since they can then make a larger fraction of their calls at the cheaper on‐net rate. In other words, the market will exhibit (positive) network effects. As is well known, unless network effects are so strong that the market tips to monopoly, in such markets competition is particularly fierce and profits are low.21 Firms can overturn this effect by setting a low MTM termination charge, which results in off‐net call charges which are below on‐net charges. In this case, subscribers will prefer to join a smaller network, which acts to relax competition. In this framework, mobile subscribers benefit both from high FTM and high MTM termination charges. However, the reason is quite different in the two cases. Mobile subscribers benefit from high FTM charges since there is a waterbed effect at work, and profits from this source are passed (at least partially) onto mobile subscribers. High FTM termination charges are a means of transferring surplus from fixed callers to mobile recipients (and in part to mobile networks if the waterbed effect is not complete). High MTM termination charges act to intensify competition between mobile networks due to strengthened network effects. When the MTM termination charge is raised, this imposes a direct cost on subscribers since they must pay more for off‐net calls but this is outweighed by the lower rental charge they pay. High MTM termination charges have little impact on fixed callers (unless market expansion possibilities are significant) and so a high MTM termination charge acts principally as a means by which to transfer surplus from mobile networks to their subscribers. This also implies that high MTM termination charges cannot be used as a means to cover fixed costs in the industry. 2.4.3. A linear demand example It is useful to illustrate these results with a specific numerical example. Suppose costs are normalised to zero (C = cO = cT = f = 0), the parameter v0 from the market expansion model is set equal to zero, the network differentiation parameter is t = 1/2 , and the two demand functions are and . The parameter μ measures the relative importance of FTM calls relative to MTM calls (in part, it reflects the fraction of time people want to make calls from home or office versus when they are on‐the‐go). The unregulated FTM termination charge is AM = 1/4 (which is broadly in line with Ofcom’s (2007b, para. 7.49) estimate of the unregulated FTM termination charge), while in the absence of market expansion possibilities firms will wish to choose as low a MTM termination charge as is feasible. The efficient choice of each termination charge in the absence of network expansion possibilities is AW = 0. As discussed, in the absence of regulation, firms will choose the monopoly FTM termination charge regardless of λ and μ, while the choice of the MTM charge will depend on these parameters but will always be below cost (which in turn is weakly lower than the efficient MTM termination charge). What is the size of the waterbed effect in this example and how sensitive is mobile penetration to changes in the FTM termination charge? In general, these effects will depend on μ and the choice of MTM termination charge a but for simplicity consider the case where μ = 1 so that MTM calls are absent. In this case, when λ = 0 there is no scope for market expansion and as in the benchmark model there is a 100% waterbed effect (network profits do not depend on A in expression (14)). If λ = 1/4, one can calculate using the analysis in Appendix A that mobile industry profits rise by approximately 0.017 when the FTM termination charge is increased from zero to the monopoly charge. The profit due to FTM termination rises from zero to 0.0537. It follows that around 32% of termination profits are retained by the industry and the waterbed effect is 68%.22 In the same situation, mobile subscriber numbers increase by 3.1%. Table 6 reports these comparative statics, as well as those for other values of λ. Table 6 
Impact as A Increases From Marginal Cost to Monopoly Level (μ = 1) . Increase in subscribers . Size of waterbed effect . λ=0 0 100% λ=1/8 1.6% 80% λ=1/4 3.1% 68% λ=1/2 6.1% 51% . Increase in subscribers . Size of waterbed effect . λ=0 0 100% λ=1/8 1.6% 80% λ=1/4 3.1% 68% λ=1/2 6.1% 51% Open in new tab Table 6 
Impact as A Increases From Marginal Cost to Monopoly Level (μ = 1) . Increase in subscribers . Size of waterbed effect . λ=0 0 100% λ=1/8 1.6% 80% λ=1/4 3.1% 68% λ=1/2 6.1% 51% . Increase in subscribers . Size of waterbed effect . λ=0 0 100% λ=1/8 1.6% 80% λ=1/4 3.1% 68% λ=1/2 6.1% 51% Open in new tab The welfare‐maximising pair of termination charges for this example are given in Table 7. The Table illustrates the general analysis which showed that both charges should be set above cost whenever market expansion is possible. In the cases considered, in which the waterbed effect varied between 50% and 80%, the potential for market expansion implies that the efficient FTM termination charge is between 20% and 40% of the distance between marginal cost and the unregulated monopoly charge.23 We also see that it is generally efficient, if feasible, to set different termination charges for the two kinds of traffic. However, the difference is negligible when market expansion possibilities are small (e.g., λ = 1/8). In this example at least, more significant market expansion possibilities lead to higher optimal termination charges. Table 7 
Welfare‐maximising Non‐uniform Termination Charges* . μ = 0 . μ = 1/2 . μ = 1 . λ = 0 a = 0 a = A = 0 A = 0 λ = 1/8 a = 0.051 a = 0.051, A = 0.050 A = 0.050 λ = 1/4 a = 0.072 a = 0.071, A = 0.076 A = 0.074 λ = 1/2 a = 0.085 a = 0.084, A = 0.102 A = 0.099 . μ = 0 . μ = 1/2 . μ = 1 . λ = 0 a = 0 a = A = 0 A = 0 λ = 1/8 a = 0.051 a = 0.051, A = 0.050 A = 0.050 λ = 1/4 a = 0.072 a = 0.071, A = 0.076 A = 0.074 λ = 1/2 a = 0.085 a = 0.084, A = 0.102 A = 0.099 *The choice of A is irrelevant when μ = 0 and the choice of a is irrelevant when μ = 1. Open in new tab Table 7 
Welfare‐maximising Non‐uniform Termination Charges* . μ = 0 . μ = 1/2 . μ = 1 . λ = 0 a = 0 a = A = 0 A = 0 λ = 1/8 a = 0.051 a = 0.051, A = 0.050 A = 0.050 λ = 1/4 a = 0.072 a = 0.071, A = 0.076 A = 0.074 λ = 1/2 a = 0.085 a = 0.084, A = 0.102 A = 0.099 . μ = 0 . μ = 1/2 . μ = 1 . λ = 0 a = 0 a = A = 0 A = 0 λ = 1/8 a = 0.051 a = 0.051, A = 0.050 A = 0.050 λ = 1/4 a = 0.072 a = 0.071, A = 0.076 A = 0.074 λ = 1/2 a = 0.085 a = 0.084, A = 0.102 A = 0.099 *The choice of A is irrelevant when μ = 0 and the choice of a is irrelevant when μ = 1. Open in new tab 2.4.4. A puzzle As mentioned in the Introduction, we are not aware of any regulator who has been concerned that firms will set MTM termination charges which are unduly low. Rather, the worry has been that MTM termination charges will be too high. Related to this, even when given the opportunity to do so, mobile networks do not typically set their MTM termination charges below their FTM charges.24 These observations seem to contradict the implications of this benchmark model with separate termination charges (with or without market expansion possibilities). In the next section we resolve this puzzle by taking into account constraints that force FTM and MTM termination charges to be locked together. Before turning to this analysis, though, we note other potential explanations for why networks may in practice prefer high MTM termination charges. If retail tariffs were simply linear prices, perhaps because a sufficient number of subscribers have pre‐paid mobile contracts involving only per‐minute charges, it is possible that firms would agree to set above‐cost MTM termination charges.25 Assuming linear pricing, Armstrong (1998), Laffont et al. (1998a) and Carter and Wright (1999) showed that networks prefer to set MTM termination charges above cost in order to raise the retail price of MTM calls and so undo retail competition. However, these early papers did not allow for differential on‐net and off‐net retail pricing of MTM calls, which is permitted and common in most countries (see Table 4 for the UK). Allowing for this, Laffont, Rey, and Tirole (1998b) show that firms may prefer below‐cost MTM termination charges when there is sufficient competition between mobile networks, even with linear pricing. Firms have two pricing instruments, and so committing to high off‐net prices through a high MTM termination charge may not raise profits since it makes firms compete more intensely through their on‐net charges. Berger (2004) extends this analysis to allow subscribers to obtain utility from receiving calls, and shows the tendency for networks to prefer a below‐cost termination charge is enhanced. Moreover, given that many subscribers do face monthly rental charges or other types of fixed fees, we think the use of nonlinear tariffs adopted in this article (and by most of the more recent literature) is reasonable. The fact that some subscribers pay linear prices, without on‐net and off‐net differentials, is not likely to be the principal reason networks have a bias towards high MTM termination charges. Another reason why incumbent networks might prefer high MTM termination charges is that high charges could act to deter entry or induce exit of a smaller rival. By setting above‐cost MTM termination charges, incumbent networks can induce network effects which make entry less attractive for the newcomer. With high MTM termination charges, off‐net calls will be more expensive, which particularly hurts a small network since the bulk of its subscribers’ calls will be off‐net.26 An additional effect of high off‐net call prices will be to reduce the number of calls received by a small network’s subscribers, thereby further reducing its ability to compete when call externalities are important.27 This explanation for high MTM termination charges is complementary to ours as outlined in the next Section. If incumbent networks faced the threat of entry, then the effect of their chosen termination charge on entry would be an additional factor determining the equilibrium charge, and the threat of entry presumably implies that the chosen termination charge will be higher than without this threat. At least two further explanations have been proposed in the literature for why firms may prefer high MTM termination charges, although both rely on considerably more subtle mechanisms. First, Cherdron (2002) and Gabrielsen and Vagstad (2008) consider a setting where calling patterns are biased towards peers (closed user groups). Setting above‐cost MTM termination charges is a way to differentiate the networks endogenously, so that consumers prefer to join the network that their peers join, thereby reducing the intensity of competition. Second, Höeffler (2006) argues in a dynamic model with heterogenous users that higher MTM termination charges can allow a collusive outcome to be sustained for a wider range of discount rates. This is because higher termination charges make deviating less attractive given that deviating in his model involves attracting the high‐usage consumers and facing a net outflow of calls. None of these explanations indicates why a mobile network typically sets its FTM and MTM termination charges uniformly. In our view, there is a simpler explanation for why firms prefer high MTM termination charges, which is based on networks not wanting to undercut their high FTM termination charges. We now turn to this explanation. 3. Uniform Termination Charges Until this point we have assumed that mobile networks negotiate an industry‐wide MTM termination charge, which could freely be set at a different level from their FTM charge. However, in practice FTM and MTM charges are not set at different levels. As mentioned in the Introduction, one reason why mobile networks may be forced to set a uniform charge is the ability of networks to arbitrage between significant differences in FTM and MTM termination charges.28 In addition, a network whose FTM termination charge is regulated (or facing the threat of regulation) may be unwilling to choose a lower charge for MTM termination if it suspects the regulator may use that information subsequently to tighten FTM regulation. In this Section we analyse three aspects of the choice of uniform termination charges. First, we isolate the impact of requiring the same termination charge for both kinds of traffic (Section 3.1). Here, we assume that networks negotiate their uniform charge. Since without the uniformity constraint we know that firms will set a high FTM charge and a low MTM charge, when the two charges are locked together the uniform charge will be higher than the jointly chosen MTM charge. Indeed, often the unregulated jointly‐chosen uniform charge will be above the efficient level. Thus, the assumption that charges must be uniform can by itself explain the regulatory concern that charges would be too high rather than too low. Second, in Section 3.2 we suppose that the uniform termination charge is chosen unilaterally by individual networks. When each network’s FTM and MTM termination charges are constrained to be equal, we believe it is more natural to assume that unregulated uniform termination charges are chosen unilaterally rather than in a coordinated fashion.29 When charges are chosen unilaterally, it is intuitive (and confirmed in the analysis) that networks will choose a higher charge than when the charge is jointly chosen. The reason is that a network will not internalise the negative impact on a rival’s business when it sets a high charge, while this externality is taken into account when charges are jointly determined. Thus, the assumption that unregulated charges are chosen unilaterally acts to reinforce the tendency towards higher charges already present in Section 3.1. Finally, in Section 3.3 we consider the impact of demand substitution between FTM and MTM calls. This acts to reduce a firm’s incentive to set a high termination charge, although never by enough to make the unregulated charge be too low. 3.1. Coordinated Choice of Uniform Termination Charges In the situation where the number of mobile subscribers is constant, so that a 100% waterbed effect is present, the analysis of Section 2.2 regarding the MTM charge applies precisely to the modified situation where a uniform termination charge is collectively chosen. The reason is that the level of the FTM termination charge has no impact on the industry’s equilibrium profits, and so when firms choose the termination charge collectively, profits from FTM termination have no impact on their overall profits. Therefore, the firms’ incentives are exactly as if there is only MTM traffic. In this case, the unregulated uniform termination charge would be set below the efficient level. Thus, the puzzle remains as to why regulators in practice are concerned with termination charges that are set too high. A more interesting and realistic case is where there is the potential for market expansion, so that the waterbed effect is only partial and firms benefit from high FTM termination charges. To analyse this case we use the model presented in Section 2.3, but assume networks jointly choose a uniform termination charge. (More details are found in Appendix B.) Firms will agree on a uniform termination charge which lies between the two charges they would choose if FTM and MTM termination could be priced separately. Firms will trade‐off the need to protect their FTM termination profits with the worry that a higher uniform termination charge will strengthen network effects and the intensity of retail competition. In other words, when both types of termination charges must be set at the same level, MTM termination will act as a constraint on the competitive bottleneck problem that arises with FTM termination. The larger the proportion of MTM calls, the more important is this constraint. On the other hand, the incentive to protect FTM termination profit given the waterbed effect is less than 100% implies they will set a higher charge than the MTM termination charge they would choose with separate charges. It is ambiguous whether firms will collectively agree on a uniform termination charge that is above or below cost, or above or below the efficient level. We illustrate with the linear demand example introduced in section 2.4. Table 8 reports the unregulated firms’ joint choice of uniform termination charge. Table 8 
Joint Profit‐maximising Uniform Termination Charge . μ = 0 . μ = 1/2 . μ = 1 . λ = 0 a as low as feasible a as low as feasible a is irrelevant λ = 1/8 a = −2.618 a = −0.111 a = 0.250 λ = 1/4 a = −1.465 a = 0.066 a = 0.250 λ = 1/2 a = −0.270 a = 0.153 a = 0.250 . μ = 0 . μ = 1/2 . μ = 1 . λ = 0 a as low as feasible a as low as feasible a is irrelevant λ = 1/8 a = −2.618 a = −0.111 a = 0.250 λ = 1/4 a = −1.465 a = 0.066 a = 0.250 λ = 1/2 a = −0.270 a = 0.153 a = 0.250 Open in new tab Table 8 
Joint Profit‐maximising Uniform Termination Charge . μ = 0 . μ = 1/2 . μ = 1 . λ = 0 a as low as feasible a as low as feasible a is irrelevant λ = 1/8 a = −2.618 a = −0.111 a = 0.250 λ = 1/4 a = −1.465 a = 0.066 a = 0.250 λ = 1/2 a = −0.270 a = 0.153 a = 0.250 . μ = 0 . μ = 1/2 . μ = 1 . λ = 0 a as low as feasible a as low as feasible a is irrelevant λ = 1/8 a = −2.618 a = −0.111 a = 0.250 λ = 1/4 a = −1.465 a = 0.066 a = 0.250 λ = 1/2 a = −0.270 a = 0.153 a = 0.250 Open in new tab When μ = 0 the equilibrium termination charge is below cost, since there are no FTM calls and so nothing to counter the firms’ desire to relax retail competition by means of below‐cost termination charges. Likewise, where market expansion possibilities are limited, there is a very strong waterbed effect and FTM termination profits do not act as much of a brake on the incentive to set low termination charges. However, when μ and λ are not too small, the profit‐maximising uniform charge is above cost, but below the monopoly charge AM = 1/4. Table 9 shows the corresponding welfare‐maximising charges. Table 9 
Welfare‐maximising Uniform Termination Charge . μ = 0 . μ = 1/2 . μ = 1 . λ = 0 a = 0 a = 0 a = 0 λ = 1/8 a = 0.051 a = 0.050 a = 0.050 λ = 1/4 a = 0.072 a = 0.074 a = 0.074 λ = 1/2 a = 0.085 a = 0.098 a = 0.099 . μ = 0 . μ = 1/2 . μ = 1 . λ = 0 a = 0 a = 0 a = 0 λ = 1/8 a = 0.051 a = 0.050 a = 0.050 λ = 1/4 a = 0.072 a = 0.074 a = 0.074 λ = 1/2 a = 0.085 a = 0.098 a = 0.099 Open in new tab Table 9 
Welfare‐maximising Uniform Termination Charge . μ = 0 . μ = 1/2 . μ = 1 . λ = 0 a = 0 a = 0 a = 0 λ = 1/8 a = 0.051 a = 0.050 a = 0.050 λ = 1/4 a = 0.072 a = 0.074 a = 0.074 λ = 1/2 a = 0.085 a = 0.098 a = 0.099 . μ = 0 . μ = 1/2 . μ = 1 . λ = 0 a = 0 a = 0 a = 0 λ = 1/8 a = 0.051 a = 0.050 a = 0.050 λ = 1/4 a = 0.072 a = 0.074 a = 0.074 λ = 1/2 a = 0.085 a = 0.098 a = 0.099 Open in new tab Here, the efficient termination charge is above cost whenever market expansion is possible (as in Section 2.3). When both kinds of traffic are present (μ = 1/2), the efficient uniform charge lies between the efficient charges reported in Table 7 for the two different types of calls. (Clearly, the outer columns in Tables 7 and 9 coincide.) Comparing Tables 8 and 9 shows it is ambiguous whether unregulated firms will choose too high or too low a uniform termination charge when they do so in a coordinated fashion. In broad terms, if MTM traffic is particularly significant then the danger is that firms agree to set too low a charge, while if FTM traffic is more significant the danger is that firms choose too high a termination charge. In a range of intermediate cases (e.g., μ = 1/2 and λ = 1/4 in the example), the unregulated jointly‐chosen termination charge will be reasonably close to the efficient level, in which case requiring networks to coordinate over a uniform FTM and MTM termination charge, if such a policy were feasible, may lead to a reasonably efficient outcome. 3.2. Unilateral Choice of Termination Charges As noted previously, given FTM and MTM termination charges are locked together, in the absence of regulation it is more natural to assume each network sets its own uniform termination charge. In this Section we analyse the situation in which the uniform termination charge is chosen unilaterally by each firm.30 We initially assume that the number of mobile subscribers is constant in this analysis, and then briefly discuss the impact of market expansion. Denote network i’s uniform termination charge for both FTM and MTM traffic by ai. Then following the same logic as in expression (8), firm i ’s call charges are (19) and, as in (10), its profit is (20) From (4), firm i ’s market share si satisfies and so (21) Since the MTM demand function q(·) and the FTM demand function Q(·) might differ, we define the monopoly charge aM to be the uniform charge which maximises total termination profit per subscriber, i.e., holding fixed the number of subscribers on each network. Given that there are two networks, half of the MTM calls received by each subscriber are off‐net calls. Each subscriber therefore generates termination revenue of F + M/2, and the monopoly termination charge, aM, satisfies (22) In the absence of strong market expansion possibilities, we will show that the equilibrium termination charge will be below aM, suggesting the competitive bottleneck problem is weakened in this setting. As before, we require the networks to be sufficiently differentiated so that network effects are not explosive. From (21), a simple sufficient condition to ensure this is (23) This rules out any deviation by network i (even to the point that ), when the other network sets its termination charge no higher than the monopoly level aM, as in the proposed equilibrium found below. Without such a condition, there is a possibility that network i could profitably deviate: by setting a very high termination charge, network i could make off‐net calls from the other network prohibitively expensive, so that under certain consumer expectations about which network people will join, everyone will subscribe to network i. Differentiating (20) with respect to ri and setting equal to zero yields (24) where we have written Mi = M(ai) and Fi = F(ai), and the second equality follows from (21). Substituting this value for ri into (20) shows that firm i’s profit is (25) Moreover, substituting ri as given in (24) and the corresponding expression for rj into (21) shows that (26) which gives the equilibrium market shares in terms of the two networks’ termination charges. Substituting (26) into (25), differentiating with respect to ai and imposing symmetry (a1 = a2 = a) implies the first‐order condition for the uniform, unilaterally chosen termination charge a is (27) Evaluating the derivative in (27) at a = cT yields (28) and thus the unilaterally chosen termination charge will be above marginal cost. Since the efficient termination charge is a = cT, it follows that unregulated firms will unambiguously choose too high, rather than too low, a uniform termination charge. This result was previously derived in Gans and King (2001, Proposition 1), in the case where FTM calls were absent.31 Gans and King (2001, p. 417) attribute this result to a ‘standard double marginalisation result that arises when firms set prices for complementary services independently’. Our interpretation is different. When network i raises its uniform termination charge there are three effects: (i) its profit from supplying call termination to its mobile rivals and to the fixed network increases (just as with FTM termination in the benchmark model); (ii) its rival is forced to raise its off‐net call charge, which directly increases firm i ’s market share; and (iii) it amplifies the differential pricing between on‐net and off‐net calls in the market which, as we argued in Section 2.4, intensifies competition for subscribers. Effects (i) and (ii) suggest that a firm will want to set its termination charge above the monopoly level, since (ii) gives a reason to boost the charge in addition to extracting termination profits. But (iii), which is a strategic effect working through softening retail competition, puts downward pressure on termination charges. In this model expression (28) shows that the net incentive is to set the termination charge above the efficient level. Thus, the two direct effects dominate the incentive to set a low charge to relax competition between networks. Expression (28) also shows that the existence of FTM calls increases the firm’s incentive to set its mobile termination charge above cost when its MTM and FTM termination charges must be uniform. The existence of FTM calls increases the direct effect in (i), so further raising the incentive to set above‐cost termination charges. We next compare the unregulated charge in (27) with the monopoly charge (22). Equation (27) can be rewritten as (29) where the dependence on a in (29) has been suppressed. It is straightforward to check that the left‐hand side of (29) is greater than the right‐hand side when a = cT. In contrast, at the monopoly termination charge aM the left‐hand side equals zero by definition, while the right‐hand side is positive from (23). Therefore, there is at least one a ∈ (cT, aM) where (29) holds. Moreover, under relatively mild regularity conditions, the left‐hand side is decreasing and the right‐hand side is increasing in a in the range (cT, aM).32 There is then a unique a ∈ (cT, aM) which satisfies (29). In sum, the equilibrium termination charge is below the monopoly level but above the efficient level. When networks set a uniform termination charge for FTM and MTM traffic, this model predicts unregulated firms will set a uniform charge which is too high relative to the efficient benchmark, and this justifies regulatory concerns that the charge is too high (rather than too low as our benchmark model predicted in Section 2). Nevertheless, the fact that the FTM and MTM termination charges are locked together does mitigate a network’s incentive to set monopoly termination charges, and the competitive bottleneck result emphasised in Section 2 is softened. The key insight here, as with the coordinated setting of termination charges in Section 3.1, is that mobile networks will be constrained in their choice of a uniform termination charge by the fact that setting it too high strengthens network effects, making the firms tougher rivals. To avoid this effect, a network will keep its termination charge below the monopoly level. How close is the equilibrium charge to the efficient level? Consider the linear demand example of Section 2.4. Condition (23) requires that which is satisfied in this example where t = 1/2. If μ = 1/2 and λ = 0 then the solution to (29) is a = 0.230, which is some 8% below the monopoly charge aM = 1/4. One can check that the right‐hand side of (29) is decreasing in t when a > cT. Therefore, the unique a ∈ (cT, aM) which solves (29) is increasing in t. That is to say, if the retail market is more competitive, in the sense that services are closer substitutes, then the equilibrium uniform termination charge will be closer to the efficient level. This contrasts with our benchmark model in which the two charges were set independently: there, competitive conditions at the retail level played no role in the determination of either charge. However, in our numerical example we find that the competitiveness of the retail market does not seem to play a very significant role in the determination of the equilibrium uniform charge. The change in a with respect to a change in t is greatest when all calls are MTM but even then increasing t from t =1/4 to t = 3/4 increases a only by about 2%. Another observation is that if the relative importance of FTM traffic declines, in the sense that F′(·) in (29) is reduced, the equilibrium termination charge falls too. (The left‐hand side of (29) increases with F′.) Recall from Section 1 that the relative importance of FTM traffic has fallen in recent years in the UK, and so this suggests that firms now have a reduced incentive to set very high termination charges compared to the period when regulation was first introduced. The first row of Table 10 shows the significant reduction in equilibrium termination charges induced by a reduction in the proportion of FTM calls in the case without market expansion. When MTM calls are more important, they play a greater role in mitigating a network’s incentive to set monopoly termination charges.33 Table 10 
Unilateral Choice of Uniform Termination Charge . μ = 0 . μ = 1/2 . μ = 1 . λ = 0 a = 0.197 a = 0.230 a = 0.250 λ = 1/8 a = 0.223 a = 0.241 a = 0.250 λ = 1/4 a = 0.237 a = 0.247 a = 0.250 λ = 1/2 a = 0.251 a = 0.251 a = 0.250 . μ = 0 . μ = 1/2 . μ = 1 . λ = 0 a = 0.197 a = 0.230 a = 0.250 λ = 1/8 a = 0.223 a = 0.241 a = 0.250 λ = 1/4 a = 0.237 a = 0.247 a = 0.250 λ = 1/2 a = 0.251 a = 0.251 a = 0.250 Open in new tab Table 10 
Unilateral Choice of Uniform Termination Charge . μ = 0 . μ = 1/2 . μ = 1 . λ = 0 a = 0.197 a = 0.230 a = 0.250 λ = 1/8 a = 0.223 a = 0.241 a = 0.250 λ = 1/4 a = 0.237 a = 0.247 a = 0.250 λ = 1/2 a = 0.251 a = 0.251 a = 0.250 . μ = 0 . μ = 1/2 . μ = 1 . λ = 0 a = 0.197 a = 0.230 a = 0.250 λ = 1/8 a = 0.223 a = 0.241 a = 0.250 λ = 1/4 a = 0.237 a = 0.247 a = 0.250 λ = 1/2 a = 0.251 a = 0.251 a = 0.250 Open in new tab How do these results change when subscriber participation is elastic? Appendix C derives the equilibrium conditions for this case. Table 10 reports the equilibrium termination charge for the linear demand example. This shows that market expansion possibilities induce firms to choose higher termination charges relative to the outcome with a constant market size, presumably because firms take into account the benefits in terms of market expansion of having a higher termination charge. When market expansion effects are strong enough, the chosen charge might even be above the monopoly level. In these (perhaps rather extreme) cases, a firm’s desire to choose the monopoly charge for FTM termination actually tempers their incentive to set an even higher termination charge for MTM calls. Finally, we note that the uniform termination charge is higher when it is unilaterally determined (Table 10) compared to when it is set in a coordinated fashion (Table 8). Given it is set unilaterally, the MTM termination charge is also typically higher when it is required to be set uniformly compared to when networks are free to set distinct FTM and MTM termination charges (the exception is when market expansion effects are so strong that the unilaterally set MTM termination charge is above the monopoly level, which arises in our example when λ = 1/2). As a result, at least in this linear demand example, there is still a significant gap between the efficient and unregulated termination charges (Table 9 vs. Table 10), and thus there remains a rationale for intervention. 3.3. Substitution Between FTM and MTM Calls The previous literature, including this article up to this point, has largely treated FTM callers and MTM callers as though they were disjoint groups, although this is clearly not the case. Mobile subscribers who have access to a fixed line can choose between the two types of calls. It turns out that this form of substitution has two beneficial effects on the market. First, it weakens a mobile network’s incentive to set a high termination charge. Second, it reduces the need to regulate the fixed network’s FTM call charges. To discuss these points, assume as in Section 3.2 that networks unilaterally set a uniform charge for terminating FTM and MTM traffic. Suppose there is full penetration of both the fixed and the mobile networks, so everyone potentially has access to both a fixed and a mobile phone to make calls.34 When a user has access to a fixed phone, we assume that making calls to a mobile subscriber using a mobile phone is a perfect substitute for making the call with a fixed‐line phone and a caller will use whichever method is the cheaper. Since people may have different demands for calling when they have access to a fixed‐line phone (for example, when they are at home) from when they do not (for example, when they are on‐the‐go), we allow what we previously termed the FTM demand function Q(·) to denote demand in the former situation and the MTM demand function q(·) to denote demand in the latter situation. To make the main points as cleanly as possible, suppose that the cost of making calls using the two methods is the same, i.e., C = cO. Given these assumptions, a subscriber’s utility from network i is modified from (3) to be (30) Here Pi is the charge for making FTM calls to network i. The final two terms in (30) are new compared to (3) and represent the utility of a subscriber being able to make calls either with her fixed‐line phone or her mobile phone whenever she has a choice. Previously, we ignored the possibility of such substitution and so implicitly assumed subscribers in these situations could only make these calls using their fixed‐line phone. In that case, the corresponding utility siV(Pi) + (1 − si)V(Pj) was independent of the mobile network a subscriber joins, and so did not affect network i’s market share and had no bearing on the analysis. Allowing for substitution, though, implies network i’s utility in (30) will depend on whether it allows subscribers to call other mobile subscribers more cheaply using its mobile network than they can do using their fixed‐line phone. Suppose that FTM calls satisfy (1), i.e., (31) (We will see shortly that competition with mobile networks for calls will indeed force the fixed network to price in this way.) As with and , we define and . Suppose that in the first stage the two networks have chosen their, possibly unequal, termination charges a1 and a2 , and are considering their choice of retail tariff. When choosing its call charges, a mobile network must decide whether or not to undercut the prevailing FTM call charges, which will determine which phone a subscriber uses when she has a choice. Consider the situation in which a subscriber on network i wishes to make a call to another subscriber on the same network and where the caller has access to her fixed‐line phone. If the caller uses the fixed‐line phone, she pays Pi = cO + ai and enjoys surplus , while network i obtains termination profit (ai − cT)Q(cO + ai). Thus, the joint surplus available to the two parties is On the other hand, if the network undercuts the FTM call charge by choosing pi ≤ Pi, the subscriber enjoys surplus V(pi), the firm obtains profit (pi − cO − cT)Q(pi), where this profit now comes from supplying calls rather than call termination, and the available surplus is If ai > cT, the latter strategy yields the higher joint surplus, and this second joint surplus is maximised by setting pi = cO + cT. We deduce that when ai ≥ cT firm i will set pi = cO + cT, with the result that its subscribers will always use their mobile phone to make calls to others on the same network, even when they have a choice of phone. On the other hand, if ai < cT, total surplus is reduced if the network strictly undercuts the (low) FTM call charge, and the network might as well set pi = cO + cT in order to achieve the maximum total surplus in those situations in which the subscriber cannot use her fixed‐line phone. Consider next the network’s off‐net call charge, . If the subscriber uses her fixed‐line phone to call someone on the rival network, her surplus is and network i makes nothing. If network i undercuts this FTM call charge, so that , then the surplus of the subscriber is and the network’s profit is . The joint surplus following the latter strategy is maximised by setting , in which case the joint surplus with the two strategies is identical. Therefore, the firm and subscriber are indifferent between the two strategies. As such, it is strictly optimal for the network to set , since that also maximises the joint surplus in those situations where the subscriber can only use her mobile phone. In sum, for all termination charges a1 and a2 it is optimal for each network to set its call charges to reflect its cost of making calls, so that (19) holds. Note next that, in this model with full penetration and perfect substitutes, the fixed network will be forced to set its call charges as in (31) above, even if these charges are not regulated. Given the mobile networks are setting their call charges as in (19), if the fixed network increases its price above the price in (31) consumers will always use their mobile phone, while if it decreases its price it will offer its service at a loss. Thus, an important feature of demand substitution is that it potentially eliminates the need to regulate FTM call charges. In situations where ai ≥ cT, it follows that subscriber utility in (30) becomes35 so that from (4) firm i’s market share satisfies In particular, relative to Section 3.2, network effects are more pronounced, and condition (23) now needs to be tightened to . Network i’s profit when ai ≥ cT is To understand this expression, notice that (i) network i ’s subscribers will never use a fixed‐line phone to call others on the same network, (ii) when the network’s subscribers call someone on the other network, they are indifferent about using their mobile or fixed‐line phone (when they have a choice), and the network makes no profit from these calls in either case, and (iii) when a subscriber on the rival network calls a network i subscriber and has the option to use their fixed‐line phone to make the call, they are indifferent about which type of phone to use, and network i makes the same profit F(ai) in either event (and where they do not have this option they will use their mobile phone and network i makes profit M(ai)). Following the same steps as used in Section 3.2 shows that instead of (27) the unilaterally‐chosen uniform termination charge here satisfies which can be rewritten as (32) We are interested in whether demand substitution increases or reduces a network’s incentive to set an excessive termination charge, i.e., whether the solution to (32) is above or below that corresponding to expression (29). The comparison is most transparent if we specialise the demand for calls somewhat. Suppose that the underlying demand for calls, which we denote X(·), is the same whether the caller is at home or on‐the‐go, and that each caller is at home a fraction μ of the time. In this case, Q = μ X and q = (1 − μ)X, which gives one interpretation of the different weights on the two types of demand in the example given in Section 2.4. The equilibrium termination charge in (32) does not depend on μ, since all that matters in (32) is the sum of demands Q + q, which always equals X. With this demand specification, expression (29) coincides with (32) when μ = 0, i.e., when there are no FTM calls. Then one can check that the left‐hand side of (29) is increasing in μ in this example, while the right‐hand side is decreasing with μ. It follows that the equilibrium termination charge which solves (29) is increasing with μ. We deduce that the charge a which solves (32) is below the charge a which solves (29). Moreover, the larger μ is, the greater is the difference between the two. Thus, we can conclude that demand‐side substitution between FTM and MTM calls lowers the equilibrium termination charge towards marginal cost. For instance, using the linear demand example of Section 2.4 and assuming that μ = 1/2 and λ = 0 one can show that the equilibrium termination charge which solves (32) is approximately a = 0.197, which is some 22% below the monopoly charge (rather than the 8% which applied when there was no such substitution). In sum, the fact that consumers substitute on‐net MTM calls for FTM calls whenever the former’s price is lower mitigates a mobile network’s incentive to set high mobile termination charges but it does not eliminate it altogether. The intuition for why this substitution reduces a network’s incentive to set high termination charges stems from two sources. First, the volume of FTM termination traffic falls since all calls made to people on the same mobile network are made with a mobile phone, even when a fixed‐line phone is available. As discussed in Section 3.2, when the volume of FTM traffic falls, this will reduce the equilibrium termination charge since firms put more weight on avoiding the intense competition caused by network effects induced by high termination charges. Second, for a given termination charge, network effects are more important when substitution is possible. This is because calls made from home (those with demand function Q) also have on‐net and off‐net charge differentials now. This reinforces the first effect, since firms place more weight on the danger of intensifying competition via high termination charges. This discussion has assumed that the costs of making the two types of call are equal. It may be more realistic to assume that FTM calls are more efficient than MTM calls when callers have the choice of phone, so that C < cO. However, unless the cost differential is very large, the previous analysis remains valid. A network still has an incentive to set an above‐cost termination charge and it still wishes to under‐cut the FTM call charge by setting lower on‐net MTM call charges. That is to say, a mobile firm will encourage its subscribers to use the less efficient MTM mode of communication whenever subscribers have a choice. For instance, with the call charges in Table 4 for 2001, it is plausible that mobile subscribers were making too many on‐net calls relative to FTM calls. In other words, high mobile termination charges distort competition between FTM and on‐net MTM calls; this danger provides an additional benefit of regulating the termination charge to equal cost.36 Finally, when there is the possibility of FTM and MTM call substitution, high FTM call charges and the availability of cheap on‐net MTM calls will give people an additional reason to subscribe to a mobile network, which is to avoid high FTM prices for on‐net calls. Allowing for elastic subscriber demand, this introduces another avenue by which high mobile termination charges can induce marginal subscribers to join and therefore another positive externality of above‐cost termination charges (in addition to the externalities discussed in Section 2.3). 4. Vertical Integration The analysis to this point has assumed, in line with the current UK market, that mobile networks and fixed networks are separately owned. In this Section we briefly discuss the impact of vertical integration in the context of our model. Thus, suppose now that there are two mobile networks as before but that one of these networks (say, network 1) is integrated with the (monopoly) fixed network. If the monopoly fixed network is tightly regulated, it turns out that vertical integration has little impact on our previous analysis. Suppose that the fixed network’s termination charge for receiving calls from mobile subscribers is regulated to be equal to its cost of supplying termination, so that the fixed network obtains no profit from providing this service. Second, suppose that the fixed network’s FTM call charges are regulated so that (1) is satisfied. (In the case where the MTM and FTM termination charges can be set differently, network 1’s FTM charge A1 is merely an ‘accounting’ price which is used to determine the integrated firm’s regulated price for FTM calls to network 1.) Finally, suppose that the level of the fixed network’s FTM call charge does not affect other aspects of the fixed network’s business (its demand for its other services, its number of subscribers, or the prices it is able to charge for other services). This collection of assumptions together imply that the fixed network’s profit is unaffected by decisions made by its mobile sub‐division. As such, our previous analysis applies to this situation where one mobile network is integrated with the fixed network. For example, in the case where termination charges need not be uniform (as in Section 2), the integrated mobile network still has an incentive to set its FTM termination charge at the monopoly level, since that maximises the profit obtained from making FTM calls to that network. Thus, vertical integration has no impact when the fixed network is tightly regulated. More generally, when the integrated network is tightly constrained in what it can extract from fixed subscribers, it may prefer, where possible, to shift profit to the mobile sector where it is unregulated in the retail market. In this case, the integrated firm’s incentives are likely to be similar to a vertically separated mobile firm’s incentives as modelled in this article. On the other hand, if the fixed network is not regulated at all in the charges it can set, or only in a partial or indirect way (e.g., its retail charges are subject to an average price cap), the fixed network’s profits will be negatively related to the level of FTM termination charges it has to pay to both mobile networks. In this case, the integrated firm has an incentive to set less distorted FTM prices for calls within its network (so to extract more profit from its fixed subscribers) and also to compete more aggressively against its mobile rival (so that its fixed subscribers need to make fewer expensive calls to the rival network). These results can be established in the extreme case in which the integrated firm is an unregulated monopolist in the fixed market and only FTM calls are considered. In such a setting, the integrated firm will prefer to maximise the surplus it can extract from its fixed subscribers by pricing FTM calls at cost (perceived cost in the case the calls are off‐net) and extracting its subscribers’ surplus through the fixed charge. To achieve this, the integrated firm’s (internal) FTM termination charge is set at the efficient level so that the firm can set efficient prices for FTM calls within its combined network. In contrast, the mobile‐only network will continue to set the monopoly FTM termination charge so as to maximise the subsidy it can offer its subscribers. As a result, the integrated firm will want to compete more aggressively for mobile subscribers. This is to reduce the rival’s market share and, hence, the share of FTM calls for which its fixed subscribers pay monopoly prices. This strategy allows the integrated firm to charge a higher fixed charge to its fixed subscribers. If, instead, the rival’s termination charge was regulated to be equal to cost, then competition between the integrated firm and the mobile‐only firm would not be distorted. Nevertheless, we are not aware of examples where an integrated firm has lobbied for, or has voluntarily set, a low termination charge. Thus, it may be that situations where the fixed network is tightly regulated – where our analysis with vertical separation applies – is often the more relevant one. A full analysis of the impact of vertical integration on incentives to set termination charges, as well as the incentive to integrate in the first place, will be sensitive to the details of how the fixed network is regulated, and we leave this important issue for future research. 5. Conclusions This paper provides a unifying framework for discussing policy towards mobile call termination. Initially, in section 2 we assumed that the charges for terminating FTM and MTM calls could freely be set at different levels. In this setting, a competitive bottleneck exists for FTM termination, so FTM charges are set too high (at the monopoly level), while reciprocal MTM termination charges are set too low so as to relax retail competition between mobile networks. This model accurately reflects UK experience with respect to high FTM termination charges. However, for MTM termination, the model incorrectly predicts that authorities should be concerned that unregulated charges will be too low. Section 3 resolves this puzzle by noting that wholesale arbitrage implies that a mobile network cannot sustain a FTM termination charge significantly above its MTM termination charge. Taking this supply‐side substitution into account, we provide a new analysis in which each network sets a uniform termination charge. To avoid intensifying retail competition through network effects, mobile networks choose their uniform termination charge to be below the monopoly level (at least when market expansion is not a major factor) but above the low level that they would set if MTM termination could be priced separately. The dichotomy that existed in the previous literature is resolved, with unregulated termination charges lying between the two earlier extremes. We establish this result both when networks set their uniform charges unilaterally and when they are determined collectively. In the former case, which we think is more plausible in this setting, the equilibrium termination charge is always above cost. (In the latter case, inefficiently high termination charges arise whenever FTM calls are sufficiently important.) Section 3.3 shows that demand‐side substitution, in which mobile subscribers use their mobile phones to avoid expensive FTM calls, strengthens the constraints imposed by network effects on unregulated termination charges. Assuming a uniform termination charge, such substitution only arises with respect to on‐net, not off‐net, MTM calls. On‐net MTM calls are made by a network’s own subscribers, so by setting a low price for on‐net MTM calls, a network can insulate such callers from the effects of its choice of high termination charges. As a result, a network’s FTM termination profit becomes less important and the intensification of retail competition through network effects becomes more important, in the choice of termination charges. Thus, with both supply‐side arbitrage and demand‐side substitution, the incentive to set mobile termination charges above cost is mitigated, although not eliminated. A quite different approach to overcoming the bottleneck present with mobile termination is to try to adjust the bargaining position of fixed‐line and mobile networks so that the latter no longer hold all the bargaining power.37 Key to this (without relying on the threat of regulation) would be to relax the fixed‐line network’s obligation to interconnect with any mobile network at the latter’s chosen termination charge. For instance, if there were no such obligation, then with a single fixed network and multiple mobile networks, the natural outcome is that the fixed network holds most of the bargaining power. The mobile networks might then compete in a winner‐take‐all fashion for the right to deliver the fixed network’s FTM calls (which they will want to do since their subscribers value receiving calls), resulting in low mobile termination charges. How this might work when off‐net MTM calls are also taken into account (for instance, taking into account the supply‐side arbitrage and demand‐side substitution that we consider) remains an open question. It is worth noting, however, that for such a market‐based mechanism to replace a regulatory approach would require authorities be willing to let a breakdown in interconnection between two networks to occur. It is clear in the UK they are not willing to do so at the present time (Ofcom, 2007b, paras. 5.148–5.162). In the coming years, regulation of the termination of voice calls on mobile networks may be extended to include other services. For instance, Europe has introduced regulation of wholesale termination charges for international roaming. Similarly, Ofcom has announced that it plans to review the market for SMS termination (Ofcom, Wholesale SMS Termination Market Review, 13 September 2006).38 Some of the issues discussed in this article also apply to text messages, which are an increasingly important part of the mobile market. However, networks are better able to set SMS termination charges separately from prevailing call termination charges, and so the analysis in Section 2.2 might be more relevant and networks may choose to set low termination charges. On the other hand, networks might wish to set higher SMS termination charges due to a fear that text messages could substitute for calls, in which case some of the issues in Section 3.3 could come into play. In future work it would be worthwhile to investigate these issues in more depth. Footnotes 1 " See Farrell and Klemperer (2007) for an overview. 2 " See Armstrong (2006) and Armstrong and Wright (2007) for general analyses of competitive bottlenecks. 3 " Text messaging (or SMS) termination fits into our framework of MTM call termination, although in most countries it is not yet regulated. 4 " See Laffont et al. (1998b) and Gans and King (2001). 5 " This resolution of the puzzle fits what happened in France and New Zealand, where unregulated networks initially agreed on bill‐and‐keep (or zero charges) for terminating MTM calls while at the same time setting high FTM termination charges. At some point these very different charges became impossible to sustain, and networks moved to uniform FTM and MTM termination charges but at above‐cost levels. 6 " See Competition Commission (2003, Fig. 3.46) and Ofcom (2007b, p. 7). 7 " Oftel, Mobile Access and Call Origination Services Market: Identification and Analysis of Market and Determination on Market Power (chapter 3), 4 August 2003. 8 " Another possibility is that subscribers may sign up to multiple mobile networks. This could potentially resolve the bottleneck problem if a subscriber’s callers can choose to call him on the network with the lower termination charge. However, despite relatively high off‐net MTM and FTM pricing, so far this does not seem to have been a very significant constraint on pricing. According to Ofcom (2007b, para. 3.29), in 2006 only 7% of mobile subscribers had more than one mobile phone and this was typically for reasons other than pricing, such as to separate business and personal calls. 9 " One might also consider mobile‐to‐fixed calls. However, in the simple framework presented here, these calls play no significant role in the analysis and are ignored. (A mobile network would just set the price for such calls equal to its cost of providing such calls.) 10 " An alternative assumption is that the fixed network cannot ‘price discriminate’ in this way, perhaps because callers do not always know which mobile network they are calling. Without price discrimination, the market failures identified in the following analysis are typically amplified. For instance, a small mobile network’s termination charge has only an insignificant impact on the fixed network’s average cost of providing FTM calls, and so the network has a unilateral incentive to set an extremely high termination charge, since such a charge does not cause demand for calls from the fixed network to its subscribers to fall significantly. Indeed, FTM call charges will be above the monopoly level; see Gans and King (2000) and Wright (2002). 11 " Indeed, in the UK the FTM termination charge and the FTM retail price have declined together over the period 2001–5 (Ofcom, 2006b, Fig. 3.38) although not one‐for‐one as assumed in (1): recently Ofcom (2007b, para. 3.22) concluded about two‐thirds of the reductions to termination charges had been passed through to the FTM call charge. 12 " Throughout the article it is assumed the parameters are such that an equilibrium exists. Laffont et al. (1998b) discuss the conditions for such an equilibrium to exist in the retail pricing stage, which requires that the MTM termination charge does not differ too much from cost and/or that the mobile networks are sufficiently differentiated. More generally, for equilibrium to exist at all stages of the game we require that networks be sufficiently differentiated so that t is not too small. 13 " This result holds much more generally. It holds allowing for asymmetric market shares and allowing for more than two firms. One case where the result does not apply is when subscribers care about receiving calls. Berger (2005) considers this and shows that the on‐net call charge is adjusted downwards to reflect the call externality subscribers enjoy from being called more often from others on the same network. Conversely, the off‐net call charge is adjusted upwards to reduce the number of calls subscribers on rival networks receive, thereby reducing the rival’s ability to compete. 14 " In a recent survey of 621 subscribers, Ofcom (2006c) reports that 12% knew at least one person who had subscribed to a mobile network for the first time in the previous year, and that on average they made 10 calls and received 9 calls per month from these new subscribers. 15 " The finding that with scope for market expansion the welfare‐maximising FTM termination charge is above cost but below the unregulated level has been noted in the existing literature (Armstrong, 2002; Wright, 2002, and Valletti and Houpis, 2005), but previous models have unrealistically assumed away MTM calls. 16 " During the 2002 Competition Commission enquiry, Vodafone stated that the unregulated charge was in range 17 pence per minute (ppm) to 20 ppm. At this time, the cap on O2 and Vodafone’s termination charge was 9.3 ppm, so Vodafone was suggesting that the FTM termination charge would roughly double without the charge control (Competition Commission, 2003, paras. 2.440–2.445). Ofcom (2007b, para. 7.49) estimated more recently that the unregulated FTM termination charge would be about 24 ppm, compared with its current regulated charge of about 5 ppm, and so the regulator believed there would be a five‐fold increase if regulation were abandoned. 17 " See Competition Commission (2003, paras. 2.390–2.400). 18 " Ofcom (2007b, para. A19.37) estimates that consumers will be better off by £ 3.2 billion in present value terms as a result of regulation in the period 2007–11 compared to the unregulated alternative. (Ofcom’s analysis assumed that competition was sufficiently strong so that excess profits were eliminated, and so ‘consumer’ and ‘total welfare’ standards coincide.) 19 " One reason for this regulatory approach was the fact that mobile telephone numbers were in the same numbering range as fixed network numbers, posing the obvious problem that callers could not tell they were calling a potentially much more expensive mobile number. 20 " In a recent response to proposed EU regulation to cut mobile termination rates drastically, Vodafone claimed that around 40 million European mobile subscribers would cancel their mobile service if the policy were to be implemented. See Vodafone, Comments on the Draft Commission Recommendation on the regulatory treatment of fixed and mobile termination rates in Europe, 1 September 2008, para. 1.13. Note also that the United States has significantly lower termination charges than in Europe and also has significantly lower mobile penetration than in Europe. 21 " See Farrell and Klemperer (2007) for an overview. 22 " Genakos and Valletti (2007) estimate the size of the waterbed effect empirically and conclude (p. 2): ‘although the waterbed is shown to be high, our analysis also provides evidence that it is not full: accounting measures of profits are positively related to [termination rates]. Mobile firms tend to keep part of termination rents instead of passing them on to their customers, and thus suffer from cuts in termination rates.’ 23 " Ofcom estimated that the ‘externality surcharge’ needed to reflect the benefits of market expansion was 0.30 pence per minute, or around 5% of their estimated costs of call termination (Ofcom, 2007b, p. 349). Their chosen methodology was to assume that mobile networks could, at least in part, target a subsidy to marginal subscribers. It is not clear exactly how this could be done in practice, even if mobile networks wanted to do so. Our analysis assumes that the subsidy is applied equally to all mobile subscribers, which means a higher level of subsidy will be needed, all else equal. 24 " In the UK, in the period 1998 to 2002 the two leading mobile networks’ MTM termination charge was not regulated, and in the period since 2003 to the time of writing, mobile networks’ MTM termination charges were capped (at the same level as the FTM termination charges) but networks were free to set MTM charges below this cap. At no time have networks chosen to set their MTM and FTM termination charges at different levels. 25 " Ofcom (2006a, Fig. 3.36) shows that approximately two‐thirds of UK mobile subscribers use pre‐paid service. However, many pre‐paid tariffs still incorporate nonlinear pricing of various kinds, and many involve off‐net and on‐net price differentials. In addition, subscribers on contracts generate about four times more revenue on average than pre‐pay subscribers – see Ofcom (2007a, Fig. 4.39). 26 " Calzada and Valletti (2008) demonstrate that incumbent networks can sometimes use high MTM termination charges to deter entry. A related literature examines whether a smaller or entrant network should be allowed to set a higher MTM termination charge than its larger rivals to help offset its disadvantage in this respect; see Peitz (2005). 27 " Hoernig (2007) analyses the impact of on‐net and off‐net price differentials on the profitability of small networks in the presence of call externalities. He shows that larger firms will choose greater differentials than smaller firms. 28 " In our discussions with industry representatives in various countries, this was the explanation given for why mobile operators set a uniform termination charge for both types of calls. 29 " With FTM and MTM termination charges tied together, if networks jointly choose a MTM termination charge they will also effectively be choosing their FTM termination charge jointly. This could raise price‐fixing concerns. 30 " The situation in which MTM termination charges are chosen unilaterally but not necessarily at the same level as FTM charges, is a special case of this framework in which FTM demand is set to zero. This situation was previously examined by Gans and King (2001), who assumed away FTM calls. However, without the requirement of uniformly set termination charges, it is not clear why networks would set their charges unilaterally. For instance, in the examples we know of where firms have attempted to set different charges for the two kinds of termination (France and New Zealand), firms have set a low MTM charge, which could not be an equilibrium with unilaterally‐chosen charges. Moreover, even if for some other reason networks set their charges unilaterally, this does not explain why in practice networks set FTM and MTM termination charges at the same level. 31 " Setting Q(C + cT) = 0 in (28) gives Gans and King (2001, Proposition 1). 32 " One can check that the right‐hand side is increasing in a over this range provided that M(·) is concave in a. The left‐hand side is decreasing in a over this range if the stronger condition M ′′ < q ′ holds. (A sufficient condition for this second inequality to hold is that q(·) be weakly concave.) 33 " For similar reasons, if MTM termination charges are set unilaterally but are not required to be set at the same level as FTM charges, they will also be set lower than in the uniform case (although from (28) still above the efficient level). For instance, with and λ = 0, we find the unilateral choice of a separate MTM termination charge is a = 0.199 compared to a = 0.230 for the uniform case. 34 " An alternative kind of substitution between fixed and mobile networks may take place at the subscription level rather than the per‐call level, so that some people might give up their fixed line altogether and become mobile‐only users. We do not consider this possibility here (see Hansen (2006) for an interesting analysis). 35 " It can be checked that firm i will not wish to set ai < cT. Such a policy would induce subscribers to replace on‐net calls with FTM calls whenever possible, for which firm i will incur a termination loss. Moreover, even ignoring FTM calls, the analysis in Section 3.2 shows that networks will always wish to set an above‐cost termination charge when it is set unilaterally. 36 " In the 2002 enquiry (Competition Commission, 2003, paras. 2.408 and 2.424), BT suggested that high FTM termination charges could act to distort competition in this way. 37 " Binmore and Harbord (2005) consider the bargaining process between a large fixed network and a small mobile entrant. 38 " SMS termination charges are already controlled in France and Israel, for instance. 39 " The mobile‐to‐mobile part of the following analysis is somewhat related to section 5 of Dessein (2003), although he does not allow for off‐net/on‐net call charge differentials. Like us, he shows that the unregulated MTM termination charge is below cost and the efficient MTM termination charge is above cost. 40 " Assume also that the mobile‐to‐fixed termination charge is regulated to be equal to the fixed network’s cost. Otherwise, the extra profits on the fixed network caused by terminating more traffic when the mobile market expands would need to be considered too. References Armstrong , M. ( 1998 ). ‘Network interconnection in telecommunications’ , Economic Journal , vol. 108 ( 448 ), pp. 545 – 64 . Google Scholar Crossref Search ADS WorldCat Armstrong , M. ( 2002 ). ‘The theory of access pricing and interconnection’, in ( M. Cave, S. Majumdar and I. Vogelsang, eds). Handbook of Telecommunications Economics , Vol. 1 , pp. 295 – 384 . Amsterdam: North‐Holland . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Armstrong , M. ( 2006 ). ‘Competition in two‐sided markets’ , Rand Journal of Economics , vol. 37 ( 3 ), pp. 668 – 91 . Google Scholar Crossref Search ADS WorldCat Armstrong , M. , and Wright , J. ( 2007 ). ‘Two‐sided markets, competitive bottlenecks and exclusive contracts’ , Economic Theory , vol. 32 ( 2 ), pp. 353 – 80 . Google Scholar Crossref Search ADS WorldCat Berger , U. ( 2004 ). ‘Access charges in the presence of call externalities’ , Contributions to Economic Analysis and Policy , vol. 3 (1), article 21. OpenURL Placeholder Text WorldCat Berger , U. ( 2005 ). ‘Bill‐and‐keep vs. cost‐based access pricing revisited’ , Economics Letters , vol. 86 ( 1 ), pp. 107 – 12 . Google Scholar Crossref Search ADS WorldCat Binmore , K. , and Harbord , D. ( 2005 ). ‘Bargaining over fixed‐to‐mobile termination rates: countervailing buyer power as a constraint on monopoly power’ , Journal of Competition Law and Economics , vol. 1 ( 3 ), pp. 449 – 72 . Google Scholar Crossref Search ADS WorldCat Calzada , J. , and Valletti , T. ( 2008 ). ‘Network competition and entry deterrence’ , Economic Journal , vol. 118 ( 531 ), pp. 1223 – 44 . Google Scholar Crossref Search ADS WorldCat Carter , M. , and Wright , J. ( 1999 ). ‘Interconnection in network industries’ , Review of Industrial Organization , vol. 14 ( 1 ), pp. 1 – 25 . Google Scholar Crossref Search ADS WorldCat Cherdron , M. ( 2002 ). ‘Interconnection, termination‐based price discrimination, and network competition in a mature telecommunications market’ , mimeo , Mannheim University. Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Competition Commission ( 2003 ). Vodafone, O2, Orange and T‐Mobile: Reports on References Under Section 13 of Telecommunications Act 1984 on Charges Made by Vodafone, Orange, O2 and T‐Mobile for Terminating Calls Made by Fixed and Mobile Networks . London: HMSO . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Dessein , W. ( 2003 ). ‘Network competition in nonlinear pricing’ , Rand Journal of Economics , vol. 34 ( 4 ), pp. 593 – 611 . Google Scholar Crossref Search ADS WorldCat Farrell , J. , and Klemperer , P. ( 2007 ). ‘Coordination and lock‐in: competition with switching costs and network effects’, in ( M. Armstrong and R. Porter eds). Handbook of Industrial Organization , Vol. 3 , pp. 1967 – 2072 . Amsterdam: North‐Holland . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Gabrielsen , T. S. , and Vagstad , S. ( 2008 ). ‘Why is on‐net traffic cheaper than off‐net traffic? Access markup as a collusive device’ , European Economic Review , vol. 52 ( 1 ), pp. 99 – 115 . Google Scholar Crossref Search ADS WorldCat Gans , J. , and King , S. ( 2000 ). ‘Mobile network competition, customer ignorance and fixed‐to‐mobile call prices’ , Information Economics and Policy , vol. 12 ( 4 ), pp. 301 – 28 . Google Scholar Crossref Search ADS WorldCat Gans , J. , and King , S. ( 2001 ). ‘Using ‘bill and keep’ interconnect agreements to soften network competition’ , Economics Letters , vol. 71 ( 3 ), pp. 413 – 20 . Google Scholar Crossref Search ADS WorldCat Genakos , C. , and Valletti , T. ( 2007 ). ‘Testing the ‘waterbed’ effect in mobile telephony’ , CEP Discussion Paper No. 827 . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Hansen , B. ( 2006 ). ‘Termination rates and fixed mobile substitution’ , mimeo, Norwegian School of Management . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Höffler , F. ( 2006 ). ‘Mobile termination and collusion, revisited’ , mimeo, Working Paper 2006/16, Max Planck Institute for Research on Collective Goods Bonn . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Hoernig , S. ( 2007 ). ‘On‐net and off‐net pricing on asymmetric telecommunications networks’ , Information Economics and Policy , vol. 19 ( 2 ), pp. 171 – 88 . Google Scholar Crossref Search ADS WorldCat Laffont , J.‐J. , Rey , P. and Tirole , J. ( 1998a ). ‘Network competition: I. Overview and nondiscriminatory pricing’ , Rand Journal of Economics , vol. 29 ( 1 ), pp. 1 – 37 . Google Scholar Crossref Search ADS WorldCat Laffont , J.‐J. , Rey , P. and Tirole , J. ( 1998b ). ‘Network competition: II. Price discrimination’ , Rand Journal of Economics , vol. 29 ( 1 ), pp. 38 – 56 . Google Scholar Crossref Search ADS WorldCat MMC ( 1999 ). Cellnet and Vodafone: A Report on a Reference Under Section 13 of Telecommunications Act 1984 on the Charges Made by Cellnet and Vodafone for Terminating Calls from Fixed‐Line Networks , HMSO : London. Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC OFCOM ( 2006a ). The Communications Market 2006 , London: Office of Communications . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC OFCOM ( 2006b ). Mobile Call Termination: Market Review , London: Office of Communications . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC OFCOM ( 2006c ). Mobile Call Termination: Report of Market Research Findings , London: Office of Communications . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC OFCOM ( 2007a ). The Communications Market 2007 , London: Office of Communications . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC OFCOM ( 2007b ). Mobile Call Termination: Statement , London: Office of Communications . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Peitz , M. ( 2005 ). ‘Asymmetric access price regulation in telecommunications markets’ , European Economic Review , vol. 49 ( 2 ), pp. 341 – 58 . Google Scholar Crossref Search ADS WorldCat Valletti , T. , and Houpis , G. ( 2005 ). ‘Mobile termination: what is the ‘right’ charge?’ , Journal of Regulatory Economics , vol. 28 ( 3 ), pp. 235 – 58 . Google Scholar Crossref Search ADS WorldCat Wright , J. ( 2002 ). ‘Access pricing under competition: an application to cellular networks’ , Journal of Industrial Economics , vol. 50 ( 3 ), pp. 289 – 316 . Google Scholar Crossref Search ADS WorldCat Appendix Appendix: Elastic Subscriber Participation A. Non‐uniform FTM and MTM Termination Charges The model is as described in section 2.3.39 Combining (15) with (17) shows that (33) where as before. (From (18), the denominators in (33) are positive so long as a is not too far away from cost cT.) Write (34) where for the own and cross‐price effects on subscriber numbers with respect to changes in the rental charge. Here γ is positive, μ is smaller than γ, and μ is positive when a is close to cT and (18) holds. From (16) the equilibrium rental charge is given by (35) where Substituting the value for r in (35) into the formula for profit in (16) shows that industry profit is (36) while substituting the value for r in (35) into expression (33) shows that the market size given a and A is (37) (From (18), the denominator in (37) is positive for a close to cT.) Since N in (37) is increasing with F, it follows from (36) that mobile profits are higher with a higher FTM termination profit, F(A), and the waterbed effect present in Section 2.1 is no longer complete. The impact of the MTM termination charge on profit is more complex, since a affects the terms γ and μ in Λ. For simplicity, consider the impact of a small change in a away from marginal cost cT . One can show that where the inequality follows from (18). From (37) we see that where , and so raising the MTM termination charge above cost induces network expansion. From (36) we have where the inequality follows (after some manipulation) from (18). Therefore, mobile networks wish to set a below‐cost MTM termination charge in order to relax competition for subscribers, just as in the benchmark model without market expansion. What are the socially efficient termination charges in this framework? When subscriber utility is u1 = u2 = u, aggregate consumer surplus of mobile subscribers is Φ(u) = λ u2 + u, where . (So Φ′(u) ≡ N.) In equilibrium, r is given by (35). The consumer surplus of fixed‐line subscribers from calls to mobile subscribers is , where is the consumer surplus on the fixed network for each mobile subscriber. Assume that FTM calls are charged at cost, so that expression (1) holds.40 Then total welfare is obtained by summing mobile sector profit, mobile subscriber surplus and fixed caller surplus, which is (38) When there are no market expansion possibilities (i.e., λ = 0), it is optimal to set both the FTM and MTM termination charges equal to cost cT as in our benchmark model. When λ > 0 it is socially optimal to set both termination charges above cost. To see this for the FTM termination charge, differentiate (38) with respect to A to obtain (39) where and NA > 0 is the derivative of N in (37) with respect to A when A = a = cT. Thus, raising the FTM termination charge above cost induces network expansion, and this benefits both fixed‐line callers and existing mobile subscribers (and also the mobile networks). To see the impact of raising the MTM termination charge, differentiate (38) to obtain: which is the same form as (39). Thus, setting the MTM termination charge above cost benefits both mobile and fixed‐line subscribers and overall welfare, but it harms the mobile networks. B. Jointly‐chosen Uniform Termination Charge Now suppose the FTM and MTM termination charges must be equal and firms coordinate on the choice of this uniform charge, denoted a. Then the analysis of Appendix A continues to apply except that differentiating (37) with respect to the uniform charge a now yields and The sign of this derivative is ambiguous. If Q is sufficiently large it is positive and firms will agree to set a uniform termination charge above cost. C. Unilateral Choice of Uniform Termination Charge Finally, suppose that the two networks choose their uniform termination charge unilaterally. Firm i’s profit with charge ai is modified from (16) to be Firm i has ni subscribers, where this is modified from (33) to be where Then the own‐price and cross‐price effects on subscriber numbers with respect to changes in firm i’s rental charge are The rental charge ri satisfies and so (40) where Expression (40) gives firm i’s profit as a (complicated) function of the pair of termination charges a1 and a2, and this can be differentiated with respect to ai to find the symmetric equilibrium charge when the charges are chosen unilaterally. Author notes " We are grateful to Stefan Behringer, David Harbord, Bruce Lyons, Paul Muysert, Geoffrey Myers, David Sappington, Tommaso Valletti, John Vickers, Ingo Vogelsang, Helen Weeds and to two referees for very helpful comments and corrections. Armstrong is grateful for the support of the Economic and Social Research Council (UK) and Wright gratefully acknowledges the support of the Singapore Ministry of Education AcRF Tier 1 fund under Grant No. R122000080‐101/112/133. © The Author(s). Journal compilation © Royal Economic Society 2009

Andersson,, Fredrik;Freedman,, Matthew;Haltiwanger,, John;Lane,, Julia;Shaw,, Kathryn

doi: 10.1111/j.1468-0297.2009.02277.xpmid: N/A

Abstract Innovative firms need to hire and motivate highly talented workers. This article connects the potential returns to innovation to the structure of compensation for skilled employees. We show that the software firms that operate in software sectors with high potential upside gains to innovation pay more to ‘star’ workers than do other firms that operate in stable markets. Firms operating in product domains with highly skewed positive returns pay employees more in up‐front starting salaries and offer higher compensation growth. The large estimated effects on earnings are robust to the inclusion of a wide range of controls for worker and firm characteristics. Over the past several decades, the economy has witnessed two pronounced changes. There has been a well‐known rise in the rate of growth of labour productivity, a large share of which is due directly to the rise of the highly productive information technology sector (Jorgenson et al., 2005; Oliner and Sichel, 2000). There has also been an increase in income inequality due to rising incomes at the top end of the wage distribution, which has been termed the ‘polarisation’ of earnings (Autor et al., 2007). In this article, we seek to advance our understanding of these phenomena by taking advantage of new micro‐data on worker compensation and firm product market strategies in the US software industry. The software industry is particularly appropriate for a study of how the characteristics of product markets affect compensation policies. First, product innovation in the software industry is very closely tied to the talents of the workforce. Second, the software industry is characterised by highly skewed returns: successful innovations can produce enormous payoffs to the firm, while failed products can lead to large losses. It is also characterised by a similarly skewed compensation structure. Third, the variance of product payoffs is different in different sectors of the industry. For example, the video game sector of the software industry is characterised by very high stakes product development; some games generate hundreds of millions of dollars in revenues while many languish on store shelves and make much less. Other software sectors, such as business applications software, have substantially lower potential gains to innovation. The theory that firms in the most innovative sectors pay more for talent is backed up by empirical results. Using data on software firms and their workers, we segment firms into 30 different software sectors that vary substantially in the skewness of their average revenues per worker. Matching each individual worker’s pay to his or her firm’s software sector, we show that workers who are employed by firms operating in software sectors with greater potential upside revenue gains have substantially higher earnings. Indeed, we find that firms with highly skewed potential payoffs especially reward more experienced workers. Our findings suggest that both bonuses and stock options are important components of the reward for more experienced workers. We also show that more productive firms pay higher compensation. These results come from cross‐sectional data but, because the software industry has grown dramatically over time, one might infer that there are also time series implications from our cross‐sectional results. At the heart of this empirical work is a fundamental link between the firm’s product market strategy and its choice of human resource management practices: software firms that choose to operate in sectors that have high risk payoffs will choose human resource practices that help them attract and retain higher quality workers and pay more for performance. Few empirical papers are able to make such a link between strategy and human resource practices. Past empirical research connecting a firm’s compensation policy to its product market strategy has been stymied by data limitations, since data are needed on both firm compensation strategies and the revenue payoffs of the firm. Some ‘insider’ studies that have been conducted have focused on CEO pay, which is more readily available, and have generally found evidence of a link between strategy and pay within industries (Baker and Hubbard, 2001; Chevalier and Ellison, 1999; Stern, 2004; Fallick et al., 2006; Lerner and Wulf, 2007; Wulf, 2002, 2007; Garicano and Hubbard, 2007). Some work using survey data has uncovered a similar connection (MacLeod and Parent, 1999). Up to now, though, empirical studies have yet to establish a link between product market strategy and human resource practices using data covering more than a small number of firms or a select group of employees. The foundation of our study is our use of rich new longitudinal employer–employee matched data that track both the universe of software firms as well as the universe of workers within those firms in 10 US states. We have data for these firms from the Economic Census, which contains detailed data on firm revenues for many software product classes. Using these data, we can calculate the differential revenue payoff distributions for different types of software products, which permit us to measure both potential payoffs and actual performance among firms by software product class. We also have data on earnings for all workers in these firms, where the earnings include exercised stock options and bonuses, and can calculate within‐job and between‐job earnings growth across different firms from 1992 to 2001. The article proceeds as follows. In the next Section, we provide some background facts about the software industry to help motivate our analysis. We describe in Section 2 the theory underlying the connection between product market risk and a firm’s pay policy, and we provide a detailed description of the data we use to test the predictions of the model in Section 3. In Sections 4 and 5, we present our empirical specifications and the results from these specifications. We conclude and discuss the implications of our work in Section 6. 1. Background A number of facts motivate the approach and analysis that follows. A first key point is that software workers are not only highly paid relative to other workers but the distribution of their earnings is also much more skewed. Table 1(a) documents this point using summary statistics about the distribution of income from the 2000 Decennial Census Public‐Use Microdata Sample (PUMS) for workers ages 21–44 in all industries as well as for workers specifically in the prepackaged software industry (SIC 7372). Indeed, not only are average earnings twice as high in the software industry than all industries but the variance, as measured by the standard deviation, is almost twice as high as well.1 Table 1 
Summary Earnings Statistics, Workers 21–44 . Mean . Median* . 90th* . SD . (a) 2000 Decennial Census (PUMS) Data – Total Earnings 35+ Hours/Week & 35+ Weeks/Year All Industries 40,918 31,891 70,160 183,134 Software Industry (SIC 7372) 80,787 63,782 127,563 334,906 Computer Software Engineers (Census Occupation Code 102) in the Software Industry 90,668 70,691 138,193 369,374 (b) LEHD Data for Ten States – Earning $50,000+ Annualised All workers in Software Industry Starting Earnings (Excludes Left‐Censored) 69,353 59,665 108,692 82,432 Worker Earnings at end of job spell (Censored and Uncensored) 344,268 95,508 310,644 2,051,985 Top Decile of Workers in Software Industry Starting Earnings (Excludes Left‐Censored) 107,660 80,899 184,951 142,526 Worker Earnings at end of job spell (Censored and Uncensored) 2,532,500 670,993 6,688,470 6,064,204 . Mean . Median* . 90th* . SD . (a) 2000 Decennial Census (PUMS) Data – Total Earnings 35+ Hours/Week & 35+ Weeks/Year All Industries 40,918 31,891 70,160 183,134 Software Industry (SIC 7372) 80,787 63,782 127,563 334,906 Computer Software Engineers (Census Occupation Code 102) in the Software Industry 90,668 70,691 138,193 369,374 (b) LEHD Data for Ten States – Earning $50,000+ Annualised All workers in Software Industry Starting Earnings (Excludes Left‐Censored) 69,353 59,665 108,692 82,432 Worker Earnings at end of job spell (Censored and Uncensored) 344,268 95,508 310,644 2,051,985 Top Decile of Workers in Software Industry Starting Earnings (Excludes Left‐Censored) 107,660 80,899 184,951 142,526 Worker Earnings at end of job spell (Censored and Uncensored) 2,532,500 670,993 6,688,470 6,064,204 *Average within a 10% band around the true percentile. **Annualised earnings three quarters prior to last observed full quarter. ***Includes only individuals for whom we observe a prior spell in the data. Open in new tab Table 1 
Summary Earnings Statistics, Workers 21–44 . Mean . Median* . 90th* . SD . (a) 2000 Decennial Census (PUMS) Data – Total Earnings 35+ Hours/Week & 35+ Weeks/Year All Industries 40,918 31,891 70,160 183,134 Software Industry (SIC 7372) 80,787 63,782 127,563 334,906 Computer Software Engineers (Census Occupation Code 102) in the Software Industry 90,668 70,691 138,193 369,374 (b) LEHD Data for Ten States – Earning $50,000+ Annualised All workers in Software Industry Starting Earnings (Excludes Left‐Censored) 69,353 59,665 108,692 82,432 Worker Earnings at end of job spell (Censored and Uncensored) 344,268 95,508 310,644 2,051,985 Top Decile of Workers in Software Industry Starting Earnings (Excludes Left‐Censored) 107,660 80,899 184,951 142,526 Worker Earnings at end of job spell (Censored and Uncensored) 2,532,500 670,993 6,688,470 6,064,204 . Mean . Median* . 90th* . SD . (a) 2000 Decennial Census (PUMS) Data – Total Earnings 35+ Hours/Week & 35+ Weeks/Year All Industries 40,918 31,891 70,160 183,134 Software Industry (SIC 7372) 80,787 63,782 127,563 334,906 Computer Software Engineers (Census Occupation Code 102) in the Software Industry 90,668 70,691 138,193 369,374 (b) LEHD Data for Ten States – Earning $50,000+ Annualised All workers in Software Industry Starting Earnings (Excludes Left‐Censored) 69,353 59,665 108,692 82,432 Worker Earnings at end of job spell (Censored and Uncensored) 344,268 95,508 310,644 2,051,985 Top Decile of Workers in Software Industry Starting Earnings (Excludes Left‐Censored) 107,660 80,899 184,951 142,526 Worker Earnings at end of job spell (Censored and Uncensored) 2,532,500 670,993 6,688,470 6,064,204 *Average within a 10% band around the true percentile. **Annualised earnings three quarters prior to last observed full quarter. ***Includes only individuals for whom we observe a prior spell in the data. Open in new tab Of course, the PUMS data have several drawbacks. They do not include the performance bonuses and stock options so important in the software industry and are top‐coded. In addition, the lack of employer‐level information means that the link between the variation in product revenue and compensation levels cannot be made; the lack of longitudinal data similarly renders impossible an analysis of the link between the variation in product revenue and compensation changes over time. The use of new longitudinal employer–employee matched data enables us to overcome these problems and the summary statistics regarding the distribution of pay using these data are presented in Table 1(b).2 We once again see that the earnings distribution is highly skewed so that a small subset of workers in the software industry receives extraordinarily high compensation. However, we get a more nuanced view with the longtitudinal data, which show that earnings are not highly skewed at the start of the job (starting earnings) but that they become much more highly skewed by the end of a job spell with an employer (experienced earnings).3 These numbers are depicted graphically in Figure 1. While 70% of starting earnings are below $75,000, only 29% of experienced workers earn below $75,000 (experienced workers have an average tenure of five years). Similarly, only 4% of starting earnings are above $150,000 but 21% of experienced workers earn above that amount. Since starting earnings include the earnings paid to workers new to the firm but possibly experienced within the software industry, it is clear that compensation rises markedly with tenure. Fig. 1. Open in new tabDownload slide Distribution of Starting Earnings and Experienced Earnings
Notes. Based on LEHD data ten states. Includes workers 21–44 earning $50,000 or more in SIC 7372. Starting earnings excludes left‐censored job spells. Fig. 1. Open in new tabDownload slide Distribution of Starting Earnings and Experienced Earnings
Notes. Based on LEHD data ten states. Includes workers 21–44 earning $50,000 or more in SIC 7372. Starting earnings excludes left‐censored job spells. In the bottom of Table 1(b), we look at the subsample of the most highly paid software workers – those experienced workers who are in the top 10% of the earnings distribution. The skewness of their earnings is especially pronounced. A second key point for the software industry is that some sectors within the software industry have a very skewed revenue distribution across firms. In the video games sector, for example, the top selling video game in 2007, Halo, accounted for over a quarter of all video game sales.4 In our data below, we show that other software sectors, such as enterprise resource software for large mainframe computers, have much less skewed returns. The following Section sketches a theory that links the rightward skewness of firms’ revenue distributions to firms’ compensation policies. 2. Conceptual Framework The production function for the software industry involves the use of highly skilled workers to produce innovations in software products. In formalising the process of producing innovative software products, we are guided by what we learned from company visits to fifteen software firms and seven medical device firms, each of which focus on innovation. We also appeal to many books and case studies that describe the innovations and human resource practices at firms such Microsoft, Siebel Systems, HP, Cisco, IBM, the SAS Institute, Electronic Arts, Google and YouTube.5 The fundamental characteristic of software production is the uncertainty that arises because of firms’ inability to predict whether an innovative product will pay off.6 In software innovation, there are two key groups of employees, each of which makes decisions about undertaking risky software projects. On the one hand, programmers and engineers must begin working on a new software project without knowing whether they will develop a great product. On the other hand, managers must allocate funds to research projects without knowing whether the resulting products will succeed in the market. Any theory of project selection, therefore, should pertain to both software engineers and managers. Define a ‘star’ worker as someone who is successful because he has a higher probability of producing good projects and a lower probability of producing bad projects. This ability could stem from innate talent, could be developed on the job through learning, or could arise from higher effort in response to incentives. In any event, star programmers must develop great projects and star managers must allocate resources to those projects. Both engineering and management skills are important determinants of project success in the software industry. Following Lazear (2005), our model predicts that firms in high variance payoff markets value star talent the most.7Figure 2 illustrates this by describing products with a continuous distribution of payoffs, where the ‘payoff’ is what the project is expected to pay. In the data below, the payoff will refer to the revenue stream from the project. There is a variance to project payoffs because workers can make mistakes in picking projects: they can make a false positive error, in which case they pursue a project that actually turns out to be a failure; they can also make a false negative error, in which case they reject a project that would have turned out to be a success. The bold line in Figure 2(a) shows a high variance/more skewed payoff distribution associated with projects from one possible product line, while the bold line in Figure 2(b) shows a low variance/less skewed payoff distribution associated with another product line. The dotted lines in (a) and (b) are the shift in the distributions attributable to star talent because star workers are better than average workers at picking projects. Hiring a star would shift the left tail to the right because employing such workers reduces the occurrence of false positives. It would also shift the right tail to the right because stars reduce the number of false negatives. The effect of this rightward shift of the payoff distribution is to increase the mean payoff from PA1 to PA2 in Figure 2(a). This increase represents the gain associated with paying for a star worker, which for firms that face more skewed potential payoff distributions as in (a), are likely to exceed the cost of hiring and compensating that employee. Figure 2(b) depicts a narrower and less skewed underlying project payoff distribution, representing the situation that would occur with less risky projects: those that have both smaller potential gains and losses. When a firm in this kind of product market acquires a star, this low‐risk payoff distribution also shifts to the right. The increase in the expected mean payoff is a smaller increase from PB1 to PB2. Fig. 2. Open in new tabDownload slide Shifts in the Payoff Distribution Due to Reductions in False Positive or False Negative Errors Fig. 2. Open in new tabDownload slide Shifts in the Payoff Distribution Due to Reductions in False Positive or False Negative Errors Therefore, as is evident in the Figures, the upside revenue gains to hiring stars are smaller in low‐risk, low‐skewed product markets than in high risk, high skewed product markets, as (PB2 − PB1) < (PA2 − PA1). Both the higher variance and the higher skewness in Figure 2(a) contribute to this difference. In sum, because there are larger gains (or smaller losses) to the selection of great projects in high variance, high skewed product markets, stars are more valuable in 2(a), where expected payoffs are higher, than in 2(b), where expected payoffs are lower. Primary Hypothesis. Firms operating in software sectors that exhibit higher variance and greater right skewness in potential payoffs should pay higher compensation to attract and reward stars. This higher compensation would reflect the sorting of star workers to firms that place the highest value on their skills at innovating. Very simply, stars are more valuable when there are big ‘upside gains’ to innovations. A closely related corollary hypothesis is that the positive effect of high variance and skewness on compensation should be present especially (or even only) for the most talented workers that are the project managers and/or key programmers who are critical for the decisions made on the project. To test this hypothesis below, we develop empirical proxies for the potential upside gains for various software sectors. Some sectors, such as the video game market, may have significantly greater potential upside gains than do other sectors, such as the mainframe software sector. The means by which firms in markets with potential payoffs that exhibit higher variance and greater right skewness attract, retain and motivate stars is an open question. We have presented a static theory: firms that value project selection more will value talent more and pay more for it. We know that the software industry constitutes a much more dynamic environment; for example, bonuses and stock options represent a large share of pay when firms want to retain and motivate talent. More generally, firms could devote a lot of resources to selecting star workers carefully, or alternatively they could allocate more resources to training workers on the job and providing strong incentives that reward (and sort) star workers over time as they gain experience within the firm. Our basic theory is silent about the optimal mechanisms for obtaining talent. Our empirical results will permit these mechanisms to vary. The mechanism can be inferred, to some degree, by using different measures of pay to test the primary hypothesis. Do firms in software sectors with highly skewed payoffs pay higher starting salaries, or do they pay higher salaries for experienced workers? Do firms in software sectors with highly skewed payoffs pay a higher salary or pay higher bonuses for performance? Before proceeding to the empirical analysis, it is useful to discuss briefly an obvious alternative model for the determinants of the skewed distribution of earnings in the software industry, and in particular in the gaming software product industry – namely, the ‘economics of superstars’ arguments of Rosen (1981) and the associated implications of ‘winner take all’ effects of Frank and Cook (1995). Viewed from this perspective, the skewed payoff distribution in the software gaming industry can yield associated skewed earnings differences as the returns to talent are convex in such industries. We readily accept that some fraction of the skewed earnings in the software industry reflects such effects. However, we note that the implication from this model is that high payoff firms should have high pay for the talented stars that generated these high payoffs. In what follows, we control for the actual payoffs of firms and indeed confirm that more successful firms have high earnings.8 However, even when controlling for actual payoffs, we find an important effect of the skewness and dispersion of potential payoffs that are available in the product class (or sector) for the firm. It is the latter effect that is the focus of our analysis. 3. Data In order to study the connection between the structure of firms’ product market strategies and skill demand, we require a dataset with detailed information on the earnings and employment histories of workers as well as on the product market characteristics of the firms at which these workers are employed. We take advantage of a unique employer–employee matched data set constructed and maintained by the US Census Bureau’s LEHD Programme. We further augment the LEHD data with highly detailed firm characteristics from the Economic Census and worker characteristics from the 2000 Decennial Census PUMS. 3.1. The Software Industry We test the hypotheses of our model by focusing on the prepackaged software industry, which corresponds to the four‐digit SIC 7372.9 This narrow focus has a number of key advantages. The first is that some software sectors, such as video games, have very high variance revenues, while other software sectors, such as business enterprise software, have low variance payoffs. As we show below, across software sectors, the upside gains to innovation clearly vary. The second advantage is that software firms are comprised of workers directly involved in the development and sales of R&D intensive innovations. By contrast, many traditional industries, such as automobile manufacturing, have only small numbers of workers in R&D intensive areas of the firm. Thus, in studying software firms, we are studying innovation and the knowledge workers who do it. A final advantage of studying software is the richness of the available data. The Census Bureau collects detailed product line information (described below) as well as information on the size and age of firms.10 However, while most of our data are longitudinal in nature, we have product line information only for 1997; these data are gathered only every five years in the Economic Census. 3.2. The LEHD Data The LEHD’s longitudinal wage database contains quarterly records of the employment and total earnings of individuals from Unemployment Insurance (UI) data, which are in turn matched to internal administrative records and surveys containing workers’ date of birth, race and sex.11 We have complete UI records for ten states (California, Florida, Illinois, Maryland, North Carolina, New Jersey, Pennsylvania, Texas, Washington and Wisconsin) for approximately the years 1992 to 2001 (the precise years vary slightly by state). This length of time enables us to construct sufficiently long worker employment and earnings histories to address our research questions. These data have several important advantages. First, since the scope of the LEHD data is nearly the full universe of employers and workers, we can accurately track the movements of workers through the earnings distribution within firms as well as across firms over time.12 Second, in contrast to survey‐based information, the earnings data represent the earnings that firms actually pay workers as opposed to workers’ memories of their earnings. A third benefit of using these administrative data, particularly in the context of this study and the time frame we consider, is that the earnings measures include bonuses and exercised stock options (though not fringe benefits).13 Stock options can be valued in a variety of ways; in this case, the options are valued when they are exercised, or when the employee cashes in the options. We do not have data on when options are granted to employees. However, our sense is that exercised options are the preferred measure of pay for our analysis since the value of options is quite uncertain when hired; the value depends not only on whether an employee stays with the firm until the options are vested but also on the growth of the stock price of the company.14 We match panel data on workers to their firms’ 1997 Economic Census record, so we focus on software workers’ employment spells that span 1997, beginning or ending between 1992 and 2001. Our primary results are based on two datasets, one consisting of 51,589 employment spells and one of 26,276 spells. The samples are based on a number of restrictions aimed at isolating sets of firms and workers well suited to studying the precise connection between product market strategies and compensation policies. First, we limit the data to workers between the ages 21 and 44 in order to model the demand for a fairly homogeneous collection of individuals in the prime of their careers with similar educational vintages. This reduces the sample from the universe of 83,497 spells to 67,452. Second, we limit our individual worker data to those earning more than $50,000 (in 2001 dollars) at the end of their 1997 job spell (on an annualised basis). The rationale behind the $50,000 earnings threshold is that LEHD data do not contain information on hours of work or occupation. Therefore, to limit the data to workers who are likely to be full‐time and in more highly skilled occupations, we select those making more than $50,000. That threshold is based on a close analysis of the distribution of earnings within the relevant set of software occupations (programmers, developers, engineers, and managers) using PUMS data.15 Together, the age and earnings restrictions reduce the sample to 51,589 spells. Importantly, the earnings measure used for the worker is the quarterly earnings reported in the Unemployment Insurance wage record data. In order to reduce mis‐measurement due to partial quarter employment, we employ a full quarter earnings measure. This implies, for example, that the ‘starting’ earnings for a worker at a firm in question is the earnings for the first full quarter in which we observe the worker employed by the 1997 employer and the experienced earnings is the earnings for the last full quarter in which we observe them with their 1997 employer. While most businesses in our sample of workers could be successfully matched to the Economic Census for 1997, a smaller subset has complete information for firms, including firm size, firm age, sales and detailed product line information. There are 26,276 spells for which we have complete information on firm characteristics as well as worker characteristics. All told, 688 unique software firms appear in this sample.16 For the purposes of several robustness checks, we also consider a subset of employees in high‐skilled professions based on occupational information in the 2000 Decennial Census confidential long‐form survey records. For this sample, we limit our data to those individuals in the software industry whom we can successfully match to the Census long‐form and whom we can identify as software engineers, developers, or managers (irrespective of earnings). We drop those workers in other occupations within the software industry. Because the Decennial Census is a one in six sample of the population in 2000, this sample consists of only 2,638 workers. We use this dataset to check the robustness of our main findings but, due to its small size we refer to the results using this sample only in our discussion of robustness tests. 3.3. Economic Census Data Testing the main implications of our model requires estimates of the dispersion and/or skewness of the expected payoffs of projects in the software sectors in which each firm operates. For the prepackaged software industry, the 1997 Economic Census delineates 30 detailed software sectors ranging from consumer game and entertainment software to business graphics design and layout software to vertical industry banking software to mainframe computer applications. Software firms in the Economic Census are asked to provide data on their revenue for each of the 30 software sectors and we exploit this information in order to construct a firm‐specific measure that reflects the variance of payoffs to innovation by sector. We create each firm’s ‘product payoff dispersion/skewness’ measure in two steps. First, using the data on revenue streams for all firms, we calculate the 90–50 difference of the log of revenue per worker for the 30 software sectors.17 This gives us 30 values of the upside gains to innovation. Second, to create a firm‐specific measure, we weight the sector‐specific 90–50 differences for the 30 sectors by the share of revenue that the firm derives from each software sector in which it produces. We use this single summary measure of the 90–50 difference, as this measure reflects both dispersion and skewness.18 For ease of exposition, we refer to this measure as the ‘product payoff dispersion’ measure for the remainder of the article. Values of the product payoff dispersion measure for the software sectors with the greatest and least dispersion appear in Table 2. The results suggest that there is substantial variation in the variance and skewness of revenue per worker across software sectors, implying a high degree of heterogeneity in the potential upside gains to innovation. Further, observed patterns of dispersion across different product lines are in line with expectations, with categories such as video games topping the list of software sectors with high payoff dispersion, and database and distribution software falling near the bottom. Table 2 
Software Industry By Sector Product Payoff Dispersion Software sector . Software sector description . 90–50 Difference of Software sector log revenue per worker . Product Lines in Software with Greatest Potential Payoffs/Risks 1122 Game and entertainment software 1.31 1183 Networking software 1.17 1123 Home productivity software 1.03 Product Lines in Software with Smallest Potential Payoffs/Risks 1161 Banking and finance software 0.66 1142 Distribution software 0.57 1184 Database software 0.55 Software sector . Software sector description . 90–50 Difference of Software sector log revenue per worker . Product Lines in Software with Greatest Potential Payoffs/Risks 1122 Game and entertainment software 1.31 1183 Networking software 1.17 1123 Home productivity software 1.03 Product Lines in Software with Smallest Potential Payoffs/Risks 1161 Banking and finance software 0.66 1142 Distribution software 0.57 1184 Database software 0.55 Based on a 1997 Economic Census data for a national sample of firms. Open in new tab Table 2 
Software Industry By Sector Product Payoff Dispersion Software sector . Software sector description . 90–50 Difference of Software sector log revenue per worker . Product Lines in Software with Greatest Potential Payoffs/Risks 1122 Game and entertainment software 1.31 1183 Networking software 1.17 1123 Home productivity software 1.03 Product Lines in Software with Smallest Potential Payoffs/Risks 1161 Banking and finance software 0.66 1142 Distribution software 0.57 1184 Database software 0.55 Software sector . Software sector description . 90–50 Difference of Software sector log revenue per worker . Product Lines in Software with Greatest Potential Payoffs/Risks 1122 Game and entertainment software 1.31 1183 Networking software 1.17 1123 Home productivity software 1.03 Product Lines in Software with Smallest Potential Payoffs/Risks 1161 Banking and finance software 0.66 1142 Distribution software 0.57 1184 Database software 0.55 Based on a 1997 Economic Census data for a national sample of firms. Open in new tab It is worth emphasising that the firm‐specific product payoff measure reflects each firm’s actual product mix, not its actual revenue. That is, the payoff measure reflects the variance and skewness of revenue per worker in the software sectors in which the firm operates as opposed to its actual revenue per worker. Thus, a firm with a high product payoff dispersion measure is not necessarily a high or low performing business but rather has a product mix in sectors with a more highly dispersed and skewed distribution of potential payoffs. 4. Empirical Approach Our primary goal is to test whether firms operating in software sectors with high potential upside gains to innovation, pay higher compensation. We infer that this higher compensation reflects a greater demand for talent – the demand for workers with the skills to innovate. Our secondary goal is to examine how the alternative compensation structures, and thus alternative human resource practices of firms, reflect that demand for innovative output of workers. We estimate alternative variants of the following specification: (1) where wijt is earnings for worker i at firm j at time t; Xijt is a vector of worker controls some of which vary by job; Zj is a vector of firm controls and is the product payoff dispersion measure (calculated as described in Section 3.3). The worker and firm controls are described in detail below and are intended to control for other factors that impact earnings and, in particular, relevant determinants of earnings that might be correlated with product payoff dispersion. One key control is the actual payoff (or revenue per worker) which differs from the potential product payoff dispersion measure and controls for the individual firm’s performance and possible rent‐sharing. This empirical specification enables us to test several hypotheses. First, we can test whether firms operating in software sectors that have potential payoffs that exhibit higher variance and skewness pay higher earnings (α > 0). Such a result would be consistent with firms’ need to attract and retain workers who have greater skills or put forth more effort in innovating. For example, firms that want to hire highly skilled employees are likely to expend greater resources screening job candidates and, perhaps, place greater emphasis on training, teamwork or incentive pay in efforts to boost effort. There is certainly extensive industry evidence that suggests that innovative software firms engage in very careful and deliberate hiring practices, all aimed at identifying the right talent (Hoch et al., 2000). Second, we can evaluate earnings outcomes across individuals at different points in the earnings distribution and relate earnings to differences in skill demand arising from each firm’s product market strategy as measured by its product payoff dispersion. We estimate this with a series of quantile regressions. Third, we use earnings at different points of the tenure profile and the career path to shed light on the precise structure of compensation used to attract and retain stars. We examine ‘starting earnings’ for new workers for the 1997 employer as well as the ‘earnings with previous employer’ in the prior job spell. Both of these measures should be more related to skill as opposed to skill and effort. We then go on to examine earnings at the end of the 1997 job spell with the software firm (both censored and non‐censored with appropriate controls for censoring in the statistical analysis). The end‐of spell earnings, or what we call ‘experienced earnings’, likely includes incentive pay components such as exercised stock options.19 Comparing and contrasting the patterns for experienced and starting earnings provides insight into how firms structure their compensation packages to attract and retain stars. We also examine the patterns of earnings for experienced workers using measures that strip out most incentive pay, so that we can compare salary to total compensation (with incentives). The first of these is ‘experienced earnings lagged one year prior’ to the end of workers’ observed 1997 job spells. This measure should be less affected by stock options. The second measure, which we call ‘experienced salary’ is the minimum value of each worker’s end‐of‐spell earnings and earnings one‐year lagged to the end‐of‐spell. This measure is also aimed at mitigating the effect of performance‐based pay such as bonuses or stock options on measured earnings. There are several reasons for expecting the patterns to vary across these earnings measures. Firms may learn about a worker’s capacity for innovation as tenure rises in the firm, producing better matches as workers sort within firms and across firms to the projects or firms that value innovation. If teamwork is important, it may also take time for employers to identify and reward individual worker’s skills. Overall, firms in high payoff dispersion markets may invest more heavily in the matching and human capital of their employees in light of the high returns to good project selection. In terms of the effects of differences in payoff dispersion across firms on starting versus experienced earnings, firms in high payoff dispersion markets may offer slow vesting stock options and bonuses that grow more sharply with experience (or pay level) to provide higher returns to effort with tenure. The richness of the data also permits us to estimate regressions in which the dependent variable is a growth, rather than level measure. This both sweeps out fixed effects that may impact the level regressions and permits an earnings growth comparison for workers who stay with a firm versus those who change firms.20 In all of our variants of (1), we interpret the effect of sectoral payoff variance, , on wages as reflecting differences in demand for skills. However, we recognise that there are many other factors that are relevant for the determination of earnings, so we include a rich set of controls. Since our data permit us to construct wage histories for all workers, we create a set of control variables for (1) that include quadratics of tenure at job, tenure in industry and age, fully interacted with each other and with appropriate left and right censoring dummies when the spells for the 1997 job are left or right censored. As previously noted we limit the sample to 21–44 year‐old workers earning at least $50,000 in the software publishing industry who had ongoing job spells in the software industry during 1997. This has the effect of focusing the empirical analysis on a relatively homogeneous sample of individuals who are likely to be in similar educational and occupation categories (two characteristics we do not have in our data). We also use as control variables a number of firm characteristics, including a quadratic in (log) employment; log revenue per worker; dummies for a firm’s age; a firm’s employment growth rate; and a dummy for whether a firm is located in a high density, high education, industrially diversified county and the worker churning rate.21 These variables are intended to control for a variety of different potential influences on earnings outcomes. For example, log revenue per worker and the growth rate of the firm help to control for actual firm performance. We include the worker turnover or churning rate to control for any compensating differential associated with the risk of working at a high product dispersion firm. We include location information to control for local labour supply effects. We control for firm size and age effects as these have been shown in other literature to be related to firm performance as well as the structure of compensation. We recognise that in spite of these rich controls, there remains some potential for omitted variable bias and endogeneity. In what follows, we consider alternative specifications to deal with these issues and also provide further discussion of the possible directions of any lingering biases. 5. Empirical Results 5.1. Earnings and Product Payoff Dispersion We use both OLS and quantile regressions to determine whether the data substantiate our theory of the link between compensation and product market strategies. The goal of the quantile regressions is to determine if the shape of the distribution of earnings changes with the degree of product payoff dispersion faced by firms.22 In light of our conceptual framework in Section 2, we would expect that software workers at the upper reaches of the earnings distribution (the key decision makers for projects) have the highest differential in earnings from working in software sectors characterised by greater product payoff dispersion. The main results are presented in Table 3. For ease of exposition, we report only the impact of the product payoff dispersion variable on earnings, although all specifications have the rich set of controls discussed above. Before discussing the results in the table, it is worth nothing that in all specifications, there is a large, positive and significant effect from the measures of actual payoffs: firms that actually achieve a high revenue per worker pay high earnings to workers.23 Table 3 
The Effects of Product Payoff Dispersion on Earnings . (1) . (2) . (3) . . OLS . Quantile Regression . Dependent variable Starting and Previous Employer Earnings 10thPercentile 90thPercentile Starting earnings 0.0526 −0.1848 0.2129 (0.0331) (0.0460)*** (0.0557)*** Earnings with previous employer 0.1840 0.1145 0.1406 (0.0563)*** (0.1149) (0.0722)* Experienced Earnings End of spell experienced earnings 0.3868 0.0537 0.8279 (0.0629)*** (0.0340) (0.0990)*** Experienced earnings one year prior to end of spell 0.1312 −0.1674 0.4983 (0.0518)** (0.0445)*** (0.1001)*** Eperienced salary (excludes some bonuses and options) 0.0551 −0.1251 0.3731 (0.0340) (0.0343)*** (0.0697)*** Earnings Growth Within‐job earnings growth 0.0706 −0.0060 0.1837 (0.0120)*** (0.0073) (0.0191)*** Between‐job earnings growth −0.2169 −0.2352 −0.2597 (0.0476)*** (0.0653)*** (0.0921)*** . (1) . (2) . (3) . . OLS . Quantile Regression . Dependent variable Starting and Previous Employer Earnings 10thPercentile 90thPercentile Starting earnings 0.0526 −0.1848 0.2129 (0.0331) (0.0460)*** (0.0557)*** Earnings with previous employer 0.1840 0.1145 0.1406 (0.0563)*** (0.1149) (0.0722)* Experienced Earnings End of spell experienced earnings 0.3868 0.0537 0.8279 (0.0629)*** (0.0340) (0.0990)*** Experienced earnings one year prior to end of spell 0.1312 −0.1674 0.4983 (0.0518)** (0.0445)*** (0.1001)*** Eperienced salary (excludes some bonuses and options) 0.0551 −0.1251 0.3731 (0.0340) (0.0343)*** (0.0697)*** Earnings Growth Within‐job earnings growth 0.0706 −0.0060 0.1837 (0.0120)*** (0.0073) (0.0191)*** Between‐job earnings growth −0.2169 −0.2352 −0.2597 (0.0476)*** (0.0653)*** (0.0921)*** Each reported coefficient represents the coefficient on the product payoff dispersion in different regressions, where the dependent variable is given in the left hand column for different measures of earnings and the columns represent different regression specifications (OLS or quantile regressions at different points of the distribution). All regressions include worker controls include quadratics of tenure at job, tenure in industry and age, fully interacted with each other and with appropriate left and right censoring dummies. Firm controls include a quadratic in (log) firm employment, actual revenue per worker, dummies for firm age (<6 years, 6–10, 11+ years), the net employment growth rate, firm average worker churn and a dummy for whether the firm is in a high density/high education/industrially diverse county. Controls also include time dummies for quarter of separation and/or quarter of accession as appropriate. All spells are the workers 1997 spell when we observe firm‐level data. For OLS, robust standard errors are in parenthesis, bootstrapped (50 repetitions) robust standard errors are in parentheses. * significant at 10%; ** significant at 5%; *** significant at 1%. Based on LEHD data for ten states. Open in new tab Table 3 
The Effects of Product Payoff Dispersion on Earnings . (1) . (2) . (3) . . OLS . Quantile Regression . Dependent variable Starting and Previous Employer Earnings 10thPercentile 90thPercentile Starting earnings 0.0526 −0.1848 0.2129 (0.0331) (0.0460)*** (0.0557)*** Earnings with previous employer 0.1840 0.1145 0.1406 (0.0563)*** (0.1149) (0.0722)* Experienced Earnings End of spell experienced earnings 0.3868 0.0537 0.8279 (0.0629)*** (0.0340) (0.0990)*** Experienced earnings one year prior to end of spell 0.1312 −0.1674 0.4983 (0.0518)** (0.0445)*** (0.1001)*** Eperienced salary (excludes some bonuses and options) 0.0551 −0.1251 0.3731 (0.0340) (0.0343)*** (0.0697)*** Earnings Growth Within‐job earnings growth 0.0706 −0.0060 0.1837 (0.0120)*** (0.0073) (0.0191)*** Between‐job earnings growth −0.2169 −0.2352 −0.2597 (0.0476)*** (0.0653)*** (0.0921)*** . (1) . (2) . (3) . . OLS . Quantile Regression . Dependent variable Starting and Previous Employer Earnings 10thPercentile 90thPercentile Starting earnings 0.0526 −0.1848 0.2129 (0.0331) (0.0460)*** (0.0557)*** Earnings with previous employer 0.1840 0.1145 0.1406 (0.0563)*** (0.1149) (0.0722)* Experienced Earnings End of spell experienced earnings 0.3868 0.0537 0.8279 (0.0629)*** (0.0340) (0.0990)*** Experienced earnings one year prior to end of spell 0.1312 −0.1674 0.4983 (0.0518)** (0.0445)*** (0.1001)*** Eperienced salary (excludes some bonuses and options) 0.0551 −0.1251 0.3731 (0.0340) (0.0343)*** (0.0697)*** Earnings Growth Within‐job earnings growth 0.0706 −0.0060 0.1837 (0.0120)*** (0.0073) (0.0191)*** Between‐job earnings growth −0.2169 −0.2352 −0.2597 (0.0476)*** (0.0653)*** (0.0921)*** Each reported coefficient represents the coefficient on the product payoff dispersion in different regressions, where the dependent variable is given in the left hand column for different measures of earnings and the columns represent different regression specifications (OLS or quantile regressions at different points of the distribution). All regressions include worker controls include quadratics of tenure at job, tenure in industry and age, fully interacted with each other and with appropriate left and right censoring dummies. Firm controls include a quadratic in (log) firm employment, actual revenue per worker, dummies for firm age (<6 years, 6–10, 11+ years), the net employment growth rate, firm average worker churn and a dummy for whether the firm is in a high density/high education/industrially diverse county. Controls also include time dummies for quarter of separation and/or quarter of accession as appropriate. All spells are the workers 1997 spell when we observe firm‐level data. For OLS, robust standard errors are in parenthesis, bootstrapped (50 repetitions) robust standard errors are in parentheses. * significant at 10%; ** significant at 5%; *** significant at 1%. Based on LEHD data for ten states. Open in new tab The earnings for new employees, or the ‘starting earnings’, rise significantly with product payoff dispersion for workers whose earnings are at the 90th percentile of the distribution of starting earnings (row 1, column 3). The OLS results suggest, on the other hand, that the impact of product payoff dispersion measure on starting earnings is positive but not significant at the mean (row 1, column 1). Turning to the ‘earnings with previous employer’ as the dependent variable, these rise significantly with the product payoff dispersion at both the mean and 90th percentile measured for the current employer (row 2). We interpret earnings with the previous employer as a better measure of ex ante skills than starting earnings since the initial salary may reflect other factors like the structure of compensation at high payoff dispersion firms (i.e., higher payoff dispersion firms may pay more of the compensation in the form of bonuses and stock options that are not realised at the beginning of the job spell, which as we will see, turns out to be the case). The earnings for experienced workers rises very significantly with the firm’s potential product pay off dispersion. The first row second panel of Table 3 shows that workers at the end of their job spell have much higher pay if they are employed at a high product payoff dispersion firm. The estimated coefficients at both the mean and 90th percentile are positive and significant. Comparing the coefficients to the first panel, the impact of product payoff dispersion on experienced earnings is much larger than the impact on starting earnings; in fact the effect on experienced earnings is four times larger than the effect on starting earnings for workers at the 90th percentile. The quantitative implications of these differences for the level of earnings are examined below. Further results show that highly skilled workers tend to earn more in software sectors at firms in software with high potential upside gains even when bonus and stock option effects are stripped out of the dependent variable.24 In the second row of the second panel, we use earnings lagged one year prior to the end of workers’ observed job spells as the dependent variable, which is less likely to include exercised stock options than end‐of‐spell earnings. The same basic results hold, although the effects are smaller.25 In the third row we use an alternative specification of the dependent variable that represents the minimum of the measures in rows 1 and 2 of the second panel (earnings at the end of the spell and earnings lagged one year from the end of the spell). Again, this measure is aimed at stripping out the potential effects of bonuses and stock options. We find even for this measure that at the 90th percentile workers at high product payoff dispersion firms have significantly higher earnings. The results from the quantile regressions in Table 3 illustrate that earnings outcomes across individuals at different points in the earnings distribution appear to be related to differences in each firm’s product market strategy as measured by its product payoff dispersion. Since this result lies at the heart of our theory, we graph the implied wage distributions in 15Figure 3, where each point in the wage distribution is calculated for firms that operate in software sectors with the ‘minimum product market risk’ (the solid line) and the ‘maximum product market risk’ (the dashed line). An analysis of the graphs suggests that the right tail of the earnings distribution is substantially thicker for firms operating in the riskiest product markets notwithstanding which measure of earnings one uses. Indeed the upper tail is thicker for each measures of earnings, including starting salaries (Figure 3(a)), experienced earnings (Figure 3(b)), and experienced earnings one year prior to the end of the spell (Figure 3(c)).26,27 Fig. 3. Open in new tabDownload slide Predicted Starting Earnings, Experienced Earnings, and Experienced Salary at Minimum and Maximum Product Market Payoff Dispersion
Notes. Based on LEHD data ten states. Includes workers 21–44 earning $50,000 or more in SIC 7372. Independent variables in the regression include product payoff dispersion, log revenue per worker, a quadratic in log employment, the employment growth rate, firm age dummies (<6 years, 6–10 years, 11+ years), a dummy for whether the firm is in a high density/high education/industrially diverse county, and time dummies. Variables set at means for the figure include log revenue per worker, log employment, and the employment growth rate. Figure assumes firm age 6–10 years, that the firm is located in a high density/high education/industrially diverse county and that the worker accessed in the second quarter. Fig. 3. Open in new tabDownload slide Predicted Starting Earnings, Experienced Earnings, and Experienced Salary at Minimum and Maximum Product Market Payoff Dispersion
Notes. Based on LEHD data ten states. Includes workers 21–44 earning $50,000 or more in SIC 7372. Independent variables in the regression include product payoff dispersion, log revenue per worker, a quadratic in log employment, the employment growth rate, firm age dummies (<6 years, 6–10 years, 11+ years), a dummy for whether the firm is in a high density/high education/industrially diverse county, and time dummies. Variables set at means for the figure include log revenue per worker, log employment, and the employment growth rate. Figure assumes firm age 6–10 years, that the firm is located in a high density/high education/industrially diverse county and that the worker accessed in the second quarter. The results in the first two panels of Table 3 focus on earnings levels for workers with different amounts of tenure at their respective firms. As such, based on these results alone, we can make some inferences about the earnings–tenure profile. However, it is of interest to estimate the effects for the growth of earnings directly. Our data permit us to examine an individual earnings growth rate during the period the worker is within the firm, as well as when s/he moves between firms. We find that the within firm earnings growth is positively and significantly related to product payoff dispersion at the mean and at the 90th percentile (the first row of panel 3 in Table 3). In contrast, between‐job earnings growth is lower for workers moving to high potential payoff firms. The earnings losses from moving to high product dispersion effects may stem from multiple sources. Our interpretation is that high payoff product dispersion firms structure compensation to reward loyalty. That is, consistent with our primary hypothesis, they offer high earnings levels relative to alternatives but especially for those that stay. The gains for ‘loyalty’, or staying on the job are substantial: the average lost income with job change (the coefficient of 0.22) will not be offset by gains to tenure until after about four years (the coefficient of 0.07). However, for workers at the upper end of the distribution, the gains to loyalty (0.18) are sizable and can more quickly offset any losses associated with job transitions.28 In sum, firms appear to structure compensation packages so that when software workers change jobs, they must stay with the firm a number of years before their compensation rises. In this sense, loyalty pays. Moreover, the firms that reward loyalty the most are the very firms that operate in high‐risk, high potential payoff software sectors. We cannot assess the reasons for this result but it is clear that loyalty in the software industry pays and pays disproportionately among firms that face the riskiest software sectors. Firms in these dynamic sectors, therefore, structure compensation not only to select the most talented workers, but also to ensure firms motivate and retain them. 5.2. Discussion We show that workers at software firms operating in sectors that have high potential payoffs tend to have higher earnings and experience faster earnings growth. Even after controlling for high pay due to the firm’s actual revenue per worker, the firm’s potential upside payoff is a critical determinant of pay. These results are robust to using different measures of earnings and, as we find in unreported regressions, to using different specifications with varying sets of control variables. The results are also basically unchanged when we estimate regressions using the sub‐sample of workers who we know are software engineers or managers using the linked occupational information from the Decennial Census (thus dropping administrative staff). Firms in software sectors that appear to demand innovative workers pay higher wages for these skills. We interpret our results as being consistent with the primary hypothesis – high product payoff dispersion firms have a greater demand for the most talented software workers. Of course, there could be alternative underlying causes for the striking positive correlation between individuals’ earnings and potential payoffs in the software industry. One possibility is that high earnings might be due to the payment of a positive compensating differential for entering a risky software sector, although the work of Prendergast (2000, 2002) suggests that for such high skilled workers in high tech sectors, the risk or insurance effects are likely to be of limited importance. We do, however, include worker churn at the firm level as a proxy for job security to control for this possibility. Another possibility is that our results purely reflect incentive pay, even though the theoretical relationship between product market payoff dispersion and the use of incentive pay is unclear.29Table 3 shows that there is a positive and significant relationship between product payoff dispersion and starting earnings. While it is difficult to argue that this finding in particular is driven by differences in the propensity to use incentive pay mechanisms for high payoff dispersion firms, our findings of large within firm earnings growth and the associated high earnings of experienced workers indicate that such mechanisms are potentially part of the story.30 Still, part of our point is that even if incentive pay is being used, it is connected to the product market strategy of the firms. It might also be argued that our results documenting a link between high wages and certain segments of the software industry reflect a supply side, rather than a demand side story. This is unlikely given recent research that suggests that much of the dramatic rise in returns to education in the US over the last 20 years is attributable to a rising demand for skills as opposed to a reduced supply (Autor et al., 2008). But we are able to examine this directly because of the different supply factors in our two key occupations, software engineers and managers. Although one could argue that the lack of science graduates in the US produces a supply shortage of software engineers, managerial skills are unlikely to be in short supply since the number of graduates from MBA programmes rose steadily over the sample period. Indeed, when we examine our data by occupation, we find that the returns to being a software manager as well as a software engineer are higher in those sectors with greater payoff dispersion. We conclude from this that our observed compensation patterns are not consistent with a labour supply shortage hypothesis. Finally, since we are modelling the software industry during a period in which stock options became popular, the results might reflect variation in returns to options across firms and over time. Indeed, some end‐of‐spell earnings gains are so large that they must be due to exercised stock options. Therefore, it is important to point out that when we drop workers’ last year of earnings and look instead at earnings in a prior period (when stock options are exercised much less), all our qualitative results remain, suggesting that firm’s potentially endogenous use of different forms of performance‐based pay is not responsible for the observed connection between product market strategies and compensation. 6. Conclusion The process of innovation in the US economy is fundamentally dependent on firms employing and rewarding highly skilled workers. This article draws a link between the product market strategies of firms and the structure of compensation. We suggest that one way in which innovation can occur is when workers create or select new projects. Therefore, firms that operate in software sectors that have high potential returns to innovation should select talent carefully and pay workers highly for these skills. To examine this link between software product market innovation and skill demand, we assemble panel data on individuals as they move across firms and link that data to information on firms’ product market strategies. We show that software firms that operate in software sectors with highly skewed returns to innovation, or high upside gains to innovation, are more likely to attract and pay for highly talented workers. Such firms do so first by paying more up‐front in starting salaries to attract skilled employees and second by rewarding workers handsomely for experience or loyalty. These effects are robust to the inclusion of a wide range of controls for both worker and firm characteristics, including variables capturing the actual payoffs of firms. As such, our results are not driven by the fact that successful firms pay high earnings. Using alternative measures of earnings that strip out most performance‐based pay, like bonuses and stock options, we also show that our results are not driven by the high stock options of the 1990s; indeed, starting salaries and experienced salaries are higher in firms operating in high potential payoff software sectors. Though we focus on the software industry, our model and findings should generalise to other industries in which firms employ knowledge workers and face uncertainty in the probability of success on any given project. Our results documenting a link between income variance and innovation also complement the literature on changing skill demand and income inequality. Recent research suggests that the returns to skills have been increasing within as well as across occupations and industries, and furthermore that increases in earnings inequality in recent decades have been driven largely by rising pay differentials in the upper tail of the income distribution and by greater performance pay (Autor et al., 2003; Autor et al., 2007, 2008; Lemieux, 2006). In addition, researchers show that rewarding highly skilled workers with performance pay accounts for much of the increase in income inequality at the top end of the pay distribution (Lemieux, et al., 2007). While we have not conducted any time series analysis of overall wage trends in this article, it is clear that innovative industries are growing as a share of total US employment. Thus, high‐technology firms that pay a premium for productivity at innovation could be contributing to an increasing positively skewed distribution of earnings. We cast this rising income inequality in a positive light, showing that high variance in earnings goes hand‐in‐hand with innovative activity in dynamic markets. To the extent that these markets have been and will continue to be a source of growth in the economy, our research makes contributions to our understanding of not only firms’ human resource practices and product market strategies but also patterns of income inequality and economic growth. Footnotes 1 " The results hold for both total earnings and wage and salary earnings. 2 " The details of this dataset are described in much greater detail below. 3 " ‘Starting earnings’ reflect starting pay for new hires to the firm (measured as annualised earnings at the start of the observed job spell with their current employer), while ‘experienced earnings’ capture ending period pay for experienced workers (measured as annualised earnings at the end of the observed job spell with current employer). 4 " http://www.shacknews.com/onearticle.x/50809 5 " See, for example, Cusumano and Selby (1995), Hoch et al. (2000), Pfeffer (1998), Saloner et al. (2001) and Stross (1997). 6 " There are other related forms of uncertainty about product market payoffs that fit within the framework of our theory. Suppose, for example, that a component of the uncertainty relates to whether workers implement a new idea effectively. In this case, the talented programmers may be those that implement the idea well (e.g., without problematic bugs or other product market features that would have an adverse impact on the returns from the product). 7 " Further discussion of the ideas underlying this conceptual framework can be found in Lazear and Shaw (2007). 8 " A closely related idea is that the highest labour productivity firms are those that have the most advanced technology and, given some complementarity between technology and skills, demand the best workers. Our control for actual payoff controls for this effect as well. The role of technology–skill complementarity in accounting for between firm variation in earnings has been explored by Abowd et al. (2007), Bresnahan et al. (2002), Doms et al. (1997), Dunne et al. (2004), Machin and Van Reenan (1998) and Van Reenen (1996). 9 " The Census defines SIC 7372 as ‘establishments primarily engaged in designing and developing prepackaged software, including operating, utility, and applications programmes. These establishments may also prepare software documentation for the user, install software for the user, and train the user in the use of the software. Establishments primarily engaged in buying and selling prepackaged software are classified in Wholesale or Retail Trade. Custom computer software services, including computer code authors, are classified in Industry 7371.’ 10 " We thank Ron Jarmin for sharing information on firm age with the LEHD Programme for this project. 11 " Because of the sensitivity of these data, they are anonymised before they are used in any Census Bureau projects; all standard identifiers and names are stripped and replaced by a unique ‘Protected Identification Key’. Only Census Bureau employees or individuals who have Special Sworn Status are permitted to work with the data, and there are serious penalties for disclosing the identity of an individual or business. Any research must be for statistical purposes only and must be reviewed by the Census Bureau and other data custodians. Under Title 13 of the US code, any breach of confidentiality can result in prosecution in which violators are subject to a $250,000 fine and/or 5 years in jail. 12 " There are important exceptions. Most federal employment as well as some agricultural and nonprofit employment is not covered. Independent contractors and self‐employed individuals are also not covered. See Stevens (2002) for a full discussion of coverage issues. 13 " To our knowledge, no previous studies have included stock options data for such a wide range of workers across firms. The nature of our data permits us to exploit the fact that in most employment contracts, employees must exercise all options within 90 days of leaving the firm. We are able to track the earnings of employees for those 90 days and we can thus capture the value of all exercised options. For the laws surrounding the reporting of options, see the example from the California Employment Development Department at http://www.edd.ca.gov/taxrep/de231sk.pdf. For an analysis of options granted and data available on option values, see Oyer and Schaeffer (2002). 14 " Indeed, as Oyer and Schaefer (2002) point out, it generally takes about four years for stock options to be fully vested. Further, as Russell (2005) notes, for a typical software company, options are worth nothing for an employee’s first two years and then are vested at a rate of 2% per month for the remaining three years. 15 " The primary occupations on which we focused included Census industry occupation codes 100 (Computer and Information Scientists, Research), 101 (Computer Programmers) and 102 (Computer Software Engineers, Applications and Systems Software), as well as 001‐043 (managerial occupations). In the PUMS data, two‐thirds of all software workers and four‐fifths of software engineers have total earnings of at least $50,000. Indeed, the mean of total earnings for software engineers in SIC 7372 earning at least $50,000 is $103,881, only slightly higher than the $90,668 reported in Table 1 for workers at all earnings levels. It is also worth noting that the $50,000 represents the worker’s earnings when we last observed him or her in the data; 36% of those earning $50,000 or more when we last observe them have starting salaries less than $50,000. Fortunately, the results in Table 1 (as well our robustness analysis discussed in more detail below) indicate that by using a relatively simple income cutoff, we can identify the software developers and managers in the administrative data. That is, focusing on workers earning more than $50,000 annually in constant 2001 dollars yields workers that are well identified as software developers and managers. 16 " Throughout this article, when we refer to a firm, we are referring to a firm defined at the State Employer Identification Number (the SEIN, or UI account number), which is the unit of observation in the UI‐Wage data. It is an 11‐digit number used for reporting taxes at the state level. For single‐unit firms, this reflects the entire firm but, for multi‐unit firms, the SEIN reflects activity of the firm within a given state. We are able to match the workers to information in to the Economic Censuses since the UI files also include the federal Employer Identification Number (the EIN is on the ES‐202 data that is part of the related administrative data system). The EIN is a nine‐digit number assigned by Internal Revenue Service (IRS) and used for federal tax purposes by employers, sole proprietors, corporations, partnerships, non‐profit organisations, trusts, estates of decedents, government agencies, certain individuals and other business entities. 17 " We have the actual revenue from 688 firms for 1997 that we use to calculate the dispersion of revenue per worker across firms. However, because firms report their revenues by software sectors and since most firms produce in more than one software sectors, we treat each revenue stream as though it were a separate firm and thus have far more than 688 revenue streams that we use in the calculation of the revenue dispersion for the 30 sectors. Treating each revenue stream as though it is a separate firm is consistent with our model of product innovation by sector. 18 " We note that it might be of interest to explore the lower tail of the distribution as well but the latter is more challenging in terms of measurement. In the data, revenues are truncated at zero for firms going out of business. The lack of complete data on the lower tail, in part motivates our focus on the 90‐50 difference. Furthermore, although in our theory we emphasise both the value of avoiding losses and creating gains, in the software industry, it is reasonable to argue that the key objective of the firm is to produce big wins, since downside losses are typically much smaller in magnitude than the upside gains when a product is successful in the market. See Lazear (1998) and Baron and Kreps (1999) for examples of industries in which downside losses can be huge because the entire brand value of the firm is lost if workers make mistakes (as with an oil spill or a product recall due to customer injuries). 19 " Case study evidence suggests that some firms offer such contracts. Russell (2005) finds that a larger percentage of a given workers’ pay is performance‐based as the worker’s skill level rises. Our end‐of‐spell measure captures the earnings of workers leaving the firm as well as right‐censored job spells and it potentially contains exercised stock options. By law, workers must exercise options prior to separation or typically within 90 days following separation. By end of spell we mean the last full quarter of earnings for the observed spell. 40% of the end‐of‐spell earnings are censored and we control for this with a full set of interacted censoring dummies. 20 " Within‐job earnings growth is defined as log annualised end‐of‐spell earnings less log annualised beginning‐of‐spell earnings, divided by the number of full quarters that a worker was on the job. We measure between‐job earnings growth as the difference between earnings in the first quarter of the worker’s software job spell that is ongoing in 1997 and the last full quarter of his or her prior job. 21 " The dummy for being located in a high density, high education, industrially diverse county takes a value of one if the county in which a firm is located has employment density (employment per square mile), average educational attainment levels (share of population with a college degree) and industial diversity (inverse Herfindahl‐Hirschman index) each greater than the median calculated based on all counties in the ten sample states, and zero otherwise. We measure churning as the worker accession rate plus the separation rate less the absolute value of net employment growth. 22 " In an analogous approach, Hallock et al. (2004) use CEO data to also show that ‘higher ability managers [would have] higher pay for performance incentives than low ability managers’ (7) due to the lower cost of effort for high ability managers. Buchinsky (1994) shows that the returns to education are higher at high wage quantiles while the returns to experience are lower at high wage quantiles. 23 " For coefficient values, see the longer working paper, Andersson et al. (2007). 24 " As discussed above, the end‐of‐spell earnings likely include exercised stock options. 25 " For example, at the 90th percentile the estimated coefficient is about 0.50 compared to the coefficient of about 0.83 for end‐of‐spell experienced earnings 26 " Appendix Figures A1(a) to (c) corroborate these results by plotting the difference between the expected earnings for high and low payoff firms. The standard errors around the coefficients at each quantile are also small, as displayed in Figure A2. 27 " In interpreting the results, it is important to emphasise that the reported effects from the quantile regressions yield the implied effect of the variable in question on the conditional quantile distribution. By the conditional quantile distribution, we mean the distribution of earnings taking into account all of the other explanatory variables including the controls. Thus, the coefficients should not be interpreted to yield inferences about the impact of variables on the unconditional distribution of earnings. For our purposes, the focus on the conditional distribution of earnings is appropriate, since we are interested precisely in the impact of product payoff distribution holding the impact of all other factors constant. For further discussion of these issues, see Buchinsky (1994). 28 " The raw data corroborate the regression‐based findings that loyalty pays in software. Of the approximately 4% of the sample who earned over $1 million in the last observed period in the data, over 95% of their earnings growth arose within firms, and less than 5% from movement between firms. By contrast, among software workers who earned $50,000–$75,000 in the last observed period in the data, the final pay is achieved by a combination of changing jobs (25%) and by earnings growth when they stay within a firm and experience earnings increases (75%). For evidence on job hopping in software, see Fallick et al. (2006). 29 " In a tournaments model of incentive pay, increasing the amount of noise or luck reduces the use of incentive pay (Lazear and Rosen, 1981). In contrast, Prendergast (2000, 2002) points out that higher risk environments may have more performance‐based pay because the cost of determining what inputs to monitor in such environments is greater. Moreover, in the software industry, talented workers may seek out high‐risk, high‐payoff jobs because they have the talent to control the outcomes; risk in this case is not random noise or luck. 30 " For related empirical models of risk‐pay incentive relationships, see Baker and Hall (2004), Core et al. (2003), Ittner et al. (2003), Murphy (1986), Jensen and Murphy (1990), Schaefer (1998) and Wulf (2007). For excellent reviews of the literature on the risk‐pay relationship for CEOs in particular, see Hallock and Murphy (1999) and Murphy (1999). References Abowd , J. , Haltiwanger , J., Lane , J., McKinney , K. and Sandusky , K. ( 2007 ). ‘ Technology and skill: an analysis of within and between firm differences ’, NBER Working Paper No. 13043. Andersson , F. , Freedman , M., Haltiwanger , J., Lane , J. and Shaw , K. ( 2007 ) ‘ Reaching for the stars: who pays for talent in innovative industries? ’, NBER Working Paper No. 12435. Autor , D. , Katz , L. and Kearney , M. ( 2007 ). ‘The polarization of the U.S. labor market’ , American Economic Review Papers and Proceedings , vol. 96 ( 2 ), pp. 189 – 94 . Google Scholar Crossref Search ADS WorldCat Autor , D. , Katz , L. and Kearney , M. ( 2008 ). ‘Trends in U.S. wage inequality: reassessing the revisionists’ . Review of Economics and Statistics , vol. 90 ( 2 ), pp. 300 – 32 . Google Scholar Crossref Search ADS WorldCat Autor , D. , Levy , F. and Murname , R. ( 2003 ). ‘Skill content of recent technological change: an empirical exploration’ , Quarterly Journal of Economics , vol. 118 ( 4 ), pp. 1279 – 333 . Google Scholar Crossref Search ADS WorldCat Baker , G. and Hall , B. ( 2004 ). ‘ CEO incentives and firm size’ , Journal of Labour Economics , vol. 22 ( 4 ), pp. 767 – 98 . Google Scholar Crossref Search ADS WorldCat Baker , G. and Hubbard , T.N. ( 2001 ). ‘Empirical strategies in contract economics: information and the boundary of the firm’ , American Economic Review , vol. 91 ( 2 ), pp. 189 – 94 . Google Scholar Crossref Search ADS WorldCat Baron , J. and Kreps , D. ( 1999 ). Strategic Human Resources , New York: John Wiley & Sons . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Bresnahan , T. , Brynjolfsson , E. and Hitt , L.M. ( 2002 ). ‘Information technology, workplace organization and the demand for skilled labor: firm‐level evidence’ , Quarterly Journal of Economics , vol. 117 ( 1 ), pp. 339 – 76 . Google Scholar Crossref Search ADS WorldCat Buchinsky , M. ( 1994 ). ‘Changes in the U.S. wage structure 1963‐1987: application of quantile regression’ , Econometrica , vol. 62 ( 2 ), pp. 405 – 58 . Google Scholar Crossref Search ADS WorldCat Chevalier , J. and Ellison , G. ( 1999 ). ‘Career concerns of mutual fund managers’ , Quarterly Journal of Economics , vol. 114 ( 2 ), pp. 389 – 432 . Google Scholar Crossref Search ADS WorldCat Core , J.E. , Guay , W. and Larcker , D.F. ( 2003 ). ‘Executive equity compensation and incentives: a survey’ , Federal Reserve Bank of New York Economic Policy Review , vol. 9 ( 1 ), pp. 27 – 50 . OpenURL Placeholder Text WorldCat Cusumano , M.A. and Selby , R.W. ( 1995 ). Microsoft Secrets: How the World’s Most Powerful Software Company Creates Technology, Shapes Markets, and Manages People , New York: Free Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Doms , M. , Dunne , T. and Troske , K. ( 1997 ) ‘Workers, wages, and technology’ , Quarterly Journal of Economics , vol. 112 ( 1 ), pp. 253 – 90 . Google Scholar Crossref Search ADS WorldCat Dunne , T. , Foster , L., Haltiwanger , J. and Troske , K. ( 2004 ). ‘Wage and productivity dispersion in U.S. manufacturing: the role of computer investment’ . Journal of Labor Economics , vol. 22 ( 2 ), pp. 397 – 429 . Google Scholar Crossref Search ADS WorldCat Fallick , B. , Fleischman , C.A. and Rebitzer , J.B. ( 2006 ). ‘Job‐hopping in silicon valley: some evidence concerning the micro‐foundations of a high‐technology cluster’ . Review of Economics and Statistics , vol. 88 ( 3 ), pp. 472 – 81 . Google Scholar Crossref Search ADS WorldCat Frank , R. and Cook , P. ( 1995 ). The Winner‐Take‐All Society: Why the Few At The Top Get So Much More Than The Rest Of Us , New York: Free Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Garicano , L. and Hubbard , T. ( 2007 ). ‘ Managerial leverage is limited by the extent of the market: hierarchies, specialization and the utilization of lawyers’ human capital’ , Journal of Law and Economics , vol. 50 ( 1 ), pp. 1 – 43 . Google Scholar Crossref Search ADS WorldCat Hallock , K. , Madalozzo , R. and Reck , C. ( 2004 ). ‘ Uncovering heterogeneity in managerial pay: firm performance relationships using quantile regression ’, Cornell University ILR School Working Paper. Hallock , K. and Murphy , K.J. ( 1999 ). The Economics of Executive Compensation, Volumes I and II , Cheltenham: Edward Elgar . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Hoch , D.J. , Roeding C.R., Purkert , G. and Lindner , S.K. ( 2000 ). Secrets of Software Success: Management Insights from 100 Software Firms around the World , Boston: Harvard Business School Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Ittner , C.D. , Lambert , R.A. and Larcker , D.F. ( 2003 ). ‘The structure and performance consequences of equity grants to employees of new economy firms’ , Journal of Accounting and Economics , vol. 34 , pp. 89 – 127 . Google Scholar Crossref Search ADS WorldCat Jensen , M.C. and Murphy , K.J. ( 1990 ). ‘Performance pay and top‐management incentives’ , Journal of Political Economy , vol. 98 ( 2 ), pp. 225 – 64 . Google Scholar Crossref Search ADS WorldCat Jorgenson , D. , Ho , M.S. and Stiroh , K.J. ( 2005 ). Information Technology and the American Growth Resurgence , Cambridge MA: MIT Press ,. Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Lazear , E. ( 1998 ). Personnel Economics for Managers , New York: John Wiley & Sons . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Lazear , E. ( 2005 ). ‘Output‐based pay: incentives or sorting?’ in ( S.W. Polachek, ed.), Research in Labor Economics: Accounting for Worker Well‐Being , pp. 1 – 25 , New York, JAI Press. Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Lazear , E. and Rosen , S. ( 1981 ). ‘Rank‐order tournaments as optimum labor contracts’ , Journal of Political Economy , vol. 89 ( 5 ), pp. 841 – 64 . Google Scholar Crossref Search ADS WorldCat Lazear , E. and Shaw , K. ( 2007 ), ‘Personnel economics: the economist’s view of human resources’ , Journal of Economic Perspectives , vol. 21 ( 4 ), pp. 91 – 114 . Google Scholar Crossref Search ADS WorldCat Lemieux , T . ( 2006 ). ‘ Post‐secondary education and increasing wage inequality’ , American Economic Review , vol. 96 ( 2 ), pp. 195 – 99 . Google Scholar Crossref Search ADS WorldCat Lemieux , T. , Macleod , W.B. and Parent , D. ( 2007 ). ‘ Performance‐pay and wage inequality . NBER Working Paper No. 13128. Lerner , J. and Wulf , J. ( 2007 ). ‘ Innovation and incentives: evidence from corporate R&D’ , Review of Economics and Statistics , vol. 89 ( 4 ), pp. 634 – 44 . Google Scholar Crossref Search ADS WorldCat Machin , S. and Van Reenen , J. ( 1998 ). ‘Technology and changes in skill structure: evidence from seven OECD countries’ , Quarterly Journal of Economics , vol. 113 ( 4 ), pp. 1215 – 44 . Google Scholar Crossref Search ADS WorldCat MacLeod , W.B. and Parent , D. ( 1999 ). ‘Job Characteristics and the form of compensation’ , Research in Labour Economics , vol. 18 , pp. 177 – 242 . Google Scholar Crossref Search ADS WorldCat Murphy , K.J. ( 1986 ). ‘Incentives, learning, and compensation: a theoretical and empirical investigation of managerial labour contracts’ , RAND Journal of Economics , vol. 17 ( 1 ), pp. 59 – 76 . Google Scholar Crossref Search ADS WorldCat Murphy , K.J . ( 1999 ). ‘Pay, performance and executive compensation’, in ( O. Ashenfelter and D. Card, eds). Handbook of Labour Economics , vol. 3 , pp. 1463 – 555 , Netherlands: Elsevier . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Oliner , S.D. and Sichel , D.E. ( 2000 ). ‘The resurgence of growth in the late 1990s: is information technology the story?’, Journal of Economic Perspectives , vol. 14 , pp. 3 – 22 . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Oyer , P. and Schaeffer , S. ( 2005 ). ‘ Why do some firms give stock options to all employees? An empirical examination of alternative theories’ , Journal of Financial Economics , vol. 76 (1) , pp. 99 – 133 . Google Scholar Crossref Search ADS WorldCat Pfeffer , J . ( 1998 ). ‘ SAS Institute (A): a different approach to incentives and people management practices in the software industry ’, Stanford GSB Case HR6A. Prendergast , C. ( 2000 ). ‘What trade‐off of risk and incentives?’ , American Economic Review , vol. 90 ( 2 ), pp. 421 – 5 . Google Scholar Crossref Search ADS WorldCat Prendergast , C. ( 2002 ). ‘The tenuous trade‐off between risk and incentives’ , Journal of Political Economy , vol. 110 ( 5 ), pp. 1071 – 102 . Google Scholar Crossref Search ADS WorldCat Rosen , S. ( 1981 ). ‘The economics of superstars’ , American Economic Review , vol. 71 ( 5 ), pp. 845 – 58 . OpenURL Placeholder Text WorldCat Russell , K.D. ( 2005 ). ‘ The relationship between deferred compensation and employee retention ’, Stanford University Dissertation. Saloner , G. , Shepard , A. and Podolny , J. ( 2001 ). Strategic Management , New York: John Wiley & Sons . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Schaefer , S. ( 1998 ). ‘The dependence of pay‐performance sensitivity on the size of the firm’ , Review of Economics and Statistics , vol. 80 ( 3 ), pp. 436 – 43 . Google Scholar Crossref Search ADS WorldCat Stern , S. ( 2004 ). ‘Do scientists pay to be scientists?’ , Management Science , vol. 50 ( 6 ), pp. 835 – 53 . Google Scholar Crossref Search ADS WorldCat Stevens , D . ( 2002 ). ‘ Employment that is not covered by state unemployment insurance laws ’, LEHD Technical Paper TP 2002‐16, United States Census Bureau. Stross , R. E. ( 1997 ) The Microsoft Way: The Real Story of How the Company Outsmarts Its Competition , New York: Perseus Books Group . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Van Reenen , J. ( 1996 ). ‘The creation and capture of rents: wages and innovation in a panel of U.K. companies’ , Quarterly Journal of Economics , vol. 111 ( 1 ), pp. 195 – 226 . Google Scholar Crossref Search ADS WorldCat Wulf , J. ( 2002 ) ‘Internal capital markets and firm‐level compensation incentives for division managers’ , Journal of Labour Economics , vol. 20 ( 2 ), pp. S219 – 62 . Google Scholar Crossref Search ADS WorldCat Wulf , J. ( 2007 ). ‘Authority, risk, and performance incentives: evidence from division manager positions inside firms’ , Journal of Industrial Economics , vol. 55 ( 1 ), pp. 169 – 96 . Google Scholar Crossref Search ADS WorldCat Appendix A1 [ Difference in Predicted Starting Earnings, Experienced Earnings, and Experienced Salary at Observed Maximum and Minimum Product Market Risk Levels
Notes. Based on LEHD data ten states. Includes workers 21–44 earning $50,000 or more in SIC 7372. Independent variables in the regression include product payoff dispersion, log revenue per worker, a quadratic in log employment, the employment growth rate, firm age dummies (<6 years, 6–10 years, 11+ years), a dummy for whether the firm is in a high density/high education/industrially diverse county, and time dummies. Variables set at means for the Figure include log revenue per worker, log employment, and the employment growth rate. Figure assumes firm age 6–10 years, that the firm is located in a high density/high education/industrially diverse county and that the worker accessed in the second quarter. ] A2 [ Coefficients on Product Payoff Dispersion across the Experienced Earnings Distribution
Notes. Based on LEHD data ten states. Includes workers 21–44 earning $50,000 or more in SIC 7372. Other independent variables in the regression include quadratics in tenure at job, tenure in industry, and age, fully interacted with each other and with appropriate left and right‐censoring dummies. Firm controls include a quadratic in (log) firm employment, firm age dummies (<6 years, 6–10 years, 11+ years), the employment growth rate, and a dummy for whether the firm is in a high density/high education/industrially diverse county. Controls also include time dummies for the quarter of separation and/or the quarter of accession as appropriate ] Author notes " Thanks to George Baker, Tim Bresnahan, Charlie Brown, Ben Campbell, Robert Gibbons, Britta Glennon, Erica Groshen, Phil Hardiman, Edward Lazear, Alex Mas, Paul Oyer, Julie Wulf and seminar participants at Stanford University, Harvard University, MIT, University of Illinois, UC Berkeley, University of Michigan, University of Toronto, Yale University, Washington State University, UC Santa Barbara, NBER, and the Society of Labour Economics meetings, for their helpful comments. This document reports the results of research and analysis undertaken by the US Census Bureau staff. It has undergone a Census Bureau review more limited in scope than that given to official Census Bureau publications. This document is released to inform interested parties of research and to encourage discussion. This research is a part of the US Census Bureau’s Longitudinal Employer‐Household Dynamics Programme (LEHD), which is partially supported by the National Science Foundation Grants SES‐9978093 and SES‐0427889 to Cornell University (Cornell Institute for Social and Economic Research), the National Institute on Aging Grant R01 AG018854‐02, and the Alfred P. Sloan Foundation. The views expressed on statistical issues are those of the author(s) and not necessarily those of the US Census Bureau, its programme sponsors or data providers. Some or all of the data used in this article are confidential data from the LEHD Programme. The US Census Bureau supports external researchers’ use of these data through the Research Data Centers (see http://www.ces.census.gov). For other questions regarding the data, contact US Census Bureau. © The Author(s). Journal compilation © Royal Economic Society 2009

Lancaster,, Tony

doi: 10.1111/j.1468-0297.2009.02281.xpmid: N/A

‘Another brilliant Manski book.’ This will be a common, and reasonable, reaction to the latest output of his prolific research agenda to study the limits of identifiability. This book follows his earlier Identification Problems in the Social Sciences (1995) and two technical monographs1 and, like the earlier book, it is ‘intended to be broadly accessible to students and researchers in the social sciences’. It is meant, in the author’s phrase, to describe ‘the next generation’ of work on the general problem of missing data leading to lack of point identification or to ‘partial identification’. It also re‐presents his methodological views or ‘fresh worldview to guide empirical researchers in the social sciences’, as he puts it in the preface. Manski deals with a very broad range of topics, not all which are ordinarily thought of as involving missing data, and he almost always has interesting and novel insights, making his book a rich and absorbing read. His topics include non‐response in surveys; decomposition of mixtures; response‐based sampling; selection problems; treatment response; decision making with unknown objective functions; measuring expectations; and even the simultaneous equations problem that started the identification hare running. And although the book is about identification he does not entirely omit discussion of inference. After all, if you have shown that something is (partially) identified, why not explain how to estimate it? The basic idea is that the economist observes a frequency distribution that differs in some systematic way – the observed data are incomplete or some is missing – from the distribution that he would like to observe. The question is what, if anything, can be said about the latter distribution knowing the former, without making incredible assumptions. Before offering some comments it is worth making a remark about the relation between the research agenda of this book and the traditional identification agenda. ‘In many fields the object of the investigator’s inquisitiveness is not just a “population” in the sense of a distribution of observable variables, but a physical structure projected behind this distribution, by which the latter is thought to be generated.’ (Koopmans and Reiersol, 1950, p. 165) This is how Koopmans and Reiersol thought back in 1950; identification consisted of going behind the observed distribution of data to the physical structure that is assumed to have generated it. By physical structure it seems reasonable to suppose they had principally in mind an economic model. Although you can cast Koopmans and Riersol’s perspective as one of coping with missing data – for example if only you could observe quantities demanded and supplied at alternative not necessarily equilibrium prices then you can work out demand and supply curves which provide the physical structure – it is not perhaps the natural way of looking at it. The present book places much less emphasis on the ‘physical mechanism’, and much more on missingness mechanisms. The typical problem is that of going from one frequency distribution not to a physical structure but to another frequency distribution. (I shall illustrate this with two examples shortly.) In this it looks more like the large and increasing number of recent books by statisticians and biostatisticians on the subject of missing data (Little and Rubin, 2002; Daniels and Hogan, 2008) than it does a traditional econometric identification text. Indeed, going from one frequency distribution, that of the data in your sample, to another, that in the ‘population’ from which your data is presumed to originate, is the essence of the problem of statistical inference. It has nothing particularly to do with econometrics. Two examples, both mentioned in Manski’s book, may make our discussion concrete. The first is the Piliavin and Sosin example discussed on p. 41. At a certain date there were 106 homeless men, of whom 64 could be located six months later; and of these 64, 21 were no longer homeless. The observed data are the 64 men out of 106 of whom 21 had found a home. The desired data are all 106 men and whether each had found a home. Interest is in the fraction of homeless men who are housed six months later – a measure of the rate of leaving homelessness. But we do not observe this number, all we can see is the fraction of the men who could be traced who found a home. A second example is the recent study of wages by Blundell et al. (2007) The desired distribution is that of the wages that each of 187,467 individuals in the UK between 1978 and 2000 could have earned had they worked in the paid labour market; the observed distribution is that of the wages of those who did work. Interest is in changes in the distribution of wages, as measured by the interquartile range, that cannot be attributed to changes in the composition of the work force. As is by now well known, Manski’s point of view depends on what he formalises as The Law of Decreasing Credibility. This states that: ‘The credibility of inference decreases with the strength of the assumptions maintained’ (p. 3). And since he wants credible inference he wants weak assumptions, preferably using only the empirical evidence.2 How would Manski’s agenda work out in the case of the two examples sketched above? Out of the 106 homeless men 42 could not be found after six months, so there are 43 possibilities. Either none of the missing men had found a home or one of them had, or two of them had, ....or all 42 had. So the fraction of all 106 men exiting homelessness was either 21/106 or 22/106 or … or 63/106. One can bound the true fraction to an interval. That is all one can say using the empirical evidence alone. But note that one can be more precise if one asserts that the fraction of the missing men who had found a home was the same as the fraction of the observed men who had found a home – a sort of missingness at random. In this case the fraction of all 106 men who found a home would be exactly 21/64. Inference becomes much more precise than without this assertion – indeed we have point identification – but at the expense of a possibly incredible (to whom?) assumption. The wage example is much more interesting to an econometrician, because the very distribution of interest, that of the potentially counterfactual wages – what she would have earned, in equilibrium, had she worked – is defined by economic theory. Anyone who finds the theory, hallowed by time and tradition though it is, incredible would not leave the starting post with this paper and although praised by Manski it is a very long way from inference using only ‘the empirical evidence’. The basic idea here is that if you observe 100 people of whom 90 work and 10 do not, so you have 90 observed wages, you can bound some of the quantiles of all 100 wages. One extreme possibility is that all 10 unobserved wages lie below the 90 observed ones. This implies that you can never bound from below the 10th percentile of the full distribution – it could be anywhere below the smallest observed wage – but you can bound from below the 11th percentile – it cannot lie below the smallest observed wage, being equal to it if all unobserved wages are below the observed ones and greater than it otherwise. The authors exploit this statistical type of reasoning to bound interquartile ranges of the full distribution, and they then consider the additional precision gained by further more or less (in)credible assumptions, such as that the median wage of those who work is greater than that of those who do not. But, unlike missingness at random in the homelessness example, these further assumptions tend to have strong economic content and one can imagine much of our profession agreeing on their credibility or not though outsiders may be sceptical. Moreover they put some of these extra assumptions to the test, in so far as they are able. This paper is a far cry from letting the data speak for themselves. Turning to the Law of Decreasing Credibility itself we might ask what the author means by credible. He supplies the answer, ‘An assumption is credible to the degree that someone thinks it so’ (p. 48). This raises the question of who are to be the judges of credibility? Who will reject an article because its assumptions are incredible? This criterion appears to be a recipe for methodological conservatism and seems to introduce an alarming degree of subjectivity into an econometric analysis. Reliance on the opinions of unspecified judges of credibility sits rather awkwardly with a philosophy which emphasises using only the, objective, empirical evidence. This methodological point of view also, of course, clashes sharply with the recommendations of Friedman, as Manski is well aware. Friedman (1953) would say that the credibility of assumptions is of little relevance and that what matters is the quality of predictions from a model. ‘Truly important and significant hypotheses will be found to have “assumptions” that are wildly inaccurate descriptive representations of reality, and, in general, the more significant the theory, the more unrealistic the assumptions…’ Friedman argues in his famous essay on The Methodology of Positive Economics. But for Manski the credibility of assumptions is central, and predictions seem to play little role in judging a model, notwithstanding their appearance in his book’s title. To adapt an example of Friedman, Manski would apparently reject the formula to explain the speed at which apples fall from a tree, even though it predicts well, on the grounds that the assumption of a vacuum is clearly incredible. Nowhere is this view clearer than in Manski’s take on rational expectations, of which he disapproves heartily. ‘It is not credible to suppose, without substantiation, that expectations are either literally or approximately rational.’ (p. 277) One consequence of this is that the econometrics of dynamic stochastic general equilibrium (DSGE) models, currently an important and active area of applied econometrics, fails to get a mention in his book. More generally, what used to be known as structural modelling is not a part of the Manski agenda, which could be said to be reduced form with a vengeance. It seems strange that the one‐time distinguished (and latterly Bayesian) statistician Milton Friedman advocated a methodology in which possibly unlikely economic assumptions play the central role; while economist Manski advocates a methodology in which the central role is played by a purely statistical device (bounding) and economic theoretical assumptions have, at best, a walk‐on part. And to add to the paradox Friedman’s greatest early piece of applied econometrics, A Theory of the Consumption Function (1957), dealt with a missing data problem. He wanted to observe permanent income and permanent consumption but could only observe measured income and measured consumption, permanent income was missing data. He dealt with this not by producing bounds on the consumption function but by adopting strong economic hypotheses, such as the proportionality of permanent income and consumption, and seeing whether the implications of these were consistent with a wide variety of economic data. The reader will have noticed that I have misdescribed the two examples and written that, in the homelessness example for instance, the object of interest is the fraction of all 106 men who had found a home within six months. Actually, the object of interest is the fraction of homeless people in some presumably larger ‘study population’ that find a home within six months. This number is interpreted as the probability of finding a home within six months by those, including Manski, who believe in objective probabilities. And the end points of the identified set, namely {21/106, 63/106} are regarded as estimates of theoretical bounds on the probability of interest, estimates that will vary over repeated samples from the study population. My reason for ignoring this version of the problem is that the extension of the problem to fractions in imaginary repeated samples from ill‐defined (how do you randomly sample the homeless?) study populations, though it is the traditional way of thinking of course, violates the law of decreasing credibility for this reader. You do not need to believe in objective probabilities to find the algebra of set identification useful, as both my examples demonstrate. I note also that in the wage example of Blundell et al. the discussion of bounds and possible restrictions on the relation between observed and unobserved wages could be easily thought of as exactly analogous to my discussion above of the 106 homeless men, in which the objects of interest are actual frequency distributions in a given real population, not 106 men in Minneapolis but 187,467 adults in the UK over a specific interval of time. It does not really appear to add anything to their analysis to think of their wage distributions as probabilities. Even the section of the paper that deals with inference about bounds relies largely on the bootstrap in the context of a multinomial model, a procedure which has a well‐known Bayesian interpretation, (Lancaster, 2004). Indeed, so far as I can see the wage paper looks like a fine piece of subjective Bayesian econometrics in which the authors conduct a subtle analysis of the effect of alternative prior beliefs about the unobserved wages. One of the most interesting parts of the book is the later chapters which deal with his newer work on the interpretation and use of reported expectations, intentions and beliefs. Manski rightly remarks that economists have traditionally been sceptical of using the statements of individual agents about their beliefs and behaviour, preferring instead to use only objective data about prices, incomes, purchases and the like. This scepticism goes back at least to the 1940s when Machlup (1946) and others criticised Hall and Hitch’s (1939) work that asked businessmen how they set their prices and received some apparently disconcerting answers. Manski documents that this aversion has recently decreased and that there is a growing volume of work in which agents are interviewed and asked ‘expectations’ questions. He leaves the impression that work taking seriously the answers to such questions is a recent development, but he neglects at least one earlier paper. In this work, Lancaster and Chesher (1983), unemployed people gave answers to the questions ‘what wage do you expect to make in a new full time job?’ and ‘what is the smallest wage you would accept in such a job?’. If you maintain the simplest version of the optimal job search model and interpret these questions as asking for the reservation wage, ξ, and the mean wage given that it exceeds ξ, then Chesher and I showed that one can deduce parameters of the underlying physical mechanism, in Koopmans and Riersol’s phrase. This paper takes seriously the answers to expectational questions, as Manski urges us to do but also shows the power of theory, when combined with these subjective opinions, to inform us about structure. Manski argues that ‘economists have hardly begun to use probabilistic expectations data in econometric analysis of decision making’. But, as this paper shows, that agenda began at least a quarter of a century ago. Praiseworthy as Manski’s plans for serious econometric analysis of expectations and intentions data are, economists continue to display a curious passivity about data collection (though some economists do get to add questions to existing surveys). I would have expected to find them applying for funding to ask the questions suggested by our theories; to ask the unemployed person what is the smallest wage he would accept in a full time job, as the sociologists and economists of Political and Economic Planning did in collecting the data used by Lancaster and Chesher in the 1970s; to ask the woman how many hours she would work at a given wage or if child care was available at a given price and so on. But it seems that much of the data analysed by Manski and his colleagues is gathered by non‐economists and for a variety of different reasons. This may be cheap but it can lead to apparent absurdities like the question, discussed by Manski in chapter 14, asked by the Current Population Survey of the US Bureau of the Census, namely ‘Looking ahead, do you expect to have any (more) children?’; with answers ‘Yes, No, Uncertain’. The word expect suggests that the arrival of children is comparable to a thunderstorm or other acts of nature, partially predictable maybe but wholely out of anyone’s control. The questions seems comparable to, say, ‘Do you expect it to rain tomorrow?’, though it is described as providing information about intentions. Economists would surely want to ask instead about people’s plans for children, plans which, of course, may or may not be realised. Perhaps political considerations to do with the acceptability of family planning motivate this choice of word by a government agency but it does not make life easy for an econometrician. And this remark points to an omission in the author’s generally extensive literature survey. As is well known, subjective Bayesians analyse data by showing how prior beliefs are changed by evidence into new or posterior beliefs. An important part of this process is finding out what the prior beliefs of your client or subject are, a problem that goes under the heading of elicitation of beliefs. ‘Elicitation is the process of formulating a person’s knowledge and beliefs about one or more uncertain quantities into a (joint) probability distribution for those quantites,’ (Garthwaite et al., 2005, p. 680). The paper from which this quotation is taken is one of several recent surveys of a large and growing literature dealing with this problem of eliciting beliefs, which is relevant to econometric work but is ignored in this book. If economists wish to collect data about beliefs then the Bayesian literature would appear relevant. To summarise, this is indeed a brilliant book and a rewarding read for any econometrician, full of insights and thought‐provoking remarks. While it will be clear from the tone of my comments that I find Manski’s new ‘worldview to guide empirical researchers’ to be unhelpful, a less grandiose interpretation of his contribution, in which the emphasis is on the construction of bounds from incomplete data, has clearly proved to be a valuable development for frequentist econometricians.3 Many will enjoy this book as much as I did. Footnotes 1 " Partial Identification of Probability Distributions (2003) and Social Choice with Partial Knowledge of Treatment Response (2005). 2 " ‘Inference using the empirical evidence alone sacrifices strength of conclusions in order to maximise credibility.’ (p. 62) 3 " Bounds are an automatic by‐product of an analysis based on likelihoods. References Blundell R. , Gosling , A., Ichmiura , H. and Meghir , C. ( 2007 ). ‘ Changes in the distribution of male and female wages accounting for employment composition using bounds ’, Econometrica , vol. 75 ( 2 ) (March), 323 – 63 . Google Scholar Crossref Search ADS WorldCat Daniels M. J. and Hogan J.W. ( 2008 ). Missing Data in Longitudinal Studies , Boca Raton, Flda: Chapman and Hall . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Friedman M. ( 1953 ). The Methodology of Positive Economics, in Essays in Positive Economics , Chicago: University of Chicago Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Friedman M. ( 1957 ). A Theory of the Consumption Function , Princeton: Princeton University Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Garthwaite P.H. , Kadane , J.B. and O’Hagan , A. ( 2005 ). ‘Statistical methods for eliciting probability distributions’ , Journal of the American Statistical Association , vol. 100 ( 470 ), pp. 680 – 700 . Google Scholar Crossref Search ADS WorldCat Hall R.L. and Hitch , C.J. ( 1939 ). ‘Price theory and business behaviour’ , Oxford Economic Papers , vol. 2 , pp. 12 – 45 . Google Scholar Crossref Search ADS WorldCat Koopmans T.C. and Reiersol , O. ( 1950 ). ‘ The identification of structural characteristics ’, The Annals of Mathematical Statistics , vol. 21 ( 2 ) (June), pp. 165 – 81 . Google Scholar Crossref Search ADS WorldCat Lancaster T. ( 2004 ). An Introduction to Modern Bayesian Econometrics , Oxford: Blackwells . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Lancaster T. and Chesher , A.D. ( 1983 ). ‘An econometric analysis of reservation wages’ , Econometrica , vol. 51 ( 6 ), pp. 1661 – 76 . Google Scholar Crossref Search ADS WorldCat Little R.J.A. and Rubin D.B. ( 2002 ). Statistical Analysis with Missing Data , Chichester: John Wiley and Sons . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Machlup F. ( 1946 ). ‘Marginal analysis and empirical research’ , American Economic Review , vol. 36 ( 4 ), pp. 519 – 54 . OpenURL Placeholder Text WorldCat Manski C.F. ( 1995 ). Identification Problems in the Social Sciences , Cambridge, Mass: Harvard University Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Manski C.F. ( 2003 ). Partial Identification of Probability Distributions , New York: Springer‐Verlag . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Manski C.F. ( 2005 ). Social Choice with Partial Knowledge of Treatment Response , Princeton: Princeton University Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC © The Author(s). Journal compilation © Royal Economic Society 2009

doi: 10.1111/j.1468-0297.2009.02284.xpmid: N/A

The list of thesis titles has been compiled by Amanda Wilman, the Administrator of the Royal Economic Society. It contains the titles for theses for higher degrees taken entirely by dissertation and it relates to degrees awarded in the academic years 2007/08 and early 2008/09. The subject classification is that used by this journal for Book Notes and is derived from the classification devised by the American Economic Association and used in the Journal of Economic Literature since March 1991. (It will be appreciated that assigning some titles to a single category is somewhat arbitrary.) Other lists of thesis titles have appeared in the March issue of the Economic Journal, beginning in 1973 (from 1991, the May issue and in 2000 in the June issue). Indication of availability (i) Abstract available through inter‐library loan (ii) Abstract available from university library (iii) Thesis available through inter‐library loan (iv) Copy of thesis available from university library (v) Thesis available at university library for consultation only (vi) Thesis not available until after publication (vii) Micro‐film copy of thesis available from British Library (Document Supply Centre, Boston Spa) C. MathematicalandQuantitativeMethods C.0 General Franz, J., Quantification of Environmental Health Effects in Central Asia: linking agriculture and child mortality in Central Asia, Ph.D., June 2007, St Andrews, (iv, v). Arunsawadiwong, S., Productivity Trends in the Thai Manufacturing Sector: pre‐ and post‐crisis evidence relating to the 1997 economic crisis, Ph.D., November 2007, St Andrews, (iv, v). Massacci, D., Econometric Analysis of Financial Contagion, Ph.D., January 2008, Cambridge, (v). C.1. Econometric and statistical methods: general Humpe, A., Can Macroeconomic Variables Explain Stock Market Movements?, Ph.D., June 2008, St Andrews, (iv, v). Ujjual, V., High Technology Performance, Innovation, and Networks – an applied econometric analysis of firms in Scottish high technology clusters, Ph.D., June 2008, St Andrews, (iv, v). C.5 Econometric modelling Reade, J.J., Macroeconomic Modelling and Forecasting in the Face of Non‐Stationarity, D.Phil., August 2007, Oxford‐St.Cross, (iv) Carceres, C., Asymptotic Properties of Tests for Mis‐Specification, D.Phil., February 2008, Oxford‐Nuffield (iv) Fawcett, N., Essays in Econometrics and Forecasting, D.Phil., June 2008, Oxford‐New, (iv) C.6 Mathematical models and programming Mutemererwa, A.M.., A Nonparametric Approach to Productive Efficiency Measurement: An application of bootstrap DEA to gold mining, Ph.D., January 2008, Hull, (iii,v). C.7 Game theory and bargaining theory Wallace, P., Pre‐emption and Attrition: Essays on Timing Games, D.Phil., May 2008, Oxford‐St. Edmund, (iv). D. Microeconomics D.0 General Firia‐Cuchra, M., On Securitisation of Assets, D.Phil., May 2007, Oxford‐Wadham, (iv). Tse, M., Essays on Information and Communication, D.Phil, July 2007, Oxford‐Nuffield (iv). Kawamura, K., Cheap Talk Communication and Information Sharing in Organisations, D.Phil., July 2007, Oxford‐Nuffield (iv). Rohner, D., Essays on the Economics of Conflict and Political Violence, Ph.D., March 2008, Cambridge, (v). Eterovic, D., Essays in Political Economics, Effects of Institutions on Policy Outcomes, Ph.D., May 2008, Cambridge, (v). Qiao, Y., On the Determinants of Initial Public Offering Underpricing, Ph.D., November 2008, St Andrews, (iv, v). D.1. Household behaviour and family economics Malesovich, L., Inflation, Growth and Happiness: Assessing Croatia’s stabilisation policy, Ph.D., December 2007, Staffordshire, (i). Menon, M., Theory and Models of Family Behaviour: Labour Participation, Household Production and Fertility Choices, Ph.D., May 2008, York, (iv, vii). D.4. Market structure and pricing Kushi, E.C., Information Asymmetry, Quality and Prices in the Tourism Market: An application to Albanian holiday hotels., Ph.D., June 2008, Staffordshire, (i). Xu, Z., The Growth of Privately Owned Firms in China: An entrepreneurial analysis, Ph.D., June 2008, St Andrews, (iv, v). D.5 General equilibrium and disequilibrium Alshiekh, A., A Dynamic Computable General Equilibrium Analysis of the Saudi Arabian Economy, Ph.D., January 2009, Manchester, (v). D.6 Economic welfare Dzarasov, R., Insider Control and Investment Behaviour of Russian Corporations, Ph.D., May 2008, Staffordshire, (i). D.8. Information and uncertainty Galizzi, M., Essays on Thin Markets, Networks, Bargaining: Theory and Experiments, Ph.D., May 2008, York, (iv, vii). E. MacroeconomicsandMonetaryEconomics E.0 General Trew, A., Towards the Microfoundations of Finance and Growth, Ph.D., June 2007, St Andrews, (iv, v). Kuralbayeva, K., Macroeconomic Adjustment to Resource Shocks in a Growing Two‐Sector Economy, D.Phil., September 2007, Oxford‐Lincoln (iv). Ofoghi, R., Essays on the Macroeconomic Impact of Insurance Market Integration, Ph.D., March 2008, Southampton, (i, ii, iv, vii) Alper, K., The Monetary Transmision Mechanism in Emerging Market Economies: The Turkish Case, Ph.D., April 2008, Manchester, (v). Chudik, A.,Global Macroeconomic Modelling, Ph.D., June 2008, Cambridge, (v). Newby, E., Sustainable Monetary Policy – Lessons and Evidence from the Bank Suspension Period 1797–1821, Ph.D., June 2008, St Andrews, (iv, v). Pfajfar, D., Heterogeneous Expectations in Macroeconomics, Ph.D., December 2008, Cambridge, (v). Alptekin, A., Military Expenditure, Institutions and Conflict, Ph.D., 2008, Surrey, (iv, v). E.1 General aggregative models Alshehabi, O., Macroeconomic Modelling of the Labour Market, D.Phil., 2008, Oxford‐Pembroke, (iv). Saroliya, G., Persistence in Heterogeneous New‐Keynesian Economics, Ph.D., May 2008, York, (iv, vii). E.2 Consumption, saving, production, employment and investment Liu, Y., Three Essays in New Open Economy Macroeconomics with Multiple Countries, Non‐tradable Goods, Capital Accumulation and Distribution Sector, Ph.D., January 2009, Southampton, (i, ii, iv, vii). E.3 Prices, business fluctuations and cycles Malone, S., Essays on Emerging Market Macroeconomics: Inequality and Resource Booms, Dealing with Debt Overhang, and External Volatility and Country Risk, D.Phil., May 2007, Oxford‐Balliol, (iv). Sheedy, K., Essays on Price Stickiness, Ph.D., May 2008, Cambridge, (v). Anderson, R., Essays on US Consumer Inflation Expectations, Ph.D., November 2008, Manchester, (v). E.4 Money and interest rates Horvarth, M., Optimal Monetary and Fiscal Policy in Economies with Multiple Distortions, Ph.D., June 2008, St Andrews, (iv, v). E.5 Monetary policy, central banking, and the supply of money and credit Hesse, H., Essays on Banking and Monetary Policy, D.Phil., February 2007, Oxford‐Nuffield, (iv). James, J. G., Union Objectives, Monopolistic Competition and Macroeconomic Externalities, Ph.D., February 2007, Swansea, (iii). Stehn, S.J., The Interactions Between Optimal Monetary Policy and Optimal Fiscal Policy, D.Phil., September 2007, Oxford‐Brasenose, (iv). Cruz, A., Exchange Rate Regimes, Fiscal Stance and Currency Crises, Ph.D., 2007, Surrey, (iv, v). Syrrakos, D., European Monetary Unification: progressive evolution or a discrete change? An analysis of intra‐European Union exchange rate linearities, Ph.D., January 2008, Manchester, (iii, vii). Besimi, F., Monetary and Exchange Rate Policy in Macedonia During the Process of Accession to the European Union, Ph.D., May 2008, Staffordshire, (i). Maziad, S., Monetary Frameworks in Developing Countries: Central Bank Independence and Exchange Rate Arrangements, Ph.D., June 2008, St Andrews, (iv, v). Wang, Q., Determinancy and Learning Stability for Monetary Policy in Two‐Country Models, Ph.D., July 2008, Cambridge, (v). Wong, M.S., Monetary Policy Analysis in a Fixed Exchange Rate Economy with Credit Market Imperfections – A Theoretical and Empirical Exploration for Malaysia, Ph.D., September 2008, Manchester, (v). Shijaku, H., An Assessment of Monetary Policy in Albania since 1990, Ph.D., November 2008, Staffordshire, (i). Simic, V., An Assessment of Monetary Policy in Croatia with Reference to Euroisation, Ph.D., November 2008, Staffordshire, (i). Sangduan, P., The Dynamics of Fiscal Sustainability, Ph.D., 2008, Surrey, (iv, v). E.6 Macroeconomic aspects of public finance, macroeconomic policy and general outlook Polito, V., Vector Autoregressive Analysis of Macroeconomic Policy, Ph.D., December 2008, York, (iv, vii). F. InternationalEconomics F.0 General Arpac, O., The Implementation of IMF Supported Programmes: An empirical investigation using complementary methodologies, Ph.D., 2007, Surrey, (iv, v). Xu, Y., The Impact of Regulation on the Formation of Entry Strategy of Multinational Banks, a Case of Foreign Investment into the Chinese Banking Industry, Ph.D., June 2008, Cranfield (i, ii, iii). Boakye, A., Empirical analysis of monetary and growth aspects of the idcs: hipcs vs non‐hipcs, D.Phil., October 2008, Dundee, (iv). Van, C.L., The Effects of Foreign Direct Investment on Economic Growth and Income Convergence in ASEAS Countries, D.Phil., October 2008, Dundee, (iv). F.1 Trade Ismail, N., Economic Integration and Intra Regional Trade: The Evidence from the ASEAN Free Trade Area, Ph.D., October 2007, Southampton, (i, ii, iv, vii) Bournakis, I., Competitiveness, Trade and Productivity: With Special Reference to Greece, Ph.D., March 2008, Kent, (iv). Lee, X., Essays on International Trade and Regional Economic Integration in East Asia, Ph.D., June 2008, Southampton, (i, ii, iv, vii). F. 2 International factor movements Dada, W., Financial Aspects of Monetary Union, Ph.D., June 2008, Cambridge, (v). F.3 International finance Nogueira Jr., R.P., Essays on Inflation Targeting and Exchange Rate Pass‐Through, Ph.D., October 2007, Kent, (iv). F.4 Macroeconomic aspects of international trade and finance Nawi, A., Essays on Total Factor Productivity, International Trade, Business Cycles and Mark‐up, Ph.D., November 2007, Southampton, (i, ii, iv, vii). Coric, B., Globalisation, Capital Market Imperfections and Decline in Short‐Run Output Volatility of World Economies, Ph.D., December 2008, Staffordshire, (v). G. FinancialEconomics G.0. General Ball, S., Self‐Insurance and Public Insurance over the Life‐Cycle, Ph.D., November 2008, Cambridge, (v). Grisse, C., Essays on Expectations and Financial Markets, Ph.D., November 2008, Cambridge, (v). Lentzas, G.A., Essays in Financial Economics, D.Phil., 2008, Oxford‐Balliol, (iv). G.1. General financial markets Kartsaklas, A., Long Memory, Structural Breaks and the Volatility‐Volume Relationship, Ph.D., April 2008, York, (iv, vii). Liu, Z., Essays on Two‐Country Macro‐Finance Models, Ph.D., May 2008, York, (iv, vii). Kelly, B., Sunk Cost Accounting and Entrapment in Corporate Acquisitions and Financial Markets: An experimental study, Ph.D., June 2008, St Andrews, (iv, v). Zhu, Y., Essays on the Chinese Financial Markets, Ph.D., Southampton, (i, ii, iv, vii). Kostakis, A., Essays on Dynamic Asset Allocation and Performance Measures, Ph.D., August 2008, York, (iv, vii). Evans, K. P., Macroeconomic News Announcement Effects on Financial Markets, Ph.D., 2008, Swansea (iii). Bienkowska, A., Dynamics of an Electronic Stock Exchange: Modelling and Forecasting Joint Evolution of liquidity on Ask and Bid Sides of Electronic Limit Order Book, M.Litt., 2008, Oxford‐Nuffield, (iv). Gang, J., Volatility Analysis on Macroeconomy and Financial Market, Ph.D., January 2009, York, (iv, vii). G.2 Financial institutions and services Scarfiglieri, G., Essays on Universal Banking, Economies of Scale and Scope, M.Phil., December 2008, York, (iv, vii). Chow, J., The Study of the Causal Relationship of Financial Development and Economic Growth, Ph.D., 2008, Surrey, (iv, v). G.3 Corporate finance and governance Borges Teixeira dos Santos, F., Corporate governance in economic development. A micro‐macro interaction perspective for a new approach to the theory of the firm in developing countries, Ph.D., September 2008, Cambridge, (v). Dong, M., Institutional Investors, Corporate Financial Decisions and Performance in UK Firms, Ph.D., April 2008, York, (iv, vii). H. PublicEconomics H.0 General Powell, J., Public Policy, Bureaucratic Corruption and Economic Development, Ph.D., June 2008, Manchester, (v). Toci, V., Financial Development in Transition Economies with Special Reference to South East Europe and Kosova, Ph.D., August 2008, Staffordshire, (i). Yilmaz, D. S., Productive Spending, Fiscal Policy and Growth, Ph.D., November 2008, Manchester, (v). Wallis, A., A ‘Mixed Method Appraisal Framework for Public Sector Project and Programme Ex‐Ante Appraisal in the Regeneration Domain, M.Phil., January 2009, Manchester, (v). I. Health, Education andWelfare I.1 Health Gossage, D., Empirical Essays in Health Economics, D.Phil., May 2007, Oxford‐Queen’s, (iv). Ghislandi S., Essays in the Economics of Pharmaceuticals, D.Phil., June 2008, Oxford‐Wolfson, (iv). Robone, S., Essays in Applied Health Economics: Evidence on Health and Health Care in Italy and UK, Ph.D., October 2008, York‐Bologna, (iv, vii). I.2. Education Fagernas, S., Empirical Essays on Labour Markets, Governance and Institutions in India, Ph.D., October 2008, Cambridge, (v). Jewell, S., Human Capital Acquisition and Labour Market Outcomes in UK Higher Education, Ph.D., December 2008, Reading, (v). I.3 Welfare and poverty Lugo, M.A., On Multidimensional Distributions of Well‐Being, D.Phil., May 2008, Oxford‐St. Antony’s, (iv). Zarazua, M.N., The impacts of microcredit on income poverty, labour and well‐being: A quasi‐experimental study in urban Mexico, Ph.D., April 2008, Sheffield, (v). Valente, C., The Land Transfer Programme in South Africa: Impact on beneficiary welfare, Ph.D., May 2008, Sheffield, (v). J. LabourandDemographicEconomcis J.0. General Behar, A., Are Skilled and Unskilled Labour Complements or Substitutes, D.Phil., July 2007, Oxford‐Nuffield (iv). Kanellopoulos, K., Short‐term Pay and Self‐employment Dynamics in European Union Countries in the 1990s, Ph.D., June 2008, Manchester, (v). Lopez Boo, F., Essays on Income Inequality, the Skill Premium and Schooling Decisions: Evidence from Argentina, D.Phil., October 2008, Oxford‐St.Antony’s (iv). Slattery, D., The Impact of Regulation on the UK Retail Pension Market, Ph.D., November 2008, Cranfield, (i, ii, iii). J.1 Demographic economics Omtzigt, D‐J., Demographic Change, Individual Decision Making and the Feasibility and Fairness of Policy Options, D.Phil., October 2008, Oxford‐Exeter, (iv). J.2 Time allocation, work behaviour and employment determination Allan, S., The Labour Supply and Retirement of Older Workers; an empirical analysis, Ph.D., October 2008, Kent, (iv). J.3. Wages, compensation and labour costs Li, Y., The Accumulation of Human Capital: Evidence from a UK Longitudinal Study, Ph.D., November 2008, York, (iv, vii). J.4 Particular labour markets Grazier, S. J., Empirical Essays on Occupational Health and Safety, Ph.D., January 2008, Swansea, (iii). Jones, K.J., An Investigation into the Impact of Disability on Labour Market Outcomes in the UK, Ph.D., February 2008, Swansea, (iii). J.5 Labour‐management relations: trade unions and collective bargaining Chrysanthou, G., Estimating Trade Union Membership Determinants and Wage Effects Using Alternative Specifications, Ph.D., October 2008, York, (iv, vii). Anyalezu, N.K.G., An Empirical Examination of Technology Shocks and Employment in the UK, Ph.D., 2007, Surrey, (iv, v). J.6 Mobility, unemployment and vacancies Yalonetsky, G., Essays on Economic Mobility, D.Phil., October 2008, Oxford‐Oriel, (iv). K. LawandEconomics K.0 General Qiao, Y., Funding Arrangements in the Modern Market for Legal services, Ph.D., 2008, Surrey, (iv, v). L. IndustrialOrganisation L.0 General Shaukat Khan, T., Essays on the Dividend Behaviour and Ownership Structure of Firms in the UK, D.Phil., November 2007, Oxford‐Nuffield, (iv). L.1 Market structure, firm strategy and market performance Alvi, I., Behavioural Price Discrimination and Personalisation Strategies, D.Phil., February 2007, Oxford‐Balliol (iv). Simpson, H., Essays on Firm Location Decisions, D.Phil., November 2007, Oxford‐Nuffield, (iv). Pezzino, M., Horizontal and Vertical Product Differentiation; Market Equilibrium, Welfare and Quality, Ph.D., May 2008, Manchester, (v). Pinch, D., Three Essays on Competition Policy and Innovation, Ph.D., May 2008, Southampton, (i, ii, iv, vii). Demirel, P., Firm Growth, Innovation and Implications for Market Selection: The Pharmaceutical Industry, Ph.D., June 2008, The Open University, (i, ii). Rudkin, S., The Impact of Supermarkets on Prices, Consumer Behaviour and Welfare: theoretical and empirical issues, Ph.D., November 2008, Manchester, (v). L.2. Firm objectives, organisation and behaviour Ormanidhi , O., Porter’s Model of Generic Competitive Strategies: A study of three industries in Albania, M.Phil., October 2007, Staffordshire, (i). Wilson, L., Essays on the Financial Governance of Firms, D.Phil., November 2007, Oxford‐St.Cross, (iv). L.5. Regulation and industrial policy Pollock, R., Should we give every cow its calf? Monopoly, Competition and Translations Costs in the Promotion of Innovation and Creativity, Ph.D., February 2008, Cambridge, (v). Adeyemi, O.I., Modelling OECD Industrial Energy Demand, Ph.D., 2008, Surrey, (iv, v). L.6. Industry studies: manufacturing Elshamy, H.M., Productivity growth in the small firm sector in UK manufacturing, Ph.D., 2008, Surrey, (iv, v). Saravia, S., Institutional Change, Transaction Costs and Efficiency: The Case of the Nicaraguan Coffee Industry, Ph.D., September 2008, Sheffield, (v). L.9. Industry studies: utilities and transport Keetharuth, D., Privatisation and liberalism of the telecommunication sector: A case study of Mauritius, Ph.D., February 2008, Sheffield, (v). N. EconomicHistory N.1 Macroeconomics: growth & fluctuation Bateman, V., Market Integration and Growth in Europe: The Early Modern Era, D.Phil., June 2007, Oxford‐Jesus, (iv). N.6 Manufacturing and Construction Barley, S., Hand Tool Manufacture During the Industrial Revolution: Saw Making in Sheffield c.1750‐c.1830, Ph.D., July 2008, Sheffield, (v). O. EconomicDevelopment O.0 General Laho, E., An Analysis of the IMF Supported Programmes in Transition Economies with Special Reference to Albania, Ph.D., May 2008, Staffordshire, (i). Vazquez‐Guzman D., Measurement of Income Inequality in Mexico: Methodology, Assessment and Empirical Relationship with Poverty and Human Development, Ph.D., September 2008, Stirling, (i, ii, iii, iv, v, vi, vii). Frangie, S., The Good Governance Agenda, Weak States and Economic Development: The Potential Economy of Consensus in Post‐Civil War Lebanon, Ph.D., December 2008, Cambridge, (v). O.1 Economic development Calvo, C., Essays on the Links between Risk and Poverty, D.Phil., May 2007, Oxford‐Wadham, (iv). Bold, T., Economics of Informal Insurance Arrangements, D.Phil., October 2007, Oxford‐Nuffield, (iv). Bokosi, F.F.K., Essays on Poverty and Trade Liberalisation in Malawi (1995–2002), Ph.D., December 2007, Kent, (iv). Khonsoomboon, P., Empirical Essay on Development and Trade: Case Study in Thailand, Ph.D., March 2008, Southampton, (i, ii, iv, vii). Bhattacharya, V., Poverty and School Enrolment: A Study of Two Provinces in India, M.Phil., October 2008, Southampton, (i, ii, iv, vii). Sandefur, J., Essay on Labour and Credit Markets in Africa, D.Phil, October 2008, Oxford‐Queen’s, (iv). Rijkers, B., Small Enterprise Performance and Labour Market Outcomes in Ethiopia, D.Phil., November 2008, Oxford‐Oriel, (iv). Mukiza, C., Macroeocnomic Policies, Growth, Inequality and Poverty Reduction in Uganda, Ph.D., December 2008, Southampton, (i, ii, iv, vii). Yang, B., DSGE Modelling, Application to Developed and Developing Economies, Ph.D., 2008, Surrey, (iv, v). O.2 Development Planning and Policy Segal, P., Economic Structures and Long‐Run Distribution of Income in Argentina, D.Phil, May 2007, Oxford‐Nuffield, (iv). Leon, G., Institutions and Economic Policy in the Developing World, D.Phil., January 2008, Oxford‐St. John’s, (iv). Chriipanhura, B., Labour market dynamics and economic policy in Zimbabwe 1980–2005, Ph.D., February 2008, Sheffield, (v). O.3 Technological change Pavon‐Villamayor, V., Economics of Technological Convergence, D.Phil., January 2008, Oxford‐Hertford, (iv). O.4 Economic growth and aggregate productivity Dixon, P, Empirical Essays on Inputs, Institutions and Economic Growth, D.Phil, June 2007, Oxford‐Nuffield, (iv). Voitchovsky, S., Inequality and Growth, D.Phil, September 2007, Oxford‐Exeter, (iv). Ritzmann, D., Economic Growth, Convergence and Disparities in Productivity: A World Production Frontier Approach, D.Phil., October 2007, Oxford‐Pembroke, (iv). Panahi, H., Geographical Proximity and Economic Growth, Ph.D., January 2008, Hull, (iii,v). Navajas, A. R., Socio‐Political Determinants of Economic Growth, Ph.D., October 2008, Manchester, (v). Ibrahim, S., Essays on the East Asian Economic Integration, Ph.D., November 2008, York, (iv, vii). P. EconomicSystems P.0 General Majidov, T., Regional Entrepreneurship and Business in Uzbekistan: Historical Perspective and Current Obstacles, Ph.D., March 2008 (i, ii, iii, iv, v, vi, vii, STORE). Q. Agriculture andNaturalResourceEconomics Q.0 General Di Corato, L, Essays on Information Gathering and the Use of Natural Resources Ph.D., York‐Padua, December 2008, Stirling, (v, vii). Q.3 Non‐renewable resources and conservation Hatcher, A., ITQ Markets with Non‐compliance and Market Power, Ph.D., 2008, Surrey, (iv, v). Q.4 Energy Chontanawat, J., Causality Between Energy Consumption and Economic Growth: evidence from over 100 countries, Ph.D., 2007, Surrey, (iv, v). Tuthill, L.., Emissions Policy and the US Electricity Generating Industry: Capital Investment Fuel Use and Cost Efficiency, D.Phil., October 2008, Oxford‐St. Anne’s (iv). Mahmud, H, Oil, Institutions and Growth, Ph.D., 2008, Surrey, (iv, v). R. Urban, Rural andRegionalEconomics R.1 General spatial economics Barde, S., A Comparative Economic Analysis of Agglomeration Theory, Ph.D., July 2007, Kent, (iv). R.4 Transport Broadstock, D.C., Traffic Demand and Land‐use in the UK: an econometric analysis using the TRICS database., Ph.D., 2008, Surrey, (iv, v). © Royal Economics Society. Journal compilation © Royal Economic Society 2009

Weinberg, Bruce, A.

doi: 10.1111/j.1468-0297.2009.02275.xpmid: N/A

Abstract The twentieth century saw dramatic international shifts in scientific leadership. Despite these dramatic shifts Britain’s position has been remarkably stable and strong. I study these changes using data on Nobel laureates in Chemistry, Medicine, and Physics. Raw data show a slight decline in British science, mainly in physics but once one accounts for the tremendous increase in the US, British science actually shows strong growth. I show that raw data and data that adjust for population and gross domestic product (per capita or total), consistently rank Britain as one of the top scientific performers. The twentieth century saw large changes in the scientific leadership. At the beginning of the twentieth century, Europe dominated science. Today the US does. Beyond the rise of the US to scientific leadership, the twentieth century saw the dramatic decline of Germany; see Weinberg (2007) for an analysis of these trends. This article argues that despite these large shifts, at least one feature of international science has shown a remarkably degree of stability – Britain. When the century opened, Britain was one of the strongest players, which is where it stands today. Thus, science in Britain does not follow the dramatic growth trajectory of the US but nor does it follow the tragic decline of Germany. My data show that Britain has been and remains an important scientific centre, ranking second after only the US. Although the trends are small, there is some evidence that Britain’s scientific position increased during the early part of the twentieth century and that it may be declining slightly today, with most of the decline in physics. Britain does not perform better largely because of the exceptional experience of the US. Once one accounts for the tremendous increase in the US, British science shows remarkable growth. I also show that overall and adjusting for population and gross domestic product (per capita or total), Britain is consistently one of the top scientific performers. Indeed the English‐speaking countries as a whole perform consistently well in science. Others have sought to provide quantitative assessments of the scientific competitiveness of nations. Nelson and Wright (1992) study the growth in science in the US. Cole and Phelan (1999) and Aizenman and Noy (2007) present cross‐national analyses. There is also a literature focusing on science in Britain and largely arguing that British science is in decline; see Irvine et al. (1985), Smith et al. (1986) and Martin (1994). Kealey (1991) presents an opposing view. Hunter et al. (2009) focus particularly on Britain and the brain drain from Britain. Most of the evidence I present comes from recipients of the Nobel Prizes in Chemistry, Medicine and Physics. While others have employed similar strategies, I believe that my data, which both draw on and build on data by Jones (forthcoming) and Stephan and Levin (1993), are superior to that used in previous studies in two important ways. First, they contain information on when people did the work for which they received the Nobel Prize. Most work has used information on when people received the Nobel Prize, which is, on average, a decade after they did the work for which they received the Prize. The data also contain complete histories of scientists’ institutional affiliations. Most work has studied where people were born or where they were when they received the Nobel Prize but scientists, especially eminent scientists, are highly mobile and these snapshots miss much of the mobility over people’s careers. To probe for potential biases due to the Nobel selection process and ensure that my results for the Nobel fields (chemistry, medicine and physics) are representative of science more broadly, I have also obtained data on highly‐cited researchers from the Institute for Scientific Information’s Highly Cited database. Using these data, I study the countries where the roughly 250 most cited people in 21 disciplines are affiliated. These results corroborate the strong position found for Britain in the data on Nobel Prizes. The strength of national scientific communities is particularly important today, as science increasingly provides a foundation for industrial progress. While the prizes studied here are for basic research, in many cases the awarded works have lead directly or indirectly to important new technologies. For instance, the discovery of X‐rays, the understanding of semi‐conductors and the fabrication of quinine were all awarded Nobel Prizes and had direct technological impacts. The discovery of the structure of DNA and advancements in quantum physics, for instance, have also had tremendous economic and military impact, although in these cases the connection is less immediate. There is also evidence that new technologies tend to be commercialised where scientific innovations take place; see Zucker et al. (1998) for evidence on biotechnology. A study of scientific strength is particularly timely as Europe, US and the rest of the developed world increasingly look to innovation as a way of preserving economic strength, especially in the face of rising economic strength in Asia; see, for instance, HM Treasury (2004, 2006) and National Academies (2007). 1. Data My data contain two components: (1) the institutional affiliations of each Nobel Laureate in each year of his or her career and (2) the years in which each Nobel laureate did his or her Nobel Prize‐winning work. The latter data combine and build on data that were assembled by Benjamin Jones and Paula Stephan and Sharon Levin. These two data sets differ in terms of their constructs. Jones’s (forthcoming) data gives the year in which each laureate did his or her Prize‐winning work. Stephan and Levin’s (1993) data gives the year in which each laureate started (and ended) the research agenda for which he or she received the Nobel Prize. I focus on the former construct here. The data also contain a variety of other background information, including the institutions that the laureates attended for any bachelors, masters or doctoral work and their institutional affiliations in every year of their career. My use of data on Nobel laureates warrants some discussion. I regard Nobel laureates as a sample of important innovators. In focusing on this sample, I do not seek to imply that these scientists are the most important scientists in their fields, only that they are important scientists. The focus on important innovators is particularly interesting insofar as innovation is one of the economically important activities in which small numbers of important contributions are most significant. An advantage of studying Nobel laureates is that considerable data becomes available on people when they receive the Nobel Prize making it possible to reconstruct life histories of people active in the 19th century. To address concerns with the representativeness of Nobel laureates and the fields that the Nobel Prizes cover, I use data on highly‐cited researchers from the Institute for Scientific Information’s Highly Cited. These data comprise the roughly 250 most cited researchers in each of 21 fields as well as their current institutional affiliation. I also use data on population and per capita gross domestic product to control for the size and wealth of countries. These data were taken from the Penn World Tables, Version 6.2. As discussed below, I use data from 1970, which are available for all the countries I consider but Russia. Gross domestic product (total and per capita) are expressed in 1970 dollars converted according to purchasing power parity. While my focus is on Britain, my analyses implicitly or explicitly compare Britain to other parts of the world. For these analyses, I define Western European offshoots to be Australia, Canada, New Zealand and the US. The ‘developed world’ is defined as Western Europe, its offshoots and Japan. 2. Results Figure 1 presents a first cut through the data. It plots the share of Nobel Prizes awarded for work done in Britain by the year that the work was done, pooling data for Chemistry, Medicine and Physics. This graph and the ones that follow show a series of dots for whether the work was done in Britain. In these graphs, a 1 indicates a prize awarded for work done in Britain and a 0 indicates a prize awarded for work done outside Britain. A small share of people were in more than one location in the year in which they did their prize‐winning work and these people’s prizes have been pro‐rated among the countries where they were located. The graphs also show a solid curve. For each year, this curve gives the probability that work done in that year will be done in Britain. These curves are from kernel regressions. This technique is attractive because it captures movements in the data very flexibly but the estimates can get noisy where there is relatively little data, especially at the beginning and end of the period studied. To provide some indication of the precision of the estimates, the grey region shows a 95% confidence interval. Viewing the Nobel Prize data as a sample of important scientific contributions, in another large sample of comparable contributions, the share that would be done in Britain would be in this shaded range 95% of the time.1 Consistent with expectations, the size of this confidence interval widens considerably at either end of the period. Fig. 1. Open in new tabDownload slide The Probability that Prize Winning Work Was Done in Britain
Note. A circle at 1 indicates prize‐winning work done in Britain in a given year and a circle at 0 indicates prize‐winning work done outside of Britain in that year. People who were in multiple locations in the year that they did their prize‐winning work were pro‐rated across locations. The curve gives flexible, non‐parametric (kernel) estimates of the probability that prize‐winning work done in a given year was done in Britain. The shaded region gives a 95% confidence interval for the probability that prize‐winning work is done in Britain in a given year. Fig. 1. Open in new tabDownload slide The Probability that Prize Winning Work Was Done in Britain
Note. A circle at 1 indicates prize‐winning work done in Britain in a given year and a circle at 0 indicates prize‐winning work done outside of Britain in that year. People who were in multiple locations in the year that they did their prize‐winning work were pro‐rated across locations. The curve gives flexible, non‐parametric (kernel) estimates of the probability that prize‐winning work done in a given year was done in Britain. The shaded region gives a 95% confidence interval for the probability that prize‐winning work is done in Britain in a given year. A quick appraisal of the graph shows some increase at the beginning of the period and some decline towards the end of the period but a more careful analysis indicates that the increase and decrease are driven by small numbers of observations and that they are not statistically significant at conventional levels. Abstracting from the noise in the series, the graph shows a remarkable level of stability for Britain at about 15% of Nobel Prizes. While one might be concerned based on the most recent years, which show a downturn in Britain, the datas are sparse at that point and any conclusions must be tentative. To study the overall features of the graph, I regressed the probability that a Nobel Prize was received for work done in Britain on the year in which the work was done.2 The estimates are (1) The intercept of 0.148 implies than in 1944, roughly the (prize‐weighted) midpoint of the data, slightly under 15% of Nobel Prizes were awarded for work done in Britain. The slope coefficient of −0.00043 indicates that the share of Nobel Prizes awarded for work done in Britain is trending downward at a small (and statistically insignificant rate). Given the imprecision of the estimate, any extrapolations based on it must be taken with caution but, if taken literally, the estimate of −0.00043, implies that over the course of a decade, the probability that any given Nobel Prize will be awarded for work done in Britain declines by less then half of a percentage point. While small increases or declines in science in Britain cannot be ruled out based on these estimates, the estimates are sufficiently precise to rule out large changes, making it clear that Britain’s scientific position has remained remarkably stable over the course of the twentieth century. 2.1. Comparisons to the United States and Germany Germany held the strongest scientific position at the beginning of the twentieth century, while the US does today. This section explores Britain’s position relative to these two countries. Figure 2 (a) is similar to Figure 1 except that it focuses on Prizes for work done in either Britain or the US. Thus, a point at 1 continues to indicate Prize‐winning work done in Britain in a given year but a point at 0 now indicates Prize‐winning work done in the US in a given year. Prizes awarded for work done outside Britain and the US are excluded.3 The curve shows the best estimate of the probability that work was done in Britain as opposed to the US. The graph is striking. At the beginning of the period, close to 4 prize‐winning contributions were made in Britain for every 1 made in the US, but the probabilities have now more than reversed. Table 1 reports estimates from linear regressions of the probability that work done in either Britain or the US was done in Britain. In keeping with the figure, the estimates indicate that the probability that prize‐winning work is done in Britain falls relative to the US and the decline is sharpest in the early years. Extrapolating from the linear model in column 1, the probability that prize‐winning work done in either of the countries is done in Britain falls by 6% per decade. Table 1 
The Probability that Prize‐Winning Work Was Done in Britain Versus the US and Germany . (1) . (2) . (3) . (4) . . Probability that Work Was Done in Britain Versus the US . Probability that Work Was Done in Britain Versus Germany . Year Work Was Done −0.00610*** −0.349** 0.00484*** 0.448** (0.00111) (0.151) (0.00146) (0.194) (Year Work Was Done)2 8.80e‐05** −0.000115** (3.88e‐05) (5.00e‐05) Intercept 12.17*** 345.6** −8.813*** −438.0** (2.173) (147.0) (2.821) (187.3) Observations 293 293 137 137 R‐squared 0.094 0.110 0.076 0.110 . (1) . (2) . (3) . (4) . . Probability that Work Was Done in Britain Versus the US . Probability that Work Was Done in Britain Versus Germany . Year Work Was Done −0.00610*** −0.349** 0.00484*** 0.448** (0.00111) (0.151) (0.00146) (0.194) (Year Work Was Done)2 8.80e‐05** −0.000115** (3.88e‐05) (5.00e‐05) Intercept 12.17*** 345.6** −8.813*** −438.0** (2.173) (147.0) (2.821) (187.3) Observations 293 293 137 137 R‐squared 0.094 0.110 0.076 0.110 Note. Standard errors in parentheses (*** denotes statistically significant at the 1% level; ** denotes statistically significant at the 5% level; * denotes statistically significant at the 10% level). The regressions show coefficients from regressions of whether prize‐winning work was done in Britain versus the comparison country (the US in columns (1) and (2) or Germany in columns (3) and (4)) on the year in which the work was done. People who were in multiple locations in the year that they did their prize‐winning work were pro‐rated across locations. Locations outside of Britain and the comparison country were excluded. Open in new tab Table 1 
The Probability that Prize‐Winning Work Was Done in Britain Versus the US and Germany . (1) . (2) . (3) . (4) . . Probability that Work Was Done in Britain Versus the US . Probability that Work Was Done in Britain Versus Germany . Year Work Was Done −0.00610*** −0.349** 0.00484*** 0.448** (0.00111) (0.151) (0.00146) (0.194) (Year Work Was Done)2 8.80e‐05** −0.000115** (3.88e‐05) (5.00e‐05) Intercept 12.17*** 345.6** −8.813*** −438.0** (2.173) (147.0) (2.821) (187.3) Observations 293 293 137 137 R‐squared 0.094 0.110 0.076 0.110 . (1) . (2) . (3) . (4) . . Probability that Work Was Done in Britain Versus the US . Probability that Work Was Done in Britain Versus Germany . Year Work Was Done −0.00610*** −0.349** 0.00484*** 0.448** (0.00111) (0.151) (0.00146) (0.194) (Year Work Was Done)2 8.80e‐05** −0.000115** (3.88e‐05) (5.00e‐05) Intercept 12.17*** 345.6** −8.813*** −438.0** (2.173) (147.0) (2.821) (187.3) Observations 293 293 137 137 R‐squared 0.094 0.110 0.076 0.110 Note. Standard errors in parentheses (*** denotes statistically significant at the 1% level; ** denotes statistically significant at the 5% level; * denotes statistically significant at the 10% level). The regressions show coefficients from regressions of whether prize‐winning work was done in Britain versus the comparison country (the US in columns (1) and (2) or Germany in columns (3) and (4)) on the year in which the work was done. People who were in multiple locations in the year that they did their prize‐winning work were pro‐rated across locations. Locations outside of Britain and the comparison country were excluded. Open in new tab The second panel of the Figure shows a comparable analysis for work done in Britain versus Germany. Here the figure is dramatically different. Whereas considerably more prize‐winning work is done in Germany than Britain at the beginning of the period, the probability that prize‐winning work is done in Britain increases so that by the middle of the twentieth century four Nobel Prize winning contributions are done in Britain for every one done in Germany. In recent years, German science has recovered, accounting for the downturn in Britain relative to Germany. Nevertheless at the end of the series slightly more prize‐winning work was being done in Britain than Germany, a country larger in both population and output than Britain. Columns 3 and 4 of Table 1 report regression estimates of the probability that prize‐winning work that is done in one of the countries is done in Britain. The model in column 4 shows that Britain’s share of prize‐winning contributions done in either Britain or Germany peaks in the middle of the twentieth century. Over the century as a whole, Britain’s share of the prize‐winning contributions from both countries increases by just under 5% per decade. The results for Britain versus the US differ from those for Britain versus Germany because of the markedly different trends in the US and Germany. While it is clear that Britain has not kept up with the tremendous progress in the US, it is equally clear that it has not suffered the unfortunate fate of Germany science. Explaining the trends in the US and Germany lies beyond the scope of the current study but it seems fair to assume that what happened in both the US and Germany was exceptional – over this period the US went from underperforming to its current leadership role whereas Germany experienced a series of tragedies; see Weinberg (2007) for an analysis. Thus, while it would be foolish to take comfort from Britain’s improvement relative to Germany, it would be equally foolish to fret over Britain’s decline relative to the US. 2.2. Comparisons to the Rest of the World Given the exceptional experiences of the US and Germany, this Section benchmarks British science against a variety of other groups of countries. Figure 3 follows the structure of Figure 2, showing for a variety of country groups Britain’s share of prize‐winning work done in Britain versus that country group. Fig. 2. Open in new tabDownload slide The Probability that Prize‐Winning Work Was Done in Britain Versus the US and Germany
Note. A circle at 1 indicates prize‐winning work done in Britain in a given year and a circle at 0 indicates prize‐winning work done in the comparison country (US in Panel (a) or Germany in Panel (b)) in that year. People who were in multiple locations in the year that they did their prize‐winning work were pro‐rated across locations. Locations outside of Britain and the comparison country were excluded. The curves give flexible, non‐parametric (kernel) estimates of the probability that prize‐winning work done in a given year was done in Britain as opposed to the comparison country. The shaded region gives a 95% confidence interval for the probability that prize‐winning work is done in Britain as opposed to the comparison country in a given year. Fig. 2. Open in new tabDownload slide The Probability that Prize‐Winning Work Was Done in Britain Versus the US and Germany
Note. A circle at 1 indicates prize‐winning work done in Britain in a given year and a circle at 0 indicates prize‐winning work done in the comparison country (US in Panel (a) or Germany in Panel (b)) in that year. People who were in multiple locations in the year that they did their prize‐winning work were pro‐rated across locations. Locations outside of Britain and the comparison country were excluded. The curves give flexible, non‐parametric (kernel) estimates of the probability that prize‐winning work done in a given year was done in Britain as opposed to the comparison country. The shaded region gives a 95% confidence interval for the probability that prize‐winning work is done in Britain as opposed to the comparison country in a given year. Fig. 3. Open in new tabDownload slide The Probability that Prize‐Winning Work Was Done in Britain Versus Various Regions
Note. A circle at 1 indicates prize‐winning work done in Britain in a given year and a circle at 0 indicates prize‐winning work done in the comparison region (indicated for that panel) in that year. People who were in multiple locations in the year that they did their prize‐winning work were pro‐rated across locations. Locations outside Britain and the comparison region were excluded. The curves give flexible, non‐parametric (kernel) estimates of the probability that prize‐winning work done in a given year was done in Britain as opposed to the comparison region. The shaded region gives a 95% confidence interval for the probability that prize‐winning work is done in Britain as opposed to the comparison region in a given year. The developed world is defined as Western Europe, Australia, Canada, Japan, New Zealand and the US. Western European offshoots defined as Australia, Canada, New Zealand and the US. Fig. 3. Open in new tabDownload slide The Probability that Prize‐Winning Work Was Done in Britain Versus Various Regions
Note. A circle at 1 indicates prize‐winning work done in Britain in a given year and a circle at 0 indicates prize‐winning work done in the comparison region (indicated for that panel) in that year. People who were in multiple locations in the year that they did their prize‐winning work were pro‐rated across locations. Locations outside Britain and the comparison region were excluded. The curves give flexible, non‐parametric (kernel) estimates of the probability that prize‐winning work done in a given year was done in Britain as opposed to the comparison region. The shaded region gives a 95% confidence interval for the probability that prize‐winning work is done in Britain as opposed to the comparison region in a given year. The developed world is defined as Western Europe, Australia, Canada, Japan, New Zealand and the US. Western European offshoots defined as Australia, Canada, New Zealand and the US. Panel (a) uses all countries other than the US and Germany as the comparison group. Here Britain’s share increases over the early part of the century, reaching a plateau in the middle of the century before its recent but imprecisely estimated downturn. The estimates are striking. At its peak, close to 40% of prize‐winning contributions made outside the US and Germany were made in Britain! Because many of the countries included in the comparison group are developing countries, which may experience very different trajectories than a country like Britain, Panel (b) reports estimates taking all developed countries other than the US and Germany as a comparison. Britain’s share increases but the basic pattern is unchanged. Japan, whose economy grew dramatically relative to Britain is included among the developed countries. Panel (c) benchmarks Britain against all other Western European countries and their offshoots (Canada, Australia and New Zealand) other than the US (an offshoot) and Germany. Again the results are quantitatively and qualitatively similar. Lastly, one might want to compare Britain to other Western European countries, whose war‐time experiences as well as economies may be more similar to Britain’s. Panel (d) benchmarks Britain against Western European countries other than Germany, obtaining estimates that are again similar to the others. Table 2 reports regression estimates that correspond to each of the preceding analyses. In all cases, Britain shows strong growth that peaks in the middle of the century. From the regressions it is clear that Britain’s position improved at a decreasing rate but the extent of the decline is less clear. Table 2 
The Probability that Prize‐Winning Work Was Done in Britain Versus Various Regions . (1) . (2) . (3) . (4) . (5) . (6) . (7) . (8) . . Britain Versus World Other than US and Germany . Britain Versus Developed World Other than US and Germany . Britain Versus Western Europe and Offshoots Other than US and Germany . Britain Versus Western Europe Other than Germany . Year Work Was Done 0.00219* 0.280* 0.00211* 0.395** 0.00333*** 0.415** 0.00264** 0.382** (0.00112) (0.153) (0.00118) (0.162) (0.00125) (0.172) (0.00121) (0.168) (Year Work Was Done)2 −7.19e‐05* −0.000101** −0.00011** −9.79e‐05** (3.94e‐05) (4.18e‐05) (4.44e‐05) (4.35e‐05) Intercept −3.905* −273.2* −3.717 −383.9** −6.032** −404.3** −4.729** −371.3** (2.164) (147.8) (2.278) (156.6) (2.419) (166.2) (2.336) (163.0) Observations 222 222 201 201 180 180 194 194 R‐squared 0.017 0.032 0.016 0.044 0.038 0.069 0.024 0.050 . (1) . (2) . (3) . (4) . (5) . (6) . (7) . (8) . . Britain Versus World Other than US and Germany . Britain Versus Developed World Other than US and Germany . Britain Versus Western Europe and Offshoots Other than US and Germany . Britain Versus Western Europe Other than Germany . Year Work Was Done 0.00219* 0.280* 0.00211* 0.395** 0.00333*** 0.415** 0.00264** 0.382** (0.00112) (0.153) (0.00118) (0.162) (0.00125) (0.172) (0.00121) (0.168) (Year Work Was Done)2 −7.19e‐05* −0.000101** −0.00011** −9.79e‐05** (3.94e‐05) (4.18e‐05) (4.44e‐05) (4.35e‐05) Intercept −3.905* −273.2* −3.717 −383.9** −6.032** −404.3** −4.729** −371.3** (2.164) (147.8) (2.278) (156.6) (2.419) (166.2) (2.336) (163.0) Observations 222 222 201 201 180 180 194 194 R‐squared 0.017 0.032 0.016 0.044 0.038 0.069 0.024 0.050 Note. Standard errors in parentheses (*** denotes statistically significant at the 1% level; ** denotes statistically significant at the 5% level; * denotes statistically significant at the 10% level). The regressions show coefficients from regressions of whether prize‐winning work was done in Britain versus the comparison region on the year in which the work was done. People who were in multiple locations in the year that they did their prize‐winning work were pro‐rated across locations. Locations outside of Britain and the comparison region were excluded. Open in new tab Table 2 
The Probability that Prize‐Winning Work Was Done in Britain Versus Various Regions . (1) . (2) . (3) . (4) . (5) . (6) . (7) . (8) . . Britain Versus World Other than US and Germany . Britain Versus Developed World Other than US and Germany . Britain Versus Western Europe and Offshoots Other than US and Germany . Britain Versus Western Europe Other than Germany . Year Work Was Done 0.00219* 0.280* 0.00211* 0.395** 0.00333*** 0.415** 0.00264** 0.382** (0.00112) (0.153) (0.00118) (0.162) (0.00125) (0.172) (0.00121) (0.168) (Year Work Was Done)2 −7.19e‐05* −0.000101** −0.00011** −9.79e‐05** (3.94e‐05) (4.18e‐05) (4.44e‐05) (4.35e‐05) Intercept −3.905* −273.2* −3.717 −383.9** −6.032** −404.3** −4.729** −371.3** (2.164) (147.8) (2.278) (156.6) (2.419) (166.2) (2.336) (163.0) Observations 222 222 201 201 180 180 194 194 R‐squared 0.017 0.032 0.016 0.044 0.038 0.069 0.024 0.050 . (1) . (2) . (3) . (4) . (5) . (6) . (7) . (8) . . Britain Versus World Other than US and Germany . Britain Versus Developed World Other than US and Germany . Britain Versus Western Europe and Offshoots Other than US and Germany . Britain Versus Western Europe Other than Germany . Year Work Was Done 0.00219* 0.280* 0.00211* 0.395** 0.00333*** 0.415** 0.00264** 0.382** (0.00112) (0.153) (0.00118) (0.162) (0.00125) (0.172) (0.00121) (0.168) (Year Work Was Done)2 −7.19e‐05* −0.000101** −0.00011** −9.79e‐05** (3.94e‐05) (4.18e‐05) (4.44e‐05) (4.35e‐05) Intercept −3.905* −273.2* −3.717 −383.9** −6.032** −404.3** −4.729** −371.3** (2.164) (147.8) (2.278) (156.6) (2.419) (166.2) (2.336) (163.0) Observations 222 222 201 201 180 180 194 194 R‐squared 0.017 0.032 0.016 0.044 0.038 0.069 0.024 0.050 Note. Standard errors in parentheses (*** denotes statistically significant at the 1% level; ** denotes statistically significant at the 5% level; * denotes statistically significant at the 10% level). The regressions show coefficients from regressions of whether prize‐winning work was done in Britain versus the comparison region on the year in which the work was done. People who were in multiple locations in the year that they did their prize‐winning work were pro‐rated across locations. Locations outside of Britain and the comparison region were excluded. Open in new tab All of these analyses tell the same story, once the exceptional cases of the US and Germany are excluded, Britain performs extremely strongly relative to a variety of benchmarks. While it seems unlikely that Britain is still gaining against other regions, it continues to have a strong scientific position. 2.3. Specific Fields Figure 3 looks at British performance by field. The symbols at the top of the graph indicate when Nobel Prize‐winning work was done in each of the 3 fields in Britain. The symbols at the bottom indicate when Prize‐winning work was done in each of the 3 fields outside Britain. The three curves show the probability that Prize‐winning work at a given point in time was done in Britain in each of the three fields. Over the period studied as a whole there is little difference across fields in the probability that any given Prize is awarded for work done in Britain. The probability for Physics (of 14.2%) is only slightly lower than that for Medicine (at 14.9%) or Chemistry (15.8%). The essentially equal probability that Nobel Prize winning work was done in Britain both across fields and, as shown above, over time masks different trends within the fields. In particular, until World War II Britain did better in Physics than in Chemistry or Medicine in that the probability that prize‐winning work was done in Britain was higher for physics than chemistry or medicine. Since World War II, Britain has done better in both Chemistry and Medicine than Physics. While the probability that a prize in Chemistry or Medicine is awarded for work done in Britain has increased, albeit very slightly over the period studied, the probability that a Physics Prize is awarded for work done in Britain has declined. To show this, I regress the probability that a Physics Nobel Prize was received for work done in Britain on the year in which the work was done. The estimates are: (3) The intercept of 0.140 implies that around 1944, the (prize‐weighted) midpoint of the data, 14% of Physics Nobel Prizes were awarded for work done in Britain. The slope coefficient of 0.0024 indicates that the share of Physics Nobel Prizes awarded for work done in Britain is trending downward by just under two and a half percentage points per decade. One might speculate that the reason for this decline of physics in Britain relative to chemistry and medicine is that the US is doing better in physics than the other fields. This turns out not to be the case. The same relative decline in physics in Britain holds true even if one only looks at Nobel Prizes awarded for work done outside the US. 2.4. Other Data The preceding estimates, based as they are on Nobel laureates, may reflect some particulars of the process for awarding Nobel Prizes. Moreover, the Nobel Prize only covers 3 fields, which may be unrepresentative. To investigate these possibilities, I have obtained contemporaneous data on highly‐cited researchers from the Institute for Scientific Information’s Highly Cited. As indicated, this website lists the roughly 250 most cited researchers in each of 21 fields as well as their current institutional affiliation. These data contain current affiliations as opposed to affiliations when people did their important works, nor do they indicate when people did their important works, both defects that our data on Nobel laureates addresses. I nevertheless believe that these data on highly‐cited researchers provide a valuable check on the previous results. While still focusing on science’s elite, this analysis focuses less on the extreme right tail. Overall tabulations of these data place Britain solidly in second among all countries, with 7.6% of highly‐cited researchers. This figure is well beneath that for the US, which has two‐thirds of highly cited researchers but well above third‐ranked Germany and fourth‐ranked Japan, which each have slightly over 4% of highly cited researchers. For each of the 21 fields for which data are available, Table 3 lists the share of highly‐cited researchers that are in Britain and Britain’s ranking in terms of the number of highly cited researchers out of all countries. The results are striking and remarkably consistent with the previous ones. Across the 21 fields, Britain is ranked second in 12; third in 8; and fourth in one. The only country to do better is the US, which stands out, being ranked first in all fields. By comparison, Germany is ranked second in 5 fields and Japan is ranked second in 2 fields. Table 3 
United Kingdom’s Share of Highly‐cited Researchers and Rank, by Field Field . Share . Rank . Agricultural Science 0.075 3 Biology & Biochemistry 0.052 3 Chemistry 0.085 3 Clinical Medicine 0.110 2 Computer Science 0.030 3 Ecology/Environment 0.076 2 Economics/Business 0.041 2 Engineering 0.044 3 Geosciences 0.085 2 Immunology 0.049 3 Material Science 0.066 3 Mathematics 0.073 2 Microbiology 0.069 2 Molecular Biology and Genetics 0.052 3 Neuroscience 0.124 2 Pharmacology 0.176 2 Physics 0.062 4 Plant & Animal Science 0.128 2 Psychology/Psychiatry 0.085 2 Social Sciences, General 0.032 2 Space Sciences 0.092 2 Field . Share . Rank . Agricultural Science 0.075 3 Biology & Biochemistry 0.052 3 Chemistry 0.085 3 Clinical Medicine 0.110 2 Computer Science 0.030 3 Ecology/Environment 0.076 2 Economics/Business 0.041 2 Engineering 0.044 3 Geosciences 0.085 2 Immunology 0.049 3 Material Science 0.066 3 Mathematics 0.073 2 Microbiology 0.069 2 Molecular Biology and Genetics 0.052 3 Neuroscience 0.124 2 Pharmacology 0.176 2 Physics 0.062 4 Plant & Animal Science 0.128 2 Psychology/Psychiatry 0.085 2 Social Sciences, General 0.032 2 Space Sciences 0.092 2 Note. Share gives the share of highly cited researchers in the UK in each field. Rank gives the ranking of the United Kingdom among all nations in the number of highly cited researchers in each field. Open in new tab Table 3 
United Kingdom’s Share of Highly‐cited Researchers and Rank, by Field Field . Share . Rank . Agricultural Science 0.075 3 Biology & Biochemistry 0.052 3 Chemistry 0.085 3 Clinical Medicine 0.110 2 Computer Science 0.030 3 Ecology/Environment 0.076 2 Economics/Business 0.041 2 Engineering 0.044 3 Geosciences 0.085 2 Immunology 0.049 3 Material Science 0.066 3 Mathematics 0.073 2 Microbiology 0.069 2 Molecular Biology and Genetics 0.052 3 Neuroscience 0.124 2 Pharmacology 0.176 2 Physics 0.062 4 Plant & Animal Science 0.128 2 Psychology/Psychiatry 0.085 2 Social Sciences, General 0.032 2 Space Sciences 0.092 2 Field . Share . Rank . Agricultural Science 0.075 3 Biology & Biochemistry 0.052 3 Chemistry 0.085 3 Clinical Medicine 0.110 2 Computer Science 0.030 3 Ecology/Environment 0.076 2 Economics/Business 0.041 2 Engineering 0.044 3 Geosciences 0.085 2 Immunology 0.049 3 Material Science 0.066 3 Mathematics 0.073 2 Microbiology 0.069 2 Molecular Biology and Genetics 0.052 3 Neuroscience 0.124 2 Pharmacology 0.176 2 Physics 0.062 4 Plant & Animal Science 0.128 2 Psychology/Psychiatry 0.085 2 Social Sciences, General 0.032 2 Space Sciences 0.092 2 Note. Share gives the share of highly cited researchers in the UK in each field. Rank gives the ranking of the United Kingdom among all nations in the number of highly cited researchers in each field. Open in new tab Looking at individual fields, Britain does particularly well in Pharmacology (with 17.6% of highly‐cited researchers); Plant and Animal Science (with 12.8%); Neuroscience (with 12.4%); and Clinical Medicine (with 11%). Britain’s lowest share of highly‐cited researchers is in Computer Science, where it has a share of 3%. This lower performance may be because Computer Science is a field that appears quite competitive, with many developing countries doing quite well. At the same time, Britain is ranked third in computer science. Although I are not aware of studies of trends in the importance of various fields over time, my sense is that the scientific community has seen some shift to more engineering‐oriented fields. Medicine also seems to be particularly salient. Looking at these fields, Britain does quite well in the many medical fields, but perhaps somewhat less well in more engineering‐oriented fields, including Computer Science. In addition to being of interest in their own right, these rankings provide a valuable check on the Nobel Prize data. Among the Nobel fields, Britain ranks fourth in Physics (the one field in which Britain ranks fourth) and third in Chemistry. Medicine comprises a range of fields, with Britain ranking second or third in all of them. (4) [ The Probability that Prize‐Winning Work Was Done in Britain, by Field
Note. A mark near 1 indicates prize‐winning work done in Britain in a given year and a mark near 0 indicates prize‐winning work done outside of Britain in that year. People who were in multiple locations in the year that they did their prize‐winning work were pro‐rated across locations and are included in the non‐parametric estimates, but are not shown. The curves give flexible, non‐parametric (kernel) estimates of the probability that prize‐winning work done in each field in a given year was done in Britain. ] Despite the many differences, I see no reason to question the results based on Nobel Prizes based on these results. In keeping with the Nobel Prize data, Britain ranks second in highly‐cited researchers overall. The Nobel fields are, if anything, ones in which Britain tends to do somewhat worse than average. Even among the Nobel fields, the comparisons are remarkably consistent with the data on Nobel Prizes, with Physics being somewhat lower than the other fields. Overall, I conclude that Britain is indeed an exceptionally strong scientific contributor, surpassed only by the US and surpassing Germany and Japan, all larger countries than Britain. 2.5. Migration There is considerable interest in the effect of brain drains on scientific performance; Beine et al. (2001) provide general theory and evidence and Hunter et al. (2009) focus on the recent experience in Britain. My data on Nobel Laureates contain information on where people were born in addition to where they did their Prize‐winning work, permitting me to analyse the extent to which Britain experiences a brain drain. Of the Nobel laureates considered, 12.1% were born in Britain. As indicated, 14.8% of the Nobel laureates did their prize‐winning work in Britain. The gap (of just under 3%) indicates that Britain has historically been a net importer of scientific talent. Figure 5 analyses inward and outward migration. The top two curves show the probability that prize‐winning work done in any given year was done in Britain (the solid curve, which is reproduced from Figure 1) and probability that prize‐winning work done in any given year was done by someone born in Britain (the dashed curve). As indicated, Britain has been a net importer of Nobel laureates, so the curve for the probability that prize‐winning work is done in Britain is above the one for the probability that someone born in Britain does prize‐winning work. Fig. 5. Open in new tabDownload slide Analysis of Migration
Note. The top curves shows the probability that prize‐winning work is done in Britain or by someone born in Britain in a given year. The shaded regions above the horizontal axis show the probability that someone migrates from the regions shown to Britain to do their prize‐winning work in a given year. The shaded regions below the horizontal axis show the probability that someone migrates from Britain to the regions shown to do their prize‐winning work in a given year. (Note that no people move from Britain to the developing world to do prize‐winning work.) All curves are flexible, non‐parametric (kernel) estimates. People who were in multiple locations in the year that they did their prize‐winning work were pro‐rated across locations and are included in the non‐parametric estimates. Fig. 5. Open in new tabDownload slide Analysis of Migration
Note. The top curves shows the probability that prize‐winning work is done in Britain or by someone born in Britain in a given year. The shaded regions above the horizontal axis show the probability that someone migrates from the regions shown to Britain to do their prize‐winning work in a given year. The shaded regions below the horizontal axis show the probability that someone migrates from Britain to the regions shown to do their prize‐winning work in a given year. (Note that no people move from Britain to the developing world to do prize‐winning work.) All curves are flexible, non‐parametric (kernel) estimates. People who were in multiple locations in the year that they did their prize‐winning work were pro‐rated across locations and are included in the non‐parametric estimates. The shaded areas beneath these curves break up this gap into inward migration and outward migration. The three shaded regions above the horizontal axis show the probability that someone born in each of three regions – the developing world; the US; and the rest of the developed world (i.e other than Britain and the US) did prize‐winning work in Britain in each year. While few people born in the US have done prize‐winning work in Britain, a considerable number of people born in the developing world and the rest of the developed world have done prize‐winning work in Britain. The shaded regions below the horizontal axis show the (negative of the) probability that someone born in Britain did prize‐winning work in each of the three regions in each year. Note that no one born in Britain has moved to the developing world to do prize‐winning work. A small share of people born in Britain have moved to the US or the rest of the developed world to do their prize‐winning work. The gap between the two upper curves is given by the difference between the two shaded areas. Thus, the shaded areas above the horizontal axis are larger than the shaded areas beneath the horizontal axis, because more people born outside Britain have migrated there to do their prize‐winning work than people born in Britain have left to do their prize‐winning work. Although the various series in the graph increase and/or decrease, it is important to recall that the number of observations in the series is relatively small, so the series are noisy. Regressions of the various series on a time trend with or without a squared term in time indicate that none of the trends is statistically significant. Thus, while the share of net imports is low, making it hard to identify small changes, overall Britain has been a net importer of important scientists and there are no discernable trends in either in‐migration or out‐migration of Nobel laureate scientists. 2.6. Adjusting for Size and Income The trends I have shown indicate that by the late twentieth century, the US was performing substantially better than Britain which, in turn, was out‐performing Germany and Japan. Of course, Britain is the smallest of these countries in terms of population and gross domestic product but, at the same time, there are other strong performers that are smaller than Britain. Here I take a simple approach to adjusting for size and income when comparing the performance of Britain, the US and other countries – I divide the number of prizes awarded for work done in each country in the post‐war period (defined as in or after 1950) by population, gross domestic product, and per capita gross domestic product. Economic and demographic data are drawn from 1970.4 Table 4 lists the countries where prize‐winning work was done in the post‐war period sorted by the amount of prize‐winning work done in the country (as above, people in multiple locations in the year in which they did their prize winning work were pro‐rated across locations). In the post‐war period, prize‐winning work was done in 16 countries. Of these countries, the US dominates the list, with close to 150 prizes (60% of the total), followed by Britain with 30 (12% of the total), Germany with 14 (6%) and Switzerland with 11 (5%). The Table also shows the tremendous disparities in population and income. In 1970, the population of the US was close to 4 times the population of Britain and the gross domestic product of the US was over 5 times that of Britain. Table 4 
Prizes per Country Since 1950 Country . Number of Prizes . Share of Prizes . 1970 Population (Millions) . 1970 GDP Per Capita (Thousand $) . 1970 GDP (Billion $) . US 147.333 0.601 210.111 4.878 1025.000 UK 30.167 0.123 54.832 3.518 192.915 Germany 14.000 0.057 78.169 3.782 295.645 Switzerland 11.000 0.045 6.187 5.698 35.253 France 8.833 0.036 50.772 3.782 192.002 Canada 6.000 0.024 21.717 3.864 83.904 Russia 5.000 0.020 130.246 2.046 266.506 Sweden 5.000 0.020 8.043 4.521 36.365 Japan 4.667 0.019 104.331 3.257 339.790 Denmark 4.333 0.018 4.929 4.537 22.363 Australia 3.000 0.012 12.728 4.129 52.561 The Netherlands 2.000 0.008 13.039 4.277 55.766 Italy 1.333 0.005 53.822 3.286 176.835 Argentina 1.000 0.004 23.962 2.688 64.405 New Zealand 1.000 0.004 2.820 3.859 10.880 Pakistan 0.333 0.001 65.706 0.359 23.557 Country . Number of Prizes . Share of Prizes . 1970 Population (Millions) . 1970 GDP Per Capita (Thousand $) . 1970 GDP (Billion $) . US 147.333 0.601 210.111 4.878 1025.000 UK 30.167 0.123 54.832 3.518 192.915 Germany 14.000 0.057 78.169 3.782 295.645 Switzerland 11.000 0.045 6.187 5.698 35.253 France 8.833 0.036 50.772 3.782 192.002 Canada 6.000 0.024 21.717 3.864 83.904 Russia 5.000 0.020 130.246 2.046 266.506 Sweden 5.000 0.020 8.043 4.521 36.365 Japan 4.667 0.019 104.331 3.257 339.790 Denmark 4.333 0.018 4.929 4.537 22.363 Australia 3.000 0.012 12.728 4.129 52.561 The Netherlands 2.000 0.008 13.039 4.277 55.766 Italy 1.333 0.005 53.822 3.286 176.835 Argentina 1.000 0.004 23.962 2.688 64.405 New Zealand 1.000 0.004 2.820 3.859 10.880 Pakistan 0.333 0.001 65.706 0.359 23.557 Open in new tab Table 4 
Prizes per Country Since 1950 Country . Number of Prizes . Share of Prizes . 1970 Population (Millions) . 1970 GDP Per Capita (Thousand $) . 1970 GDP (Billion $) . US 147.333 0.601 210.111 4.878 1025.000 UK 30.167 0.123 54.832 3.518 192.915 Germany 14.000 0.057 78.169 3.782 295.645 Switzerland 11.000 0.045 6.187 5.698 35.253 France 8.833 0.036 50.772 3.782 192.002 Canada 6.000 0.024 21.717 3.864 83.904 Russia 5.000 0.020 130.246 2.046 266.506 Sweden 5.000 0.020 8.043 4.521 36.365 Japan 4.667 0.019 104.331 3.257 339.790 Denmark 4.333 0.018 4.929 4.537 22.363 Australia 3.000 0.012 12.728 4.129 52.561 The Netherlands 2.000 0.008 13.039 4.277 55.766 Italy 1.333 0.005 53.822 3.286 176.835 Argentina 1.000 0.004 23.962 2.688 64.405 New Zealand 1.000 0.004 2.820 3.859 10.880 Pakistan 0.333 0.001 65.706 0.359 23.557 Country . Number of Prizes . Share of Prizes . 1970 Population (Millions) . 1970 GDP Per Capita (Thousand $) . 1970 GDP (Billion $) . US 147.333 0.601 210.111 4.878 1025.000 UK 30.167 0.123 54.832 3.518 192.915 Germany 14.000 0.057 78.169 3.782 295.645 Switzerland 11.000 0.045 6.187 5.698 35.253 France 8.833 0.036 50.772 3.782 192.002 Canada 6.000 0.024 21.717 3.864 83.904 Russia 5.000 0.020 130.246 2.046 266.506 Sweden 5.000 0.020 8.043 4.521 36.365 Japan 4.667 0.019 104.331 3.257 339.790 Denmark 4.333 0.018 4.929 4.537 22.363 Australia 3.000 0.012 12.728 4.129 52.561 The Netherlands 2.000 0.008 13.039 4.277 55.766 Italy 1.333 0.005 53.822 3.286 176.835 Argentina 1.000 0.004 23.962 2.688 64.405 New Zealand 1.000 0.004 2.820 3.859 10.880 Pakistan 0.333 0.001 65.706 0.359 23.557 Open in new tab Table 5 lists the countries according to the amount of prize‐winning work done in them relative to population. It is worth noting that the number of prizes is for the entire post‐war period, while the population figures are for 1970. Consequently, the figures reported for people per prize give the number of people per prize‐winning work done over a 40‐ or 50‐year period. If one wanted to obtain people per prize per year, the reported figures would have to be multiplied by 40 or 50. (The same principle applies to the dollars per prize estimates discussed below.) Table 5 
Prizes per Country Since 1950, Adjusting for Population Country . Number of Prizes . Share of Prizes . 1970 Population (Millions) . Prizes per Million People (in 1970) . 1970 Population (in Millions) per Prize . Switzerland 11.000 0.045 6.187 1.778 0.562 Denmark 4.333 0.018 4.929 0.879 1.137 US 147.333 0.601 210.111 0.701 1.426 Sweden 5.000 0.020 8.043 0.622 1.609 UK 30.167 0.123 54.832 0.550 1.818 New Zealand 1.000 0.004 2.820 0.355 2.820 Canada 6.000 0.024 21.717 0.276 3.619 Australia 3.000 0.012 12.728 0.236 4.243 Germany 14.000 0.057 78.169 0.179 5.584 France 8.833 0.036 50.772 0.174 5.748 The Netherlands 2.000 0.008 13.039 0.153 6.519 Japan 4.667 0.019 104.331 0.045 22.357 Argentina 1.000 0.004 23.962 0.042 23.962 Russia 5.000 0.020 130.246 0.038 26.049 Italy 1.333 0.005 53.822 0.025 40.366 Pakistan 0.333 0.001 65.706 0.005 197.118 Country . Number of Prizes . Share of Prizes . 1970 Population (Millions) . Prizes per Million People (in 1970) . 1970 Population (in Millions) per Prize . Switzerland 11.000 0.045 6.187 1.778 0.562 Denmark 4.333 0.018 4.929 0.879 1.137 US 147.333 0.601 210.111 0.701 1.426 Sweden 5.000 0.020 8.043 0.622 1.609 UK 30.167 0.123 54.832 0.550 1.818 New Zealand 1.000 0.004 2.820 0.355 2.820 Canada 6.000 0.024 21.717 0.276 3.619 Australia 3.000 0.012 12.728 0.236 4.243 Germany 14.000 0.057 78.169 0.179 5.584 France 8.833 0.036 50.772 0.174 5.748 The Netherlands 2.000 0.008 13.039 0.153 6.519 Japan 4.667 0.019 104.331 0.045 22.357 Argentina 1.000 0.004 23.962 0.042 23.962 Russia 5.000 0.020 130.246 0.038 26.049 Italy 1.333 0.005 53.822 0.025 40.366 Pakistan 0.333 0.001 65.706 0.005 197.118 Open in new tab Table 5 
Prizes per Country Since 1950, Adjusting for Population Country . Number of Prizes . Share of Prizes . 1970 Population (Millions) . Prizes per Million People (in 1970) . 1970 Population (in Millions) per Prize . Switzerland 11.000 0.045 6.187 1.778 0.562 Denmark 4.333 0.018 4.929 0.879 1.137 US 147.333 0.601 210.111 0.701 1.426 Sweden 5.000 0.020 8.043 0.622 1.609 UK 30.167 0.123 54.832 0.550 1.818 New Zealand 1.000 0.004 2.820 0.355 2.820 Canada 6.000 0.024 21.717 0.276 3.619 Australia 3.000 0.012 12.728 0.236 4.243 Germany 14.000 0.057 78.169 0.179 5.584 France 8.833 0.036 50.772 0.174 5.748 The Netherlands 2.000 0.008 13.039 0.153 6.519 Japan 4.667 0.019 104.331 0.045 22.357 Argentina 1.000 0.004 23.962 0.042 23.962 Russia 5.000 0.020 130.246 0.038 26.049 Italy 1.333 0.005 53.822 0.025 40.366 Pakistan 0.333 0.001 65.706 0.005 197.118 Country . Number of Prizes . Share of Prizes . 1970 Population (Millions) . Prizes per Million People (in 1970) . 1970 Population (in Millions) per Prize . Switzerland 11.000 0.045 6.187 1.778 0.562 Denmark 4.333 0.018 4.929 0.879 1.137 US 147.333 0.601 210.111 0.701 1.426 Sweden 5.000 0.020 8.043 0.622 1.609 UK 30.167 0.123 54.832 0.550 1.818 New Zealand 1.000 0.004 2.820 0.355 2.820 Canada 6.000 0.024 21.717 0.276 3.619 Australia 3.000 0.012 12.728 0.236 4.243 Germany 14.000 0.057 78.169 0.179 5.584 France 8.833 0.036 50.772 0.174 5.748 The Netherlands 2.000 0.008 13.039 0.153 6.519 Japan 4.667 0.019 104.331 0.045 22.357 Argentina 1.000 0.004 23.962 0.042 23.962 Russia 5.000 0.020 130.246 0.038 26.049 Italy 1.333 0.005 53.822 0.025 40.366 Pakistan 0.333 0.001 65.706 0.005 197.118 Open in new tab On a per capita basis Switzerland stands out, with close to 2 Nobel Prizes per million people. Denmark, the US, Sweden and Britain fall in a cluster with a Nobel Prize per every 1 to 2 million people. Not surprisingly, countries with relatively small populations tend to be high on the list. In addition to those mentioned, the next group of countries with roughly a Nobel Prize per every 3 or 4 million people, are New Zealand, Canada and Australia. Table 6 lists the countries according to the amount of prize‐winning work done in them relative to per capita gross domestic product. Here the US and Britain and, to a lesser extent, Germany, stand out, each at a separate level, with the other countries distributed beneath them. Pakistan and Russia both have considerably lower incomes than the other countries – Russia’s (imputed) per capita GDP is about half of that in the other countries, while Pakistan is less than one tenth. Consequently both do well by this metric. Table 6 
Prizes per Country Since 1950, Adjusting for Per Capita Gross Domestic Product Country . Number of Prizes . Share of Prizes . 1970 GDP Per Capita (Thousand $) . Prizes per Thousand $ in Output Per Capita in 1970 . Per Capita Output (in Thousand $ in 1970) per Prize . US 147.333 0.601 4.878 30.201 0.033 UK 30.167 0.123 3.518 8.574 0.117 Germany 14.000 0.057 3.782 3.702 0.270 Russia 5.000 0.020 2.046 2.444 0.409 France 8.833 0.036 3.782 2.336 0.428 Switzerland 11.000 0.045 5.698 1.930 0.518 Canada 6.000 0.024 3.864 1.553 0.644 Japan 4.667 0.019 3.257 1.433 0.698 Sweden 5.000 0.020 4.521 1.106 0.904 Denmark 4.333 0.018 4.537 0.955 1.047 Pakistan 0.333 0.001 0.359 0.930 1.076 Australia 3.000 0.012 4.129 0.726 1.376 The Netherlands 2.000 0.008 4.277 0.468 2.139 Italy 1.333 0.005 3.286 0.406 2.464 Argentina 1.000 0.004 2.688 0.372 2.688 New Zealand 1.000 0.004 3.859 0.259 3.859 Country . Number of Prizes . Share of Prizes . 1970 GDP Per Capita (Thousand $) . Prizes per Thousand $ in Output Per Capita in 1970 . Per Capita Output (in Thousand $ in 1970) per Prize . US 147.333 0.601 4.878 30.201 0.033 UK 30.167 0.123 3.518 8.574 0.117 Germany 14.000 0.057 3.782 3.702 0.270 Russia 5.000 0.020 2.046 2.444 0.409 France 8.833 0.036 3.782 2.336 0.428 Switzerland 11.000 0.045 5.698 1.930 0.518 Canada 6.000 0.024 3.864 1.553 0.644 Japan 4.667 0.019 3.257 1.433 0.698 Sweden 5.000 0.020 4.521 1.106 0.904 Denmark 4.333 0.018 4.537 0.955 1.047 Pakistan 0.333 0.001 0.359 0.930 1.076 Australia 3.000 0.012 4.129 0.726 1.376 The Netherlands 2.000 0.008 4.277 0.468 2.139 Italy 1.333 0.005 3.286 0.406 2.464 Argentina 1.000 0.004 2.688 0.372 2.688 New Zealand 1.000 0.004 3.859 0.259 3.859 Open in new tab Table 6 
Prizes per Country Since 1950, Adjusting for Per Capita Gross Domestic Product Country . Number of Prizes . Share of Prizes . 1970 GDP Per Capita (Thousand $) . Prizes per Thousand $ in Output Per Capita in 1970 . Per Capita Output (in Thousand $ in 1970) per Prize . US 147.333 0.601 4.878 30.201 0.033 UK 30.167 0.123 3.518 8.574 0.117 Germany 14.000 0.057 3.782 3.702 0.270 Russia 5.000 0.020 2.046 2.444 0.409 France 8.833 0.036 3.782 2.336 0.428 Switzerland 11.000 0.045 5.698 1.930 0.518 Canada 6.000 0.024 3.864 1.553 0.644 Japan 4.667 0.019 3.257 1.433 0.698 Sweden 5.000 0.020 4.521 1.106 0.904 Denmark 4.333 0.018 4.537 0.955 1.047 Pakistan 0.333 0.001 0.359 0.930 1.076 Australia 3.000 0.012 4.129 0.726 1.376 The Netherlands 2.000 0.008 4.277 0.468 2.139 Italy 1.333 0.005 3.286 0.406 2.464 Argentina 1.000 0.004 2.688 0.372 2.688 New Zealand 1.000 0.004 3.859 0.259 3.859 Country . Number of Prizes . Share of Prizes . 1970 GDP Per Capita (Thousand $) . Prizes per Thousand $ in Output Per Capita in 1970 . Per Capita Output (in Thousand $ in 1970) per Prize . US 147.333 0.601 4.878 30.201 0.033 UK 30.167 0.123 3.518 8.574 0.117 Germany 14.000 0.057 3.782 3.702 0.270 Russia 5.000 0.020 2.046 2.444 0.409 France 8.833 0.036 3.782 2.336 0.428 Switzerland 11.000 0.045 5.698 1.930 0.518 Canada 6.000 0.024 3.864 1.553 0.644 Japan 4.667 0.019 3.257 1.433 0.698 Sweden 5.000 0.020 4.521 1.106 0.904 Denmark 4.333 0.018 4.537 0.955 1.047 Pakistan 0.333 0.001 0.359 0.930 1.076 Australia 3.000 0.012 4.129 0.726 1.376 The Netherlands 2.000 0.008 4.277 0.468 2.139 Italy 1.333 0.005 3.286 0.406 2.464 Argentina 1.000 0.004 2.688 0.372 2.688 New Zealand 1.000 0.004 3.859 0.259 3.859 Open in new tab Table 7 lists the countries according to the amount of prize‐winning work done in them relative to (total) gross domestic product. Relative to the size of its economy, Switzerland stands out, with close to 1 Nobel Prize for every 3 billion dollars in output. Adjusting for output (as opposed to population) pushes Britain into the third position, behind Denmark but above the US and Sweden because its (per capita) GDP in 1970 GDP was relatively low. All three countries produce a Nobel Prize for every 6 or 7 billion dollars in production. Population varies much more than gross domestic product – the coefficient of variation for population exceeds 4 while the coefficient of variation in per capita GDP is less than 1.4. Consequently variation in population dominates these rankings but Pakistan and Russia do somewhat better using these rankings than those based purely on population while Japan and Italy do worse. Table 7 
Prizes per Country Since 1950, Adjusting for (total) Gross Domestic Product Country . Number of Prizes . Share of Prizes . 1970 GDP (Billion $) . Prizes per Billion $ in Output in 1970 . 1970 Output (in Billion $) per Prize . Switzerland 11.000 0.045 35.253 0.312 3.205 Denmark 4.333 0.018 22.363 0.194 5.161 UK 30.167 0.123 192.915 0.156 6.395 US 147.333 0.601 1025.000 0.144 6.957 Sweden 5.000 0.020 36.365 0.137 7.273 New Zealand 1.000 0.004 10.880 0.092 10.880 Canada 6.000 0.024 83.904 0.072 13.984 Australia 3.000 0.012 52.561 0.057 17.520 Germany 14.000 0.057 295.645 0.047 21.118 France 8.833 0.036 192.002 0.046 21.736 The Netherlands 2.000 0.008 55.766 0.036 27.883 Russia 5.000 0.020 266.506 0.019 53.301 Argentina 1.000 0.004 64.405 0.016 64.405 Pakistan 0.333 0.001 23.557 0.014 70.672 Japan 4.667 0.019 339.790 0.014 72.812 Italy 1.333 0.005 176.835 0.008 132.626 Country . Number of Prizes . Share of Prizes . 1970 GDP (Billion $) . Prizes per Billion $ in Output in 1970 . 1970 Output (in Billion $) per Prize . Switzerland 11.000 0.045 35.253 0.312 3.205 Denmark 4.333 0.018 22.363 0.194 5.161 UK 30.167 0.123 192.915 0.156 6.395 US 147.333 0.601 1025.000 0.144 6.957 Sweden 5.000 0.020 36.365 0.137 7.273 New Zealand 1.000 0.004 10.880 0.092 10.880 Canada 6.000 0.024 83.904 0.072 13.984 Australia 3.000 0.012 52.561 0.057 17.520 Germany 14.000 0.057 295.645 0.047 21.118 France 8.833 0.036 192.002 0.046 21.736 The Netherlands 2.000 0.008 55.766 0.036 27.883 Russia 5.000 0.020 266.506 0.019 53.301 Argentina 1.000 0.004 64.405 0.016 64.405 Pakistan 0.333 0.001 23.557 0.014 70.672 Japan 4.667 0.019 339.790 0.014 72.812 Italy 1.333 0.005 176.835 0.008 132.626 Open in new tab Table 7 
Prizes per Country Since 1950, Adjusting for (total) Gross Domestic Product Country . Number of Prizes . Share of Prizes . 1970 GDP (Billion $) . Prizes per Billion $ in Output in 1970 . 1970 Output (in Billion $) per Prize . Switzerland 11.000 0.045 35.253 0.312 3.205 Denmark 4.333 0.018 22.363 0.194 5.161 UK 30.167 0.123 192.915 0.156 6.395 US 147.333 0.601 1025.000 0.144 6.957 Sweden 5.000 0.020 36.365 0.137 7.273 New Zealand 1.000 0.004 10.880 0.092 10.880 Canada 6.000 0.024 83.904 0.072 13.984 Australia 3.000 0.012 52.561 0.057 17.520 Germany 14.000 0.057 295.645 0.047 21.118 France 8.833 0.036 192.002 0.046 21.736 The Netherlands 2.000 0.008 55.766 0.036 27.883 Russia 5.000 0.020 266.506 0.019 53.301 Argentina 1.000 0.004 64.405 0.016 64.405 Pakistan 0.333 0.001 23.557 0.014 70.672 Japan 4.667 0.019 339.790 0.014 72.812 Italy 1.333 0.005 176.835 0.008 132.626 Country . Number of Prizes . Share of Prizes . 1970 GDP (Billion $) . Prizes per Billion $ in Output in 1970 . 1970 Output (in Billion $) per Prize . Switzerland 11.000 0.045 35.253 0.312 3.205 Denmark 4.333 0.018 22.363 0.194 5.161 UK 30.167 0.123 192.915 0.156 6.395 US 147.333 0.601 1025.000 0.144 6.957 Sweden 5.000 0.020 36.365 0.137 7.273 New Zealand 1.000 0.004 10.880 0.092 10.880 Canada 6.000 0.024 83.904 0.072 13.984 Australia 3.000 0.012 52.561 0.057 17.520 Germany 14.000 0.057 295.645 0.047 21.118 France 8.833 0.036 192.002 0.046 21.736 The Netherlands 2.000 0.008 55.766 0.036 27.883 Russia 5.000 0.020 266.506 0.019 53.301 Argentina 1.000 0.004 64.405 0.016 64.405 Pakistan 0.333 0.001 23.557 0.014 70.672 Japan 4.667 0.019 339.790 0.014 72.812 Italy 1.333 0.005 176.835 0.008 132.626 Open in new tab Taken as a whole, these results indicate that adjusting for size and income Britain’s performance was among the very best in the world during the post‐war period. Beyond this strong performance for Britain, these results point to another remarkable result. Adjusting for population or (total) gross domestic product, the English‐speaking countries – Australia, Canada and New Zealand as well as Britain and the US – are exceptionally strong in science. In fact, of the Nobel Prizes awarded for work done in the post‐war period, over 76% were awarded for work done in one of these six countries, which accounted for less than 10% of the world’s population and less than a third of world output. 3. Conclusion The twentieth century was a turbulent time for scientific leadership. When it began, Europe dominated science. Today, the US does. Germany went from being a leader to playing a lesser role today. Despite these changes, Britain’s position has remained remarkably stable. Using data on Nobel laureates in Chemistry, Medicine and Physics, I show that British science has declined slightly, with most of the decline in physics. Once one accounts for the tremendous increase in the US, however British science shows remarkable growth, especially relative to the rest of the world. The raw data and data that adjust for population and gross domestic product (per capita or total) indicate that Britain is consistently one of the top scientific performers. Given the consistently strong position of British science, it would be a mistake to think about a crisis. There may, however be lessons that can be learned from the growth of science in the US. To some extent the growth of the US in science is a product of economic growth – the US grew tremendously relative to Europe and Britain. Weinberg (2007) argues the scientific success of the US is also due to its investments in science, and a highly competitive scientific market at a variety of levels. Footnotes 1 " The estimates are generated from linear models because of the (somewhat) continuous dependent variable. The standard error bands are truncated at 0 but do not account for this truncation. 2 " To explore the statistical significance of the increase in the early years and subsequent decline, I have also estimated a model that includes a quadratic term yielding, (2) This specification implies a hump‐shaped pattern in when prize‐winning work is done in Britain, with a peak in 1956 but the linear coefficient is statistically insignificant and the squared term has a p‐value of 0.072. 3 " For work done in multiple location, all locations outside of Britain and the US are excluded and the Figure shows Britain’s share of the locations in Britain and the US. 4 " Among the prizes awarded for work done in the post‐war period, the average year in which the work was done was 1967. Thus, 1970 is close the mid‐point of the prizes considered and is the first year for which economic data are available for all countries (other than Russia). Economic data on Russia only become available in 1990. Gross domestic product (total or per capita) for Russia in 1970 was calculated using Russia’s per capita gross domestic product in 1990 assuming that the Russian economy grew at the same rate per capita between 1970 and 1990 as the other countries where prize‐winning work was done. Insofar as the Russian economy grew more slowly, the results will overstate Russian performance. References Aizenman , J. and Noy , I. ( 2007 ). ‘Prizes for basic research: human capital, economic might and the shadow of history’ , Journal of Economic Growth , vol. 12 , pp. 261 – 82 . Google Scholar Crossref Search ADS WorldCat Beine , M , Docqier , F. and Rapoport , H. ( 2001 ). ‘Brain drain and economic growth: theory and evidence’ , Journal of Development Economics , vol. 64 , pp. 275 – 89 . Google Scholar Crossref Search ADS WorldCat Cole , S. and Phelan , T. ( 1999 ). ‘The scientific productivity of nations’ , Minerva , vol. 37 , pp. 1 – 23 . Google Scholar Crossref Search ADS WorldCat HM Treasury. ( 2004 ). Science and Innovation Investment Framework 2004‐2014 , London: The Stationery Office . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC HM Treasury. ( 2006 ). Science and Innovation Investment Framework 2004‐2014: Next Steps . London: The Stationery Office . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Hunter , R. , Oswald , J. and Charlton , B. ( 2009 ). ‘ The elite brain drain ’, Economic Journal , this issue. OpenURL Placeholder Text WorldCat Irvine , J. , Martin , B. Peacock , T. and Turner , R. ( 1985 ). ‘Charting the decline in British science’ . Nature , vol. 316 , pp. 587 – 90 . Google Scholar Crossref Search ADS WorldCat Jones , B. (forthcoming). ‘Age and great invention’ , Review of Economics and Statistics . OpenURL Placeholder Text WorldCat Kealey , T. ( 1991 ). ‘The growth of British science’ , Nature , vol. 350 , p. 370 . Google Scholar Crossref Search ADS PubMed WorldCat Martin , B. ( 1994 ). ‘British science in the 1980s – has the relative decline continued?’ Scientometrics , vol. 29 , pp. 27 – 56 . Google Scholar Crossref Search ADS WorldCat National Academies. ( 2007 ). Rising Above the Gathering Storm: Energizing and Employing America for a Brighter Economic Future , Washington DC: National Academies Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Nelson , R. and Wright , G. ( 1992 ). ‘The rise and fall of American technological leadership: the postwar era in historical perspective’ , Journal of Economic Literature , vol. 30 , pp. 1931 – 64 . OpenURL Placeholder Text WorldCat Smith , D. , Collins P., Hicks , D. and Wyatt , S. ( 1986 ). ‘National performance in basic research’ , Nature , vol. 323 , pp. 681 – 4 . Google Scholar Crossref Search ADS WorldCat Stephan , P. and Levin , S. ( 1993 ). ‘Age and the Nobel Prize revisited’ , Scientometrics , vol. 28 , pp. 387 – 99 . Google Scholar Crossref Search ADS WorldCat Weinberg , B. ( 2007 ). ‘ Scientific leadership ’, Working Paper, Ohio State University. Zucker , L. , Darby , M. and Brewer , M. ( 1998 ). ‘Intellectual human capital and the birth of U.S. biotechnology enterprises’ , American Economic Review , vol. 88 , pp. 290 – 306 . OpenURL Placeholder Text WorldCat Author notes " I am grateful to the editor and referee for helpful comments; to Ben Jones, Sharon Levin and Paula Stephan for sharing their data and to Melanie Bynum, Damian Hruszkewycz, Tom LaPille and Andrew Morbitzer for excellent research assistance. I am grateful to the John Templeton Foundation, the National Institutes of Health and Ohio State University for financial support. I take responsibility for all errors. © The Author(s). Journal compilation © Royal Economic Society 2009

Goodhart,, C.

doi: 10.1111/j.1468-0297.2009.02279.xpmid: N/A

Ed Nelson claims that the main cause of the improvement in monetary policy during the 1990s and 2000s [until 2007?] was a ‘changed view of the transmission mechanism’, rather than ‘shifts in policy maker objectives or to changing views about the long‐run inflation/unemployment trade‐off’, (p. F334). I agree with most of his comments on the changing views about the transmission mechanism but I dispute his suggestions about contemporary ideas on the unemployment/inflation trade‐off. In particular, Nelson goes so far as to argue that in the ‘old doctrine’ there was ‘a unit weight on inflation expectations in the inflation equation’ (p. F340, also equation (2) on p. F336). Moreover he further claims that ‘In fact, it seems appropriate, conditional on their incorrect model [the ‘old doctrine’], to attribute rational expectations from that model to UK policy makers’, (p. F342, also see equation (6), on p. F343). On what does Nelson base his claims? He does so on the basis of five quotes, from George Brown in 1965 (p. F343), Ted Heath in 1969 (p. F352), Harold Wilson in 1972 and 1975 (p. F353) and Ken Berrill in 1974 (p. F356). Ken Berrill was, I believe, commenting on both the non‐linearity and the lack of stability of the Phillips curve. In particular his phrase ‘Then there is a large flat band’ is hardly consistent with a (medium or long‐run) vertical Phillips curve. Wilson and Heath both denied any inverse relationship between inflation and unemployment. Remember, however, that the underlying monetary regime was a pegged exchange rate, until the end of Bretton Woods, and thereafter a fragile exchange rate (1973–6). Harold Wilson struggled to prevent devaluation. In this context faster domestic inflation (than abroad) meant loss of competitiveness, balance of payments pressure, exchange rate weakness, and restrictive fiscal and monetary policies (the stop part of ‘stop/go’). Heath/Wilson were trying to get Trade Unions to moderate wage claims by threatening them with future unemployment, not because (rationally expected) future inflation entered the Phillips curve with a unit coefficient but because of the policy reaction to sterling weakness. The one quote that deals specifically with inflation expectations was that by George Brown in 1965 (p. F343) where he states that firms will ‘pass on [cost] increases with just that little more added… because they expect more increases to happen in the future’. A most percipient remark but it does not equate to an assumption of a unit coefficient on rationally expected inflation in a Phillips curve. I was actually working in George Brown’s Department of Economic Affairs at the time (1965–6), and I participated, at a distance, in the internal (primarily HMT) forecasting process. When I left LSE for the Bank of England (1968–1985), I continued to be privy to that forecasting process. During these years the Chancellor of the Exchequer set both fiscal and monetary policy (interest rates) (with the agreement of the Prime Minister and, for the latter, on the advice of the Bank of England). These were the glory years of the development of large‐scale Keynesian structural models, before Lucas and repeated model ‘failures’ took away our innocence. Whereas Chancellors often publicly professed to be sceptical of model forecasts, in practice they usually followed them slavishly. What else did they have to go on, personal intuition, anecdotes from business contacts? Moreover HMT, and outside forecasters, such as the National Institute (NIESR), tended to use similar models, with closely aligned point projections (no fan charts until 1993). So, were HMT forecasts for inflation based on a vertical Philipps curve with a unit coefficient on forward looking rational expectations? Of course not. Under the 30 year rule, HMT forecasts up till 1978 should now be publicly accessible. NIESR forecasts, which until recent years were similar in methodology, are available throughout. What Nelson ought to have done is to go back to these forecasts and see how they were actually done. I should have done so also but, owing to age and other commitments (on account of the current financial crisis), I am going to be lazy and rely on fallible memory. My recollection is that the Phillips curve was generally estimated in the 1960s with a lagged dependent variable with a coefficient significantly less than unity. Then, sometime during the 1970s, after the Phelps/Friedman analytical break‐throughs at the end of the 1960s, there was a shift in HMT, and elsewhere, to imposing a unit coefficient on the lagged dependent variable, though this constraint was soon then also found to be accepted by the data. Others who were more closely involved, such as Wynne Godley, may have a clearer memory. Thus, by the end of the 1970s, a vertical (medium and long‐run) Phillips curve had become common‐place in forecasting and was filtering into political consciousness. But, by construction, it related to backwards‐looking variables, not to forwards‐looking (rational?) expectations. Although the Lucasian rational expectations DSGE models may have taken the academic world by storm, they met resistance amongst those (British) officials who actually had the task of preparing forecasts on which actual policy was to be based. It was not until, as I recall, the 1990s that expectations in such forecasts were made model consistent and even since then such expectations forecasts generally have remained partly backwards looking and part model‐consistent forwards looking. To be blunt, the idea that British policy (implicitly) incorporated a unit coefficient in rationally expected future inflation in the Phillips curve is just plain wrong. A subtext of the paper (p. F357) is that the weights placed on unemployment and inflation remained constant in ‘the policy maker objective function’ from the late 1960s onwards. I find it hard to believe that the objectives of Harold Wilson and of Margaret Thatcher were identical. Moreover, politicians follow the lead of the populace, so that the weights tend to vary depending on which of these two evils the electorate has recently found more disturbing. But that is rather a minor quibble. For the rest of Nelson’s historical account, I have far less objection, indeed considerable agreement. Let me illustrate this by going one‐by‐one through the basic postulates of the old and modern doctrines that he sets out (on pp. F335–6), appending my own comments in each case. OD1. Aggregate demand is insensitive to short‐term interest rates. Insofar as short rates matter at all, they enter the IS equation as a first‐differenced nominal rate. Mostly agree, with two comments. First, Nelson consistently underestimates the importance of external factors, the current account balance and the exchange rate on policy. Short term interest rates were often raised to protect the exchange rate and allowed to decline when sterling was strong. In turn movements in the exchange rate would affect net exports and, hence, aggregate demand. Second, one of the reasons why the change in the short rate mattered lay in the nature of housing finance at the time. Housing mortgages were provided by building societies with rates set oligopolistically by the Building Societies Association, which preferred to adjust their recommended rates slowly. Consequently every time Bank rate was raised sharply, there was a mortgage famine; and a glut followed falls. Then, as now, the housing and mortgage markets were important. OD2. Fiscal policy has powerful effects on aggregate demand. The expansionary impact of deficit spending can be reduced if the deficit is financed by issuing long‐term debt; financing by Treasury bills, on the other hand, amounts to monetisation of the deficit. Mostly agree. The other instrument that was considered powerful in the 1960s was required variations in the terms of hire‐purchase (consumer) credit. The second sentence reflects a view commonly held in the Bank of England. It was not shared by HMT who generally regarded the financing of the deficit as a second‐order issue. OD3. Aggregate demand may depend to a small degree on (real) long‐term interest rates. But this has no implications for short‐term interest rate policy, as the authorities can fix the long‐term rate directly for a given path of short rates and the monetary base. Remember that housing finance in the UK is done on a variable rate basis, so short term rates are more (and long term rates less) potent in the UK than in the US. The second sentence is not generally correct. It is a possible interpretation of the views of the Radcliffe Committee but the latter’s views on this subject were idiosyncratic and not generally shared amongst policy makers. Amongst academics, there were adherents of both the expectations theory and the Culbertson segmentation view. Amongst policy makers, the Bank of England took the lead on debt management, and it had a momentum theory of debt management, in which long gilt sales could only be done in periods of expected falling rates. The conclusion of that was that official short rates should be raised in large jumps at the onset of bad news and then, when a bottom had been found, lowered slowly. The Radcliffe Committee and the Bank of England disagreed profoundly over this issue. OD4. Cost‐push forces come from a variety of sources and can produce substantial and sustained inflation without any monetary accommodation. Agreed. OD5. This is so because inflation is insensitive to negative output gaps. Positive output gaps can add to inflationary pressure but a given amount of economic slack does not remove inflationary pressure. Mostly agreed. But I would add a proviso. Most of those professing the OD would have accepted that at some point unemployment could get so high that it would depress inflation. But it was thought that that level was incalculable in advance and likely to be very high. Accordingly the use of this route, so long as other alternatives such as prices and incomes policies remained viable, was held to be immoral. What gets forgotten now is that the Chicago monetarist school was not only regarded as mistaken but also heartless and unethical. OD6. Inflation, on the other, does depend in a symmetric manner on the change in the output gap. Agreed. Nelson has done a good job of reminding us of this facet of policy‐making. MD1. Aggregate demand is highly interest‐elastic, with the levels of both real short‐term and real long‐term interest rates appearing in the IS equation. I would not agree with the word ‘highly’. Corporate expenditure, investment and inventories, does not appear to be interest sensitive. Views on the interest elasticity of exchange rates and net exports remain much the same as ever. Deregulation of financial markets has made consumer expenditure and residential investment less sensitive to direct credit controls (now abandoned) and more sensitive to interest rates. MD2. Pure fiscal policy – that is, deficit spending not accompanied by base money creation – does not have substantial effects on aggregate demand. Apart from some tentative flirtations in 1979–83, no British policy maker has ever worried about ‘base money’. Fiscal policy is still thought to have substantial effects. However, the lags are too long and uncertain to use for normal counter‐cyclical purposes but fiscal policy will still be used in a severe downturn (1990–92, 2008–9). Under more normal circumstances, and with a vertical Phillips curve, the MPC sets interest rates to steer the economy and fiscal policy can be based on longer‐term considerations (e.g. Golden Rules and Prudence). MD3. The long‐term interest rate and other key asset prices are highly sensitive to the short‐term interest rate and expectations of its path. Agreed. MD4. The degree of monetary accommodation is critical in determining whether cost‐push forces produce sustained inflation. MD5. This is so because inflation is sensitive to both positive and negative output gaps. The level of the output gap enters the Phillips curve. MD6. Inflation does not depend on the change in the output gap. Agreed. OD7 & MD7. A low single‐digit rate of inflation is desirable and makes it easier for aggregate demand to proceed along the path of potential GDP; in addition, low inflation ensures that productive potential is not damaged by inflation. OD8 & MD8. The real structure of the economy produces a potential output path that policy makers should use as their output reference value. Associated with this sustainable output value is an unemployment level which can vary with structural changes. Agreed. It is nice to be able to end this note in a tone of agreement, because I did agree with the vast majority of what he wrote. But this is billed as a rejoinder and the purpose of such an exercise is to highlight points of difference. © The Author(s). Journal compilation © Royal Economic Society 2009

Hunter, Rosalind, S.;Oswald, Andrew, J.;Charlton, Bruce, G.

doi: 10.1111/j.1468-0297.2009.02274.xpmid: N/A

Abstract We collect data on the movement and productivity of elite scientists. Their mobility is remarkable: nearly half of the world’s most‐cited physicists work outside their country of birth. We show they migrate systematically towards nations with large R & D spending. Our study cannot adjudicate on whether migration improves scientists’ productivity, but we find that movers and stayers have identical h‐index citations scores. Immigrants in the UK and US now win Nobel Prizes proportionately less often than earlier. US residents’ h‐indexes are relatively high. We describe a framework where a key role is played by low mobility costs in the modern world. Where scientific enquiry is stunted, the intellectual life of a nation dries up, which means the withering of many possibilities of future development...
Einstein (1934, p.30) This study is an analysis of the international movement and productivity levels of elite research scientists. We begin with data on Nobel Prizes. We then construct a data set on the world’s most highly‐cited physicists. Although our sample, of 158 people, is inevitably a small one, the individuals covered within it seem of particular interest. The article discusses the conceptual implications of the observed empirical patterns. We attempt to address questions such as the following (and give brief answers in parentheses): How mobile are the world’s top research scientists, and do they migrate disproportionately to the richest countries? (Very; yes) Are elite movers more productive, on average, than elite stayers? (No) How – in the spirit of Freeman (2006)– might the new world of globalisation be expected to influence nations’ across‐person productivity distributions? (To make them more similar) Are physicists who migrate to the US more productive than home‐grown US physicists? (No) While there is a large literature on the brain drain,1 few researchers have looked at migration among world‐class scientists. To anticipate results to come, it is shown that nearly half of the elite physicists in our sample no longer work in the country in which they were born, that the major per capita importers are the US and Switzerland, and that (at least within this sample of unusually highly‐cited people) migrants and non‐migrants have similar productivities as measured by citations h‐indexes. Those in the sample who move from Europe to the US go on to be neither more nor less distinguished than American‐born elite physicists. One way to make sense of all these patterns, we argue, drawing partly upon a simple formal model, is that in a world with low mobility costs the distribution of talent can be expected to be similar across different countries. More broadly, this article relates to policy issues discussed in, for example, a recent 2007 editorial in Nature and a literature on the competitiveness of European and US economics and science (Levin and Stephan, 1999; Machin and Oswald, 2000; Regets, 2001; Royal Society, 1963; Oswald, 2007b, 2009; Summers, 2007; UUK, 2007). Those persuaded by a sanguine view of international brain ‘circulation’, rather than drain, may wish to know that our later data paint largely a one‐way picture and one disproportionately towards the US. 1. Theoretical Issues Consider a world in which scientists vary in their innate ability and productivity. Let a person’s productivity be q, which for simplicity here is defined to lie between 0 and 1. We can think of Nobel Prize winners, say, as having a level of q that is close to the upper bound of unity. Assume the talent distribution is described by a density function f(q). Among ‘elite’ scientists, define a cut‐off minimum threshold of quality, given by q*. Assume that such scientists can choose whether or not to move to a new country. This receiving country is rich, by assumption, and will pay a percentage wage premium, p, compared to the home country. There is a cost of movement, c. The model will be a timeless one but this cost could be thought of as a continuing one per‐unit‐of‐time. It might, for instance, be viewed as, in part, capturing any continuing cultural and personal cost caused by living outside one’s nation of birth. The net utility levels of individuals are taken to be given by a simple additive form: (1) (2) so that an individual will therefore choose to move if (3) The productivities, on average, of the movers and stayers can then be calculated. The average productivity of elite migrants is (4) By contrast, the average productivity of stayers (in the country to which the migrants are moving), who are drawn partially from a different segment of the talent distribution, is (5) The difference in mean productivities is therefore (6) and it can be checked that D(c,p,q*) is an increasing function of the mobility cost c and, equivalently, a decreasing function of the premium p, so that for example: (7) If the cost of mobility and the premium are both positive, then in a large class of cases: (8) and the quality of movers, on average, will exceed the quality of stayers. This is because the return from moving is biggest for the most able people. Put into words, if it is extremely costly to leave one’s country, only absolutely outstanding scientists will find it worth their while. Such people will then stand out in ability among those of their adopted nation. When the costs of international mobility are sufficiently low, however, elite migrants and elite non‐migrants come from approximately the same section of the underlying talent distribution and they will therefore have similar observed productivity levels. In this case, as c declines, the difference D approaches zero. Kwok and Leland (1982) also allow for distributions of ability in the two countries; but they assume the existence of asymmetric information and conclude that in equilibrium the average productivity of movers will always exceed that of those workers who stay. We later try to check that empirically. New work on the theory of the brain drain, and how an optimising government should act, includes Egger et al (2007) and Ionescu and Polgreen (2009). 2. Earlier Evidence Levin and Stephan (1991) gathered data on the age and publishing productivity of PhD scientists in American institutions. The authors examined six scientific specialities.2,3 They used four different measures for productivity4 and found that in five out of the six specialities – particle physics being the exception – an increase in age significantly reduced the level of productivity when controlling for ability and motivation etc. Weinberg and Galenson (2005) analysed the optimal productivity age of Nobel Prize winning economists. They concluded that experimental and applied economists peaked later in life, possibly due to accrued knowledge over time. In contrast, theoretical economists did their best work early. Carayol and Matt (2005) combined both individual and collective factors to analyse influences on productivity. They studied more than a thousand faculty members at Louis Pasteur University. Dietz and Bozeman (2005) looked at scientists’ inter‐sector job movements within the US. They surveyed 1,200 scientists, from various fields, with 5,490 career moves between them. The authors showed that job transfers were associated with higher productivity. However, they could not prove a causal relationship. Laudel (2003) used bibliometric methods (analysing patterns in publications) to investigate the movement of elite scientists. Laudel (2005) extended her previous paper, examining the brain drain of elite scientists. She studied two different specialities, angiotensin and vibrational spectroscopy, and generated three important conclusions. First, micro‐level studies may identify brain drain effects which macro studies do not, because migration flows can counteract each other. Second, some specialities have much higher levels of migration than others. Finally, migration generally occurs young and before scientists have gained ‘elite’ status. In the work closest to our own, Ioannidis (2004) took a cross‐section of 1,523 scientists, including 46 physicists, from the Institute of Scientific Information’s (ISI) Highly Cited Researcher scientist lists. Although the database had many missing observations (because of non‐completion by the majority of individuals), the author was able to analyse the patterns in scientists’ countries of birth and their current country affiliations. There was great variation across scientific fields. On an aggregate level he found that approximately one third had migrated and that three‐quarters of this third had migrated to the US. Ioannidis argues that although migration may be good for science, because it exposes scientists to new ways of thinking, when one country experiences the majority of the net in‐flows the other countries experience a damaging brain drain. He argues that for most fields, excluding those with highly specialised expensive equipment such as nuclear physics, keeping a small network of scientists, a critical mass, in a country is important if the field is not to stagnate. The previous literature identifies factors correlated with scientists’ productivity. Our article is a (retrospective) study of scientists’ mobility. It is difficult to say what in the counter‐factual case would have happened to the productivity of each physicist who migrated/remained. Hence this article cannot conclusively address the question: how does migration affect a person’s productivity? Nor can we measure directly how scientists create externalities upon colleagues, although common sense and sources such as Laband and Tollison (2003) suggest they will. Yet Waldinger (2008), using a natural experiment, finds spillover effects only on to co‐authors rather than mere departmental colleagues. This article provides new data and inquires into the nature of the productivity distribution – across different scientists within a nation – that we would expect to see under alternative assumptions. 3. Data on US and UK Winners of Nobel Prizes There is an anecdotal view that in the modern world the US acts as a giant funnel of scientific talent. We document facts consistent with that claim. We begin by recording the extent of the decline in the ratio of UK/US Nobel Prizes in science. The results are presented in Table 1. By using biographies and autobiographies on the official Nobel Prize web pages (http://nobelprize.org/nobel_prizes/lists/all), we initially examined data since World War II on all Nobel prize‐winners in science broadly defined (physics, chemistry, physiology/medicine; and economics since 1969). To do so, we treated the national affiliation of each laureate in the same way as the Nobel committee – which assigns the working address of the laureate at the time the prize is awarded. However, official biographies are of varying clarity and completeness; sometimes it was not possible to be sure of a laureate’s educational experience, or to allocate the national provenance of a prize (for example, when the laureate was retired at the time of award, worked in several countries or worked at an international laboratory such as CERN). Such laureates are omitted from the tabulation. As a referee has pointed out, our approach assumes that the selection criterion for winning Nobels has remained the same through time. Table 1 
The Number of Science Nobel Prizes Won by the United States Relative to the United Kingdom (1947–2006) . 1947–66 . 1967–86 . 1987–06 . Proportion of UK‐based Nobels as a % of US 44 28 8 Nobels won (US:UK) 45: 20 85: 24 112: 9 Immigrants after their first degree 1: 1 3: 1 7: 0 Immigrants after their doctorate 10: 2 22: 5 14: 0 Percentages of laureates who were immigrants 24: 15 29: 25 19: 0 UK migrants to US who won a Nobel 0 5 5 . 1947–66 . 1967–86 . 1987–06 . Proportion of UK‐based Nobels as a % of US 44 28 8 Nobels won (US:UK) 45: 20 85: 24 112: 9 Immigrants after their first degree 1: 1 3: 1 7: 0 Immigrants after their doctorate 10: 2 22: 5 14: 0 Percentages of laureates who were immigrants 24: 15 29: 25 19: 0 UK migrants to US who won a Nobel 0 5 5 Notes. (i) US relative to UK numbers are expressed in this Table as X: Y (ii) The population of the US is now approximately fivefold that of the UK, so on arithmetical grounds in the current era the expected ratio is 5:1. Source. Own calculations from http://www.nobelprize.org. Charlton (2007) uses different definitions and categories, and does not study immigrants. Open in new tab Table 1 
The Number of Science Nobel Prizes Won by the United States Relative to the United Kingdom (1947–2006) . 1947–66 . 1967–86 . 1987–06 . Proportion of UK‐based Nobels as a % of US 44 28 8 Nobels won (US:UK) 45: 20 85: 24 112: 9 Immigrants after their first degree 1: 1 3: 1 7: 0 Immigrants after their doctorate 10: 2 22: 5 14: 0 Percentages of laureates who were immigrants 24: 15 29: 25 19: 0 UK migrants to US who won a Nobel 0 5 5 . 1947–66 . 1967–86 . 1987–06 . Proportion of UK‐based Nobels as a % of US 44 28 8 Nobels won (US:UK) 45: 20 85: 24 112: 9 Immigrants after their first degree 1: 1 3: 1 7: 0 Immigrants after their doctorate 10: 2 22: 5 14: 0 Percentages of laureates who were immigrants 24: 15 29: 25 19: 0 UK migrants to US who won a Nobel 0 5 5 Notes. (i) US relative to UK numbers are expressed in this Table as X: Y (ii) The population of the US is now approximately fivefold that of the UK, so on arithmetical grounds in the current era the expected ratio is 5:1. Source. Own calculations from http://www.nobelprize.org. Charlton (2007) uses different definitions and categories, and does not study immigrants. Open in new tab Table 1 shows the proportions of US relative to UK laureates for each of three 20‐year segments. The data are for the period 1947–2006. For the first third of the period, 1947–66, the UK was a successful Nobel prize‐winning nation. It gained nearly half the number of prizes of the US. Over the past 60 years the population of the US has approximately doubled from 150 million to 300 million while the UK population has only increased about 20% from 50 million to 60 million. But UK success in winning Nobel science prizes has sharply declined both in relative US: UK terms over the whole period and in absolute numbers of UK laureates over the past 20 years. A much fuller analysis of UK is provided in the new work of Weinberg (2009). It may be assumed that a dominant scientific nation will attract high‐quality scientists from other countries. This can be studied for the US and UK by looking at scientists who did their university education elsewhere, then migrated either to the US or UK, where they eventually received a Nobel prize. The row in Table 1‘Immigrant after first degree’ gives the number of US or UK laureates who moved to the US or UK after they did their first college degree or equivalent; while the row ‘Immigrated after doctorate’ shows the US or UK laureates who had come to the country, where they later won the Nobel prize after completing their doctorate (PhD or an equivalent such as a medical degree). ‘Proportion of immigrant laureates’ is the total immigrant laureates (both after college and doctorate) expressed as a percentage of the total number of laureates. The proportion of immigrant laureates represents an approximate measure of a country’s power to attract the best (potential Nobel‐prize‐winning) scientists. These data reveal that in the past 20 years the UK has lost its previous ability to attract future Nobel‐prize‐winning scientists from elsewhere. There is also evidence of a decline in the percentage of immigrant laureates in the US: immigrants are now only 19% of total laureates. Considering the overwhelming dominance of the US in winning Nobel prizes during 1987–2006, the number of immigrant laureates might have been expected to increase. That this has not happened may indicate signs of increasing parochialism in US science, or perhaps increasing bureaucratic barriers preventing the easy movement of top class scientists into the US. It should be said, however, that the lags between doing the work and receiving the prize make it difficult to say. Table 1’s row ‘UK to US migration’ shows the number of scientists during a 20‐year segment who were educated in the UK but then migrated to the US and eventually were awarded a Nobel prize. (The reverse situation did not happen during the past 60 years – i.e. by the above definitions there were no UK laureates who had migrated from the US). This number may be an approximate measure of the greater attractiveness of the US compared to the UK as a place of residence and work for the highest quality scientists. This suggests that the US has become more attractive to UK‐educated scientists over the past 60 years. In 1987–2006, for example, five out of fourteen of all UK‐educated laureates had moved to the US by the time they won the Nobel prize. 4. Data on Highly‐cited Physicists We now draw upon data on elite physics researchers. In our analysis, the numbers of physicists in each country at first degree and the number currently affiliated are measured; the net gain for each country can then be established. This is then normalised for population size (source: OECD Statistics). Country of first degree is used rather than country of birth. Each country’s net gain is then compared to measures of its wealth to assess if there is a correlation. There are several ways to calculate a scientist’s career productivity. They include the total number of publications, average citations per paper and total citations. This article uses a particular citations measure, the h‐index. The h‐index was proposed by Hirsch (2005), who is a physicist by profession, as an attempt to ‘quantify the cumulative impact and relevance’ of an individual’s scientific research output (Hirsch, 2005, pp. 16 and 568). The measure incorporates a flavour of both quality and quantity of publications. Inevitably, the use of a single measure has disadvantages as well as advantages (Henrekson and Waldenstrom, 2007). By definition, an h‐index of x means that a scientist has x number of papers with x or more citations. This is calculated by ranking a physicist’s papers from the most‐cited to the least‐cited and then descending down the list until the rank of a paper becomes greater than or equal to the number of citations to that paper. The rank of this marginal paper is the h‐index. Like other measures of productivity, the h‐index has drawbacks. First, it is affected by career length. Second, although a high h‐index typically signifies a high‐quality scientist, the reverse is not always true (Hirsch, 2005); a scientist with only a few highly cited papers may have a fairly low h‐index no matter how important the papers. Third, citations may not always capture a physicist’s true impact if there is a bias towards English‐language journals (Van Leeuwen et al. 2001). The h‐index also does not account for the number of co‐authors on each paper. Hirsch (2005, pp. 16 and 574) suggests a normalisation of the h‐index for the number of co‐authors; however, as discussed by Laudel (2003, p. 221), there is no easy way of establishing the relative levels of contribution for each co‐author. Normalisation would underestimate the output of those who gave a high proportion of the input but with a large number of co‐authors and vice versa. Accordingly, the h‐index here is not adjusted. 5. The Sample of Physicists Our main sample is drawn from http://www.isihighlycited.com. We take the ISI list of physicists, which contained at the time of data collection the names of the 272 most‐cited scientists writing in physics journals between 1981 and 1999. Laudel (2003, p.219) argues that the ISI’s subject groupings are not sufficiently broken down into specialities and, therefore, that in‐depth analysis of ‘cause and consequences’ of migration cannot be analysed. However, data on these factors, such as R & D funding, do not have sufficient coverage over physics, let alone its specialities, for that depth of analysis to be undertaken. The data‐collection process took some time. We searched for biographical and bibliometric information on each of the 272 listed highly‐cited physicists. We particularly wished to determine career movements and overall career productivity. For each person, their year and place of birth, of first degree, and of PhD, were recorded. So was country of current affiliation. Data were initially gathered from the ISI website and then from physicists’ own web‐pages. This was followed by a further search of the internet. To gather further information beyond what was available through the web, emails were sent to 146 physicists where their email addresses could be identified. Of these, 63 replied. In this way, we eventually compiled a data set on 158 highly‐cited physicists. However, we obtained data on their first degrees for only 150 of them. Other aggregate data, on countries of origin and of current affiliation, were collected from OECD Statistics. The data for variables such as GDP were averaged between 1970 and 2006 to cover the main period during which the physicists were active. Data were available for 21 countries.5 In order to maintain consistency, data for the missing countries6 were not collected from other sources Our physicists currently live in 16 different countries. This produced some language difficulties for us. We could read websites well only in English or Italian. We used some online translators. Emails were sent in English. To examine a possible bias towards English speaking countries, the proportions of the final 158 physicists can be compared to those of the original 272. The extent of bias seems small. The US, however, appears to be overrepresented and Japan to be underrepresented. There is no clear way to solve this problem – our response rate (43%) is similar to those of previous studies (Laudel, 2003, p.224) – although it is considered later. The next issue was how to calculate productivity levels. We decided to focus on citations rather than numbers of publications.7 The ISI Web of Knowledge was used to calculate the h‐index. This required us to identify each physicist’s publication list, which can be problematic when some physicists have the same surname. We decided to consider each individual separately. In many cases, initial inspection showed no problems; physicists had identified how many papers they had published. However, sometimes further examination of the names used on published papers and the institutions worked for had to be undertaken. In two cases, we had such difficulties distinguishing names that the physicists had to be removed from the sample. One advantage of working with the h‐index is that the probability of a second physicist with the same surname and initials appearing within the relatively small selection of papers which affect an h‐index score is lower than occurs when using data on an entire list of publications. Of the sample of 158 physicists, 1 is female, and 8 have won Nobel prizes. A referee has pointed out that 8 seems a small number given this distinguished group but presumably some of these scientists will win the prize in the future. The majority, 61.4%, have worked in multiple countries and 97.5% have worked in multiple institutions. Currently 76% are affiliated to a university; 17% to other types of public institutions; and 7% are in private institutions. Regarding the span of their careers, 96% have, at some point, worked in academia since their PhD; 54% have experienced another type of public institution; and 47% have spent a period in the private sector. The mean number of institutions worked in is 6.03. The mean number of countries worked in is 2.41. 6. Migration and Productivity These physicists were born in 32 different countries. They studied for their first degree in 30 different countries; they did PhDs in 22 countries; and they are presently located in only 16 countries. Hence the data show a kind of ‘funnelling’ effect of approximately 50% from birth: people from 32 nations now reside in half that number. The percentage of physicists present in each country shows a gradual funnelling effect towards the US (Table 2 and Figure 1). At birth, 29.7% of physicists are in the US. This increases to 43.4% at first degree, to 55.1% at PhD, and to 67.1% at present. The proportion in the 2nd‐ and 3rd‐ranked countries falls by approximately 3 percentage points from birth to present day, with the share through time in the rest of the world falling dramatically from 56.4% at birth to only 19.6% at present. Fig. 1. Open in new tabDownload slide The Funnelling of Elite Physicists Towards the US. Sample Size: 158. Country of birth missing: 20 (12.7%); BSc country missing: 7 (4.4%); PhD country missing: 0 (0%); Current country missing: 0 (0%) Fig. 1. Open in new tabDownload slide The Funnelling of Elite Physicists Towards the US. Sample Size: 158. Country of birth missing: 20 (12.7%); BSc country missing: 7 (4.4%); PhD country missing: 0 (0%); Current country missing: 0 (0%) Table 2 
The International Distribution of Highly‐cited Physicists at Each Career Stage from Birth to the Present (% shares) . At birth (32 countries) . At BSc. (30 countries) . At PhD (22 Countries) . Now (16 countries) . 1st US (29.7) US (43.0) US (55.1) US (67.1) 2nd UK (10.9) Germany (8.6) UK (8.9) Germany (7.6) 3rd Germany (9.4) UK (7.9) Germany (8.2) Switzerland (5.7) Others 50.0 40.2 27.8 19.614 . At birth (32 countries) . At BSc. (30 countries) . At PhD (22 Countries) . Now (16 countries) . 1st US (29.7) US (43.0) US (55.1) US (67.1) 2nd UK (10.9) Germany (8.6) UK (8.9) Germany (7.6) 3rd Germany (9.4) UK (7.9) Germany (8.2) Switzerland (5.7) Others 50.0 40.2 27.8 19.614 Notes. The top left‐hand number of 29.7% means that at the point of birth the US was home to 29.7% of those in our sample who would go on to become the world’s most distinguished physicists. Today, as shown in the top right‐hand number, 67.1% of the 158 live and work in the US.14The UK was ranked 5th with 3.8% of the physicists, after Japan which was 4th with 4.4%. Open in new tab Table 2 
The International Distribution of Highly‐cited Physicists at Each Career Stage from Birth to the Present (% shares) . At birth (32 countries) . At BSc. (30 countries) . At PhD (22 Countries) . Now (16 countries) . 1st US (29.7) US (43.0) US (55.1) US (67.1) 2nd UK (10.9) Germany (8.6) UK (8.9) Germany (7.6) 3rd Germany (9.4) UK (7.9) Germany (8.2) Switzerland (5.7) Others 50.0 40.2 27.8 19.614 . At birth (32 countries) . At BSc. (30 countries) . At PhD (22 Countries) . Now (16 countries) . 1st US (29.7) US (43.0) US (55.1) US (67.1) 2nd UK (10.9) Germany (8.6) UK (8.9) Germany (7.6) 3rd Germany (9.4) UK (7.9) Germany (8.2) Switzerland (5.7) Others 50.0 40.2 27.8 19.614 Notes. The top left‐hand number of 29.7% means that at the point of birth the US was home to 29.7% of those in our sample who would go on to become the world’s most distinguished physicists. Today, as shown in the top right‐hand number, 67.1% of the 158 live and work in the US.14The UK was ranked 5th with 3.8% of the physicists, after Japan which was 4th with 4.4%. Open in new tab Overall, 44% of scientists have moved since birth, 33% since their first degree and 27% since their PhD. These proportions are in fact only a little different from those of Ioannidis (2004) who, on a much smaller sample, found 50% of physicists had moved since birth. We have data on 158 physicists compared to 46 in Ioannidis’s work. The summary statistics for the individuals’ h‐index scores can be seen in Table 3. The mean h‐index over the sample is 58.97. The minimum and maximum values are 22 and 115 respectively. Table 3 
Summary Data on Physicists’ h‐index Scores Number of observations . 158 . Mean 58.97 Standard Deviation 13.52 Minimum 22 Maximum 115 Median 57 Number of observations . 158 . Mean 58.97 Standard Deviation 13.52 Minimum 22 Maximum 115 Median 57 Open in new tab Table 3 
Summary Data on Physicists’ h‐index Scores Number of observations . 158 . Mean 58.97 Standard Deviation 13.52 Minimum 22 Maximum 115 Median 57 Number of observations . 158 . Mean 58.97 Standard Deviation 13.52 Minimum 22 Maximum 115 Median 57 Open in new tab In order to examine the effect of co‐authorship, the number and countries of the co‐authors of ten randomly selected physicists in the sample were examined. The average number of co‐authors for each of the ten varies enormously and the number of affiliated countries from 1 to 7.25. Although there is a tendency for those with more co‐authors to have higher h‐indexes, the evidence is not substantial. This gives us some reassurance in the decision to not try to adjust h‐indexes for co‐authorship. For our sample, we compared these h‐index results with the physicists’ total number of published papers, total citations and average citation count per‐paper. For the h‐index, there is no correlation with the last of these, average citations per article. But there is a significant positive correlation of 0.40 with total number of published papers, and of 0.54 with total citations. People with a high h‐index also score highly on these two criteria. Figure 2 shows that those currently in the US have an h‐index which is on average 5.71 higher than those in non‐US institutions. This difference – one that continues to hold weakly when we adjust the data using regression equations – is close to statistically significant at the 5% level. There are several possible reasons for this, which we discuss later in the article. Fig. 2. Open in new tabDownload slide Mean h‐index of Physicists by Current Geographical Location
Note. This difference is not (quite) statistically significant at 5% on a 2‐tailed test. Fig. 2. Open in new tabDownload slide Mean h‐index of Physicists by Current Geographical Location
Note. This difference is not (quite) statistically significant at 5% on a 2‐tailed test. We now separate the sample into those who have migrated and those who have not. Whether we work with the periods since birth, BSc. or PhD, we find no statistically significant difference in productivity as measured by an h‐index (Table 4). There is no way of measuring the productivity levels in the alternative situation. However, Figure 2 suggests that there are country‐specific effects. Table 4 
Productivity Levels (as measured by physicists’ h‐index levels) Between Those Who Moved Country and Those Who Did Not Move Stage . Stayers
Average h‐index if not moved country since the stage indicated . Movers
Average h‐index if moved country since that stage . Statistically different? . Birth 60.69 57.66 No, t = −1.24 BSc. 60.04 59.21 No, t = −0.36 PhD. 59.19 58.38 No, t = 0.33 Stage . Stayers
Average h‐index if not moved country since the stage indicated . Movers
Average h‐index if moved country since that stage . Statistically different? . Birth 60.69 57.66 No, t = −1.24 BSc. 60.04 59.21 No, t = −0.36 PhD. 59.19 58.38 No, t = 0.33 Open in new tab Table 4 
Productivity Levels (as measured by physicists’ h‐index levels) Between Those Who Moved Country and Those Who Did Not Move Stage . Stayers
Average h‐index if not moved country since the stage indicated . Movers
Average h‐index if moved country since that stage . Statistically different? . Birth 60.69 57.66 No, t = −1.24 BSc. 60.04 59.21 No, t = −0.36 PhD. 59.19 58.38 No, t = 0.33 Stage . Stayers
Average h‐index if not moved country since the stage indicated . Movers
Average h‐index if moved country since that stage . Statistically different? . Birth 60.69 57.66 No, t = −1.24 BSc. 60.04 59.21 No, t = −0.36 PhD. 59.19 58.38 No, t = 0.33 Open in new tab A natural question for economists is how effectively the rich countries draw in others’ top scientists. Figure 3 demonstrates this for Switzerland and the US. Figure 4 reveals a relationship between the net gain in physicists and the R & D adjusted GDP per capita. The correlation coefficient is then 0.49, significant at the 5% level. Physicists migrate toward richer countries, although the definition of rich should be adjusted to mean rich in R & D funding. Data on the level of physics funding would be still better, but we found it too hard to obtain data consistently across nations. Fig. 4. Open in new tabDownload slide The Relationship Between R & D Expenditure Per Capita*
Notes. Each dot is a separate country. The extreme top dot in the north‐east of the diagram is Switzerland.
*R & D expenditure as a percentage of GDP, GDP measured in US dollars, current prices and PPPs. and Net Gain in Physicists Fig. 4. Open in new tabDownload slide The Relationship Between R & D Expenditure Per Capita*
Notes. Each dot is a separate country. The extreme top dot in the north‐east of the diagram is Switzerland.
*R & D expenditure as a percentage of GDP, GDP measured in US dollars, current prices and PPPs. and Net Gain in Physicists Fig. 3. Open in new tabDownload slide Brain‐Drain Gains and Losses of Highly‐cited Physicists by Nation (Data scaled by 1,000) Fig. 3. Open in new tabDownload slide Brain‐Drain Gains and Losses of Highly‐cited Physicists by Nation (Data scaled by 1,000) Figures 5(a) and (b) depict the productivity levels between continental8 migratory groups, looking at differences between country of first degree and present affiliation. Table 5 gives fuller data. Those individuals who emigrated from Europe to North America emerge as the most productive with an average h‐index of 63.1. Those who remained in North America are the second most productive group: their h‐index is 61.2. The final three major groups – those who remained in Europe, remained in Asia, and moved from Asia to North America – have average h‐index scores of 56.1, 56.1 and 55.5, respectively. The only statistical significance at the 5% level here is between those who remained in North America and remained in Europe. Fig. 5. Open in new tabDownload slide The h‐index Scores of Movers and Stayers (a) In Europe and North America (b) In Asia and North America Fig. 5. Open in new tabDownload slide The h‐index Scores of Movers and Stayers (a) In Europe and North America (b) In Asia and North America Table 5 
Migratory Groups and Average h‐index Scores Country BSc . Country Now . Number . Average h‐index . Asia Asia 10 56.1 Asia Europe 1 59 Asia North America 6 55.5 Europe Europe 31 56.1 Europe North America 8 63.1 North America North America 91 61.2 Oceania Europe 1 57 Oceania Oceania 1 54 South America Europe 1 55 South America North America 2 52 Country BSc . Country Now . Number . Average h‐index . Asia Asia 10 56.1 Asia Europe 1 59 Asia North America 6 55.5 Europe Europe 31 56.1 Europe North America 8 63.1 North America North America 91 61.2 Oceania Europe 1 57 Oceania Oceania 1 54 South America Europe 1 55 South America North America 2 52 Open in new tab Table 5 
Migratory Groups and Average h‐index Scores Country BSc . Country Now . Number . Average h‐index . Asia Asia 10 56.1 Asia Europe 1 59 Asia North America 6 55.5 Europe Europe 31 56.1 Europe North America 8 63.1 North America North America 91 61.2 Oceania Europe 1 57 Oceania Oceania 1 54 South America Europe 1 55 South America North America 2 52 Country BSc . Country Now . Number . Average h‐index . Asia Asia 10 56.1 Asia Europe 1 59 Asia North America 6 55.5 Europe Europe 31 56.1 Europe North America 8 63.1 North America North America 91 61.2 Oceania Europe 1 57 Oceania Oceania 1 54 South America Europe 1 55 South America North America 2 52 Open in new tab There are other migratory groups, within the sample, with only a couple of representatives. This means that meaningful averages could not be constructed. It seems interesting to note that no physicist left North America nor remained in South America. The European physicists who moved to North America have productivity levels more in line with those of the natives than those left in Europe. The two Asian values are statistically the same. Some caution should be shown when looking at these results. First, the sub‐samples are small, especially for movers. Only 8 Europeans migrated to North America while 91 remained in North America. To this point, we have shown only raw patterns in the data. We now turn, in Table 6, to regression equations. The dependent variable here is at first, in the upper part of the Table, the logarithm of the scientists’ h‐indexes. In the lower half of the Table, it is the log of total citations to their work. Table 6 
Regression Equations on Physicists’ h‐indexes and Total Citations (t‐statistics in parenthesis; * indicates significant at the 5% level) Dependent Variable: Log of h‐index Constant 3.834* 3.828* 3.804* 3.737* 3.739* (53.23) (48.39) (48.71) (44.92) (44.62) Years since PhD 0.006* 0.006* 0.006* 0.007* 0.007* (3.10) (2.73) (2.55) (2.96) (2.94) US Born 0.037 −0.049 −0.063 −0.071 (0.83) (−0.88) (−1.14) (−1.10) US PhD 0.131* 0.058 0.051 (2.56) (0.95) (0.75) Now in US 0.121* 0.137 (2.13) (1.59) BSc outside US × Now in US −0.019 (−0.25) R2 0.057 0.060 0.104 0.134 0.134 R 0.051 0.046 0.084 0.107 0.101 Number of observations 158 138 138 138 138 Dependent Variable: Log of total citations Constant 9.161* 9.150* 9.120* 9.023* 9.027* (67.83) (62.40) (62.17) (57.33) (56.95) Years since PhD 0.015* 0.014* 0.014* 0.015* 0.015* (3.79) (3.45) (3.32) (3.60) (3.56) US Born 0.073 −0.034 −0.054 −0.071 (0.88) (−0.33) (−0.52) (−0.58) US PhD 0.163 0.058 0.044 (1.70) (0.50) (0.34) Now in US 0.175 0.208 (1.63) (1.27) BSc outside US × Now in US −0.038 (−0.27) R2 0.084 0.090 0.109 0.127 0.127 R 0.078 0.077 0.090 0.101 0.094 Number of observations 158 138 138 138 138 Dependent Variable: Log of h‐index Constant 3.834* 3.828* 3.804* 3.737* 3.739* (53.23) (48.39) (48.71) (44.92) (44.62) Years since PhD 0.006* 0.006* 0.006* 0.007* 0.007* (3.10) (2.73) (2.55) (2.96) (2.94) US Born 0.037 −0.049 −0.063 −0.071 (0.83) (−0.88) (−1.14) (−1.10) US PhD 0.131* 0.058 0.051 (2.56) (0.95) (0.75) Now in US 0.121* 0.137 (2.13) (1.59) BSc outside US × Now in US −0.019 (−0.25) R2 0.057 0.060 0.104 0.134 0.134 R 0.051 0.046 0.084 0.107 0.101 Number of observations 158 138 138 138 138 Dependent Variable: Log of total citations Constant 9.161* 9.150* 9.120* 9.023* 9.027* (67.83) (62.40) (62.17) (57.33) (56.95) Years since PhD 0.015* 0.014* 0.014* 0.015* 0.015* (3.79) (3.45) (3.32) (3.60) (3.56) US Born 0.073 −0.034 −0.054 −0.071 (0.88) (−0.33) (−0.52) (−0.58) US PhD 0.163 0.058 0.044 (1.70) (0.50) (0.34) Now in US 0.175 0.208 (1.63) (1.27) BSc outside US × Now in US −0.038 (−0.27) R2 0.084 0.090 0.109 0.127 0.127 R 0.078 0.077 0.090 0.101 0.094 Number of observations 158 138 138 138 138 Open in new tab Table 6 
Regression Equations on Physicists’ h‐indexes and Total Citations (t‐statistics in parenthesis; * indicates significant at the 5% level) Dependent Variable: Log of h‐index Constant 3.834* 3.828* 3.804* 3.737* 3.739* (53.23) (48.39) (48.71) (44.92) (44.62) Years since PhD 0.006* 0.006* 0.006* 0.007* 0.007* (3.10) (2.73) (2.55) (2.96) (2.94) US Born 0.037 −0.049 −0.063 −0.071 (0.83) (−0.88) (−1.14) (−1.10) US PhD 0.131* 0.058 0.051 (2.56) (0.95) (0.75) Now in US 0.121* 0.137 (2.13) (1.59) BSc outside US × Now in US −0.019 (−0.25) R2 0.057 0.060 0.104 0.134 0.134 R 0.051 0.046 0.084 0.107 0.101 Number of observations 158 138 138 138 138 Dependent Variable: Log of total citations Constant 9.161* 9.150* 9.120* 9.023* 9.027* (67.83) (62.40) (62.17) (57.33) (56.95) Years since PhD 0.015* 0.014* 0.014* 0.015* 0.015* (3.79) (3.45) (3.32) (3.60) (3.56) US Born 0.073 −0.034 −0.054 −0.071 (0.88) (−0.33) (−0.52) (−0.58) US PhD 0.163 0.058 0.044 (1.70) (0.50) (0.34) Now in US 0.175 0.208 (1.63) (1.27) BSc outside US × Now in US −0.038 (−0.27) R2 0.084 0.090 0.109 0.127 0.127 R 0.078 0.077 0.090 0.101 0.094 Number of observations 158 138 138 138 138 Dependent Variable: Log of h‐index Constant 3.834* 3.828* 3.804* 3.737* 3.739* (53.23) (48.39) (48.71) (44.92) (44.62) Years since PhD 0.006* 0.006* 0.006* 0.007* 0.007* (3.10) (2.73) (2.55) (2.96) (2.94) US Born 0.037 −0.049 −0.063 −0.071 (0.83) (−0.88) (−1.14) (−1.10) US PhD 0.131* 0.058 0.051 (2.56) (0.95) (0.75) Now in US 0.121* 0.137 (2.13) (1.59) BSc outside US × Now in US −0.019 (−0.25) R2 0.057 0.060 0.104 0.134 0.134 R 0.051 0.046 0.084 0.107 0.101 Number of observations 158 138 138 138 138 Dependent Variable: Log of total citations Constant 9.161* 9.150* 9.120* 9.023* 9.027* (67.83) (62.40) (62.17) (57.33) (56.95) Years since PhD 0.015* 0.014* 0.014* 0.015* 0.015* (3.79) (3.45) (3.32) (3.60) (3.56) US Born 0.073 −0.034 −0.054 −0.071 (0.88) (−0.33) (−0.52) (−0.58) US PhD 0.163 0.058 0.044 (1.70) (0.50) (0.34) Now in US 0.175 0.208 (1.63) (1.27) BSc outside US × Now in US −0.038 (−0.27) R2 0.084 0.090 0.109 0.127 0.127 R 0.078 0.077 0.090 0.101 0.094 Number of observations 158 138 138 138 138 Open in new tab Going from left to right, the columns of Table 6 gradually build up from a simple to a fuller regression specification. Older people tend to be more cited; this is to be expected merely because career length affects the period over which citations can be accrued. Having ten additional years after the year of the PhD increases a physicist’s h‐index by 6%. Being born in the US has a statistically insignificant effect. Residing in the US, however, does have a positive coefficient; it is associated9 with an h‐index approximately 13% higher. Nevertheless, this effect loses statistical significance by the final column of Table 6. In these regression equations, the adjusted R‐squared values are fairly low. Table 6 includes a simple direct test for an interaction effect. Are those who were educated initially outside the US but now reside in the US more productive (in the sense of having higher h‐index scores)? No, not than Americans. In the final column of Table 6, it can be seen that the coefficient on the variable ‘BSc outside the US × Now in US’ is approximately −0.02 with a t‐statistic of −0.25. Hence it is essentially zero. Migrant elite physicists into the US do not have an h‐index that differs, ceteris paribus, from the h‐index of home‐grown elite physicists. The tenor of these conclusions is replicated for the lower panel, using instead Log of Total Citations as the dependent variable, in Table 6. Table 7, following a referee’s suggestion, does a further check. It breaks the data down by the time point of migration to the US. Interestingly, whatever the stage at which someone migrated, their h‐index is approximately the same as that of those physicists born in the US. The null of equality of the various key coefficients, as in column 1 of Table 7, cannot be rejected at conventional significance levels. Here the size of effect in the h‐index regression equation is fractionally larger than in Table 6’s estimates; it varies from 14% to 19%. Table 7 
Alternative Regression Equations on Physicists’ h‐indexes (t‐statistics in brackets; * indicates significant at the 5% level) Dependent Variable: Log of h‐index Constant 3.712* 3.835* 3.732* 3.731* (43.28) (53.23) (47.75) (47.35) Year since PhD 0.007* 0.006* 0.007* 0.007* (3.14) (3.10) (3.34) (3.38) US Born – left US (1 person) −0.115 – – – (−0.50) US Born – stayed (40 people) 0.143* – 0.111* – (2.28) – (2.38) – US Born – all (41 people) – – – 0.103* (2.19) Migrated to US at BSc (28 people) 0.185* – 0.155* 0.153* (2.78) – (2.96) (2.89) Migrated to US at PhD (15 people) 0.159* – 0.129* 0.126 (2.05) – (1.96) (1.91) Migrated to US post‐PhD (22 people) 0.151* – 0.121* 0.119* (2.13) – (2.10) (2.05) Visited US otherwise (33 people) 0.053 – – (0.82) R2 0.133 0.057 0.127 0.122 Adj. R2 0.092 0.051 0.098 0.093 Dependent Variable: Log of h‐index Constant 3.712* 3.835* 3.732* 3.731* (43.28) (53.23) (47.75) (47.35) Year since PhD 0.007* 0.006* 0.007* 0.007* (3.14) (3.10) (3.34) (3.38) US Born – left US (1 person) −0.115 – – – (−0.50) US Born – stayed (40 people) 0.143* – 0.111* – (2.28) – (2.38) – US Born – all (41 people) – – – 0.103* (2.19) Migrated to US at BSc (28 people) 0.185* – 0.155* 0.153* (2.78) – (2.96) (2.89) Migrated to US at PhD (15 people) 0.159* – 0.129* 0.126 (2.05) – (1.96) (1.91) Migrated to US post‐PhD (22 people) 0.151* – 0.121* 0.119* (2.13) – (2.10) (2.05) Visited US otherwise (33 people) 0.053 – – (0.82) R2 0.133 0.057 0.127 0.122 Adj. R2 0.092 0.051 0.098 0.093 Notes. • Sample size is again 138 • There are 19 in the never been to US category Open in new tab Table 7 
Alternative Regression Equations on Physicists’ h‐indexes (t‐statistics in brackets; * indicates significant at the 5% level) Dependent Variable: Log of h‐index Constant 3.712* 3.835* 3.732* 3.731* (43.28) (53.23) (47.75) (47.35) Year since PhD 0.007* 0.006* 0.007* 0.007* (3.14) (3.10) (3.34) (3.38) US Born – left US (1 person) −0.115 – – – (−0.50) US Born – stayed (40 people) 0.143* – 0.111* – (2.28) – (2.38) – US Born – all (41 people) – – – 0.103* (2.19) Migrated to US at BSc (28 people) 0.185* – 0.155* 0.153* (2.78) – (2.96) (2.89) Migrated to US at PhD (15 people) 0.159* – 0.129* 0.126 (2.05) – (1.96) (1.91) Migrated to US post‐PhD (22 people) 0.151* – 0.121* 0.119* (2.13) – (2.10) (2.05) Visited US otherwise (33 people) 0.053 – – (0.82) R2 0.133 0.057 0.127 0.122 Adj. R2 0.092 0.051 0.098 0.093 Dependent Variable: Log of h‐index Constant 3.712* 3.835* 3.732* 3.731* (43.28) (53.23) (47.75) (47.35) Year since PhD 0.007* 0.006* 0.007* 0.007* (3.14) (3.10) (3.34) (3.38) US Born – left US (1 person) −0.115 – – – (−0.50) US Born – stayed (40 people) 0.143* – 0.111* – (2.28) – (2.38) – US Born – all (41 people) – – – 0.103* (2.19) Migrated to US at BSc (28 people) 0.185* – 0.155* 0.153* (2.78) – (2.96) (2.89) Migrated to US at PhD (15 people) 0.159* – 0.129* 0.126 (2.05) – (1.96) (1.91) Migrated to US post‐PhD (22 people) 0.151* – 0.121* 0.119* (2.13) – (2.10) (2.05) Visited US otherwise (33 people) 0.053 – – (0.82) R2 0.133 0.057 0.127 0.122 Adj. R2 0.092 0.051 0.098 0.093 Notes. • Sample size is again 138 • There are 19 in the never been to US category Open in new tab The results provide a little evidence that some groups are more productive than others, and in particular that everyone working in the US tends to have a higher h citations score. However, the reasons for this are more ambiguous. As previously mentioned, there is no clear causal relationship between migration and productivity. Put into context, migrating to the US may make you more productive (direction one) but also the most productive people may be those who are offered jobs in the US (direction two). Another, and a more sociological, possibility is that those in the American scientific circuit simply cite each other 13%−19% more. All three arguments are plausible, perhaps jointly play some role and have implications for discussions on the brain drain. Yet with these data it is not possible to distinguish among them. If direction one is the overriding influence then the world body of scientific knowledge gains from the migration because of the increase in productivity. However, there are further implications for the country left if the field stagnates or disappears. If direction two is the overriding influence, then the US is taking the best physicists, although it already had a large portion of them. It is the success of this home‐grown group of US physicists that suggests it is the resources in the US that make their physicists the most‐cited. Even so, there are other explanations for this apparent difference in productivity, not related to aiding priority, namely that there are possible biases in the productivity measure. There are suggestions that US‐based scientists disproportionately cite other US‐based scientists (Leimu and Koricheva 2005). However, Wong and Kokko (2005) indicate that this is in fact a location bias: Europeans also disproportionately cite other Europeans. A second issue, regarding language, arises: the publications used for citation counts by the ISI being substantially in English. This may reduce the citation counts of those who publish in languages other than English (Van Leuween et al., 2001), although articles in local languages would also be cited less as they are accessible to fewer readers. Van Leuween et al. argue that the research impact of countries such as Germany and France would increase if more foreign publications were accounted for. To investigate whether h‐index productivity relates to other measures of productivity, we compared each country’s h‐index rank to its patent productivity rank and its general productivity growth rates. The results are that h‐index rank positively correlates (0.66) with patent rank and with overall productivity for the majority of the countries (0.66, excluding Finland and Italy). Hence those countries with the highest average h‐indexes also tend to have higher productivity measured in other ways. The patent productivity finding seems particularly relevant as it also measures R & D in which physicists are involved. This will, at least in part, be affected by the same incentives. Such an analysis suggests that the higher level of productivity shown in the US data is not substantially due to a US/English bias. A natural question is whether the article’s finding of approximately equal productivities for migrants and non‐migrants holds for elite scientists in other disciplines. The only other data available to us are for bio‐scientists. Again, these use ISI highly‐cited scientists as the sample. Appendix Table A4 shows that, as with physics, it appears that movers and stayers have similar h‐indexes. Further discussion for bio‐sciences is contained in the unpublished report by Warwick University (2007). Finally, a high level of elite mobility has also recently been reported for young economists in Oswald and Ralsmark (2007). Figure A1 reveals that in the top‐10 US departments of economics approximately 75% of assistant professors did their first degree outside the US. This is consistent with, although necessarily not precisely comparable to, the findings from our data on senior physicists. 7. Conclusions This article attempts to contribute to knowledge about the nature of the elite brain drain. It draws five conclusions. First, the UK currently wins fewer Nobel Prizes in science than it used to,10 and the US garners many more. What is less widely known is that, in both the UK and the US, immigrant scientists win the Prize less often, proportionately, than in earlier decades. Second, by charting the careers of a group of distinguished physicists, we show that they are strikingly mobile. Almost half of the highly‐cited scientists in our sample are migrants: our 158 physicists were born in 32 countries but now live in only 16. Approximately 30% migrated after their first degrees and went predominantly to the US. Third, among highly‐cited physicists the average productivity (as measured by a citations h‐index) of movers is not different from that of stayers.11 We are unable, with our data, to say whether migration itself causally increases a scholar’s productivity but it might be argued that our results are consistent with Waldinger’s (2008) finding that – except for co‐authors – there are no strong externality effects among senior scholars. Fourth, international flows of physicists between first degree and the present day demonstrate that top scholars head to countries with high levels of R&D spending. Switzerland and the US are the world’s large importers, per capita, of elite physicists. CERN in Switzerland must play some role here but, because of difficulties caused by multiple affiliations, we have not attempted to separate out those scientists. Fifth, we find evidence, from regression equations in Tables 6 and 7, that among elite physicists a current affiliation in the US is associated with a 13%–19% higher h‐index. This may be a genuine productivity difference, or reflect some form of pro‐US citations bias, or some mixture of the two. How, conceptually, can we make sense of the data? One way to view the findings on physicists is as supporting a theoretical model in which in the modern globalised world the costs of migration are low. Intuitively, the idea is the following. Consider a world with very high costs – whether because of cultural differences across societies, or costly travel, or poor communication – of switching between countries. Then only the very best workers will migrate. This is because they alone are the ones who will make a big enough return from international labour mobility to outweigh the high costs. In that case, migrants will be disproportionately from the top end of the ability distribution. They will be outstanding scientists with, in our terminology, particularly large h‐indexes. Now contrast this with the case of low mobility costs. In that case, elite scientists of more average kinds of abilities, like the norm within the country into which they migrate, will find it rational to choose to switch nations. Hence mobile incoming scientists will be of similar quality to the average of those in the receiving nation, and most of these newcomers will not go on to win science prizes in the way that happened in an older world – think of an early twentieth century setting of ocean liners and telegrams – where mobility costs were high.12 Any increases through time in the wage premium (p in our earlier notation) earned by distinguished scientists in the rich receiving countries will act to reinforce these tendencies. Footnotes 1 " Space constraints mean that it is not possible to summarise the literature here but valuable papers in economics include Beine et al. (2001, 2008), Bhagwati and Hamada (1974), Johnson and Regets (1998), Kanbur and Rapoport (2005), Saint‐Paul (2004), Schiff (2005) and Grossman and Stadelmann (2008). Like us, Stephan and Levin (2001) and Constant and D’Agosto (2008) focus on truly high‐skill individuals. Commander et al (2004) is a helpful survey, and Zimmermann (1995) and Stephan (1996) review parts of this literature. See also Bekhradnia and Sastry (2005), Kuhn and McAusland (2006), and Zaiceva and Zimmerman (2008). 2 " Three in physics (solid state/condensed matter physics, particle physics, and atomic and molecular physics) and three in Earth Science (oceanography, geophysics and geology). 3 " Defined by Laudel (2003, p. 218) as ‘a community of scientists who directly or indirectly interact in the production of new knowledge about a common subject matter’. 4 " Publication counts over two years; these publication counts adjust for co‐authorship and then adjust for journal impact factors (a measure of quality) and are finally adjusted for both co‐authorship and journal impact factors. 5 " Australia, Austria, Canada, Denmark, Finland, France, Germany, Greece, Hungary, Ireland, Italy, Japan, Korea, New Zealand, Poland, Spain, Sweden, Switzerland, Turkey, UK and US 6 " Israel, Argentine, Chile, China, India, Russia, Brazil, Iran, Taiwan 7 " We had to distinguish between good scientists and outstanding ones. Citations, although not perfect, therefore seemed the best metric. As pointed out by Starbuck (2005), Oswald (2007a) and others, even prestigious journals publish large numbers of papers that make little impact. See Van Raan (2000) for more on the use of citations. 8 " United Nation’s Statistics Division country classification; Russia is included in Europe and Turkey in Asia. 9 " The exact effects are calculated from: 100[exp(β)−1]. 10 " Nevertheless, as a referee has emphasised to us, the UK still does well in most sciences by the standards of other European countries. 11 " This is somewhat against the spirit of, for example, Pierson and Cotgreave (2000), who focus, arguably a little strangely, on citations per paper rather than on total citations (their data show that stayers write more papers than movers but the authors do not discuss this fact). 12 " Although lower mobility costs alone do not imply that the absolute numbers of immigrant Nobel winners would fall. References Bekhradnia , G. and Sastry , T. ( 2005 ). ‘Brain drain: migration of academic staff to and from the UK’ , HEPI Paper No. 19, Higher Education Policy Institute , Oxford. Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Beine , M. , Docquier , F. and Rapoport , H. ( 2001 ). ‘Brain drain and economic growth: theory and evidence’ , Journal of Development Economics , vol. 4 ( 1 ), pp. 275 – 89 . Google Scholar Crossref Search ADS WorldCat Beine , M. , Docquier , F. and Rapoport , H. ( 2008 ). ‘Brain drain and human capital formation in developing countries: winners and losers’ , Economic Journal , vol. 118 ( 528 ), pp. 631 – 52 . Google Scholar Crossref Search ADS WorldCat Bhagwati , J.N. and Hamada , K. ( 1974 ). ‘The brain drain, international integration of markets for professionals and unemployment: a theoretical analysis’ , Journal of Development Economics , vol. 1 ( 1 ), pp. 19 – 42 . Google Scholar Crossref Search ADS WorldCat Carayol , N. and Matt , M. ( 2005 ). ‘Individual and collective determinants of academic scientists’ productivity’ , Information Economics and Policy , vol. 18 , pp. 55 – 72 . Google Scholar Crossref Search ADS WorldCat Charlton , B.G. ( 2007 ). ‘Scientometric identification of elite ‘‘revolutionary science’’ research institutions by analysis of trends in Nobel prizes 1947‐2006’ , Medical Hypotheses , vol. 68 , pp. 931 – 4 . Google Scholar Crossref Search ADS PubMed WorldCat Commander , S. , Kangasniemi , M. and Winters , L.A. ( 2004 ). ‘The brain drain: curse or boon? A survey of the literature’ , in (R.E. Baldwin and L.A. Winters, eds), Challenges to Globalization: Analyzing the Economics, pp. 235 – 78 , Chicago: University of Chicago Press . Constant , A.F. and D’Agosto , E. ( 2008 ). ‘Where do the brainy Italians go? ’, IZA Discussion Paper No. 3325. Dietz , J. and Bozeman , B. ( 2005 ). ‘Academic careers, patents, and productivity: Industry experience as scientific and technical human capital’ . Research Policy , vol. 34 , pp. 349 – 67 . Google Scholar Crossref Search ADS WorldCat Editorial ( 2007 ). ‘In praise of the brain drain’ . Nature , vol. 446 , p. 231 . OpenURL Placeholder Text WorldCat Egger , H. , Falkinger , J. and Grossmann , V. ( 2007 ). ‘Brain drain, fiscal competition, and publifc education expenditure’ , IZA Discussion Paper No. 2747. Einstein , A. ( 1934 ). The World as I See it , English translation, Carol Publishing Group Edition (1999) . US: Carol Publishing Group . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Freeman , R.B. ( 2006 ). ‘People flows in globalization’ , Journal of Economic Perspectives , vol. 20 , (Spring), pp. 145 – 70 . Google Scholar Crossref Search ADS WorldCat Grossman , V. and Stadelmann , D. ( 2008 ). ‘International mobility of the highly skilled, endogenous R & D, and public infrastructure investment’ , IZA Discussion Paper No. 3366. Henrekson , M. and Waldenstrom , D. ( 2007 ). ‘Should research performance be measured unidimensionally? Evidence from rankings of academic economists’ , Working Paper, Research Institute of Industrial Economics , Stockholm. Hirsch , J. E. ( 2005 ). ‘An index to quantify an individual’s scientific research output’ . Proceedings of National Academy Science , vol. 102 ( 46 ), pp. 6569 – 72 . Google Scholar Crossref Search ADS WorldCat Hunter , R. ( 2007 ). Undergraduate dissertation , Department of Economics, Warwick University . Ioannidis , J.P.A. ( 2004 ). ‘Global estimates of high‐level brain drain and deficit’ . The Journal of the Federation of American Societies for Experimental Biology , vol. 18 , pp. 936 – 9 . Google Scholar Crossref Search ADS WorldCat Ionescu , A. and Polgreen , L. ( 2009 ). ‘A theory of brain drain and public funding for higher education in the US’ , American Economic Review, May 2009 (Papers & Proceedings) , forthcoming. OpenURL Placeholder Text WorldCat Institute for Scientific Information (ISI) . Highly Cited Physicists [online] . Available from: http://www.isihighlycited.com [originally accessed October 2006]. Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Johnson , J. M. and Regets , M. ( 1998 ). ‘International mobility of scientists and engineers to the United States – brain drain or brain circulation?’ , National Science Foundation (June), NSF 98–316. Kanbur , R. and Rapoport , H. ( 2005 ). ‘Migration selectivity and the evolution of spatial inequality’ , Journal of Economic Geography , vol. 5 ( 1 ), pp. 43 – 57 . Google Scholar Crossref Search ADS WorldCat Kuhn , P.J. and McAusland , C. ( 2006 ). ‘The international migration of knowledge workers: when is brain drain beneficial?’ , IZA Discussion Paper No. 2493. Kwok , V. and Leland , H.E. ( 1982 ). ‘An economic model of the brain drain’ , American Economic Review , vol. 72 ( 1 ), pp. 91 – 100 . OpenURL Placeholder Text WorldCat Laband , D. and Tollison , R.D. ( 2003 ). ‘Good colleagues’ , Journal of Economic Behavior and Organization , vol. 52 ( 4 ), pp. 505 – 12 . Google Scholar Crossref Search ADS WorldCat Laudel , G. ( 2003 ). ‘Studying the brain drain: can bibliometric methods help?’ , Scientometrics , vol. 57 ( 2 ), pp. 215 – 37 . Google Scholar Crossref Search ADS WorldCat Laudel , G. ( 2005 ). ‘Migration currents among the scientific elite’ , Minerva , vol. 43 ( 4 ), pp. 377 – 95 . Google Scholar Crossref Search ADS WorldCat Leimu , R. and Koricheva , J. ( 2005 ). ‘What determines the citation frequency of ecological papers?’ , Trends in Ecology and Evolution , vol. 20 ( 1 ), pp. 28 – 32 . Google Scholar Crossref Search ADS PubMed WorldCat Levin , S.G. and Stephan , P.E. ( 1991 ). ‘Research productivity over the life‐cycle: evidence for academic scientists’ . American Economic Review , vol. 81 ( 1 ), pp. 14 – 32 . OpenURL Placeholder Text WorldCat Levin , S.G. and Stephan , P.E. ( 1999 ). ‘Are the foreign‐born a source of strength for US science?’ , Science , vol. 285 , pp. 213 – 4 . Google Scholar Crossref Search ADS WorldCat Machin , S. and Oswald , A.J. ( 2000 ). ‘UK economics and the future supply of academic economists’ , Economic Journal , vol. 110 , pp. 334 – 49 . Google Scholar Crossref Search ADS WorldCat OECD ( 2000 ) Science, Technology and Industry Outlook 2000 . Paris: OECD . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC OECD . OECD Statistics v.4.2 [online] , available from: http://stats.oecd.org/WBOS/ [accessed 18/02/2007]. Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Oswald , A.J. ( 2007a ). ‘An examination of the reliability of prestigious scholarly journals: evidence and implications for decision‐makers’ , Economica , vol. 74 , pp. 21 – 31 . Google Scholar Crossref Search ADS WorldCat Oswald , A.J. ( 2007b ). ‘Comment on Summers on life sciences’ , Financial Times , January 29th. OpenURL Placeholder Text WorldCat Oswald , A.J. ( 2009 ). ‘World‐leading research and its measurement’ , Warwick Economics Research Paper No. 887. Oswald , A.J. and Ralsmark , H. ( 2007 ). ‘Some evidence on the future of economics’ , Royal Economic Society Newsletter , July. OpenURL Placeholder Text WorldCat Pierson , A.S. and Cotgreave , P. ( 2000 ). ‘Citation figures suggest that the UK brain drain is a genuine problem’ , Nature , vol. 407 , pp. 13 . Google Scholar Crossref Search ADS PubMed WorldCat Regets , M. ( 2001 ). ‘Research and policy issues in high‐skilled international migration: a perspective with data from the United States’ , IZA Discussion Paper No. 366. Royal Society ( 1963 ). ‘Emigration of scientists from the United Kingdom: report of a committee appointed by the Council of the Royal Society’ , London: Royal Society . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Saint‐Paul , G. ( 2004 ). ‘The brain drain: some evidence from European expatriates in the United States’ , IZA Discussion Paper No. 1310. Schiff , M. ( 2005 ). ‘Brain gain: claims about its size and impact on welfare and growth are greatly exaggerated’ , IZA Discussion Paper No. 1599. Starbuck , W.H. ( 2005 ). ‘How much better are the most‐prestigious journals? The statistics of academic publication’ , Organization Science , vol. 16 , pp. 180 – 200 . Google Scholar Crossref Search ADS WorldCat Stephan , P.E. ( 1996 ). ‘The economics of science’ , Journal of Economic Literature , vol. 34 ( 3 ), pp. 1199 – 235 . OpenURL Placeholder Text WorldCat Stephan , P.E. and Levin , S.G. ( 2001 ). ‘Exceptional contributions to US science by the foreign‐born and foreign‐educated’ , Population Research and Policy Review , pp. 59 – 79 . OpenURL Placeholder Text WorldCat Summers , L.H. ( 2007 ). ‘America must not surrender its lead in life sciences’ , Financial Times , January 29th. OpenURL Placeholder Text WorldCat Universities UK ( 2007 ). ‘Talent wars: the international market for academic staff’ , Policy Briefing, July. Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Van Leeuwen , T.N. , Moed , H.F, Tijssen , R.J.W, Visser , M.S and Van Raan , A.F.J. ( 2001 ). ‘Language biases in the coverage of the Scientific Citation Index and its consequences for interpretational comparisons of national research performance’ . Scientometrics , vol. 51 ( 1 ), pp. 335 – 46 . Google Scholar Crossref Search ADS WorldCat Van Raan , A.F.J. ( 2000 ). ‘The Pandora’s box of citation analysis: measuring scientific excellence – the last evil?’ , in (B. Cronin and H.E. Atkins, eds) The Web of Knowledge – A Festschrift in Honor of Eugene Garfield , pp. 301 – 19 , Medford, New Jersey: Information Today, Inc . OpenURL Placeholder Text WorldCat Waldinger , F. ( 2008 ). ‘Peer effects in science: evidence from the dismissal of scientists in Nazi Germany’ , Working Paper, London School of Economics. Warwick University ( 2007 ). ‘Elite universities and the global brain drain’ , conference paper presented at the Second International Conference on World‐Class Universities , Shanghai. Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Weinberg , B.A. ( 2009 ). ‘An assessment of British science over the 20th century’ , Economic Journal, this issue. Weinberg , B.A. and Galenson , D.W. ( 2005 ). ‘Creative careers: the life cycles of Nobel laureates in economics’ , NBER Working Paper No. 11799. Wong , B. and Kokko , H. ( 2005 ). ‘Is science as global as we think?’ , Trends in Ecology and Evolution , vol. 20 ( 9 ), pp. 475 – 6 . Google Scholar Crossref Search ADS PubMed WorldCat Zaiceva , A. and Zimmerman , K.F. ( 2008 ). ‘Scale, diversity, and determinants of labour migration in Europe’ , Oxford Review of Economic Policy , vol. 24 ( 3 ), (Fall), pp. 428 – 52 . Google Scholar Crossref Search ADS WorldCat Zimmermann , K.F. ( 1995 ). ‘Tackling the European migration problem’ , Journal of Economic Perspectives , vol. 9 ( 2 ) (Spring), pp. 45 – 62 . Google Scholar Crossref Search ADS WorldCat Appendix A1 [ Measurement of the Brain Drain among Elite US‐Based Assistant Professors of Economics
Source. Oswald and Ralsmark (2007).
Note. This draws upon data on 112 assistant professors in the top‐10 US departments of economics. ] Table A1 
Net Losses and Gains in Physics Researchers by Country Country . Number after undergraduate BSc. . Number present day . Net gain . Net gain normalised by population . Australia 4 1 −3 −0.183 Austria 1 1 0 −0.128 Brazil 1 0 −1 −0.006 Canada 5 2 −3 −0.112 China 2 1 1 −0.001 Denmark 1 1 0 0.000 Finland 2 2 0 0.000 France 4 2 −2 −0.035 Germany 12 12 0 −0.028 Greece 2 1 −1 −0.099 Hungary 1 0 −1 −0.095 India 2 0 −2 −0.002 Ireland 1 0 −1 −0.286 Italy 2 2 0 0.000 Japan 8 7 −1 −0.008 Rep. of Korea 1 0 −1 −0.024 New Zealand 1 0 −1 −0.029 Poland 1 0 −1 −0.027 Russia 2 0 −2 −0.013 Spain 1 1 0 −0.026 Sweden 1 0 −1 −0.011 Switzerland 3 9 6 0.599 Turkey 1 0 −1 −0.018 UK 12 6 −6 −0.104 US 65 106 41 0.158 Country . Number after undergraduate BSc. . Number present day . Net gain . Net gain normalised by population . Australia 4 1 −3 −0.183 Austria 1 1 0 −0.128 Brazil 1 0 −1 −0.006 Canada 5 2 −3 −0.112 China 2 1 1 −0.001 Denmark 1 1 0 0.000 Finland 2 2 0 0.000 France 4 2 −2 −0.035 Germany 12 12 0 −0.028 Greece 2 1 −1 −0.099 Hungary 1 0 −1 −0.095 India 2 0 −2 −0.002 Ireland 1 0 −1 −0.286 Italy 2 2 0 0.000 Japan 8 7 −1 −0.008 Rep. of Korea 1 0 −1 −0.024 New Zealand 1 0 −1 −0.029 Poland 1 0 −1 −0.027 Russia 2 0 −2 −0.013 Spain 1 1 0 −0.026 Sweden 1 0 −1 −0.011 Switzerland 3 9 6 0.599 Turkey 1 0 −1 −0.018 UK 12 6 −6 −0.104 US 65 106 41 0.158 Open in new tab Table A1 
Net Losses and Gains in Physics Researchers by Country Country . Number after undergraduate BSc. . Number present day . Net gain . Net gain normalised by population . Australia 4 1 −3 −0.183 Austria 1 1 0 −0.128 Brazil 1 0 −1 −0.006 Canada 5 2 −3 −0.112 China 2 1 1 −0.001 Denmark 1 1 0 0.000 Finland 2 2 0 0.000 France 4 2 −2 −0.035 Germany 12 12 0 −0.028 Greece 2 1 −1 −0.099 Hungary 1 0 −1 −0.095 India 2 0 −2 −0.002 Ireland 1 0 −1 −0.286 Italy 2 2 0 0.000 Japan 8 7 −1 −0.008 Rep. of Korea 1 0 −1 −0.024 New Zealand 1 0 −1 −0.029 Poland 1 0 −1 −0.027 Russia 2 0 −2 −0.013 Spain 1 1 0 −0.026 Sweden 1 0 −1 −0.011 Switzerland 3 9 6 0.599 Turkey 1 0 −1 −0.018 UK 12 6 −6 −0.104 US 65 106 41 0.158 Country . Number after undergraduate BSc. . Number present day . Net gain . Net gain normalised by population . Australia 4 1 −3 −0.183 Austria 1 1 0 −0.128 Brazil 1 0 −1 −0.006 Canada 5 2 −3 −0.112 China 2 1 1 −0.001 Denmark 1 1 0 0.000 Finland 2 2 0 0.000 France 4 2 −2 −0.035 Germany 12 12 0 −0.028 Greece 2 1 −1 −0.099 Hungary 1 0 −1 −0.095 India 2 0 −2 −0.002 Ireland 1 0 −1 −0.286 Italy 2 2 0 0.000 Japan 8 7 −1 −0.008 Rep. of Korea 1 0 −1 −0.024 New Zealand 1 0 −1 −0.029 Poland 1 0 −1 −0.027 Russia 2 0 −2 −0.013 Spain 1 1 0 −0.026 Sweden 1 0 −1 −0.011 Switzerland 3 9 6 0.599 Turkey 1 0 −1 −0.018 UK 12 6 −6 −0.104 US 65 106 41 0.158 Open in new tab Table A2 
Comparison with Ioannidis (2004) Finding . This paper . Ioannidis – physics . Ioannidis – overall . Sample size 158 46 1523 Moved since birth 44% 50% 33% (approx.) US‐born scientists who migrated 0% – 2% Females 0.63% – 3‐4% Finding . This paper . Ioannidis – physics . Ioannidis – overall . Sample size 158 46 1523 Moved since birth 44% 50% 33% (approx.) US‐born scientists who migrated 0% – 2% Females 0.63% – 3‐4% Open in new tab Table A2 
Comparison with Ioannidis (2004) Finding . This paper . Ioannidis – physics . Ioannidis – overall . Sample size 158 46 1523 Moved since birth 44% 50% 33% (approx.) US‐born scientists who migrated 0% – 2% Females 0.63% – 3‐4% Finding . This paper . Ioannidis – physics . Ioannidis – overall . Sample size 158 46 1523 Moved since birth 44% 50% 33% (approx.) US‐born scientists who migrated 0% – 2% Females 0.63% – 3‐4% Open in new tab Table A3 
Summary Statistics – Highly‐cited Physicists Variable . Observations . Mean . Std. Dev. . Minimum . Maximum . Presently in the US hindex 105 61.04 13.00 30 115 results 105 383.82 208.25 56 1248 avcite 105 57.52 36.56 15.74 216.84 totalcite 105 18458.52 13727.75 7409 135831 nocount 105 2.04 1.56 1 12 noinst 105 5.91 4.32 2 25 yborn 102 1947 9.928 1920 1976 ybsc 101 1968 9.39 1941 1986 yphd 105 1974 9.36 1947 1991 Presently not in the US hindex 53 54.89 13.72 22 85 results 53 445.60 270.99 82 1314 avcite 53 47.08 34.33 4.95 179.48 totalcite 53 15591.47 6757.03 2355 32714 nocount 53 3.15 1.89 1 10 noinst 53 6.26 4.44 1 21 yborn 50 1945 8.31 1926 1974 ybsc 44 1968 7.86 1951 1983 yphd 53 1972 7.99 1955 1987 Variable . Observations . Mean . Std. Dev. . Minimum . Maximum . Presently in the US hindex 105 61.04 13.00 30 115 results 105 383.82 208.25 56 1248 avcite 105 57.52 36.56 15.74 216.84 totalcite 105 18458.52 13727.75 7409 135831 nocount 105 2.04 1.56 1 12 noinst 105 5.91 4.32 2 25 yborn 102 1947 9.928 1920 1976 ybsc 101 1968 9.39 1941 1986 yphd 105 1974 9.36 1947 1991 Presently not in the US hindex 53 54.89 13.72 22 85 results 53 445.60 270.99 82 1314 avcite 53 47.08 34.33 4.95 179.48 totalcite 53 15591.47 6757.03 2355 32714 nocount 53 3.15 1.89 1 10 noinst 53 6.26 4.44 1 21 yborn 50 1945 8.31 1926 1974 ybsc 44 1968 7.86 1951 1983 yphd 53 1972 7.99 1955 1987 Here, hindex is h‐index; results is the number of papers; avcite is average cites per paper; total cite is total lifetime citations; nocount is missing; noinst is missing institution; yborn is year born; ybsc is year of BSc degree; yphd is year of PhD degree. Further details are available in Hunter (2007) or on request. Open in new tab Table A3 
Summary Statistics – Highly‐cited Physicists Variable . Observations . Mean . Std. Dev. . Minimum . Maximum . Presently in the US hindex 105 61.04 13.00 30 115 results 105 383.82 208.25 56 1248 avcite 105 57.52 36.56 15.74 216.84 totalcite 105 18458.52 13727.75 7409 135831 nocount 105 2.04 1.56 1 12 noinst 105 5.91 4.32 2 25 yborn 102 1947 9.928 1920 1976 ybsc 101 1968 9.39 1941 1986 yphd 105 1974 9.36 1947 1991 Presently not in the US hindex 53 54.89 13.72 22 85 results 53 445.60 270.99 82 1314 avcite 53 47.08 34.33 4.95 179.48 totalcite 53 15591.47 6757.03 2355 32714 nocount 53 3.15 1.89 1 10 noinst 53 6.26 4.44 1 21 yborn 50 1945 8.31 1926 1974 ybsc 44 1968 7.86 1951 1983 yphd 53 1972 7.99 1955 1987 Variable . Observations . Mean . Std. Dev. . Minimum . Maximum . Presently in the US hindex 105 61.04 13.00 30 115 results 105 383.82 208.25 56 1248 avcite 105 57.52 36.56 15.74 216.84 totalcite 105 18458.52 13727.75 7409 135831 nocount 105 2.04 1.56 1 12 noinst 105 5.91 4.32 2 25 yborn 102 1947 9.928 1920 1976 ybsc 101 1968 9.39 1941 1986 yphd 105 1974 9.36 1947 1991 Presently not in the US hindex 53 54.89 13.72 22 85 results 53 445.60 270.99 82 1314 avcite 53 47.08 34.33 4.95 179.48 totalcite 53 15591.47 6757.03 2355 32714 nocount 53 3.15 1.89 1 10 noinst 53 6.26 4.44 1 21 yborn 50 1945 8.31 1926 1974 ybsc 44 1968 7.86 1951 1983 yphd 53 1972 7.99 1955 1987 Here, hindex is h‐index; results is the number of papers; avcite is average cites per paper; total cite is total lifetime citations; nocount is missing; noinst is missing institution; yborn is year born; ybsc is year of BSc degree; yphd is year of PhD degree. Further details are available in Hunter (2007) or on request. Open in new tab Table A4 
The h‐indexes of Highly‐cited Bio‐scientists Currently Working in the United States Sample size: 163 . Mean . Lower bound (95%) . Upper bound (95%) . Birth to BSc Moved to US 88.60 70.51 106.69 Remained 89.67 82.80 96.54 BSc to PhD Moved to US 83.38 78.13 88.62 Remained 88.63 82.38 94.87 . Mean . Lower bound (95%) . Upper bound (95%) . Birth to BSc Moved to US 88.60 70.51 106.69 Remained 89.67 82.80 96.54 BSc to PhD Moved to US 83.38 78.13 88.62 Remained 88.63 82.38 94.87 Source.Warwick University (2007) Open in new tab Table A4 
The h‐indexes of Highly‐cited Bio‐scientists Currently Working in the United States Sample size: 163 . Mean . Lower bound (95%) . Upper bound (95%) . Birth to BSc Moved to US 88.60 70.51 106.69 Remained 89.67 82.80 96.54 BSc to PhD Moved to US 83.38 78.13 88.62 Remained 88.63 82.38 94.87 . Mean . Lower bound (95%) . Upper bound (95%) . Birth to BSc Moved to US 88.60 70.51 106.69 Remained 89.67 82.80 96.54 BSc to PhD Moved to US 83.38 78.13 88.62 Remained 88.63 82.38 94.87 Source.Warwick University (2007) Open in new tab Author notes " For many helpful suggestions, we thank three referees and Robin Ball, Sandra Chapman, Joseph Falkinger, Amanda Goodall, Simon Hands, Steve Machin, Robert May, Robert MacKay, David Spiegelhalter, Richard Tol, Radu Vranceanu, Fabian Waldinger and co‐authors on the unpublished paper by Warwick University (2007), especially Showkat Ali and Hilda Ralsmark, for allowing us to draw upon some of the results described there. We are extremely grateful to all the physicists who replied to our emails. Oswald’s work was supported by an ESRC research professorship. © The Author(s). Journal compilation © Royal Economic Society 2009

Linton, Oliver, B.

doi: 10.1111/j.1468-0297.2009.02282.xpmid: N/A

This book is an update on the author’s previous books and brings the reader up to date with various recent developments in the field including work of the author on treatment effects and decision analysis. The core material here is concerned with the new way of thinking about economics data that are incomplete or have missing components. The main idea is based on the law of total probability where z is an indicator of observability. The object of interest is the conditional c.d.f. P(y | x) but we can only estimate P(y | x,z = 1), P(z = 1 | x), and P(z = 0 | x) from available data. The traditional approach is to specify parametric or semiparametric models for the unknown quantities such that one can consistently estimate parameters of P(y | x). The author explains how such approaches are doomed when those model assumptions are incorrect and, since we often have no way of verifying whether the model assumptions hold, the results thereby obtained are not credible broadly. However, he shows that without any assumptions the object of interest P(y | x) lies in the set The point is that H is a strict subset of [0,1] whenever P(z = 0 | x) ∈ (0,1). This shows that even making only very weak assumptions can lead us to a quantifiable restriction on the data. This idea has made a great impact in econometrics and there has been much recent work developing this methodology in a number of directions following Manski’s pioneering earlier work. Amongst other things, this book reviews such developments. There is some discussion of estimation of nonparametric regression and conditional c.d.f.1 and what is meant by consistency in the context where a set is the estimation target but mostly the book, as suggested by the title, is concerned with identification. The second section applies these ideas further in the context of treatment response. The value of monotonicity assumptions such as monotone treatment selection is described. In particular, they can also provide sharp bounds on parameters of interest such as mean treatment effect. Another chapter in this section introduces decision theory into the mix as a way of defining a best treatment rule. Several decision principles are discussed including the author’s favoured minimax regret rule, in which the decision maker chooses an action that minimises the maximum loss to welfare that results from not knowing the objective function. The next chapter takes this further and gives an application of decision theory to using a randomised experiment to evaluate an innovation. The third section of the book reviews recent developments in choice theory including discussions of Samuelson’s revealed preference, McFadden’s random utility models and more recent Kahneman and Tversky’s ideas. Many chapters have a detailed discussion of an application. In Chapter 1, this is predicting criminality and predicting High School graduation. In Chapter 2, we learn about wage regressions and the reservation‐wage model of labour supply. In Chapter 5 there is a discussion of the income distribution in the US obtained from the CPS. In Chapter 6, there is a section on using administrative records to infer AFDC transition rates. In Chapter 7, there is a section on sentencing and recidivism. These applications are very well integrated into the text and enhance the value of this monograph considerably. One can be interested in other quantities like conditional mean and conditional median and here some further issues arise. Specifically, for the conditional mean, one may not obtain any reduction in the set of feasible values H when the support of the dependent variable is the whole real line. In this case, one can only say that nothing is known about such a parameter.2 However, for conditional quantiles one can obtain a reduction. However, what about average derivatives? This is a very central parameter in applied econometrics but, unfortunately, one cannot obtain a useful bound on this quantity even for the case of a conditional c.d.f., since ∂P(y | x,z = 0)/∂x can take any value. These facts are explained early on in the book and subsequent material directs attention to the case where there is a reduction in the feasible set or rather to certain parameters for which this pertains, in particular, to parameters that respect stochastic dominance. Is this book the way forward for the empirical researcher? As Chairman Mao replied when asked about whether he thought the French Revolution was a good thing, ‘it is too early to tell’. However, it is a major event in the econometrics of the last 20 years. Manski is one of the most creative and original econometricians of his generation. His book is incredibly well written and presented and interesting and will be a useful resource for a large set of microeconometric researchers. I conclude my review with some quibbling comments that mostly arise from my personal envy of Manski’s big impact on the profession. It seems to me that while it is necessary to study first identification it is not by itself sufficient. Nor should it in my view be a subject separated off from inference. Just as foreplay may be necessary and enjoyable, it is not by itself much use in reproducing the species. Consistency of an estimator is not by itself terribly useful because it does not restrict the range of values the parameter can take for a given dataset, even approximately. The further results associated with inference provide much more precise description of the uncertainty in going from sample to population. This is true for both the point identified case and the set identified case. Providing rules of inference in the set identified case is an area of much recent work reflecting the intellectual challenge and importance. The author argues against the use of modelling, parametric or otherwise, on the grounds of credibility. He introduced earlier The Law of Decreasing Credibility: the credibility of inference decreases with the strength of the assumptions maintained. I agree with the general principle behind this but see some issues in always therefore adopting the most general setting.3 I come to my main point. In the absence of missing data, the author would presumably be advocating estimation of the conditional distribution P(y | x) and functionals of it regardless of the dimensionality of x. There is no identification issue here but there are plenty of practical issues to bridge the gap between what it is theoretically possible to learn if we have a specific type of data and what one can realistically hope to learn with the amount of data available in practice. Consider the extension to a time series setting where we want to predict a time series yt from its own past and the present and past of a vector of covariates. In this case, the relevant conditioning set is infinite dimensional. Although it is possible to estimate such conditional distributions consistently (Linton and Sancetta, forthcoming), in practice such results are meaningless. This approach is difficult to justify for well‐known reasons. First, how do we understand the shape of a hundred or thousand dimensional function or sets of such functions? Second, the curse of dimensionality means that estimating such models is pointless because the confidence interval will include not just the whole real line but perhaps also the complex plane. The point is that in most problems the most general starting point is too general. In most empirical studies there are many plausible covariates. The applications considered in the book all have at most one covariate. The author forcefully argues that just including other variables does not necessarily better control for endogeneity or other problems (conditioning is not controlling), which parallels the non‐nestedness of the assumptions of independence and conditional independence. But in many cases one is interested in the effects of many covariates and, as discussed above, one then runs into serious practical issues of presentation and interpretation of empirical results. As an alternative, most econometricians and statisticians would advocate some kind of model selection procedure where irrelevant and less relevant covariates are filtered out by a procedure that works out‐of‐sample and the set of models considered are restricted in terms of their effective dimensionality. In many practical time series it is well known that better out of sample predictions can be obtained from simpler models either in terms of number of covariates or in terms of functional form. However, the principles behind the author’s approach can be applied to these restricted settings and I look forward to seeing future developments along these lines. Finally, the author criticises Milton Friedman for advocating the Popperian approach to scientific methodology in economics, whereby a single strong hypothesis should be advocated and tested: ‘I do not see why a scientist must choose one hypothesis to hold’. This seems to be a different point. The usual statistical approach is of course consistent with ambiguity, I mean one usually has parameters θ = (α,β), real valued or function valued and hypotheses about a subset of parameters α, say. For example α = 0 or α ≤ 0. The nuisance parameters β are not restricted by the theory. This is called a composite hypothesis and means that the null hypothesis is that θ ∈ Θ0 Θ a set of values. There is nothing incompatible with the author’s ideas here.4 Furthermore, just because a source of data can only partially reveal P(y | x), does this mean we have to formulate our scientific hypothesis in this way? It seems a dangerous idea to advocate not testing ones theories, although it should be said that statistical hypothesis testing is a very heavily criticised idea in a number of fields. Working with the weakest possible assumptions under which the parameter of interest is identified usually means that this hypothesis is untestable. I think one needs to take a starting point somewhere less ambitious. Footnotes 1 " Strictly speaking, the discussion on non‐parametric regression in Section 1.5 is incorrect or at least confusing. First, it is not necessary for consistency that the conditional variance be bounded, we as usual only require bounded first moments for consistency. Second, it is not even necessary that the regression function be continuous, one can have consistent estimation of the regression function under a cadlag assumption (i.e., a finite number of discontinuities) provided one uses one‐sided kernels. 2 " As discussed in the book sometimes shape restrictions like monotonicity can reduce the identified set. 3 " As usual there is some ambiguity implicit in such a general principle. It is generically the case that consistent estimation of a parameter from sample data requires stronger assumptions than are required for the parameter itself to be well defined. To conduct statistical inference about such a parameter will require even stronger assumptions and these assumptions may have to be strengthened more in some contexts than others. So any discussion about the strength of assumptions depends on whether identification, consistency, or inference results are being considered. Furthermore, the set of assumptions used to justify inference is often only partially ordered. For example, the assumptions used for inference about the median are non‐nested with those used for inference about the mean. Likewise, the assumptions used for inference in linear models are non‐nested with those used for nonparametric regression. Specifically, nonparametric regression requires increasing information locally while linear regression requires increasing global information and these two requirements can be quite different. For estimation of a slope coefficient there is no requirement that the covariates have a density or have compact support, whereas these assumptions are frequently made in the context of nonparametric regression. Incidentally, I do not know what are the necessary and sufficient conditions for consistent estimation of the conditional c.d.f. Unlike the c.d.f. these conditions could vary depending on the metric used and the assumptions about the support of the conditioning variables. 4 " For example, demand is downward sloping in price. Demand could also depend on other variables whose effect is not restricted. This is one of the central predictions of economics recognisable by the laymen and it seems reasonable to try to test whether this restriction is true. This is a restriction that a certain function D(x,P), where P represents price, satisfies ∂D(x,P)/∂P ≤ 0. Is this a single hypothesis or a set of hypotheses? Reference Linton , O.B. and Sancetta , A. (forthcoming). ‘ Consistent estimation of a general nonparametric regression function in time series ’, Journal of Econometrics . OpenURL Placeholder Text WorldCat © The Author(s). Journal compilation © Royal Economic Society 2009

Nelson,, Edward

doi: 10.1111/j.1468-0297.2009.02278.xpmid: N/A

Abstract The inflation targeting regime prevailing in the UK is not the result of a change in policy maker objectives. Analysis of UK policymakers’ statements demonstrates that objectives have been essentially unchanged over five decades. Instead, the crucial underpinning of UK inflation targeting is an overhaul of doctrine. This overhaul involves changes in policymakers’ views regarding key IS and Phillips curve parameters. They particularly have involved whether levels terms (of the real interest rate and the output gap) appear in the curves. Contrary to conventional wisdom, changing views on the expected‐inflation term in the Phillips curve do not play a role. The substantial improvement in the UK’s economic performance during the 1990s and 2000s, manifested in lower mean inflation and reduced variability for key macroeconomic series, is well known. But if one takes for granted (as I do) that successful monetary policy was a major reason for that improvement, this leaves unanswered the reasons for the inappropriate monetary policy of prior decades. For example, did economic outcomes in the 1960s and 1970s reflect attempts by UK policymakers to select from a menu of inflation/unemployment combinations, only to find themselves at an unexpected – and unpalatable – combination not on the original menu? Alternatively, did UK policymakers simply not value price stability, with better recent performance reflecting a belated recognition of the costs of inflation? Or are still other explanations superior? The repeated revisiting of the 1960s and 1970s in articles and speeches by UK policymakers, e.g., King (2005, 2007) suggests that the controversy associated with these questions is not dying out. Understanding the reason for the changes in UK macroeconomic behaviour is a matter of global interest. The US has been the focus of the discussion of the Great Moderation in studies such as Stock and Watson (2002) and Bernanke (2004). But the changes in policy regime and in data behaviour in the last quarter century have been greater in the UK than in the US. Consequently, as stressed by Benati (2008), Blake and Markovic (2007) and others, the UK serves as a particularly valuable laboratory for testing hypotheses regarding the Great Moderation. This article argues, based on my own investigation of the record, that the inflation targeting regime prevailing in the UK does not reflect a change in policy maker objectives. These objectives have been essentially unchanged; instead, the crucial underpinning of UK inflation targeting has been an overhaul of doctrine – a changed view of the transmission mechanism. This overhaul can be understood in terms of changes in policymakers’ views on the values of a few key parameters in their specifications of the economy’s IS and Phillips curves. The changed views pertain to the issues of whether interest rates enter the IS equation and the extent of policy maker influence on those rates; whether a speed‐limit term matters for inflation dynamics; and whether the level of the output gap appears in the Phillips curve when the gap is negative. My focus on these issues produces a factually defensible account of UK policy changes, in contrast to approaches which attribute regime changes to shifts in policy maker objectives or to changing views about the long‐run inflation/unemployment trade‐off. An implication of my analysis is that inflation targeting since 1992 does not reflect a further overhaul of doctrine but instead a refinement within the same basic doctrinal framework prevailing since 1979. While building on my previous studies of the UK,1 the present article has the specific aim of providing an analogue to Romer and Romer’s (2004) study of the views of successive Federal Reserve Chairmen. I take into account two institutional features about the UK. First, due to the lack of Bank of England independence until 1997, the principal monetary policymakers for most of the years studied were the Prime Minister and other senior government members such as the Chancellor of the Exchequer, not the Bank of England Governor. Therefore, for the pre‐1997 period, my focus is on the economic doctrines subscribed to by the Prime Minister and ministerial colleagues. Second, the lack of a clear UK analogue to FOMC Minutes before the 1990s and the absence of any UK analogue to FOMC Transcripts for any period, lead me to assemble an alternative base of material for studying policymakers’ views. I use a database of policy maker statements that substantially exceeds that used for prior studies of the UK. This database consists primarily of newspaper articles and parliamentary proceedings. At the outset let me state what I am not saying. I am not saying that pre‐1979 views of the economy were correct for their time and that therefore the old doctrine was appropriate. On the contrary, I believe that modern doctrine should have been used in the 1950s, 1960s and 1970s, and that the result would have been a far better UK economic performance. My position is that economic policies in those decades were based on an internally consistent but incorrect set of beliefs. This article proceeds as follows. In Section 1, I state the basic propositions underlying old and modern doctrine, and give equations that characterise each doctrine. Section 2 discusses how the old doctrine developed from the 1940s to 1964. Section 3 discusses advantages of my methodology; Section 4 considers the old doctrine over 1964‐79, while Section 5 considers other types of evidence on the 1960s and 1970s. Section 6 turns to modern doctrine and Section 7 concludes. 1. Statements of Old and Modern Doctrine In this Section I state the basic postulates underlying old and modern doctrine, and summarise each doctrine in terms of its implied IS and Phillips curve relations. 1.1. Basic Postulates of Old and Modern Doctrine Both the old doctrine (prevailing up to 1979) and modern doctrine (applying since 1979) consist of eight propositions, which I have divided into three categories: aggregate demand behaviour; inflation behaviour; and policy objectives. I first give the postulates of the old doctrine (OD). Aggregate demand behaviour OD1. Aggregate demand is insensitive to short‐term interest rates. Insofar as short rates matter at all, they enter the IS equation as a first‐differenced nominal rate. OD2. Fiscal policy has powerful effects on aggregate demand. The expansionary impact of deficit spending can be reduced if the deficit is financed by issuing long‐term debt; financing by Treasury bills, on the other hand, amounts to monetisation of the deficit. OD3. Aggregate demand may depend to a small degree on (real) long‐term interest rates. But this has no implications for short‐term interest rate policy, as the authorities can fix the long‐term rate directly for a given path of short rates and the monetary base. Inflation behaviour OD4. Cost‐push forces come from a variety of sources and can produce substantial and sustained inflation without any monetary accommodation. OD5. This is so because inflation is insensitive to negative output gaps. Positive output gaps can add to inflationary pressure but a given amount of economic slack does not remove inflationary pressure. OD6. Inflation, on the other hand, does depend in a symmetric manner on the change in the output gap. The preceding six postulates have contrasting analogous versions in modern doctrine (MD), prevailing since 1979. Aggregate demand behaviour MD1. Aggregate demand is highly interest‐elastic, with the levels of both real short‐term and real long‐term interest rates appearing in the IS equation. MD2. Pure fiscal policy – that is, deficit spending not accompanied by base money creation – does not have substantial effects on aggregate demand. MD3. The long‐term interest rate and other key asset prices are highly sensitive to the short‐term interest rate and expectations of its path. Inflation behaviour MD4. The degree of monetary accommodation is critical in determining whether cost‐push forces produce sustained inflation. MD5. This is so because inflation is sensitive to both positive and negative output gaps. The level of the output gap enters the Phillips curve. MD6. Inflation does not depend on the change in the output gap. I do not find it useful to attribute policy regime change to shifts in objectives. On the contrary, I will show that there has been considerable continuity in objectives. Thus I append both lists above with the same views about objectives. Policy objectives OD7 & MD7. A low single‐digit rate of inflation is desirable and makes it easier for aggregate demand to proceed along the path of potential GDP; in addition, low inflation ensures that productive potential is not damaged by inflation. OD8 & MD8. The real structure of the economy produces a potential output path that policymakers should use as their output reference value. Associated with this sustainable output value is an unemployment level which can vary with structural changes. 1.2. Representative Equations Old and modern doctrine may be represented in terms of IS and Phillips curve equations. Old doctrine saw log real aggregate demand (yt) as determined according to: (1) where b1 > 0, Δ is the first‐difference operator, Rt is the short‐term nominal interest rate and ‘t.i.p.’ denotes ‘terms independent of policy’ as defined by Woodford (2003), i.e., it covers all variables that are not sensitive to open‐market policy (the policy that delivers the short rate Rt). Note that, as in postulate (OD1), no level term for the interest rate (real or nominal) enters. The old doctrine’s hypothesised relation for inflation (πt) was where ξ > 0, δ > 0, yt* is log potential output, ut is a cost‐push shock and is an expected inflation term (whose specification is discussed below). Since this Phillips curve is asymmetric with respect to the output gap,2 it may be written more compactly as: (2) where Dt = 1 if (yt−yt*) ≥ 0; 0 otherwise. Reflecting postulates (OD5) and (OD6) of old doctrine, inflation in (2) depends continuously on the change in the output gap but is insensitive to negative levels of the output gap. The modern doctrine’s IS curve is: (3) where b2 > 0 b3 > 0 and rlt is the real long‐term interest rate. In line with postulate (MD1), the levels of both the real short‐term interest rate and the real long‐term interest appear as influences on aggregate demand in (3). The modern doctrine’s Phillips curve is the standard relation: (4) The remainder of this article justifies the above characterisation of UK official doctrine by documenting the prevalence of the two doctrines in UK policy circles and their influence on UK economic policy. 2. Development of Old Doctrine: 1945–64 In this Section I sketch how the old doctrine developed up to 1964. I consider developments in thinking by 1950s and 1960s policymakers, as well as contributions by participants in the economic debate who would enter policymaking from 1964 onward. 2.1. The Basic Developments Aggregate demand behaviour Views about the behaviour of aggregate demand (the IS equation) developed in several increments: Level of short‐term interest rate does not matter. By the mid‐1950s, UK policy experience convinced some observers that the level of short rates did not matter at all for aggregate demand behaviour. For example, Anthony Crosland, a Labour politician and influential writer on economics who later served in Cabinet, said in 1955: ‘Is there any sign that the Bank Rate is moderating excess demand? There is none whatsoever, and the Chancellor knows it perfectly well.’ (House of Commons Debates (hereafter HCD), April 20, 1955, col. 239.) Similar views were expressed in policy and banking circles in the 1950s and 1960s; in 1963 a leading banker observed that the ‘influence of changes in Bank Rate upon industry is often questioned’ (in Financial Times (FT), January 21, 1963) while a 1964 financial column observed, ‘interest rates have long ago and authoritatively been acknowledged to have little or no effect on the decisions of industry’ (Yorkshire Post, January 20, 1964).3 First difference of short‐term interest rate matters. Harold Wilson, the Labour Party’s economics spokesman, believed that the ‘effect of an increased Bank Rate is an impact effect, a temporary effect’ and that ‘monetary policy only seemed to work for some little time… [with] a new impact… [requiring] pushing up Bank Rate still further.’4 This perspective treated aggregate demand as insensitive to short‐term rates but private spending as constrained by the availability of bank loans: specifically, funds for consumption and firms’ working capital.5 A rise in Bank Rate would produce a temporary outflow of funds from banks until they raised their deposit rates by the amount of the Bank Rate increase.6 During this interval,7 spending was constrained by the sudden shortage of bank funds. The argument also implied that when Bank Rate fell, temporary inflows into banks relaxed budget constraints and meant higher flows of private spending. The upshot is that the first difference of the nominal interest rate enters the IS relation, as in (1). Long‐term interest rates matter. The influential Radcliffe Committee on monetary issues endorsed the idea that domestic spending was insensitive to short‐term interest rates (e.g. 1959, paras. 450, 464), adding that the asset prices crucial for aggregate demand depended on a ‘liquidity’ aggregate. Open market operations simply changed the composition of liquidity, leaving the aggregate unchanged. But the Radcliffe Committee did acknowledge a non‐zero elasticity of investment spending (and so the domestic component of yt) with respect to long‐term interest rates. Its Report stressed that the elasticity was low, and the view that long rates mattered was not universally held among those shaping the Report. But the long‐term rate’s relevance for investment was accepted by Prime Ministers Harold Macmillan and Harold Wilson, and in a statement by the Bank of England that a ‘fall in the cost of finance for industrial investment might assist expenditure by British industry’ (Bank of England Quarterly Bulletin (hereafter QB), June 1962, p. 88). Therefore, it is appropriate to include in the IS equation characterising the old doctrine an explicit (real) long‐term interest rate (rl) term:8 (5) The particular long‐term interest rates thought to enter the IS equation were the rates on long‐term corporate securities (debentures); these rates in turn were acknowledged as being closely linked to rates on long‐term government bonds (gilts).9 The Radcliffe Committee’s acknowledgement of a role for long‐term interest rates did not overturn its bottom‐line message that investment was not necessarily sensitive to monetary policy actions, in the sense of Bank Rate policy. Its Report argued that the authorities, via debt‐management operations that left Bank Rate and the monetary base unchanged, could manipulate the long‐term interest rate. This implied that the expansionary impact of deficit spending could be reduced by financing the deficit through long‐term debt, thereby pushing up long‐term interest rates and producing some downward pressure on investment spending. More generally, the Report held that the long‐term interest rate could be utilised as one instrument of demand management. This position of the Report was, of course, a rejection of standard views about term‐structure determination. To be more specific, the Radcliffe Committee felt that expectations‐theory type relations might dominate long‐rate behaviour if the authorities were ‘fatalistic’ (1959, para. 552) but that a sufficiently vigorous debt‐management policy would allow them to overwhelm this term‐structure relationship and install a regime in which the long‐term interest rate was a distinct policy instrument. Hence the Committee’s infamous conclusion (1959, para. 982), ‘debt management has become the fundamental domestic task of the central bank.’ The UK authorities took up this idea, with the Federation of British Industries Review (November 1961) observing, ‘In recent years the Bank has tried to strike more directly at these longer rates, which are the crucial ones for monetary policy. As manager of [the] national debt, it can readily influence the whole structure of interest rates if it chooses.’ Note that, because the long rate was seen as separable from open‐market policy, it is part of ‘t.i.p.’ in (1) and so (5) is a special case of (1). Inflation behaviour In the simplest representation of what Keynesian economics as of the 1940s had to say about inflation behaviour, potential GDP coincided with maximum feasible output; inflation was non‐existent until output hit potential, after which point any excess of nominal demand was instantaneously recorded in prices. Therefore, inflation was given by πt = 0 for yt < y*; and πt = Δxt for yt ≥ y*, where Δx is the log‐difference of nominal aggregate demand. This version of the Keynesian theory, however, was not that used in postwar UK policymaking, instead being modified in several respects. Recognising positive output gaps. The view that potential output represented an inviolable upper bound on GDP was discarded; in its place came recognition that excess demand could lead to output overshooting potential in the short run. For example, Crosland (1956, p. 398) rejected the idea of a ‘razor’s edge’ between full employment and inflation conditions, in favour of the notion that positive output gaps and inflation would be joint symptoms of excessive demand. In UK policy circles, the possibility of positive output gaps – together with the desirability of zero gaps – was recognised as early as 1947 in senior Cabinet member Herbert Morrison’s description of current conditions as ‘overfull employment’.10 Prevalence of cost‐push inflation. UK policymakers, as we shall see repeatedly, thought of cost‐push inflation as a problem that could occur even with negative output gaps, a concern reflected in Chancellor of the Exchequer Reginald Maudling’s statement in 1962: ‘The problem in previous years used to be more what is called a demand‐pull problem. Now it is a cost‐push problem…’ (HCD, November 5, 1962, col. 621.) Change in output gap matters. Phelps (1968, p. 679) characterised Keynes (1936) as taking the position that below full employment, inflation responded to the growth but not the level of the output gap. This characterisation, whatever its merits in capturing Keynes’ views, describes well the position of leading participants in the UK policy debate in the 1960s. For increases in the output gap, the claimed reaction of inflation had a ‘bottleneck’ interpretation: a rise in total spending, while not creating excess demand in aggregate, might create shortages in particular markets and provoke price increases there.11 Similarly, decreases in the output gap were said to provoke an inflation reaction because firms, whose longer‐term pricing decisions were not influenced by demand deficiencies, deviated from their longer‐term policy in order to prevent large increases in the amount of unsold goods. While the preceding motivation for a gap‐change term centred on price‐setting decisions, an alternative rationalisation was provided by appealing to labour market behaviour. According to this version, labour monopoly power makes wage bargaining insensitive to a given level of labour market slack; but periods of increasing unemployment threaten employees’ livelihood and so temporarily break union resistance to wage concessions. This view of the labour market was shared across the political spectrum: the Labour‐sympathetic Socialist Commentary (August 1966) referred to the ‘increasingly large rise in unemployment to obtain a given level of wage and price restraint’; the Conservative‐supporting magazine The Spectator claimed (April 14, 1967): ‘wage increases fall to a non‐inflationary level when unemployment is rising… [T]here is no evidence to show that inflation will not recur when unemployment is stabilised at that higher level.’ Similar views were expressed by policymakers (see Section 4 below). Whether the growth‐rate term is motivated by appeal to the labour market or the goods market, the policy implication is that insofar as restriction of demand exerts a negative impact on inflation, it does so only during the period when the slack is being created – i.e., during the transition from one negative output gap to a more negative value. The Radcliffe Report endorsed the above description of inflation; it rejected the notion of a Phillips‐type trade‐off relationship in favour of a cost‐push view of inflation.12 The resulting vision of inflation is captured by (2). Embodied in (2) is the implication that policymakers from the 1950s to the 1970s accepted that there was a unit weight on inflation expectations in the inflation equation. This specification reflects a basic fact: rejection of a long‐run inflation/unemployment trade‐off is common to both old and modern UK doctrines. Therefore, nothing is lost by attributing to policymakers in all periods the acceptance of a unit weight on expected inflation. Consistent with this characterisation of the debate, one of the proponents of the expectational Phillips curve, Milton Friedman, represented early Keynesian views by an inflation equation with a unit weight on expectations (Friedman and Schwartz, 1982, p. 60). Policy objectives A common element of old and modern doctrine is that a low single‐digit inflation rate is helpful, relative to a zero or negative rate, in allowing the economy to stay on its full‐employment growth path. Crosland (1956, pp. 401, 446) conjectured that ‘continuing mild inflation’ or ‘(gently) rising prices’– i.e., what today would be called ‘price stability’– might be more conducive to the economy staying on its full‐employment path than would literal price stability. Views such as these also appeared in official outlets. For example, the Department of Economic Affairs stated in February 1965: ‘No Western country since the war has yet succeeded for long in combining stable prices with a high level of growth…’ (Progress Report No. 2, quoted in HCD, May 11, 1965, col. 272). In 1970, Chancellor Roy Jenkins said that ‘absolute stability of prices’, i.e. a constant price level, was not compatible with growth along the full‐employment path (HCD, April 14, 1970, col. 1224) but that his own definition of price stability, 2% to 2.5%, was completely consistent with continuous full employment (The Times, January 12, 1970; FT, January 12, 1970).13 Likewise, many expositions of inflation targeting have defended inflation rates of 2% compared to a zero target, and have defined price stability so that it corresponds to this low inflation rate, e.g., King (1997a, p. 93). Underlying these discussions is the view that low inflation can facilitate relative price adjustment by reducing the required extent of absolute declines in prices. In the 1950s, this perspective may have reflected the conviction that there existed permanent obstacles to absolute declines in prices and wages. The more modern position (King, 1999) does not presume permanent nominal rigidities in either direction, but does assume temporary nominal rigidities in both directions, of which those in the downward direction are likely to be more formidable. The position in favour of low inflation corresponds to postulate (OD7)/(MD7) given in Section 1. Its lack of equivalence with a belief in a permanent Phillips curve trade‐off should be stressed. Adherents to article (OD7)/(MD7) believed that a low positive inflation rate might help the economy replicate the conditions of a no‐nominal‐rigidity economy; believers in a Phillips curve trade‐off instead held that appreciable inflation rates would deliver a permanently higher level of output than the output level associated with no nominal rigidity. Let me now consider how several issues can be fit into this characterisation of doctrine: the exchange rate; money and credit; expectations; and the policy mix. 2.2. The Exchange Rate UK policymakers have always understood that the nominal exchange rate is sensitive to the domestic short‐term nominal interest rate;14 that nominal exchange rate movements are reflected in short‐run real exchange rate movements; and that the real exchange rate matters for net export demand. So it has always been accepted (even by the Radcliffe Committee) that short rates matter for real aggregate demand at least through an open‐economy channel. It might therefore seem desirable to augment (1) with an explicit log real exchange rate (qt) term, and to note the dependence of qt on the short rate Rt. But this would make no difference to the analysis because for the 1949–67 period the nominal exchange rate was constant and so the part of qt that Rt immediately influences15 was constant. It can therefore be suppressed. From 1967, exchange rates did fluctuate; but it is from precisely this period onward that the authorities finally acknowledged influences of the short rate on domestic spending; i.e., from the late 1960s policymakers moved from (OD1) to (MD1) even though they continued to subscribe to the remainder of the old doctrine. If (3) is thought of as describing policymakers’ views from 1968 onward, and the interest elasticities in this equation are understood as inclusive of the open‐economy channel, then no explicit exchange rate term is needed. Import price shocks were one source of cost‐push pressure invoked by pre‐1979 policymakers. Pre‐1979 policymakers thought of interest‐rate increases as a way of reducing the size of this form of the ut shocks, i.e., by engineering an exchange rate appreciation. The ut series was therefore not regarded as completely endogenous with respect to monetary policy. Most ut variation was, however, regarded as exogenous, so I make that assumption in my presentation of (2), while taking note below of pre‐1979 monetary policy responses to perceived import price shocks. 2.3. Money and Credit The absence of money from the IS equation is not an area of dispute. For example, the Radcliffe Committee (1959, para. 397) said that ‘the structure of interest rates’, not money, was what mattered for aggregate demand; while in the monetarist literature monetary policy affects aggregate demand via asset prices (Friedman and Schwartz, 1982, p. 58). Therefore, the money stock is absent from my representation of both the old and modern doctrine’s IS equations. What is not common ground is the susceptibility of yields beyond short rates to influence by open market operations. In particular, in modern doctrine a definite term‐structure relation connects long rates to short rates, so that rlt cannot be included in the ‘t.i.p.’ term in the IS equation. With respect to credit, most policymakers before 1979 did believe that direct controls of financial institutions – imposed on their aggregate assets and/or liabilities, or on specific types of lending – contribute to aggregate demand control for a given setting of short‐term interest rates. Policymakers’ vision of these controls’ effects can be represented by including intercept dummy variables in the IS equation that take non‐zero values when the controls are in effect. Their inclusion would not change (1) because the control dummies would be part of the ‘t.i.p.’ term; recall that the ‘t.i.p.’ term includes any measures independent of the short‐rate/money base path. 2.4. Specifying Expectations A leading account of past policy mistakes in the US, Sargent (1999), focuses on policymakers’ assumption of adaptive instead of rational expectations in specifying private sector behaviour. For studying UK policy developments, this approach seems inappropriate. My contention is that UK policymakers had a fundamentally incorrect specification of the economy’s structural equations; this error is not the same as having a correct model but misspecified expectations. In fact, it seems appropriate, conditional on their incorrect model, to attribute rational expectations from that model to UK policymakers. It would not be appropriate, by contrast, to assume instead that pre‐1979 policymakers took πe in (2) as well represented by lagged inflation, πt−1. The reason is that substitution of πt−1 for πe implies (after integrating (2)) that inflation is a positive function of the output gap at all levels of the gap – something UK policymakers emphatically denied. In being specific about πe, one should also incorporate policymakers’ view that expectations had some inertia.16 Policymakers further recognised that inflation was forward‐looking. For example, Economic Affairs Secretary George Brown said in 1965 that it was becoming standard for firms ‘to pass on [cost] increases with just that little more added… because they expect more increases to happen in the future’ (HCD, May 11, 1965, col. 297). These considerations make it desirable to follow Rotemberg and Woodford (1997) in assuming that the expectation in the Phillips curve refers to πt+1, but that the expectation is based on lagged information. A concrete variant of (2) that uses a special case of Rotemberg and Woodford’s timing assumptions is: (6) Similarly, the modern doctrine’s Phillips curve can be written as: (7) Relative to Rotemberg and Woodford (1997, p. 316) and, as in Clarida et al, (1999), this specification includes a cost‐push term (realised in period t) and a long‐run verticality restriction on the Phillips curve (i.e., a unit weight on inflation expectations). In addition to the differences in specification across (6) and (7), old and modern doctrine differ in their specification of ut. The old doctrine’s cost‐push view of inflation allows ut to have arbitrary persistence (since, for much of the time, ut is inflation’s basic forcing process). The importance assigned to monetary accommodation in modern doctrine instead implies that ut cannot be very persistent.17 2.5. Monetary Accommodation and the Policy Mix I have characterised old and modern doctrine in terms of policy objectives and propositions about private economic behaviour – not in terms of implications for the appropriate mix of policy instruments. It is worth confronting some specific examples of policy maker views on the appropriate policy mix. I consider these views because they might seem to contradict my characterisation of old doctrine but, in fact, do not; some also seem to contradict one another but, in fact, are compatible. And some statements sound modern, but on further inspection contradict modern doctrine. These statements concern the interrelated issues of the role of incomes policy in the policy mix and the role of monetary accommodation in producing inflation. I report the statements, then consider them together. First, some statements by 1960s policymakers rejected the notion that incomes policy was a substitute for aggregate demand actions: One central part of any policy of restraining price increases must be the control of demand… [But] we cannot have price stability without an incomes policy… (Douglas Jay, President of the Board of Trade, HCD, May 11, 1965, cols. 394, 398.)
 A firm and effective incomes policy is not a substitute for fiscal action as an instrument of demand management… Such a policy is, however, of crucial importance in relation to costs and prices. (Chancellor Jenkins, HCD, March 19, 1968, col. 263.) Second, other pre‐1979 policy maker statements seemed to suggest that demand measures and incomes policy were, in fact, substitutes. For example: [I]f all of us are determined to extract from the market the maximum possible for ourselves, the result must either be inflation or Government policies designed to restrict demand… (Chancellor Maudling, HCD, April 14, 1964, col. 263.)
[I]f wages rise beyond the limits… the Government will be compelled to take offsetting steps to curtail demand. (Chancellor Denis Healey, HCD, November 12, 1974, col. 249.) Third, the 1970s witnessed some modern‐sounding statements on the role of monetary accommodation in the inflation process: The monetary policy appropriate to our present circumstances is one that does not passively provide the amount of money that is needed to underwrite the going rate of inflation, but something less. (Chancellor Tony Barber, in FT, January 20, 1971.)
The present Government have made it clear that they are not prepared to finance inflation by printing money. (Chancellor Healey, HCD, January 25, 1979, col. 754.) Appearances to the contrary, none of these statements contradicts one another and all are consistent with old doctrine. Consider the last pair of statements: those on accommodation. These statements do not in fact fit in with the modern view of inflation. For example, when Healey spoke in 1979 about not financing inflation, his threat was to tighten aggregate demand to offset perceived inflationary pressure from wage demands. But at the time he spoke, Healey believed that the level of output was over 10% below potential; according to modern doctrine, that amount of pre‐existing slack should be more than adequate to rule out any danger of financing inflation. The old doctrine was internally consistent in characterising incomes policy as both a complement to and substitute for demand restriction; for according to that doctrine (and as implied by (2)), when output is above potential, demand restriction is necessary but incomes policy is still needed to bear down on the ut shocks; while, starting from output at or below potential, restriction of demand can be a short‐term substitute for incomes policy, in the sense that demand restriction produces downward pressure on inflation that ends once the gap settles at a new, more negative level. That is clear both from the way the output gap enters only as a first difference in (2) when yt < yt*, and from the above statements that implied incomes policy was preferable to the demand‐restriction alternative. So according to old doctrine, while it is important that demand policy ensures that output not exceed potential, persistent inflationary forces can occur even with a zero output gap and can only be temporarily offset by moving demand below potential. Incomes policy, by contrast, is – per old doctrine – both a lasting solution to this type of inflation, and a measure that avoids the need to restrict demand. According to modern doctrine, aggregate demand measures – specifically, monetary policy actions – are neither a substitute for nor a complement to incomes policies; instead, when it comes to reducing inflation, there are no alternatives and no supplements to monetary policy actions. And modern doctrine states that a monetary policy that preserves a zero output gap will keep inflation low. The difference in conclusions between old and modern doctrine may be summed up as follows: in old doctrine, demand management was seen as only a necessary element for inflation control, whereas in modern doctrine it is both necessary and sufficient, and the sole demand‐management device required is monetary policy. 3. General Approach In this Section I describe the approach used in this article to look at successive leading policymakers since 1964. I also consider the representativeness of the material used, and the relation of this article to previous literature. 3.1. Methodology and Sources I draw out successive policymakers’ views of the economy largely from their public statements. This does not constitute a great difference from the approach of Romer and Romer (2004) who, though drawing on FOMC Minutes, also made considerable use of Federal Reserve Chairmen’s speeches and writings (before and after entering office). But the lack of central bank independence in the United Kingdom before 1997 means that my focus must be on leading members of the government – so I break up the analysis according to successive Prime Ministers. A stand must be taken on how to obtain public statements by Prime Ministers other than those given in Parliament. The Thatcher Foundation has placed on its website a digital archive of public statements by Margaret Thatcher that is intended to be exhaustive for the 1945–90 period. But the prospect of similar electronic databases becoming available for her predecessors, Harold Wilson, Edward Heath, and James Callaghan, is far off.18 I therefore rely extensively on newspaper material to recover statements by these policymakers and to retrieve statements by Thatcher not available in the online digital archive. The most central components of the UK financial press are perhaps the Financial Times, The Times, and The Economist. Times and Economist back issues are digitally searchable over the twentieth century, while I have obtained older material from the Financial Times by a microfilm search for the entire run of the 1970s and for most of the 1960s. It would be a mistake, however, to rely on these publications to the exclusion of other press material – including mass‐circulation London newspapers, non‐London UK newspapers, and newspapers of Ireland and Hong Kong (both of which had close financial ties to the United Kingdom over much of the postwar period). Several factors motivate my use of a wide set of material. First, an examination of reports across newspapers of a public figure’s speeches, television interviews, or press conferences helps recover more of what was said than simply consulting a single newspaper’s report. Second, leaders of political parties often gave exclusive interviews to particular outlets (for example, the Callaghan interview with the Glasgow Herald quoted below), or contributed signed guest articles to a particular newspaper (for example, the Heath and Wilson ‘op‐ed’ articles I cite). Third, policymakers’ speeches at non‐London events may be more extensively reported in the regional press than in the London press (examples include the Wilson appearance in Cambridge and the Heath appearance in Scotland, both quoted in Section 4). Leading financial newspapers therefore fail as an adequate source for ascertaining policymakers’ views. Indeed, previous studies have suffered from factual error by concentrating on a narrowly defined financial press. For example, Parsons (1989, p. 172) stated that ‘stagflation’ was ‘a term apparently introduced to Britain by The Economist’. This is doubly erroneous: the word ‘stagflation’ is of UK origin, having been coined by politician Iain Macleod and, when it comes to publications that used the term as a regular word (i.e., other than in simply quoting Macleod), the Daily Mail predated The Economist.19 3.2. Representativeness I rely heavily on material from the public record. Is my analysis therefore vulnerable to the criticism that I am not finding out what policymakers believed; only what they wanted the public to think they believed? The answer is no. The UK authorities had no plausible motive to misinform. On certain tactical issues, like the timing of interest‐rate changes, there may be an incentive to withhold information on any given day. But the doctrine that concerns me is strategic– views on how the economy works on a quarter‐to‐quarter basis. Secrecy about economic strategy is logistically impossible in the UK parliamentary system; so also is any attempt to communicate a strategy to the public different from that decided upon privately. It is infeasible to retrieve every public statement on economic matters by policymakers; even if it were feasible, no single study could report all of them. Therefore, the quotations I use are necessarily only a sample. There are reasons to be confident that this sample conveys accurately and representatively the views of the policymakers I study. Most important is the fact that, while space considerations prevent me from using all the quotations I have recovered, if I lengthened this article to include all statements assembled, my account of doctrine would stand up.20 One reflection of the voluminous material available in support of my characterisation of doctrine is that the quotations in Sections 3, 5 and 7 do not overlap with those used in my previous work on UK policymaking but are consistent with them. In addition, I provide specific evidence of consistency between privately articulated and public statements on doctrine. I further show the consistency of my characterisation of official doctrine with policy decisions– such as the decision to cut interest rates in 1975 in the face of rising inflation. These general observations about the representativeness of the statements quoted are confirmed by evidence that I present in Section 5. 3.3. Related Literature Two studies overlapping somewhat in sample period with mine, but not in content, aim or source material, are Cobham (2002) and Pepper and Oliver (2001). Neither study gives the comprehensive picture of policymakers’ views provided here. Cobham focuses on policymakers’ analysis of the current state of the economy, not on their underlying views of the transmission mechanism; in addition, his coverage effectively does not begin until 1979,21 while his focus on the Bankof England Quarterly Bulletin as a source comes at the expense of other, highly relevant material. Pepper and Oliver (2001) study certain policymakers from the 1980s, seeking to ascertain whether they had ‘monetarist’ views. Their study concentrates on the issue of whether policymakers paid attention to the money supply. Such an approach does not distinguish between the belief that monetary policy is important for aggregate demand from the further belief that monetary policy is central for inflation control. This is a crucial distinction that separates 1970s policymakers from policymakers after the 1970s, as I show. Bernanke et al. (1999) have a chapter on 1990s UK monetary policy and its objectives but the views of the key policy figures upon whom I focus my discussion – John Major and Mervyn King – are not considered; indeed, the text of their chapter does not even mention either Major or King.22 4. The Era of the Old Doctrine This Section shows that the three 1964–79 Prime Ministers adhered to the old doctrine. 4.1. Harold Wilson (1964–70 Administration) Aggregate demand behaviour Wilson’s views in the early 1960s are in line with the old doctrine’s postulates on aggregate demand behaviour. Output demand was not elastic with respect to the level of the short rate; and though the long‐term rate mattered for investment, it could be treated as a separate policy instrument. This attitude came through in a 1964 speech in which Wilson indicated his willingness to use short‐term interest rates to staunch any flow of short‐term capital so as to safeguard our sterling area reserves. But in saying this, I make one essential condition... [W]e should not… force on the nation a structure of high long‐term rates with all that means for investment. (Wilson speech, Swansea, January 25, 1964; quoted in Sunday Times (ST), January 26, 1964.) Shortly after Wilson took office, his Government indeed increased Bank Rate as a sterling‐protection measure and characterised the effects of the increase in a manner consistent with Wilson’s views. For example, Economic Affairs Secretary George Brown said that the Bank Rate decision would have a ‘mainly short‐term’ effect and ‘should not disturb productive investment’ (in The Guardian (TG), December 10, 1964). In 1965–70 Wilson used direct credit controls frequently. This reflected his lack of faith in Bank Rate for demand management and his related belief that direct controls could deliver ‘any desired degree of tightness’ for a given Bank Rate (MG, October 25, 1957). Inflation behaviour Wilson outlined his view of inflation in 1957. While acknowledging that excess demand could produce inflation, he argued that the postwar period had witnessed ‘times, including the present, when the classical definition of demand inflation – too much money chasing too few goods – has not applied… [T]he cost inflation (or cost‐push inflation, to use the American phrase) has persisted.’ (MG, October 23, 1957). Cost‐push inflation could prevail under any demand conditions so ‘restriction of production is not the answer’ (MG, October 24, 1957). In claiming that cost‐push forces could persistently affect inflation irrespective of demand (i.e., without monetary accommodation), Wilson associated himself with the old doctrine. In harmony with Wilson’s views, Wilson’s economics ministers saw cost‐push forces as implying inflation irrespective of the output gap level. For example, when the output gap was believed to be positive in 1966, Chancellor Callaghan said that just as important as demand restriction was ‘action… to stop ‘cost‐push’’ (HCD, July 26, 1966, col. 1474); while in 1967, when the output gap was believed to be negative, Economic Affairs Secretary Shore said that the Government’s incomes policy was designed to fight cost‐push inflation (HCD, November 1, 1967, col. 302). Such views lend weight to (2) as a description of 1960s doctrine. Also consistent with (2), Wilson’s colleagues perceived a speed‐limit term in inflation dynamics, with Shore claiming that growth far above a 3% rate would result in ‘rapidly mounting costs’ (HCD, March 21, 1968, col. 617). Wilson saw the solution for inflation in a negotiated wage policy alongside price controls and a full‐employment demand policy, or as he put it, an ‘incomes policy based on rising production’ (ST, January 26, 1964). After attempts at a negotiated policy in 1964‐6, Wilson imposed a wage/price freeze in 1966–7, with further wage controls in 1967–9. Policy objectives In 1965, Wilson reaffirmed his goal of a zero output gap: We reject the doctrine that this country can solve its problems only by holding production a long way below our capacity to produce… But we shall not allow the total volume of internal demands to exceed our national resources, including an adequate provision within our national production for the rising needs of our export trade. (Wilson speech, July 24, 1965, quoted in Sunday Telegraph, July 25, 1965.) This statement makes clear that external balance was not regarded as a macroeconomic objective to be traded off against a zero output gap or price stability; rather, external balance could be achieved by appropriate resource allocation without compromising the zero‐gap goal.23 Of course, Wilson’s wish was to achieve external balance without a devaluation (or, after 1967, without a second devaluation); but this does not imply that the exchange rate was an ultimate objective; on the contrary, cost‐push theories made avoiding devaluation desirable on price stability grounds. The Government’s price stability objective was also reflected in Wilson’s claim: ‘We are the first Government that has really tackled the problem of rising prices and rising incomes.’ (Sunday Citizen, March 27, 1966.) The Government’s price stability concept corresponded to 2–3% inflation, reflected in Cabinet member Richard Crossman’s statement, when the annualised inflation rate fell into this range in late 1966, that the wage‐price freeze had cured inflation (The Times, December 19, 1966). As we have seen, a similar range was associated with price stability by Chancellor Jenkins in 1970. Later views (1968–70). In the late 1960s the Wilson Government revised its views somewhat, conceding a greater effect of monetary policy on aggregate demand. This change made itself felt in an acknowledgement that investment was sensitive to short‐term interest rates. Reflecting this, Chancellor Jenkins stated that high interest rates were ‘necessary for both external and internal reasons’ (HCD, February 17, 1970, col. 190) – a contrast to Wilson’s 1964 concentration on Bank Rate’s role in preserving the exchange rate. In other respects, however, official views on monetary policy on aggregate demand were unreformed: credit control was seen as very important and long‐term interest rates were regarded as susceptible to direct control. There was also no change in official views about monetary policy and inflation. Wilson continued to subscribe to cost‐push views and, in 1968, described monetary policy (and other demand measures) as necessary but not sufficient tools against inflation: Financial policy, budgetary policy and incomes policy are needed to prevent demand getting out of hand. Incomes policy is equally necessary to prevent costs getting out of hand… (Wilson remarks, March 6, 1968; quoted in The Scotsman, March 7, 1968.) Chancellor Jenkins reinforced this judgment almost two years later.24 Therefore, monetary policy was never seen as a necessary and sufficient tool against inflation. 4.2. Edward Heath (1970–74 Administration) Aggregate demand behaviour Heath deserves some credit for early acknowledgement that short‐term interest rates matter for aggregate demand behaviour. As Leader of the Opposition, Heath challenged the view that domestic spending was insensitive to Bank Rate (HCD, August 2, 1965, col. 1078). In addition, once in office, Heath adopted a more modern perspective on the long‐term securities market than either the 1964–70 or 1974‐9 Governments. While still claiming, contrary to modern doctrine, that official control of long rates (for a given Bank Rate) was feasible, the Government withdrew from attempting to use this claimed power; from 1971 onward, there was greater stress on market determination of long rates. Inflation behaviour Though well known for major shifts in economic policy, Heath was consistent in his basic view of the inflation process over his period in office. The shifts that he did instigate are reflected in a statement of the Government’s philosophy that Heath wrote in 1971: ‘We have already made some changes reducing the involvement of government in the affairs of industry… This process will continue.’ (Liverpool Daily Post, January 20, 1971). In fact, the process ended not long afterward and the Government’s more interventionist attitudes were reflected in the abandonment of the guidepost‐based and negotiated incomes policies of 1970‐2 in favour of compulsory wage and price controls in 1972–4. But all these changes were within the old doctrine. Everything Heath did regarding inflation can be explained using the old doctrine’s inflation equation (2). Heath said in June 1970 that cost‐push inflation was ‘not susceptible to the orthodox policies of demand management’ (FT, June 17, 1970). Thus, Heath concluded in December 1970, ‘control of the money supply is one element – an important element, but only one element – in the number of weapons that one has to use against inflation.’ (HCD, December 8, 1970, col. 240.) This view of inflation came through in Heath’s failure to distinguish between sources of aggregate and individual price movement. For example, after refusing coal price increases, Heath said it ‘cannot be denied’ that this implied lower inflation than otherwise (HCD, November 19, 1970, col. 1425); he attributed the continuing high inflation to wage‐push (FT, October 12, 1971). Throughout the 1970–3 period the Government perceived the output gap as negative, so the Dt = 0 setting of (2), i.e. an inflation expression with no gap‐level term, captures its views. In keeping with (2), Heath and his colleagues believed that the first difference of the output gap did matter for inflation; for example, Heath said that growth at ‘too fast a pace’ could produce inflation (in FT, June 17, 1970). In 1970–1 his Government attempted to make use of the gap‐change term: its aim of 3% economic growth was intended to produce Δ(yt − yt*) = 0. The underlying scheme was: demand policies would make no addition to inflation; at the same time, official attempts to influence wage and price setting would reduce inflation directly by withdrawing cost‐push pressures (i.e., reducing ut); and once cost‐push inflation was broken, Δ(yt − yt*) could then be allowed to become positive (see e.g. Daily Telegraph, April 22, 1971). Contrary to the planned policy, Δ(yt − yt*) went negative in 1970–1. When, nevertheless, inflation remained strong, the Treasury lost confidence that Δ(yt − yt*) mattered for inflation. Heath was initially inclined to believe that Δ(yt − yt*) > 0 might still add to inflation even if Δ(yt − yt*) < 0 did not subtract from inflation (FT, June 19, 1971). But this speed‐limit behaviour of inflation could, it was believed, be overcome by broad‐based incomes policies; so once the Government secured voluntary price restraint from employers in July 1971, it felt that it could expand at rates fast enough to close the output gap without inflation risks.25 These expansionary policies continued when, in 1972, the compulsory wage and price controls succeeded the voluntary measures. Romer and Romer (2004, p. 155) show that G. William Miller became Federal Reserve Chairman with his own idiosyncratic variants of cost‐push ideas. In Heath’s case, two aspects of his cost‐push analysis are especially notable: an idiosyncratic ‘proximity‐push’ view of wage inflation; and his unit‐cost justification for expanding demand. (i) Proximity‐push inflation. Heath emphasised the ease of travel in the UK: We must really try to demolish this fallacy of distance. People in the south think of Scotland as being a long way away. Nothing is further from the truth. It is easier to get there from London than it is to get to most parts of England. (Quoted in Glasgow Herald (GH), September 7, 1964.) He later observed: Part of the problem is that we are a very small country. You cannot do something in one part of it without the rest of the country knowing, and very soon they all want the same thing. It is a major difficulty in our wage negotiations. (Heath remarks, press conference, Cleveland, Ohio, August 16, 1978, quoted in The Plain Dealer, August 17, 1978.) Thus, in Heath’s view, the UK’s small size and ease of travel made it particularly vulnerable to wage‐push inflation. 
(ii) Unit costs and demand expansion. Heath always emphasised the effect demand restriction had in raising unit costs and so worsening inflationary pressure (e.g. FT, May 23, 1970).26 Belief in speed limits counterbalanced any temptation to use expansionary policies as an anti‐inflation measure. Once speed‐limit ideas were dropped in 1971, the Government embraced the idea that rapid closure of the output gap would reduce inflation. A 1972 news report confirmed the new policy: The pet economic theory that a major cause of inflation was too much money chasing too few goods was shot down by Mr. Barber at the Conservative Party Conference in Blackpool... Mr. Barber believes that by expanding the money supply he will increase the production of goods… (Irish Times, October 13, 1972.) Later views on aggregate demand and inflation. Interest rates were increased substantially in 1973 but this did not reflect acceptance of modern views about using interest rates against inflation. Instead it reflected an import‐price‐push diagnosis. When the policy rate was raised in July 1973, Chancellor Barber gave the aim as ‘sustained expansion’ of output, with the interest‐rate increase designed to prevent inflationary pressure by strengthening the exchange rate (TG, July 28, 1973).27 Indeed, from early 1973 the Government relied on direct financial controls as a means of withdrawing stimulus to demand. Without the exchange rate pressure, it is likely that Heath would have opted even more for direct controls rather than rate increases and focused on incomes policy in fighting inflation. Indeed, Heath later said he did not agree that ‘high interest rates are the right way of dealing with inflation or the money supply’ (in South China Morning Post, July 3, 1981). Policy objectives After leaving office, Heath said, ‘There are some people who say I was intent on growth at any price. This is untrue.’ (TG, September 16, 1976.) Heath’s protest is supported by the fact that there is no evidence that he consciously targeted a positive output gap. In 1966 he had said: ‘We must manage a full employment economy and not an overfull employment economy, and we must achieve cost stability and price stability.’ (HCD, July 26, 1966, col. 1462.) The expansionary measures under Heath in 1971–3 were seen as aligning aggregate demand with existing supply potential, though this potential was overestimated.28 Heath monitored the Government’s communication of its policies to make clear its zero‐gap goal. Declassified materials show that Heath, in response to a passage of the draft March 1972 Budget speech that said the UK economy ‘should move into top gear’, wrote in the margin: ‘? [This] implies overheating.’ (Cabinet Documents, March 12, 1972). The passage was consequently not included in the speech. Another sense in which Heath did not pursue expansion at all costs is that he still regarded himself as giving priority to inflation control. A 1969 speech by Heath emphasised price stability and rejected any trade‐off concept: The so‐called choice between stable prices and full employment is a mirage. There is no such choice. It is precisely the fall in the value of money today that presents the greatest threat to full employment in the years to come. The most important task facing the next Conservative Government will be to... restore the nation’s faith and trust in its currency. 
 (Heath speech, March 22, 1969, in Sunday Times (ST), March 23, 1969.) On a related note, Heath later spoke of ‘the two interlocking domestic problems of inflation and unemployment’, adding, ‘There is no doubt that the rapid rate of inflation which we inherited – which was faster than we had recognised – has caused the great increase in unemployment.’ (Evening Standard, June 1, 1972.) Increased rates of inflation over 1970–3 did not seem to produce an increase in what Heath believed was an acceptable rate: the incomes policy phase introduced in October 1973, for example, was intended to arrest domestic cost‐push forces in a manner consistent with inflation being brought down into the 4‐5% range by late 1974, from which point it could fall further (Daily Telegraph, October 9, 1973). 4.3. Harold Wilson (1974–6 administration) Aggregate demand behaviour Wilson re‐entered office accepting, as his Government had in the late 1960s, that short rates mattered for aggregate spending via an investment channel (see e.g. HCD, November 19, 1973, col. 962). But Wilson still advocated credit controls (HCD, November 19, 1973, cols. 963‐4), and direct financial controls were deployed in his new administration. Inflation behaviour In 1972 Wilson wrote that he still believed in ‘urging the distinction between cost‐push and demand inflation’ (ST, August 6, 1972). Wilson favoured a national wage agreement and compulsory price controls. In late 1973, Wilson said that the ‘ready weapon’ of indirect tax cuts and subsidies could ‘mitigate the damage threatened to Britain’ by the commodity price explosion (Cambridge Evening News, October 20, 1973). Another aspect of inflation analysis endorsed by the new Wilson Government was the speed‐limit dimension of inflation. As noted above, during 1971‐4 the Treasury believed that δ = 0 in (2). But examination of the 1973 experience convinced Denis Healey (Chancellor of the Exchequer 1974‐9) that δ > 0. Healey argued in a March 1976 paper to the National Economic Development Council that fast growth generated bottlenecks and inflationary pressure even when the economy‐wide output gap remained negative (TG, March 1, 1976).29 Wilson affirmed the speed‐limit view when in February 1976 he said that while there were ‘large parts of our productive capacity unused’ (TG, February 3, 1976), for the Government to ‘reflate now on a massive scale’ would be to generate ‘yet another inflationary boom’ (FT, February 3, 1976). As 1973 was believed (accurately) to have witnessed both excess demand levels and above‐potential growth, demand factors were granted some role in producing the inflation the Government inherited in 1974. The Government nevertheless saw inflation as primarily cost‐push, particularly since excess demand in the economy was believed to be gone by early 1974. The Government’s 1974 actions featured the nonmonetary measures Wilson had foreshadowed in 1972‐3, including an agreement (the Social Contract) with unions on wage growth, and subsidies to key prices. When inflation nevertheless rose to over 25% in 1975, the Government again cited cost‐push forces, specifically wage‐push – Chancellor Healey having said that wage growth had become the ‘most important single factor in determining the rate of inflation’ (HCD, November 12, 1974, col. 249). The incomes policy negotiated in July 1975 was subsequently cited by Wilson as having ‘cut back the rate of inflation very considerably’. (BBC1, September 1, 1982, p. 9.) In announcing the 1975 anti‐inflation programme, Wilson said that it was consistent with employment improving (HCD, July 11, 1975, col. 905) and during 1975 the Government greatly cut short‐term interest rates. The monetary easing over this period was consciously carried out in the face of high inflation, and reflected the continuing belief that inflation control was appropriately handled by incomes policy. For the US in the 1970s, Orphanides (2004) has argued that policymakers thought they were actually raising the interest rate vigorously in response to inflation. The same is clearly not true of the UK; as a report in the Wall Street Journal (April 24, 1975) noted, Britain’s Labour Party Government is trying to defy a law of capitalist economics by bringing interest rates down at a time inflation is accelerating. This interest‐rate/inflation combination is brought out in a plot of the UK Treasury bill rate and four‐quarter RPI inflation (Figure 1). The circled episodes are the periods in the first half of the 1970s when interest rates were cut in the face of high inflation (1971–2) or rising inflation (1975). Fig. 1. Open in new tabDownload slide UK Short‐term Interest Rate and Inflation Fig. 1. Open in new tabDownload slide UK Short‐term Interest Rate and Inflation Policy objectives The Government rejected Phillips curve analysis: Wilson said in 1972 that the idea that inflation and unemployment were inversely related was ‘a fantasy’ (ST, August 6, 1972), adding in 1975, ‘We cannot have a little of one and less of the other. The more inflation we have, the more unemployment we have.’ (Daily Mirror, July 8, 1975.) In 1974, Chancellor Healey gave price stability as a policy objective (HCD, November 12, 1974, col. 280). Wilson spelled out this objective as to ‘bring down the level of inflation in this country to a level comparable with that of our major competitors’, ‘keep it there’, then ‘eliminate inflation’ (FT, February 3, 1976). The Government also had a zero‐output‐gap objective: Healey said that demand management should ‘avoid the twin dangers of mass unemployment and overheating’ (HCD, November 12, 1974, col. 256). 4.4. James Callaghan (1976–9 Administration) Aggregate demand behaviour When Callaghan was Chancellor of the Exchequer, a financial column noted, ‘Mr. Callaghan, like most of his colleagues, has little regard for Bank Rate and other monetary weapons as effective economic regulators…’ (GH, February 12, 1965). But as Prime Minister, Callaghan accepted that short rates mattered for aggregate demand (see e.g. HCD, November 1, 1978, col. 52). The interest‐rate policy his Government followed was, nevertheless, not enlightened by modern standards. After big increases to contain sterling depreciation in 1976, nominal rates were greatly cut (and real rates made steeply negative) in 1977. Targets for broad money growth were in effect throughout Callaghan’s tenure but substantial use was made of banking controls to hit the targets. The Government’s outlook on long‐term bonds was also not enlightened. Officials leapt from the observation that commercial banks bought short‐term Treasury debt, to the unwarranted conclusion that deposit expansion would be reduced if fewer Treasury bills were available. Effort was therefore wasted on securing long‐term debt financing. Attempts to treat long‐term interest rates as a policy instrument resumed. In 1977 the authorities were said to be signalling the ‘return of market management by the Bank and Treasury’ for long‐term securities (Yorkshire Post, May 31, 1977). Inflation behaviour Two events in 1976 are often claimed to have marked a change in the Callaghan Government’s economic doctrine (see e.g. Smith, 1987, Ch. 5). First, monetary targeting began. But to be a genuine breakthrough, monetary targeting must entail giving up non‐monetary approaches to analysing inflation. In the UK case it did not: in 1977 Chancellor Healey contrasted the view that wage pressures require monetary accommodation to produce inflation, with what ‘[e]veryone else believes’, that wage growth automatically produces inflation (Yorkshire Post, February 19, 1977). Likewise, in 1979 Healey told Parliament, ‘I do not think that fiscal and monetary policies alone can control inflation even at the cost of heavy unemployment.’ (HCD, January 25, 1979, col. 755). Healey added in 1980 that ‘we couldn’t conceivably have got inflation down [in 1975‐8]… unless we’d had pay policy as well as a monetary policy.’ (BBC2, March 22, 1980, p. 19.) Also in 1980, Callaghan criticised ‘reliance on monetary policy as the single or sole weapon’ against inflation (HCD, February 28, 1980, col. 1588). The second 1976 event is a speech Callaghan made with the message that inflation and unemployment moved up together over longer periods (Evening Standard, September 28, 1976). This might be a turning point if pre‐1976 governments had believed in an inverse relationship but they did not – see the Wilson and Heath statements quoted above. Policy objectives The Government wanted a zero output gap, reflected in a 1976 programme for achieving full employment (TG, April 8, 1976). This was specified as implying 3% unemployment, acknowledging a structural increase in the unemployment rate since the 1960s. As for inflation, Callaghan labelled beating it ‘Essential Policy Number 1’ (TG, July 1, 1978). He stated ‘the Government’s target of reducing inflation to 5% or less within the next three years’ (GH, May 2, 1979), and added that the price stability objective was ‘common ground’ with Margaret Thatcher (HCD, February 28, 1980, col. 1588). 5. Consistency with Other Evidence My determination of official doctrine from statements by the Prime Minister and senior colleagues is justified by the fact that they were in control of monetary policy – and of non‐monetary measures against inflation – throughout the old‐doctrine period. I now consider objections that might be raised about this approach. First, it might be argued that political constraints so prevented candour that official statements ran counter to the views held internally. Second, it could be claimed that politicians’ descriptions of the economy reflected their inadequate understanding of economic policy. According to this argument, the old doctrine was not accepted at the ‘technocratic’ level of senior Treasury and Bank of England staff and, therefore, was not used as the basis for policy decisions. I refute these objections in this Section. I present evidence that (OD3) and (OD5) – the most non‐standard aspects of the old doctrine – were articulated by policymakers and officials alike; and I provide further support for my hypothesis of constant policy maker objectives, a hypothesis embodied in articles (OD7) and (OD8). One of the most non‐standard aspects of the old doctrine, item (OD3), implied that long‐term rates can be manipulated as a policy instrument independent of the short rate. As we have seen, this was a Radcliffe Committee position taken up by 1960s policymakers. The clearest confirmation that it was also believed in private is in Prime Minister Harold Macmillan’s diaries. Macmillan’s October 10, 1962, entry states, ‘the Chancellor seems against lowering the Bank Rate – at any rate for the present. But he will concentrate on trying to reduce the long‐term rate of interest.’30 Similar positions were endorsed at the technocratic level in Bank of England publications over the 1960s. For example, in 1967 the Bank stated that by varying ‘official sales of stock… the authorities encouraged a moderate fall in yields’ (Bank of England Report, July 1967) and in 1970 it described recent long‐rate movements as having been ‘allowed’ by the Bank (QB, June 1970, pp. 132‐3). Further into the 1970s, the Bank’s statement that it was issuing a new long‐term bond ‘to retain a means of influencing long‐term interest rates directly’ (QB, March 1972, p. 15) conclusively shows that the authorities perceived long‐rate control as within their capacity, and a power that they had exercised in recent years. Turning to article (OD5), let us consider the following testimony from July 1974 by the Chief Economic Adviser to the Government, Sir Kenneth Berrill:31 Berrill: I would say that we do not believe the Phillips curve over quite a large band, but starting at the top end, when you reduce unemployment, you can begin to see shortages of skilled labour, bottlenecks, and so on developing which affect the balance of payments and also earnings and prices. Then there is a large flat band. What happens at the heavy levels of unemployment we do not know because we have not had that since the 1930s. 
Question: Between the summer of 1971 and the summer of 1972 we did see a distinct downturn in the rate of inflation. You would not attribute that to the rise in unemployment which had preceded it? 
Berrill: No. Import prices were in our favour during that period. Nationalised industry prices were reduced during that period, and so on. I do not think we would draw the correlation which you have just drawn. Berrill’s testimony is fully consistent with the views of Wilson and Heath on inflation, and with my representation of these views in article (OD5) and in (2). Therefore, the economic views of leading politicians were consistent with authoritative statements made at the technocratic level. Since Berrill’s testimony was made in public, should it be discounted as insincere? Contemporaneous reporting of Berrill’s testimony helps answer this question. The Guardian’s economics correspondent during the 1970s was Frances Cairncross, the daughter of Sir Alec Cairncross, one of Berrill’s predecessors. If Berrill’s testimony reflected mere political cover and not genuine Treasury thinking, Cairncross was in an ideal position to recognise this and therefore discount his statements. But Cairncross accepted Berrill’s statements as authoritative and endorsed his characterisation of inflation behaviour. In a late 1974 column, Cairncross quoted from the Berrill testimony given above and took it as reflecting Berrill’s sincere views (TG, December 2, 1974). She herself affirmed that ‘letting unemployment rise does not reduce inflation’ (TG, December 2, 1974) and claimed that the ‘links between what happens to demand and what happens to inflation are very difficult to establish’ (TG, March 4, 1978). The National Institute of Economic and Social Research also quoted and endorsed the passage of Berrill’s testimony given above (FT, March 14, 1975). So the view of a flat Phillips relation was not a political invention – it was common ground among policymakers, policy advisors and leading commentators. Let us now consider my position that policymakers in the old‐doctrine era did not have different objectives from their successors. Goodhart and Bhansali (1970) argue on the basis of polling data up to 1968 that high unemployment was penalised much more than moderate or high inflation. After the 1960s, values of unemployment above 2% became accepted as normal; unemployment had to reach far higher values to inflict political damage. Is this evidence that policymakers’ objectives changed? No, because I characterise objectives in terms of inflation and the output gap. I claim that the objective function expressed in this form has been the same since the 1950s; in particular, priority for a price‐stability goal has been frequently stated and subject to that priority, policymakers also had a zero‐gap goal. A constant‐parameter objective function encapsulating these goals is easily reconcilable with the observation that there was an increase in the unemployment rate regarded as acceptable. The Okun’s‐Law relation between production and unemployment changed. It was widely recognised by the late 1960s, largely on the basis of shifts in the Beveridge relation between unemployment and vacancies, that the full‐employment rate of unemployment rate had risen. This recognition was stated authoritatively in both ‘technocratic’ and political forums from late 1968. A structural rise in unemployment was noted publicly by the Treasury (Economic Trends, October 1968, p. iv), the Bank of England (QB, March 1969, p. 19), and by Cabinet member Anthony Crosland, who said that the ‘apparent change in the character of unemployment’ meant that ‘unemployment figures since 1966 do not necessarily indicate the same degree of slackness as would have been assumed from the same figures previously’ (HCD, May 6, 1970, col. 434). The Goodhart‐Bhansali findings on the political costs of unemployment refer to a sample ending in 1968. They therefore have no bearing on the cost of unemployment after the widely acknowledged structural change in the labour market and are completely consistent with the policy maker objective function, written with inflation and the output gap as arguments, being constant over the 1960s and beyond. Accounts of the Heath Government sometimes point to its ‘U‐Turn’ of 1971–2 as implying a shift in objectives, toward a greater weight on the output gap relative to inflation. Logically, this claim is unsupportable. The Heath Government believed that the economy was in a zone where demand stimulus would reduce inflation; therefore, no policy dilemma was perceived, and the shift to expansionary policy is not revealing at all about a change in objectives. In addition, Heath’s memoirs provide unmistakable support for my claim, based on the contemporary record, of constant policy maker objectives: Heath (1998, p. 416) writes that his 1972 measures were ‘not a departure from our underlying aims and objectives.’ 6. The Era of Modern Doctrine I now consider how modern doctrine describes policymakers’ views since 1979. 6.1. Margaret Thatcher (1979–90 Administration) Aggregate demand behaviour Thatcher went on record acknowledging that investment was sensitive to both short‐term interest rates (e.g. HCD, November 9, 1978, col. 1160) and long‐term interest rates (e.g. Thatcher press conference, September 29, 1983). Financial controls were judged ineffective (see e.g. H.M. Treasury, 1980, p. 11) and were abolished. Though not seeing the long rate as a policy instrument, the Thatcher Government at first distinguished between long‐term and Treasury‐bill financing of fiscal deficits, viewing bill issue as monetisation. From 1981, this view dissipated; the Government’s move to base money as its main money concept lent itself to a distinction between money and securities, rather than between different security types. By the mid‐1980s, the Government fully subscribed to modern doctrine on aggregate demand. Inflation behaviour In the mid‐1970s, Thatcher still favoured a voluntary incomes policy (HCD, November 29, 1976, col. 610) and believed that incomes policy mattered for inflation (LWT, May 9, 1976). She also had a speed‐limit perspective, contending that downward pressure on wage inflation ceases when unemployment stops rising (LWT, May 9, 1976). By the time Thatcher came to office, she accepted the monetary view of inflation.32 Complementing this, her Government’s policies were clearly guided by a long‐run vertical Phillips curve view of inflation behaviour.33 For example, Chancellor Geoffrey Howe said that higher unemployment is ‘not the bill we pay for reducing inflation now. It is the bill we pay for having allowed it to continue so long in the past.’ (The Times, May 13, 1982.) Smith (1987, p. 122) argues that Thatcher abandoned the expectational Phillips curve framework in 1985, citing an interview in which she said she did not embrace the ‘natural rate of unemployment’ concept. Smith, however, overlooks the fact that Thatcher made an essentially identical statement in 1981 (HCD, March 26, 1981, col. 1074). While Thatcher did not care for the ‘natural rate’ terminology, she effectively accepted the concept (see e.g. Thames TV, February 18, 1982). The Government rejected speed‐limit views of inflation. For example, the Treasury (1980, p. 9) said that provided a ‘firm basis for expectations about future inflation’ was created, sustained growth could take place. In 1987, Thatcher described the UK economic recovery that had proceeded during the 1980s (which included several years of a narrowing output gap, something seen by old doctrine as a trigger for speed‐limit inflationary pressure) as featuring ‘durable, non‐inflationary, sustained growth’ (Sydney Morning Herald, June 11, 1987). Policy objectives The Treasury, in its ‘explanation of Government strategy’, referred to the ‘ultimate objectives of price stability and high output and employment’ (1980, pp. 8, 9), with the reference output level determined by real factors, and with price stability to be achieved over several years. Chancellor Howe had before coming to office indicated that 2–3% inflation was the long‐run goal (Liverpool Daily Post, April 16, 1979). Similarly, Thatcher said in 1987 that she could promise low but not zero inflation (BBC1, June 8, 1987). Developments in 1987–90 are considered in my discussion of John Major. 6.2. John Major (1990–97 Administration) Throughout 1987–97, John Major served in important positions bearing on monetary policy: Chief Secretary of the Treasury, 1987–89; Foreign Secretary during intense debate on entering the Exchange Rate Mechanism (ERM), July–October 1989; Chancellor of the Exchequer, 1989–90; and Prime Minister, 1990–97. Aggregate demand behaviour Major regarded investment as elastic with respect to both short‐term and long‐term real interest rates, being more sensitive to the latter (e.g. HCD, November 3, 1982, col. 68; July 6, 1989, col. 463), and rejected credit controls (HCD, October 31, 1989, col. 207). A refinement during Major’s years at the Treasury, and preserved subsequently, was stress on the consumption channel. Earlier, the Treasury (1980, p. 10) had stated that monetary policy only had ‘small effects’ on consumption, but Major said that ‘interest rates exert downward pressure on the growth of consumer credit’ (HCD, January 26, 1989, col. 1170). Thatcher herself recognised the consumption channel when she noted that interest rates had been raised ‘[t]o encourage people to spend less and save more’ (Reuters, October 14, 1988). Inflation behaviour Both Chancellors Lawson and Major acknowledged the sensitivity of inflation to output gap and inflation, so they accepted the modern Phillips curve view encapsulated in proposition (MD5).34 But both Lawson and Major wanted a change in policy regime, namely, UK participation in the ERM. With this regime choice in mind, 1987–90 interest‐rate actions sought stabilisation of the sterling/mark rate. Among the arguments they advanced for this policy, Lawson and even more so Major stressed the interconnection of exchange rate stability, low import price inflation, and low CPI inflation. For example, Lawson said in 1989, ‘The exchange rate is of particular importance in the conduct of monetary policy. The Government’s clear commitment not to accommodate increases in domestic costs by exchange rate depreciation remains a key safeguard against inflation. In this context, we will continue to work with our G7 partners to maintain the greater exchange rate stability that has been a feature of the past two years.’ (HCD, March 14, 1989, col. 296.) Major similarly observed: ‘A falling exchange rate directly raises the prices of things we buy from abroad… That can only feed inflation… I favour a firm exchange rate.’ (HCD, October 31, 1989, col. 203.) It is true that in many open‐economy models there is a distinct exchange rate/import‐price term in the Phillips curve. Thus public recognition by Lawson and Major of the exchange rate/inflation link did not imply a non‐monetary view of inflation or a qualitative break with 1979‐87 official views of the transmission mechanism. But it amounted to a quantitative break: by emphasising the exchange rate’s importance for inflation, and treating exchange rate stability as equivalent to a non‐accommodative monetary policy, Lawson and Major by implication played down the output gap channel. The result was that when the exchange rate was strong or stable over 1987–9, the authorities underestimated the stimulus to inflation coming from aggregate demand. In partial mitigation, it is true that when strong spending data arrived in 1988–9, monetary policy was greatly tightened and Major recognised that the ‘problem at the moment is excess demand’ (HCD, January 26, 1989, col. 1170) and that the ‘only effective way to slow excessive demand is to put up the cost of borrowing’ (HCD, March 15, 1989, col. 427). A symmetric error occurred during ERM membership: overemphasis on the exchange rate in the transmission mechanism produced the danger of excessively falling inflation. In 1990 Major stated, ‘A firm exchange rate is a vital part of our policy to maintain tight monetary conditions in order to reduce inflation.’ (HCD, October 15, 1990, col. 928.) The ERM did produce ‘tight monetary conditions’ for the UK but in seeing exchange rate decline (either through ERM exit or realignment) as implying inflation, Major continued to have an over‐mechanical view. Even granting – in light of his acceptance of a monetary view of inflation – that Major understood that a rise in inflation following a depreciation cannot be sustained without monetary ease, he must still have underestimated the extent to which weak domestic economic conditions were burying UK inflation expectations and so preventing even a short‐run inflation spike after ERM exit. It is appropriate to conclude that Major consistently overestimated the exchange rate channel of monetary policy. Policy objectives Bernanke et al. (1999, p. 151) note accurately that UK policymakers did not see ERM membership (1990–2) as making exchange rate stability an ultimate objective; instead, the ERM was a vehicle for pursuing domestic objectives.35 Though not cited by these authors, a more senior policy maker, Major himself, made this explicit on several occasions. In particular he stated that the purpose of entering ERM was ‘to bear down on inflation and bring it down’. (The Independent, October 8, 1990.) Reinforcing this is the revealing formulation of King (1997b, pp. 82, 84) that the ERM represented a ‘conflict between domestic and external constraints’. The reference to a domestic constraint rather than domestic objectives relays the notion that the external constraint would be tolerated if, and only if, it succeeded in delivering domestic objectives. Therefore it is appropriate to regard Major’s policy objectives as the same as those of his predecessors, i.e. a zero output gap and a stable low inflation rate (the latter to be achieved in steps when starting from high inflation).36 Major stressed the desirability of low inflation: ‘Britain has nothing to gain – not now, not ever – by tolerating a high rate of inflation or even a modest rate of inflation.’ (HCD, January 23, 1990, col. 757.) In 1990 he even said, ‘Ultimately I would like to see zero inflation.’ (Reuters, October 7, 1990.) Indeed, while the inflation target band assigned to the Bank of England by Chancellor Norman Lamont in 1992 did not include zero, it was seen as a precursor to a long‐run inflation target of 0‐2%.37 But talk of zero inflation represented, both in 1990 and 1992, a conjecture about what was desirable in the far future; on other occasions Major expressed satisfaction with getting ‘back to a low or no‐inflation economy’ (Reuters, February 9, 1991).38 Major was consistent in defining low inflation as somewhere below 4%. This definition was implied by his statement that 4.1% inflation was ‘lowish’ rather than low (Reuters, March 27, 1992), and was made explicit by the fact that the inflation target band announced in 1992 was 1–4%, accompanied by an instruction to reach the lower half of this range by 1997. 6.3. Mervyn King (UK Inflation Targeting Regime, 1992–) The Bank of England assumed a greater and more public role in policymaking once inflation targeting was introduced in 1992, and became the policymaking agency when made independent in 1997. Even before becoming Bank Governor in 2003, Mervyn King was the principal communicator of the economic analysis underlying monetary policy in the inflation targeting era. Therefore, I focus on his views. As shown below, King’s positions since the 1990s fully reflect modern doctrine. Aggregate demand behaviour King’s (1977) outline of investment behaviour is in line with modern doctrine. Investment depends on the short rate, and alternative sources of short‐term funds are closely linked to the short‐term securities market in equilibrium. By emphasising the short‐rate channel and pointing out that ‘[t]o say that we do not have a complete theory of investment is not the same thing as saying that we know nothing’ (1977, p. 229), King’s analysis amounts to a rejection of the Radcliffe Report, with its denial of a short‐rate elasticity. More recently, King has reaffirmed that investment depends on short‐term real interest rates but has added that consumption also depends on these rates (e.g. King, 1999, p. 41). He has noted that long‐term interest rates, which are sensitive to expectations of the short rate, do ‘a lot of the work’ in the monetary transmission mechanism (King, 1996, p. 79) and that credit controls are not an appropriate policy instrument (King, 1995a, p. 431). Inflation behaviour – early views In contrast to his early work on aggregate demand issues, King’s initial writings on inflation behaviour were in keeping with the old doctrine. For example, writing in The Guardian in 1973, King seemed enthusiastic about incomes policies: ‘tax policy has an important role to play in the operation of an incomes policy’ (TG, November 14, 1973). In 1977 he was even more clearly in favour: while ‘some theorists say we don’t need’ incomes policies, ‘[e]veryone concerned in a practical way with the management of the British economy agrees that we do’. (Matthews and King, 1977, p. 7.) A detailed outline of King’s views on macroeconomic policy in the 1970s is given in Matthews and King (1977). The authors firmly align themselves with the non‐monetary view of inflation, and thus the old doctrine, by characterising the UK situation as one of simultaneous prolonged slack and high inflation. They subsequently give (p. 4) a figure of minus 12% for the UK output gap for 1977, roughly the same (erroneous) estimate used in policymaking at the time. A section entitled ‘Crude Monetarism’ describes expectational Phillips curve/natural‐rate theory as ‘debatable’, and concludes that, after monetary restriction, ‘the ultimate recovery of the economy [is] problematical’ (Matthews and King, 1977, p. 3). King capped off his ‘anti‐monetarist’ phase by joining 363 other economists in signing an open letter criticising the Thatcher disinflation programme and declaring that the time had come to ‘reject monetarist policies’. The signatories included a ‘rogues’ gallery’ of hard‐line sceptics about monetary policy, including Nicholas Kaldor, Frank Hahn and Robert Neild.39 Comparing King’s 1973–81 ‘anti‐monetarist’ phase and his modern writings, it is clear that sometime after 1981 he rethought his macroeconomic views in a way that put matters such as the natural rate hypothesis and the monetary theory of inflation in a much more favourable light. By the mid‐1990s King was voicing assessments clearly informed by the monetarist canon, such as his observation that ‘we have had cause to resort to “long and variable time lags” in our description of the transmission’ (King, 1992, p. 309) and his warning that it was unwise to ‘attempt to fine‐tune in our present state of ignorance’ (King, 1995b, p. 393). In fact, I have felt that King has overcompensated for his earlier antipathy toward monetarism, in the sense that he now rarely seems to encounter an argument for looking at money that he does not like, so that (alongside sound ones) some frankly shaky arguments in favour of monetary aggregates have appeared in his writings.40 Inflation behaviour – revised position When King joined the Bank of England in 1991, the United Kingdom was a participant in the ERM; the choice of monetary regime was officially a settled matter. Nevertheless, at a conference in Sydney in July 1992, King made it clear that he did not regard other policy options as off the table; his conference discussion covered alternatives to pure discretion without even mentioning the ERM. King (1992, p. 311) even granted that a move to inflation targeting might be ‘the appropriate direction to pursue’ in the UK. His discussion made clear that he had a far less apocalyptic view of the consequences of ERM exit than what Prime Minister Major was expressing over this period. By the time King’s remarks saw print in late November 1992, the ERM experiment was over and inflation targeting had begun. Relative to Major’s perspective on inflation control in 1990–2, the main refinement under King has been much less emphasis on the role of import‐price or exchange rate behaviour in securing low UK inflation. He has made use of the fact that many of the principles that are valid in closed‐economy monetary analysis hold true for the open economy.41 King (1992, p. 309) viewed inflation dynamics as a reflection of shifting private forecasts of future monetary policy stance. Lagged terms in estimated inflation equations therefore did not reflect economic structure.42 This view implies that the growth rate of the output gap is not an element of the correct Phillips curve specification.43 Policy objectives The explicit inflation targets in the United Kingdom since 1992 are set by the Chancellor of the Exchequer. But the targets (now 2% but never far from that rate) are consistent with King’s writings on the desirable rate. For example, in 1977 King noted that ‘hardly anybody says that we must have absolutely zero inflation’, suggested that a rate of 2% or 3% was tolerable, and that higher rates damaged the economy (Matthews and King, 1977, p. 3); while King (1997a, p. 93) observed that 2% inflation was ‘often associated with price stability’. King has also indicated the desirability of a zero output gap, notably in Matthews and King (1977) and King (1997b, 1999), with potential output and the implied natural unemployment rate determined by real factors (see e.g. King, 1996, 1997b). 7. Conclusions UK postwar macroeconomic policy can be split into two eras, corresponding to two doctrines, old and modern. By viewing policy developments in light of these doctrines, it is possible to explain and reconcile the statements and actions of leading UK policymakers, both before and after 1979. The doctrinal changes in the UK have not included changes in objectives. Many have argued that UK policymakers in the 1960s and 1970s consciously chose high inflation out of belief in a permanent inflation/unemployment trade‐off. The advocates of this argument have not provided supporting documentation from UK policymakers’ statements. In fact, belief in a trade‐off was not part of UK policymaking doctrine either before or after 1979. Policymakers in the 1960s and 1970s repeatedly and explicitly rejected the idea of a trade‐off; furthermore, their policy actions were not consistent with a Phillips curve philosophy. It was an overhaul of doctrine – and not altered faith in a Phillips curve trade‐off – that set off changes in UK monetary policy, and that overhaul of doctrine remains the underpinning of inflation targeting in the UK. Footnotes 1 " In particular, Nelson (2005), Nelson and Nikolov (2004) and Batini and Nelson (2005). 2 " Indeed, at yt < yt* there is no ‘curve’ relationship between the gap and inflation, even holding inflation expectations constant. 3 " The Appendix provides bibliographical details for the newspaper articles and documents quoted in the text. 4 " From Wilson speech, HCD, April 18, 1956, in Wilson (1964, p. 38); and Wilson (1957, p. 14). 5 " Productive investment was perceived as being financed by long‐term loans outside the banking system. 6 " It was widely acknowledged that the distribution of the stock of saving between bank and non‐bank instruments was interest‐sensitive, even by those such as Wilson who felt that the choice between saving and consumption flows was interest‐insensitive. 7 " In (1), the interval is taken to be one period (a quarter). If instead the interval was (e.g.) a year, ΔRt would be replaced by Δ4Rt. 8 " In general, subscribers to the old doctrine accepted that, insofar as asset returns mattered for economic decisions, they did so as real rather than nominal rates (see e.g. Radcliffe Committee, 1959, para. 572). But there was no presumption that the monetary authorities should act on nominal rates to keep real rates positive, as it was felt that the expected‐inflation term in the Fisher equation could be manipulated by non‐monetary means. 9 " Though definitions of the long rate differed across discussions, a representative maturity would be ten years or more. In believing that there was a non‐zero, but low, long‐rate elasticity of aggregate demand, UK policymakers’ position on the IS equation in the 1950s and 1960s was similar to that of US policymakers in the late 1940s (see Friedman and Schwartz, 1963, p. 700). 10 " Manchester Guardian (hereafter MG), August 21, 1947. 11 " In keeping with the cost‐push view of inflation, price increases in one market were seen as automatically meaning a rise in aggregate prices. 12 " King (2005) cites passages from the Radcliffe Report as the basis for arguing that the Committee believed in a Phillips‐style inflation/unemployment trade‐off. But the Committee’s discussion of the Phillips relationship was in reference to evidence the Committee had heard, not views it endorsed (1959, para. 64); it immediately cited contrary testimony, and went on instead to endorse the cost‐push view of inflation. Laidler (1989) concurs that the Radcliffe Committee’s outlook was cost‐push. 13 " Similarly, Jenkins’ predecessor, James Callaghan, said shortly before taking office that ‘postwar experience has shown the difficulty of combining a fast expansion rate in the economy with steady prices’ (HCD, July 20, 1964, col. 66); while one of Jenkins’ successors, Nigel Lawson, wrote in 1967 that ‘total elimination of inflation – to prices not rising all’ would produce ‘a basic conflict with faster economic growth’ but that low inflation rates would not produce this conflict (The Spectator, February 24, 1967). 14 " At the same time, even during the era of a completely pegged exchange rate (1949–67), exchange controls gave policymakers considerable freedom in their quarter‐to‐quarter interest‐rate choices. 15 " We know of course that, via the output gap, monetary policy matters for the price level, which enters the definition of qt. But the old doctrine denied this channel – see postulates (OD1) and (OD5). 16 " Edward Heath, for example, expressed the view that, in the absence of offsetting actions, ‘a faster rate of inflation is now in‐built into the system’ (in FT, June 17, 1970). 17 " Taking unconditional means of (6) and (7), one finds that each equation collapses to an expression for E(Δπ), the change in inflation, not the absolute level E(π). Cost‐push and monetary views can be shown to differ also on the determination of this long‐run level. 18 " For example, the custodians of the Harold Wilson Papers at Oxford University report that the Papers consist of approximately two thousand boxes, most of them still to be catalogued. 19 " The first use of ‘stagflation’ by The Economist was in its August 15, 1970 issue, well behind other newspapers; see Nelson (2005). 20 " This is true both of what I attribute to policymakers (their views on objectives and the inflation process) and of the ideas that I maintain policymakers rejected (e.g., a permanent Phillips curve trade‐off). 21 " The period studied by Cobham is nominally 1975–2000 but pre‐1979 discussion is brief; Wilson and Callaghan are not mentioned, nor is the Chancellor of the Exchequer referred to by name. 22 " Although the authors’ preface thanks King for comments and an endnote thanks him for a factual clarification (1999, p. 345), they nowhere quote or reference King’s writings or speeches. 23 " This perspective continued in Wilson’s 1974‐6 Government, with Chancellor Healey describing the Government’s desired investment and balance of payments patterns as ‘[w]ithin our overall commitment to fight unemployment and inflation’ (HCD, November 12, 1974, col. 256; my emphasis). 24 " ‘I believe that it has been monetary policy buttressing or buttressed by fiscal policy and the prices and incomes policy which has enabled us to achieve the results which we have achieved, and not monetary policy on its own.’ (Chancellor Jenkins, HCD, December 17, 1969, col. 1472). 25 " Speed‐limit behaviour was regarded as a market inefficiency (i.e., the price index reacting as though there was excessive demand when no excess demand existed in aggregate), so suppressing them by controls or agreement was regarded as improving efficiency rather than distorting the signals of the market. 26 " This would mean specifying (2) so that part of ut is in per‐unit terms, i.e., ut = v1t + γ(v2t − yt), where γ > 0, with both v1t and v2t exogenous. 27 " Austerity messages in December 1973 reflected the coalmining dispute rather than acceptance that the state of demand was excessive; these measures included more financial control but no interest‐rate increase. 28 " I have not covered output gap mismeasurement in this article, but it was substantial in the UK (Nelson and Nikolov, 2004), exceeding even the US case discussed by Orphanides (2004). It is desirable to see this problem as a function of other conceptual errors affecting policymaking since, as Romer and Romer (2002) observe, the errors surely interact. In particular, cost‐push views of inflation rationalise any gap/inflation combination and so encourage policymakers to take existing gap estimates at face value. 29 " This was also the position of the Callaghan Government (see Healey’s remarks, HCD, July 6, 1976, col. 1188). 30 " In Macmillan (1973, p. 386). 31 " In House of Commons (1974, p. 136). 32 " Many accounts erroneously attribute to Thatcher monetarist views on inflation in 1975 or even earlier. An exception is Campbell (2000, p. 372), who sees her monetary view of inflation as crystallising in 1978. 33 " This was acknowledged by Goodhart (1983, p. 219); see also H.M. Treasury (1980, p. 13). 34 " Also, in line with postulate (MD6), Major implied that speed‐limit dynamics were not a feature of inflation (see HCD, January 14, 1988, col. 549; April 26, 1988, col. 215). 35 " As argued above, fixed exchange rates were also not a final objective of the 1964‐70 Wilson Government. 36 " Major’s zero‐gap objective is clear from his reference, noted above, to excess demand as a ‘problem.’ 37 " See King (1997b, p. 91). 38 " Furthermore, shortly before his defeat Major confirmed he wanted a 2.5% inflation target for 1997–2002 (The Guardian, April 2, 1997) and soon after losing office spoke of the dangers of pushing inflation below this rate (Market News International, May 14, 1997). 39 " The full list of signatories is given in The Times, March 31, 1981. 40 " To take a recent example, King (2007, p. 281) argues that the fact that ‘inflation was subdued’ in 1973 (then rising subsequently) alongside rapid money growth provides lessons about the transmission mechanism. This happens to be an especially unsuitable argument if the intention is to convince money’s critics of money’s importance. The year 1973 was the only one in the last half‐century in which UK prices were under general statutory control for the whole of the calendar year; therefore, subdued behaviour of inflation in that year is not revealing at all about the transmission mechanism. 41 " An early indication of this was King’s statement in a 1996 Inflation Report press conference that the rising pound was ‘no substitute’ for monetary policy actions (AFP, November 6, 1996). 42 " King’s stress on the distinction between price‐level shocks and inflation (e.g., King, 1995b, p. 392) similarly implies that lagged inflation does not enter the Phillips curve with a heavy weight. 43 " This contrasts with King’s early belief in speed‐limit terms, clear from Matthews and King (1977, p. 4). References Batini , N. and Nelson , E. ( 2005 ). ‘ The UK’s rocky road to stability ’, Federal Reserve Bank of St. Louis Working Paper No. 2005–020A. Benati , L. ( 2008 ). ‘ The great moderation in the United Kingdom ’, Journal of Money, Credit and Banking , vol. 40 ( 1 ), pp. 121 – 47 . Google Scholar Crossref Search ADS WorldCat Bernanke , B.S. ( 2004 ). ‘ The great moderation ’, Speech at Eastern Economic Association, Washington, D.C., February 20. Bernanke , B.S. , Laubach , T., Mishkin , F.S. and Posen , A.S. ( 1999 ). Inflation Targeting: Lessons from the International Experience , Princeton, NJ: Princeton University Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Blake , A.P. and Markovic , B. ( 2007 ). ‘ The conduct of global monetary policy and domestic stability: luck versus judgment ’, manuscript, Bank of England. Campbell , J. ( 2000 ). Margaret Thatcher, Volume 1: The Grocer’s Daughter , London: Jonathan Cape . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Clarida , R. , Galí , J. and Gertler , M. ( 1999 ). ‘The science of monetary policy: a New Keynesian perspective’ , Journal of Economic Literature , vol. 37 ( 4 ), pp. 1661 – 707 . Google Scholar Crossref Search ADS WorldCat Cobham , D. ( 2002 ). The Making of Monetary Policy in the UK, 1975‐2000 , Sussex: Wiley . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Crosland , C.A.R. ( 1956 ). The Future of Socialism , New York: Macmillan . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Friedman , M. and Schwartz , A.J. ( 1963 ). A Monetary History of the United States, 1867–1960 , Princeton, NJ: Princeton University Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Friedman , M. and Schwartz , A.J. ( 1982 ). Monetary Trends in the United States and the United Kingdom: Their Relation to Income, Prices, and Interest Rates, 1867–1975 , Chicago: University of Chicago Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Goodhart , C.A.E. ( 1983 ). ‘Review: Money and Inflation’ . Economic Journal , vol. 93 ( 369 ), pp. 217 – 9 . Google Scholar Crossref Search ADS WorldCat Goodhart , C.A.E. and Bhansali , R.J. ( 1970 ). ‘Political economy’ , Political Studies , vol. 18 ( 1 ), pp. 43 – 106 . Google Scholar Crossref Search ADS WorldCat Heath , E. ( 1998 ). The Course of My Life: The Autobiography of Edward Heath . London: Hodder and Stoughton . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC H.M. Treasury ( 1980 ). ‘Memorandum by H.M. Treasury, 20 June 1980’, in Treasury and Civil Service Committee , Memoranda on Monetary Policy , pp. 8 – 16 . London: HMSO . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC House of Commons ( 1974 ). Public Expenditure, Inflation, and the Balance of Payments: Ninth Report from the Expenditure Committee , London: HMSO . Keynes , J.M. ( 1936 ). The General Theory of Employment, Interest and Money , London: Macmillan . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC King , M.A. ( 1977 ). Public Policy and the Corporation , London: Chapman and Hall . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC King , M.A . ( 1992 ). ‘Discussion: inflation indicators and monetary policy’, in ( A. Blundell‐Wignall, ed.), Inflation, Disinflation and Monetary Policy , pp. 307 – 12 , Sydney: Ambassador Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC King , M.A . ( 1995a ). ‘Comments: monetary policy rules and financial stability’, in ( K. Sawamoto, Z. Nakajima and H. Taguchi, eds.), Financial Stability in a Changing Environment , pp. 430 – 3 , New York: St. Martin’s Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC King , M.A. ( 1995b ). ‘Do inflation targets work?’ , Bank of England Quarterly Bulletin , vol. 35 ( 4 ), pp. 392 – 4 . OpenURL Placeholder Text WorldCat King , M.A. ( 1996 ). ‘How should central banks reduce inflation? – Conceptual issues’, in Achieving Price Stability , pp. 53‐91 , Kansas City, MO: Federal Reserve Bank of Kansas City . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC King , M.A. ( 1997a ). ‘Monetary stability: rhyme or reason?’ , Bank of England Quarterly Bulletin , vol. 37 ( 1 ), pp. 88 – 96 . OpenURL Placeholder Text WorldCat King , M.A. ( 1997b ). ‘Changes in UK monetary policy: rules and discretion in practice’ , Journal of Monetary Economics , vol. 39 ( 1 ), pp. 81 – 97 . Google Scholar Crossref Search ADS WorldCat King , M.A. ( 1999 ). ‘Challenges for monetary policy: new and old’, in New Challenges for Monetary Policy , pp. 11 – 57 , Kansas City, MO: Federal Reserve Bank of Kansas City . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC King , M.A. ( 2005 ). ‘ Monetary policy: practice ahead of theory: the Mais lecture 2005 ’, Bank of England Quarterly Bulletin , vol. 45 ( 2 ), pp. 226 – 36 . OpenURL Placeholder Text WorldCat King , M.A. ( 2007 ). ‘The MPC ten years on’ . Bank of England Quarterly Bulletin , vol. 72 ( 2 ), pp. 272 – 85 . OpenURL Placeholder Text WorldCat Laidler , D. ( 1989 ). ‘Dow and Saville’s Critique of Monetary Policy– a review essay’ , Journal of Economic Literature , vol. 27 ( 3 ), pp. 1147 – 59 . OpenURL Placeholder Text WorldCat Macmillan , H. ( 1973 ). At the End of the Day, 1961–1963 , London: Macmillan . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Matthews , R.C.O. and King , M.A. ( 1977 ). ‘The British economy: problems and policies in the late 1970s’. Midland Bank Review , vol. 32 ( 1 ), pp. 1 – 10 . Nelson , E. ( 2005 ). ‘ The great inflation of the seventies: what really happened? ’, Advances in Macroeconomics, vol. 5(1), p. 1297. Nelson , E. and Nikolov , K. ( 2004 ). ‘Monetary policy and stagflation in the UK’ , Journal of Money, Credit, and Banking , vol. 36 ( 3 ), pp. 293 – 318 . Google Scholar Crossref Search ADS WorldCat Orphanides , A. ( 2004 ). ‘Monetary policy rules, macroeconomic stability, and inflation: a view from the trenches’ , Journal of Money, Credit, and Banking , vol. 36 ( 2 ), pp. 151 – 75 . Google Scholar Crossref Search ADS WorldCat Parsons , W. ( 1989 ). The Power of the Financial Press: Journalism and Economic Opinion in Britain and America , Cheltenham, UK: Edward Elgar . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Pepper , G. and Oliver , M. ( 2001 ). Monetarism Under Thatcher: Lessons for the Future , Cheltenham: Edward Elgar . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Phelps , E.S. ( 1968 ). ‘Money‐wage dynamics and labor market equilibrium’ , Journal of Political Economy , vol. 76 ( 4 ), pp. 678 – 711 . Google Scholar Crossref Search ADS WorldCat Radcliffe Committee ( 1959 ). Report of the Committee on the Working of the Monetary System, Cmnd 827 , London: HMSO . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Romer , C.D. and Romer , D.H. ( 2002 ). ‘The evolution of economic understanding and postwar stabilization policy’, in Rethinking Stabilization Policy , pp. 11 – 78 , Kansas City, MO: Federal Reserve Bank of Kansas City . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Romer , C.D. and Romer , D.H. ( 2004 ). ‘Choosing the Federal Reserve chair: lessons from history’ , Journal of Economic Perspectives , vol. 18 ( 1 ), pp. 129 – 62 . Google Scholar Crossref Search ADS WorldCat Rotemberg , J.J. and Woodford , M. ( 1997 ). ‘An optimization‐based econometric framework for the evaluation of monetary policy’ , NBER Macroeconomics Annual , vol. 12 ( 1 ), pp. 297 – 346 . Google Scholar Crossref Search ADS WorldCat Sargent , T.J. ( 1999 ). The Conquest of American Inflation , Princeton, NJ: Princeton University Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Smith , D. ( 1987 ). The Rise and Fall of Monetarism , London: Penguin . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Stock , J.H. and Watson , M. ( 2002 ). ‘Has the business cycle changed and why?’ , NBER Macroeconomics Annual , vol. 17 ( 1 ), pp. 159 – 218 . Google Scholar Crossref Search ADS WorldCat Wilson , H. ( 1957 ). Post‐War Economic Policies in Britain , London: Fabian Society . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Wilson , H. ( 1964 ). Purpose in Politics: Selected Speeches , Boston: Houghton Mifflin . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Woodford , M. ( 2003 ). Interest and Prices: Foundations of a Theory of Monetary Policy . Princeton, NJ: Princeton University Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Appendix Appendix. Details of Newspaper Articles and Documents Mentioned in Text ‘Our Labour Correspondent’, ‘Convincing World That ‘We Mean Business’– Mr. Morrison Calls For New Effort: Why Some Investment Projects Must Be Postponed’, Manchester Guardian, August 21, 1947, p. 5. Harold Wilson, ‘Remedies for Inflation – I: Contradictory Policies’, Manchester Guardian, October 23, 1957, p. 6. Harold Wilson, ‘Remedies for Inflation – II: The Unemployment School’, Manchester Guardian, October 24, 1957, p. 8. Harold Wilson, ‘Remedies for Inflation – III: What Labour Would Do’, Manchester Guardian, October 25, 1957, p. 8. Francis Cassell, ‘How the City Works: The Bank of England’, Federation of British Industries Review, November 1961, pp. 51–2. ‘Commentary: January–March 1962’, Bank of England Quarterly Bulletin, June 1962, vol. 2(2), pp. 83–92. Duncan Stirling, ‘Westminster Bank: Statement by the Chairman’, Financial Times (London), January 21, 1963, p. 2. (Also appeared as Duncan Stirling, ‘Company Meeting – Westminster Bank Limited: Year of Expansion’, New Statesman (London), January 25, 1963, p. 137.) Charles Pritchard, ‘Attracting Money to the Banks: Problem of Interest Rates on Deposits’, Yorkshire Post (Leeds), January 20, 1964, p. 8. ‘The Wilson Economic Blueprint’, Sunday Times (London), January 26, 1964, p. 4. ‘Our Own Correspondent’, ‘Mr. Heath’s Advice to Scottish Industry: Modernize and Get More Drive into Exporting’, Glasgow Herald, September 7, 1964, p. 7. Peter Jenkins, ‘Mr. Brown Says No ‘Stop‐Go’,’The Guardian (London and Manchester), December 10, 1964, p. 6. Alastair Warren, ‘Business Boom Could Be Short‐Lived’, Glasgow Herald, February 12, 1965, p. 3. ‘New Clamp on Spending Soon, Says Premier’, Sunday Telegraph (London), July 25, 1965, p. 6. ‘Harold Wilson Sums Up…’, Sunday Citizen (London), March 27, 1966, p. 13. Jonathan Radice, ‘The Pound and the Seaman’s Strike’, Socialist Commentary (London), August 1966, pp. 12–4. ‘‘Inflation Has Been Cured’: Pay Warning by Mr. Crossman’, The Times (London), December 19, 1966, p. 1. Nigel Lawson, ‘An Alternative Economic Policy’, The Spectator (London), February 24, 1967, pp. 214–6. Editorial, ‘Put Not Thy Faith in Paish’, The Spectator, April 14, 1967, p. 413. ‘Gilt‐Edged’, Bank of England Report for the Year Ended 28th February 1967, July 1967, p. 10. ‘Our Political Correspondent’, ‘Premier Hints at New Wage Curbs’, The Scotsman, March 7, 1968, p. 5. H.M. Treasury, ‘The Economic Situation’, Economic Trends, October 1968. ‘Commentary’, Bank of England Quarterly Bulletin, March 1969, pp. 3–20. ‘Weekend Speeches – Heath: I Challenge Chancellor to Save £400 Million’, Sunday Times, March 23, 1969, p. 2. Anthony Thomas, ‘Inflation May Slow to 2(1/2) pc, Says Chancellor’, The Times, January 12, 1970. David Palmer, ‘Chancellor Is Hopeful’, Financial Times, January 12, 1970, p. 1. John Bourne, ‘Heath Picks on Rise in Prices as Main Election Issue’, Financial Times, May 23, 1970, p. 1. ‘Commentary’, Bank of England Quarterly Bulletin, June 1970, pp. 119–35. Edward Heath, ‘Heath: The Choice Before the Country’, Financial Times, June 17, 1970, p. 10. Editorial, ‘Winter in the Air’, The Economist, August 15, 1970, pp. 12–3. FT Reporter, ‘Cost of Tougher Monetary Policy Severe – Barber’, Financial Times, January 20, 1971, p. 15. Edward Heath, ‘Greater Opportunity for Initiative: Foreword by the Prime Minister’, Liverpool Daily Post (Merseyside) (Financial and Commercial Review section), January 20, 1971, p. 2. City Editor, ‘Not a Flinch from Mr. Barber’, Daily Telegraph (London), April 22, 1971, p. 19. John Bourne, ‘Reflation: Heath Refuses to Set Date’, Financial Times, June 19, 1971, p. 1. Philip Rawstorne, ‘Expansion Ahead Says Heath; More Investment Needed’, Financial Times, October 12, 1971, p. 1. ‘Commentary’, Bank of England Quarterly Bulletin, March 1972, pp. 3–21. ‘Budget – Secret: Budget Speech [March] 1972: Outline’, Cabinet Documents (with March 12, 1972, handwritten comments by Edward Heath); declassified (scan available on Thatcher Foundation web site, http://www.margaretthatcher.org). Charles Wintour and Robert Carvel, ‘Edward Heath Speaking Frankly: My Style of Government’, Evening Standard (London), June 1, 1972, pp. 24–5. Harold Wilson, ‘My Cure for Inflation’, Sunday Times, August 6, 1972, p. 55. AFP and PA, ‘Barber Opposes Restrictions’, Irish Times (Dublin), October 13, 1972, p. 6. Stewart Fleming, ‘11(1/2) pc Rate – But Barber Still Seeks Growth’, The Guardian, July 28, 1973, p. 1. Clifford German, ‘Faith, Hope and Charity in Phase Three’, Daily Telegraph, October 9, 1973, p. 21. ‘Cut Tax on Petrol, Says Wilson’, Cambridge Evening News (Cambridgeshire), October 20, 1973. Mervyn King, ‘Profits: Non‐Existent Squeeze’, The Guardian, November 14, 1973, p. 21. Frances Cairncross, ‘Jobs in the Balance’, The Guardian, December 2, 1974, p. 12. William Keegan, ‘Healey Budget May Have to Be Broadly Neutral – Institute’, Financial Times, March 14, 1975, pp. 1 and 38. William Ellington, ‘British Move to Bring Down Interest Fees Causes Investors to Switch to Stocks, Gold’, Wall Street Journal, April 24, 1975, p. 13. Bryn Jones, ‘Wilson Appeals to the Miners: Give a Year for Britain’, Daily Mirror, July 8, 1975, pp. 4–5. William Keegan, ‘Anti‐Inflation Fight Will Not Slacken, Promises Wilson’, Financial Times, February 3, 1976, p. 1. Victor Keegan, ‘Wilson Strikes Bright Note as Pound Dips’, The Guardian, February 3, 1976, p. 1. Victor Keegan, ‘Healey to Crush Hopes of Swift Expansion in Demand’, The Guardian, March 1, 1976, p. 18. Frances Cairncross, ‘Healey Toys with the Supernatural’, The Guardian, April 8, 1976, p. 18. LWT, Weekend World, May 9, 1976 (transcript on Thatcher Foundation site). Terry Coleman, ‘Don’t Shoot the Pianist: He May Become PM’, The Guardian, September 16, 1976, p. 13. ‘The End of the Primrose Path – By the Prime Minister (Speaking in Blackpool Today)’, Evening Standard, September 28, 1976, p. 15. ‘Pay Scramble ‘‘Would Spell Catastrophe’’’, Yorkshire Post, February 19, 1977, p. 2. Charles Pritchard, ‘Bank of England Back in Control of Gilt‐Edged Market’, Yorkshire Post, May 31, 1977, p. 4. Frances Cairncross, ‘A Mixed Bag, Indeed’, The Guardian, March 4, 1978, p. 16. Keith Harper, ‘Callaghan Toughens Pay Line’, The Guardian, July 1, 1978, pp. 1 and 22. ‘Ex‐PM Heath Is Chatty Here’, The Plain Dealer (Cleveland, Ohio), August 17, 1978, p. 1B. ‘A Tory Tax‐Cut Promise’, Liverpool Daily Post, April 16, 1979, p. 5. Stuart Trotter, ‘My Formula for the ’80s: The Prime Minister Talks to the Herald ’, Glasgow Herald, May 2, 1979, p. 6. BBC2, Free to Choose, episode ‘How to Cure Inflation’, March 22, 1980 (transcript). ‘Universities Economics Signatories’, The Times, March 31, 1981, p. 21. AFP, ‘Maggie Blamed for Crime Rise’, South China Morning Post (Hong Kong), July 3, 1981, p. 5. Thames TV, TV Eye, February 18, 1982 (transcript on Thatcher Foundation site). Peter Norman, ‘Note of Optimism at IMF Meeting’, The Times, May 13, 1982, p. 17. BBC1, The Twentieth Century Remembered, September 1, 1982, programme 4 (transcript). Press conference, Washington DC, September 29, 1983 (transcript on Thatcher Foundation site). BBC1, Panorama, June 8, 1987 (transcript). Yvonne Preston, ‘Britons Prepare to Deliver Verdict on Thatcher’, Sydney Morning Herald, June 11, 1987, p. 15. Ralph Boulton, ‘Thatcher Vows to Cut Inflation, At Three‐Year High’, Reuters, October 14, 1988. Reuters, ‘Major Targets Zero Inflation, No Electoral Boomlet’, October 7, 1990. Anthony Bevins, ‘Chancellor Denies Seeking to Fuel Pre‐Election Economic Boom’, The Independent (London), October 8, 1990, p. 1. Reuters, ‘UK Inflation to Fall Throughout Year – Major’, February 9, 1991. Reuters, ‘United Kingdom’s Major Raises Possibility of Interest‐Rate Cut’, March 27, 1992. AFP, ‘Bank of England Points to Need for Fresh Rate Rise’, November 6, 1996. Ewan MacAskill and Rebecca Smithers, ‘Schools Top Party Agendas’, The Guardian, April 2, 1997, p. 11. Market News International, ‘UK Tories’ Major: Bank of England Will Meet Inflation Target But ‘‘Too Well’’’, May 14, 1997. Author notes " I thank the Editor and two anonymous referees for useful comments on a previous draft. An early version of this article was presented at the conference ‘On the Sources of Macroeconomic Stability’, Bank of England, September 13–14, 2007. I am indebted for comments on the conference version to conference participants (especially my discussant, Charles Goodhart) and to Amit Kara, Bennett McCallum, Paolo Surico and Michael Woodford. I thank Lurline Campbell, Chris Collins, Colin Harris, Julie Judge, Britne Rockwell, Katrina Stierholz and Julia Williams for help in obtaining archival material. Justin Hauke and Faith Weller provided research assistance. The views expressed in this article are those of the author and should not be interpreted as those of the Federal Reserve Bank of St. Louis, the Federal Reserve System, or the Board of Governors. © The Author(s). Journal compilation © Royal Economic Society 2009

O’Mahony,, Mary;Timmer, Marcel, P.

doi: 10.1111/j.1468-0297.2009.02280.xpmid: N/A

Abstract This article describes the contents and the construction of the EU KLEMS Growth and Productivity Accounts. This database contains industry‐level measures of output, inputs and productivity for 25 European countries, Japan and the US for the period from 1970 onwards. The article considers the methodology employed in constructing the database and shows how it can be useful in comparing productivity trends. Although growth accounts are the organising principle, it is argued that the database is useful for a wider range of applications. We give some guidance to prudent use and indicate possible extensions. Internationally comparable studies of the relationships between skill formation, investment, technological change and growth have been hampered up to now by the lack of a readily available standard database covering a large set of countries. As a result, researchers had often to compile their own databases, making replication and comparability of studies difficult. This article describes the construction of a new database which can serve as a useful tool for empirical and theoretical research in the area of economic growth: the EU KLEMS Growth and Productivity Accounts. This database includes measures of output and input growth, and derived variables such as multi‐factor productivity at the industry level. The input measures include various categories of capital (K), labour (L), energy (E), material (M) and service inputs (S). The measures are developed for 25 individual EU member states, the US and Japan and cover the period from 1970 to 2005. The variables are organised around the growth accounting methodology, a major advantage of which is that it is rooted in neo‐classical production theory. It provides a clear conceptual framework within which the interaction between variables can be analysed in an internally consistent way. The data series, publicly available on http://www.euklems.net, can be used by researchers employing growth accounting to consider sources of output and productivity growth in cross‐country comparisons or studies of particular industries and different time periods, such as in Jorgenson et al. (2005) for the US and van Ark et al. (2008) for Europe versus the US. Although the primary aim of the EU KLEMS database is to generate comparative productivity trends, the data collected are also useful in a large number of other contexts, as the EU KLEMS database provides many basic input data‐series. These input series are derived independently from the assumptions underlying the growth‐accounting method. Due to its wide country and industry coverage, potential applications of the database vary widely. For example, there has been considerable research on the issue of whether technical change is skill‐biased and on the impact of information and communications technology (ICT) on the demand for skilled labour (e.g. Autor et al., 1998; Machin and van Reenen, 1998). Typically researchers estimate wage share equations by labour type and include a variable measuring some aspect of technological change. EU KLEMS is likely to be useful in such exercises as the database contains information on hours and wage bill shares cross‐classified by skill, gender and age and provides a breakdown of capital into ICT and non‐ICT assets. EU KLEMS data can also be combined with data from other sources to consider relationships between competition, education, R&D, innovation and growth (examples include Griffith et al., 2004; Aghion et al., 2005; Vandenbussche et al., 2006; Aghion and Griffith, 2005; Inklaar et al., 2008). More broadly, the database allows evaluation of various monetary, tax, innovation, competition and other industrial policies. And it might be used in studies of international specialisation and outsourcing (along the lines of, for example, Harrigan (1997) and Antràs and Helpman (2004) to name but a few) as well as studies of income inequality and wage setting (Koeniger and Leonardi, 2007). Frequently, due to a lack of data, many of these studies rely on industry panels. However this raises some serious issues of interpretation since many relationships are known to vary across industries. The EU KLEMS database with its rich data across countries allows for the first time estimations industry‐by‐industry so that the cross‐section panel dimension in the data is country, rather than industry. As with any resource, users need to understand the theoretical and practical underpinnings of the database in order to optimise the research benefit. The main purpose of this article is therefore to summarise the methodology employed in constructing the database and so guide researchers on appropriate uses. Naturally this requires that we also consider the practical limitations of the database and indicate areas for further improvement. In addition we illustrate the usefulness of the database by highlighting some interesting findings on trends and levels of productivity in Europe relative to the US. The remainder of this article is organised as follows. Section 1 considers theoretical and practical measurement issues and summarises data sources. Section 2 describes trends in productivity and input use in Europe and the US and considers relative productivity levels. Section 3 is essentially a user’s guide, summarising what is in the dataset, and outlines issues related to the use of the database in both growth accounting and econometric analysis. This Section ends with some health warnings. The EU KLEMS database is a dynamic resource that will be added to and revised over time. Section 4 considers future developments, those already planned and suggestions on ways forward in the longer term. 1. Theory and Measurement Issues This Section considers the measurement of output and productivity growth both in theory and practice. It begins with an outline of the growth accounting method which is the organising principle underlying the construction of the database. However it is important to emphasise that much of EU KLEMS is a resource independent of this method as many basic input series are provided as well. 1.1. Theoretical Background To assess the contribution of the various inputs to aggregate economic growth, we apply the growth accounting framework. This methodology has been theoretically motivated by the seminal contribution of Jorgenson and Griliches (1967) and put in a more general input–output framework by Jorgenson et al. (1987). It is based on production possibility frontiers where industry gross output is a function of capital, labour, intermediate inputs and technology, which is indexed by time, T. Each industry, indexed by j, can produce a set of products and purchases a number of distinct intermediate inputs, capital and labour inputs to produce its output. The production function is given by: (1) where Y is output, K is an index of capital service flows, L is an index of labour service flows and X is an index of intermediate inputs, either purchased from domestic industries or imported. Under the assumptions of competitive factor markets, full input utilisation and constant returns to scale, the growth of output can be expressed as the cost‐share weighted growth of inputs and technological change (AY), using the translog functional form common in such analyses:1 (2) where denotes the two‐period average share of input i in nominal output defined as follows: (3) and . Each element on the right‐hand side of (2) indicates the proportion of output growth accounted for by growth in intermediate inputs, capital services, labour services and technical change as measured by multifactor productivity (MFP), respectively. It is common to define aggregate input, say labour, as a Törnqvist quantity index of individual labour types as follows:2 (4) (5) (6) where Δ ln Ll,t indicates the growth of hours worked by labour type l and weights are given by the period average shares of each type in the value of labour compensation, and similarly for K and X. As we assume that marginal revenues are equal to marginal costs, the weighting procedure ensures that inputs which have a higher price also have a larger influence in the input index. So for example a doubling of hours worked by a high‐skilled worker gets a bigger weight than a doubling of hours worked by a low‐skilled worker. For many analyses it is useful to subdivide total intermediate inputs into three groups: energy, materials and services (E, M, S) such that: (7) With the period‐average share of energy products in total intermediate input costs in industry j at t and similarly for materials and services. Input volume growth of E, M and S are defined in terms of their components as (8) (9) (10) with the period‐average share of product x in total energy costs in industry j at t, and similarly for materials and services. To analyse the separate impact of ICT and non‐ICT capital, we divide capital input growth into two groups of assets: ICT (ICT) and non‐ICT (N) assets, such that: (11) with the period‐average share of ICT assets in total capital costs in industry j at t, and similarly for non‐ICT assets. Volume growth of ICT and non‐ICT are defined as: (12) (13) with the period‐average share of ICT‐asset k in total ICT‐capital costs in industry j and similarly for non‐ICT assets. In terms of labour inputs, it is useful to split the volume growth of labour input into the growth of hours worked and the changes in labour composition in terms of labour characteristics such as educational attainment, age or gender (see below). Let Hl,jt indicate the hours worked by labour type l in industry j at time t, and Hjt total hours worked by all types (summed over l) then we can decompose the change in labour input as follows: (14) with the period‐average share of labour type l in total labour costs in industry j. The first term on the right‐hand side indicates the change in labour composition and the second term indicates the change in total hours worked.3 It can easily be seen that if proportions of each labour type in the labour force change, this will have an impact on the growth of labour input beyond any change in total hours worked.4 Using the above formulas, the EU KLEMS database provides a full decomposition of growth in gross output into eight elements as follows: (15) The contribution of each intermediate and capital input is given by the product of its share in total costs and its growth rate. The contribution of labour input is split into hours worked and changes in the composition of hours worked, and any remaining output growth is picked up by the multi‐factor productivity term A. This term is also known as total factor productivity. An example of the application of this methodology is discussed in Section 3.1. Finally the EU KLEMS database also includes estimates of productivity levels. Comparing productivity levels across countries is in many ways analogous to comparisons over time. However, while one typically compares productivity in one year with productivity in the previous year, there is no such natural ordering of countries. Therefore the comparison should not depend on the country that is chosen as the base country. There are various index number methods that can be used to make multilateral comparisons. We use the method suggested by Caves et al. (1982). This index mirrors the Törnqvist index approach used in our growth accounting, but all countries are compared to an artificial ‘average’ country (AC), defined as the simple average of all N countries in the set. Gaps in multi‐factor productivity levels can be derived by subtracting the compensation‐weighted relative inputs from relative output as follows (industry and time subscript suppressed for convenience): (16) with s the input shares in gross output averaged between country c and the average country AC. A comparison between two countries, say Germany and the US, can be made indirectly: by first comparing each country with the average country and then comparing the differences in German and US levels relative to the average country. Inklaar and Timmer (2008) provide further discussion. 1.2. Practical Implementation The EU KLEMS database has largely been constructed on the basis of data from national statistical institutes (NSIs) and processed according to harmonised procedures. These procedures were developed to ensure international comparability of the basic data and to generate growth accounts in a consistent and uniform way. Cross‐country harmonisation of the basic country data has focused on a number of areas including a common industrial classification and the use of similar price concepts for inputs and outputs but also consistent definitions of various labour and capital types. Importantly, this database is rooted in statistics from the National Accounts and follows the concepts and conventions of the System of National Accounts (SNA) framework, and its European equivalent (ESA), in many respects. As a result, the basic statistics within EU KLEMS can be related to the national accounts statistics published by NSIs, although with adjustments that vary by group of variables: output and intermediate inputs, labour input and capital input. This will be discussed in more detail below. Nominal and price series for output and total intermediate inputs at the industry level are taken directly from the National Accounts. As these series are often short (as revisions are not always taken back in time) different vintages of the national accounts were bridged according to a common link‐methodology. In cases where industry detail was missing additional statistics from censuses and surveys were used to fill the gaps. Series on intermediate inputs are broken down into energy, materials and services based on supply‐and‐use tables using a standardised product classification. To ensure consistency with the national accounts series, proportions of energy, materials and services inputs were applied to the total intermediate input series from the National Accounts. Labour service input is based on series of hours worked and wages of various types of labour. These series are not part of the core set of National Accounts statistics put out by NSIs; typically only total hours worked and wages by industry are available from the National Accounts. For these series additional material has been collected from employment and labour force statistics. We cross‐classify hours worked by educational attainment, gender and age (to proxy for work experience) into 18 labour categories (respectively 3 × 2 × 3 types). For each country covered, a choice was made of the best statistical source for consistent wage and employment data at the industry level. In most cases this was the labour force survey (LFS), which in some cases was combined with earnings surveys when wages were not included in the LFS. In other instances, an establishment survey, or social‐security database was used (Timmer et al., 2007). Care has been taken to arrive at series which are time consistent, as most employment surveys are not designed to track developments over time and breaks in methodology or coverage frequently occur. Labour compensation of self‐employed is not registered in the National Accounts, which, as emphasised by Krueger (1999), leads to an understatement of labour’s share. We make an imputation by assuming that the compensation per hour of self‐employed is equal to the compensation per hour of employees. This is especially important for industries which have a large share of self‐employed workers, such as agriculture, trade, business and personal services. Also, we assume the same labour characteristics for self‐employed as for employees when information on the former is missing. These assumptions are made at the industry level. Capital input series by industry are generally not available from the National Accounts. At best, capital stocks are estimated for aggregate investment without distinguishing various asset types. In EU KLEMS, capital input is measured as capital services, rather than stocks. It has been measured as the weighted growth of stocks of eight assets as in (11)–(13).5 The weights are based on the rental price of each asset which consists of a nominal rate of return, depreciation and capital gains.6 The nominal rate of return is determined ex post as it is assumed that the total value of capital services for each industry equals capital compensation. Capital compensation is derived as gross value added minus labour compensation. This procedure yields an internal nominal rate of return that exhausts capital income and is consistent with constant returns to scale. The nominal rate of return is the same for all assets in an industry but is allowed to vary across industries. For each individual asset, stocks have been estimated on the basis of investment series using the perpetual inventory method (PIM) with geometric depreciation profiles. Depreciation rates differ by asset and industry but have been assumed identical across countries. Appendix B provides more details on capital service calculations. The basic investment series by industry and asset have been derived from capital flow matrices and benchmarked to the aggregate investment series from the National Accounts. Although the ESA provides a classification of capital assets, it is not always detailed enough to back out investment in information and communication equipment. Additional information has been collected to obtain investment series for these assets, or assumptions concerning hardware‐software ratios have been employed. When the deflator for computers did not contain an adjustment for quality change, a harmonised deflator based on the US deflator has been used as suggested by Schreyer (2002). The EU KLEMS database provides data at a detailed industry level but also provides higher level aggregates, such as the total economy, the market economy, total market services and total goods production for all variables. All aggregations of output and input volumes across industries use the Törnqvist quantity index. For example, the growth rate of total economy capital services is given as a weighted average of capital services growth across industries as follows: (17) with the period‐average share of industry j in total economy capital compensation. Similar industry aggregations are used for labour and value added. This is akin to the ‘direct aggregation across industries’ approach as developed by Jorgenson et al. (1987, ch. 2). It is based on the assumption that value‐added functions exist for each industry but does not impose cross‐industry restrictions on either value‐added or inputs. This approach allows us to trace the source of aggregate growth to the underlying industry sources explicitly.7 Aggregations are also made across countries. To do so use is made of industry‐specific Purchasing Power Parities (PPPs) which reflect differences in output price levels across countries at a detailed industry level. The PPPs are given for the benchmark year 1997. PPPs are also needed to adjust output and inputs for differences in relative price levels between countries in levels comparisons. This price adjustment is often done by means of GDP PPPs which reflect the average expenditure prices in one country relative to another and are widely available through the work of the OECD and Eurostat. However, it is well recognised that the use of GDP PPPs, which reflect expenditure prices of all goods and services in the economy, can be misleading when used to convert industry‐level output. The EU KLEMS database makes use of a new and comprehensive dataset of industry PPPs for 1997, in combination with a benchmark set of Supply and Use tables. PPPs for value added are constructed by double deflation of gross output and intermediate inputs within a consistent input‐output framework. In addition, relative price ratios for labour and capital input are developed – for details see Inklaar and Timmer (2008). Level estimates are discussed also in Section 3 below. 1.3. Comparison with OECD STAN Empirical implementation of the growth‐accounting methodology for European countries has been scarce. Despite the publication of an OECD handbook on productivity measurement (Schreyer, 2001), which is based on the growth accounting methodology, national statistical institutes (NSIs) have been slow in adopting this methodology and, to date, only one European NSI, i.e. Statistics Denmark, has published MFP‐measures on a regular basis.8 The OECD and the Groningen Growth and Development Centre maintain MFP series for OECD economies but not at the industry level.9 Because of the lack of useful statistics, various scholars have based their analysis on the OECD Structural Analysis database (STAN) and its predecessor the International Sectoral Database (ISDB). Interestingly, these databases were never designed for productivity analysis and as a result researchers had to apply additional methods and make ad hoc adjustments, for example in the calculation of capital stocks.10 This was done mostly for the purpose of one single study, which hindered validation and replication of results by others. The OECD STAN database provides industry‐level series on output, employment and aggregate investment for OECD member states. For a limited number of countries, capital stocks are given as well. It is almost exclusively based on data published in the latest vintage of the National Accounts of each country. In addition, EU KLEMS uses additional sources such as earlier vintages of the National Accounts, industry surveys, labour force surveys and capital formation surveys. While essentially complementary in many respects, the EU KLEMS database goes beyond STAN by providing:11 Long historical time series Breakdown of industries to a common level of industry detail A breakdown of intermediate inputs into energy, materials and services A breakdown of hours worked by type of worker A breakdown of investment into various asset types Calculation of capital stocks and services using a harmonised methodology Estimates of multi‐factor productivity (MFP) based on growth accounting Productivity measures based on aggregate concepts of inputs as in STAN can be seriously biased. The EU KLEMS database shows that there is a general shift towards more skilled and more experienced workers in the labour force. As such, labour services grow faster than suggested by a crude measure of hours worked, unadjusted for changes in labour composition. Similarly, especially in the past decade, the importance of short‐lived ICT assets relative to non‐ICT assets has increased. Consequently, capital service input growth rates are higher than capital stock growth rates as ICT assets deliver more services per unit of capital stock. Not accounting for this shift in the composition of capital biases input measures downwards, and consequently MFP measures upwards. In general, one can say that the differences in labour productivity growth rates between STAN and EU KLEMS are relatively minor, underlying the complementary nature of the two databases for basic data. This is useful since STAN provides data for some OECD countries not yet in EU KLEMS plus additional information on trade flows. Differences in output and employment growth are generally negligible, although differences in estimates of hours worked might be bigger (see also Section 3.2).12 The greatest difference can be found in capital stock estimates. STAN provides aggregate stock estimates for those countries which publish these in their National Accounts. The internationally harmonised approach to capital measurement in EU KLEMS often differs from the practice used in the National Accounts of a particular country. In addition, EU KLEMS provides estimates of the changes in the composition of the capital stock and in the labour force, which cannot be derived with the STAN database. In Table 1 we illustrate the differences between input and productivity growth rates based on EU KLEMS and those based on STAN for a particular sector (distributive trade) and for those countries for which STAN has capital stock estimates (Germany, Italy, Spain and the US). Indeed, differences in labour productivity growth rates (value added per hour worked) are generally small. But differences in the capital stock estimates can be sizeable due to the different methodologies employed. The adjustments for capital and labour composition in EU KLEMS provide additional information not available in STAN.13 Table 1 
EU KLEMS and STAN Estimates of Productivity Growth, Distributive Trades, 1995–2004 (annual average volume growth rates, in %) . Germany . Italy . Spain . United States . . STAN . EUK . STAN . EUK . STAN . EUK . STAN . EUK . Labour productivity growth of which contribution by 1.9 1.9 0.9 0.9 0.2 0.6 4.6 4.4 capital stock growth 0.7 0.5 2.9 1.3 1.2 0.8 1.2 0.9 changes in capital composition – 0.1 – 0.2 – 0.3 – 0.2 changes in labour composition – 0.0 – 0.5 – 0.4 – 0.3 MFP growth (value added based) 1.2 1.3 −2.0 −1.1 −1.0 −0.9 3.4 3.1 . Germany . Italy . Spain . United States . . STAN . EUK . STAN . EUK . STAN . EUK . STAN . EUK . Labour productivity growth of which contribution by 1.9 1.9 0.9 0.9 0.2 0.6 4.6 4.4 capital stock growth 0.7 0.5 2.9 1.3 1.2 0.8 1.2 0.9 changes in capital composition – 0.1 – 0.2 – 0.3 – 0.2 changes in labour composition – 0.0 – 0.5 – 0.4 – 0.3 MFP growth (value added based) 1.2 1.3 −2.0 −1.1 −1.0 −0.9 3.4 3.1 Notes. Contributions of factor inputs are calculated as the share of input times the growth rate, in percentage points. Source. STAN estimates based on OECD, STAN database, release March 2008. EUK from EU KLEMS database, release March 2008, see Timmer et al. (2007). Figures might not add due to rounding. Open in new tab Table 1 
EU KLEMS and STAN Estimates of Productivity Growth, Distributive Trades, 1995–2004 (annual average volume growth rates, in %) . Germany . Italy . Spain . United States . . STAN . EUK . STAN . EUK . STAN . EUK . STAN . EUK . Labour productivity growth of which contribution by 1.9 1.9 0.9 0.9 0.2 0.6 4.6 4.4 capital stock growth 0.7 0.5 2.9 1.3 1.2 0.8 1.2 0.9 changes in capital composition – 0.1 – 0.2 – 0.3 – 0.2 changes in labour composition – 0.0 – 0.5 – 0.4 – 0.3 MFP growth (value added based) 1.2 1.3 −2.0 −1.1 −1.0 −0.9 3.4 3.1 . Germany . Italy . Spain . United States . . STAN . EUK . STAN . EUK . STAN . EUK . STAN . EUK . Labour productivity growth of which contribution by 1.9 1.9 0.9 0.9 0.2 0.6 4.6 4.4 capital stock growth 0.7 0.5 2.9 1.3 1.2 0.8 1.2 0.9 changes in capital composition – 0.1 – 0.2 – 0.3 – 0.2 changes in labour composition – 0.0 – 0.5 – 0.4 – 0.3 MFP growth (value added based) 1.2 1.3 −2.0 −1.1 −1.0 −0.9 3.4 3.1 Notes. Contributions of factor inputs are calculated as the share of input times the growth rate, in percentage points. Source. STAN estimates based on OECD, STAN database, release March 2008. EUK from EU KLEMS database, release March 2008, see Timmer et al. (2007). Figures might not add due to rounding. Open in new tab In addition to improving MFP estimates, detailed data on various input types are useful in their own right. They can for example be used in studies of energy efficiency, services outsourcing, skill formation and skill premia and investment in ICT assets. 2. Growth and Productivity in Europe, Japan and the US The data provided in the EU KLEMS database can be used for the study of a variety of issues. In this Section we highlight some interesting findings in an analysis of productivity growth and level differences across Europe, Japan and the US focusing on the market economy.14 Since the mid‐1990s, labour productivity growth in most European countries has significantly slowed compared to earlier decades. In contrast, labour productivity growth in the US accelerated, so that a new productivity gap has opened up. On the basis of EU KLEMS data, van Ark et al. (2008) argue that this is attributable to the slower emergence of the knowledge economy in Europe over the period 1995–2004. When looking at the growth accounts from the perspective of the emerging knowledge economy, they focus on the summed contributions of three factors: direct impacts from investments in information and communication technology, changes in labour composition mostly driven by greater demand for skilled workers and multifactor productivity growth, which includes the impact of innovation and intangible investments such as organisational changes related to the use of ICT. Table 2 reproduces the calculations underlying their findings, updated to 2005 and including some additional European Union countries and Japan. The first column of Table 2 shows the growth rate of output in the market economy of 16 European countries, the EU total, Japan and the US between 1995 and 2005. The second and third columns divide that growth into changes in hours worked and changes in output per hour – or labour productivity. Columns 4–7 divide up the growth in labour productivity into four factors: changes in labour composition, investments in ICT, other types of investments and multifactor productivity. The final column in Table 2 shows the ‘knowledge economy’ contributions as defined above and highlights the discrepancy between growth in Europe and the US noted by van Ark et al. (2008). Table 2 
Gross Value Added Growth and Contributions, 1995–2005, Market Economy . Growth rate of value added . Value contribution added from . Labour productivity contributions from . Labour productivity contribution of the knowledge economy . . Hours worked . Labour productivity . Labour composition . ICT capital per hour . Non‐ICT capital per hour . Multi factor productivity . . 1 = 2 + 3 . 2 . 3 = 4 + 5 + 6 + 7 . 4 . 5 . 6 . 7 . 4 + 5 + 7 . Austria 2.5 0.5 2.0 0.2 0.6 0.2 1.1 1.9 Belgium 2.3 0.7 1.7 0.2 1.0 0.6 0.1 1.3 Czech Republic 1.5 −0.7 2.1 0.1 1.1 1.7 −1.1 0.1 Denmark 2.2 0.7 1.6 0.2 1.1 0.4 0.0 1.3 Finland 4.3 1.1 3.2 0.1 0.7 0.2 2.5 3.2 France 2.4 0.4 2.1 0.4 0.4 0.5 0.8 1.6 Germany 1.0 −0.6 1.5 0.1 0.5 0.5 0.4 0.9 Hungary 4.3 1.6 2.7 0.1 0.0 0.1 2.7 2.9 Ireland 9.7 4.5 5.2 0.2 0.8 3.7 2.0 3.0 Italy 1.2 0.9 0.3 0.2 0.3 0.8 −0.7 −0.2 Netherlands 2.8 0.7 2.1 0.4 0.6 0.3 1.0 2.0 Portugal 3.9 1.0 3.0 −0.4 0.9 1.9 0.8 1.3 Slovenia 4.3 −1.0 5.3 0.3 0.5 2.1 1.4 2.2 Spain 3.6 3.2 0.4 0.4 0.5 1.4 −0.8 0.0 Sweden 4.4 1.2 3.2 0.1 0.8 1.5 1.0 2.0 UK 3.2 0.6 2.6 0.5 0.9 0.6 0.9 2.3 EU 2.2 0.7 1.5 0.2 0.6 0.6 0.4 1.2 Japan 1.0 −1.7 2.6 0.4 0.5 0.6 0.5 1.3 US 3.6 0.7 2.9 0.3 1.0 0.5 1.3 2.7 . Growth rate of value added . Value contribution added from . Labour productivity contributions from . Labour productivity contribution of the knowledge economy . . Hours worked . Labour productivity . Labour composition . ICT capital per hour . Non‐ICT capital per hour . Multi factor productivity . . 1 = 2 + 3 . 2 . 3 = 4 + 5 + 6 + 7 . 4 . 5 . 6 . 7 . 4 + 5 + 7 . Austria 2.5 0.5 2.0 0.2 0.6 0.2 1.1 1.9 Belgium 2.3 0.7 1.7 0.2 1.0 0.6 0.1 1.3 Czech Republic 1.5 −0.7 2.1 0.1 1.1 1.7 −1.1 0.1 Denmark 2.2 0.7 1.6 0.2 1.1 0.4 0.0 1.3 Finland 4.3 1.1 3.2 0.1 0.7 0.2 2.5 3.2 France 2.4 0.4 2.1 0.4 0.4 0.5 0.8 1.6 Germany 1.0 −0.6 1.5 0.1 0.5 0.5 0.4 0.9 Hungary 4.3 1.6 2.7 0.1 0.0 0.1 2.7 2.9 Ireland 9.7 4.5 5.2 0.2 0.8 3.7 2.0 3.0 Italy 1.2 0.9 0.3 0.2 0.3 0.8 −0.7 −0.2 Netherlands 2.8 0.7 2.1 0.4 0.6 0.3 1.0 2.0 Portugal 3.9 1.0 3.0 −0.4 0.9 1.9 0.8 1.3 Slovenia 4.3 −1.0 5.3 0.3 0.5 2.1 1.4 2.2 Spain 3.6 3.2 0.4 0.4 0.5 1.4 −0.8 0.0 Sweden 4.4 1.2 3.2 0.1 0.8 1.5 1.0 2.0 UK 3.2 0.6 2.6 0.5 0.9 0.6 0.9 2.3 EU 2.2 0.7 1.5 0.2 0.6 0.6 0.4 1.2 Japan 1.0 −1.7 2.6 0.4 0.5 0.6 0.5 1.3 US 3.6 0.7 2.9 0.3 1.0 0.5 1.3 2.7 Notes. Growth rates are annual average volume growth rates. Contributions in percentage points. Figures might not add due to rounding. European Union is (weighted) average of the old EU‐15 countries shown in the Table. Authors’ calculations based on EUKLEMS database, release March 2008, see Timmer et al. (2007). Open in new tab Table 2 
Gross Value Added Growth and Contributions, 1995–2005, Market Economy . Growth rate of value added . Value contribution added from . Labour productivity contributions from . Labour productivity contribution of the knowledge economy . . Hours worked . Labour productivity . Labour composition . ICT capital per hour . Non‐ICT capital per hour . Multi factor productivity . . 1 = 2 + 3 . 2 . 3 = 4 + 5 + 6 + 7 . 4 . 5 . 6 . 7 . 4 + 5 + 7 . Austria 2.5 0.5 2.0 0.2 0.6 0.2 1.1 1.9 Belgium 2.3 0.7 1.7 0.2 1.0 0.6 0.1 1.3 Czech Republic 1.5 −0.7 2.1 0.1 1.1 1.7 −1.1 0.1 Denmark 2.2 0.7 1.6 0.2 1.1 0.4 0.0 1.3 Finland 4.3 1.1 3.2 0.1 0.7 0.2 2.5 3.2 France 2.4 0.4 2.1 0.4 0.4 0.5 0.8 1.6 Germany 1.0 −0.6 1.5 0.1 0.5 0.5 0.4 0.9 Hungary 4.3 1.6 2.7 0.1 0.0 0.1 2.7 2.9 Ireland 9.7 4.5 5.2 0.2 0.8 3.7 2.0 3.0 Italy 1.2 0.9 0.3 0.2 0.3 0.8 −0.7 −0.2 Netherlands 2.8 0.7 2.1 0.4 0.6 0.3 1.0 2.0 Portugal 3.9 1.0 3.0 −0.4 0.9 1.9 0.8 1.3 Slovenia 4.3 −1.0 5.3 0.3 0.5 2.1 1.4 2.2 Spain 3.6 3.2 0.4 0.4 0.5 1.4 −0.8 0.0 Sweden 4.4 1.2 3.2 0.1 0.8 1.5 1.0 2.0 UK 3.2 0.6 2.6 0.5 0.9 0.6 0.9 2.3 EU 2.2 0.7 1.5 0.2 0.6 0.6 0.4 1.2 Japan 1.0 −1.7 2.6 0.4 0.5 0.6 0.5 1.3 US 3.6 0.7 2.9 0.3 1.0 0.5 1.3 2.7 . Growth rate of value added . Value contribution added from . Labour productivity contributions from . Labour productivity contribution of the knowledge economy . . Hours worked . Labour productivity . Labour composition . ICT capital per hour . Non‐ICT capital per hour . Multi factor productivity . . 1 = 2 + 3 . 2 . 3 = 4 + 5 + 6 + 7 . 4 . 5 . 6 . 7 . 4 + 5 + 7 . Austria 2.5 0.5 2.0 0.2 0.6 0.2 1.1 1.9 Belgium 2.3 0.7 1.7 0.2 1.0 0.6 0.1 1.3 Czech Republic 1.5 −0.7 2.1 0.1 1.1 1.7 −1.1 0.1 Denmark 2.2 0.7 1.6 0.2 1.1 0.4 0.0 1.3 Finland 4.3 1.1 3.2 0.1 0.7 0.2 2.5 3.2 France 2.4 0.4 2.1 0.4 0.4 0.5 0.8 1.6 Germany 1.0 −0.6 1.5 0.1 0.5 0.5 0.4 0.9 Hungary 4.3 1.6 2.7 0.1 0.0 0.1 2.7 2.9 Ireland 9.7 4.5 5.2 0.2 0.8 3.7 2.0 3.0 Italy 1.2 0.9 0.3 0.2 0.3 0.8 −0.7 −0.2 Netherlands 2.8 0.7 2.1 0.4 0.6 0.3 1.0 2.0 Portugal 3.9 1.0 3.0 −0.4 0.9 1.9 0.8 1.3 Slovenia 4.3 −1.0 5.3 0.3 0.5 2.1 1.4 2.2 Spain 3.6 3.2 0.4 0.4 0.5 1.4 −0.8 0.0 Sweden 4.4 1.2 3.2 0.1 0.8 1.5 1.0 2.0 UK 3.2 0.6 2.6 0.5 0.9 0.6 0.9 2.3 EU 2.2 0.7 1.5 0.2 0.6 0.6 0.4 1.2 Japan 1.0 −1.7 2.6 0.4 0.5 0.6 0.5 1.3 US 3.6 0.7 2.9 0.3 1.0 0.5 1.3 2.7 Notes. Growth rates are annual average volume growth rates. Contributions in percentage points. Figures might not add due to rounding. European Union is (weighted) average of the old EU‐15 countries shown in the Table. Authors’ calculations based on EUKLEMS database, release March 2008, see Timmer et al. (2007). Open in new tab One key observation to be drawn from this Table is that the main difference in labour productivity growth between individual European economies and the US is to be found in multifactor productivity, not in differences in the intensity of the production factors. For the EU as a whole, the labour productivity growth gap is 1.4 percentage points of which 0.9 can be explained by lower MFP growth in the EU. Lagging ICT investments explains only 0.4 percentage points of the growth‐gap, while growth differences in non‐ICT capital per hour and in labour services per hour are negligible. The same story holds for most European countries. Similarly, the sources of labour productivity growth in Japan are very similar to those in the EU, except that employment growth in Japan was negative. Within Europe, there is also a divergence of labour productivity growth rates. This is not so much due to differences in input growth rates. In particular, in all countries there have been positive contributions of changes in labour composition and from investment in ICT. Instead differences in multifactor productivity seem to have driven the divergence across Europe. In Belgium, Denmark and Germany, MFP growth is below 0.5% per year and in Italy, Czech Republic and Spain it is even negative. In contrast, MFP growth in Finland, Hungary and Ireland is around 3%. Note in the case of these countries, MFP also captures some element of conventional catch‐up growth as they lagged relatively far behind the rest in 1995. When analysing cross‐country patterns, growth accounts provide only a partial analysis. It is now widely accepted that understanding the pattern of cross‐country growth and productivity requires estimates of relative levels. For example, studies on the impact of differences in education, research and development or market regulation across countries rely heavily on level measures of MFP which indicate the distance to the technology frontier.15Inklaar and Timmer (2008) provide new productivity level estimates which complement the EU KLEMS growth accounts. Table 3 shows relative MFP levels for the market economy and a division of this into four sectors: ICT goods and services production, manufacturing, other goods production and market services. These estimates suggest that in 2005 the US had a significant MFP lead over all European countries. Relative MFP in the EU is only 79% of the US. This gap is mainly in market services, ICT production and other goods production (around 80%), rather than in manufacturing (90%). In fact, in manufacturing MFP levels in 2005 in some countries are close to the US or even above, including Belgium, Finland, France, Germany, Ireland and Netherlands. In contrast, in market services significant productivity gaps with the US exist in almost all European countries, in particular Italy. In Japan large gaps are to be found in both manufacturing and market services. Table 3 
Relative Levels of Multifactor Productivity, 2005 (US = 1) . Market economy . ICT production . Manufacturing . Other goods production . Market services . Austria 0.67 0.48 0.75 0.97 0.63 Belgium 0.86 0.50 1.09 1.03 0.80 Czech Republic 0.44 0.18 0.54 0.57 0.42 Denmark 0.84 0.51 0.70 0.84 1.01 Finland 0.90 1.24 1.07 0.86 0.78 France 0.80 0.93 0.97 0.75 0.80 Germany 0.85 0.80 0.97 0.75 0.84 Hungary 0.47 0.45 0.51 0.54 0.45 Ireland 0.93 0.52 1.54 0.56 0.87 Italy 0.65 0.61 0.82 0.78 0.60 Netherlands 0.92 0.53 0.96 0.81 1.03 Portugal 0.61 0.43 0.54 0.56 0.78 Slovenia 0.46 0.33 0.66 0.27 0.51 Spain 0.71 0.41 0.65 0.90 0.77 Sweden 0.84 2.58 0.85 0.95 0.76 UK 0.77 0.90 0.91 0.78 0.77 EU 0.79 0.77 0.90 0.79 0.79 Japan 0.47 0.66 0.60 0.33 0.46 US 1.00 1.00 1.00 1.00 1.00 . Market economy . ICT production . Manufacturing . Other goods production . Market services . Austria 0.67 0.48 0.75 0.97 0.63 Belgium 0.86 0.50 1.09 1.03 0.80 Czech Republic 0.44 0.18 0.54 0.57 0.42 Denmark 0.84 0.51 0.70 0.84 1.01 Finland 0.90 1.24 1.07 0.86 0.78 France 0.80 0.93 0.97 0.75 0.80 Germany 0.85 0.80 0.97 0.75 0.84 Hungary 0.47 0.45 0.51 0.54 0.45 Ireland 0.93 0.52 1.54 0.56 0.87 Italy 0.65 0.61 0.82 0.78 0.60 Netherlands 0.92 0.53 0.96 0.81 1.03 Portugal 0.61 0.43 0.54 0.56 0.78 Slovenia 0.46 0.33 0.66 0.27 0.51 Spain 0.71 0.41 0.65 0.90 0.77 Sweden 0.84 2.58 0.85 0.95 0.76 UK 0.77 0.90 0.91 0.78 0.77 EU 0.79 0.77 0.90 0.79 0.79 Japan 0.47 0.66 0.60 0.33 0.46 US 1.00 1.00 1.00 1.00 1.00 Notes. For industry classification, see Appendix Table 2. EU refers to (weighted) average of the old EU‐15 countries shown in the Table. MFP is based on value added. Based on GGDC Productivity Level database, see Inklaar and Timmer (2008). Open in new tab Table 3 
Relative Levels of Multifactor Productivity, 2005 (US = 1) . Market economy . ICT production . Manufacturing . Other goods production . Market services . Austria 0.67 0.48 0.75 0.97 0.63 Belgium 0.86 0.50 1.09 1.03 0.80 Czech Republic 0.44 0.18 0.54 0.57 0.42 Denmark 0.84 0.51 0.70 0.84 1.01 Finland 0.90 1.24 1.07 0.86 0.78 France 0.80 0.93 0.97 0.75 0.80 Germany 0.85 0.80 0.97 0.75 0.84 Hungary 0.47 0.45 0.51 0.54 0.45 Ireland 0.93 0.52 1.54 0.56 0.87 Italy 0.65 0.61 0.82 0.78 0.60 Netherlands 0.92 0.53 0.96 0.81 1.03 Portugal 0.61 0.43 0.54 0.56 0.78 Slovenia 0.46 0.33 0.66 0.27 0.51 Spain 0.71 0.41 0.65 0.90 0.77 Sweden 0.84 2.58 0.85 0.95 0.76 UK 0.77 0.90 0.91 0.78 0.77 EU 0.79 0.77 0.90 0.79 0.79 Japan 0.47 0.66 0.60 0.33 0.46 US 1.00 1.00 1.00 1.00 1.00 . Market economy . ICT production . Manufacturing . Other goods production . Market services . Austria 0.67 0.48 0.75 0.97 0.63 Belgium 0.86 0.50 1.09 1.03 0.80 Czech Republic 0.44 0.18 0.54 0.57 0.42 Denmark 0.84 0.51 0.70 0.84 1.01 Finland 0.90 1.24 1.07 0.86 0.78 France 0.80 0.93 0.97 0.75 0.80 Germany 0.85 0.80 0.97 0.75 0.84 Hungary 0.47 0.45 0.51 0.54 0.45 Ireland 0.93 0.52 1.54 0.56 0.87 Italy 0.65 0.61 0.82 0.78 0.60 Netherlands 0.92 0.53 0.96 0.81 1.03 Portugal 0.61 0.43 0.54 0.56 0.78 Slovenia 0.46 0.33 0.66 0.27 0.51 Spain 0.71 0.41 0.65 0.90 0.77 Sweden 0.84 2.58 0.85 0.95 0.76 UK 0.77 0.90 0.91 0.78 0.77 EU 0.79 0.77 0.90 0.79 0.79 Japan 0.47 0.66 0.60 0.33 0.46 US 1.00 1.00 1.00 1.00 1.00 Notes. For industry classification, see Appendix Table 2. EU refers to (weighted) average of the old EU‐15 countries shown in the Table. MFP is based on value added. Based on GGDC Productivity Level database, see Inklaar and Timmer (2008). Open in new tab Figure 1 traces the developments of relative input and productivity levels in the European market economy over the period 1980–2005 by extrapolating the level estimates with growth rates from the EU KLEMS database. Up to the mid‐1990s labour productivity in Europe caught up with the US, continuing the long‐term trend since the Second World War. This narrowing of the labour productivity gap was mainly due to higher investment in Europe than in the US. Capital intensity levels increased from about 82% of the US level in 1980 to 95% in 1995, while relative MFP levels remained more or less constant. This period of rapid capital intensification was primarily related to the high wage/rental ratios in Europe compared to the US (van Ark et al., 2008). This trend reversed in the mid‐1990s as investment in ICT in the US soared and relative cost of labour in the EU countries declined due to policies to raise the employment rate. At the same time the MFP gap between the EU and the US widened which, as implied by Table 3, is concentrated in market services. Fig. 1. Open in new tabDownload slide Productivity and Capital Intensity Levels in Europe, Market Economy, US = 100
Notes. EU refers to EU‐15. Authors’ calculations based on Inklaar and Timmer (2008). Fig. 1. Open in new tabDownload slide Productivity and Capital Intensity Levels in Europe, Market Economy, US = 100
Notes. EU refers to EU‐15. Authors’ calculations based on Inklaar and Timmer (2008). This short description of productivity in the EU compared to the US illustrates some of the uses of the database. To date, EU KLEMS data have also been used to study the role of general purpose technology on economic growth (Jalava and Pohjola, 2008), the issue of embodiment of energy‐saving technologies (Kratena, 2007), the impact of market regulation on productivity growth in services (Inklaar et al., 2008), the impact of structural change on productivity (Maudos et al., 2008) and studies of European competitiveness (European Commission, 2007). 3. A Short Description of the EU KLEMS Database This Section is a brief guide to the use of the EU KLEMS database. It begins with a description of the data series and then considers some measurement issues which might impact on researchers’ use of the data. It highlights the main issues and looks at some problems that might arise when using the data in these contexts – readers should refer to the methodology and sources document (Timmer et al., 2007) for details. 3.1. Country/Industry/Variable Coverage The second public release of the EU KLEMS database in March 2008 covers 25 EU countries, as well as Australia, Japan and the US. In general, data for 1970–2005 are available for the ‘old’ EU‐15 countries, while series from 1995 onwards are available for the new EU member states which joined the EU on 1 May 2004. Appendix Table 1 provides an overview of all the series included in the EU KLEMS database. The variables covered can be split into three main groups: (1) labour productivity variables; (2) growth accounting variables and (3) additional variables. The labour productivity series contain all the data needed to construct labour productivity (output per hour worked) and unit labour costs. These series include nominal, volume and price series of output, and volumes and prices of employment. Most series are part of the present European System of National Accounts (ESA 1995) and can be found in the National Accounts of all individual countries, at least for the most recent period. The main adjustments to these series were to fill gaps in industry detail and to link series over time, in particular in those cases where revisions were not taken back to 1970 by the NSIs. The variables in the growth accounting series are of an analytical nature and cannot be directly derived from published National Accounts data without additional assumptions. These include series of capital services, of labour services, and of multi factor productivity. The construction of these series was based on the theoretical model of production, requiring additional assumptions as spelled out in Section 1.1. Finally, additional series are given which have been used in generating the growth accounts and are informative by themselves. These include, for example, various measures of the relative importance of ICT capital and non‐ICT capital, and of the various labour types within the EU KLEMS classification.16 At the lowest level of aggregation, data were collected for 71 industries. The industries are classified according to the European NACE revision 1 classification. But the level of detail varies across countries, industries and variables due to data limitations. In order to ensure a minimal level of industry detail for which comparisons can be made across all countries, so‐called ‘minimum lists’ of industries have been used. All national datasets have been constructed in such a way that these minimum lists are met but often more detailed data are available. For output and employment, the minimum number covered is 62 industries for the period from 1995 and 48 industries pre‐1995. Growth accounts are available for 31 industries as given in Appendix Table 2. This list also includes higher level industry aggregates provided in the EU KLEMS database. Growth accounts are included for 14 EU countries (excluding Greece) and for the Czech Republic, Hungary and Slovenia from the New Member States, Australia, Japan and the US. For all other countries only labour productivity and its underlying data series are included. Appendix Table 3 provides more details on the period‐coverage for each variable. Finally, data are also provided for four institutional country groupings: EU‐25, EU‐15, EU‐10 and Euro zone.17 It is useful at this stage to present an example of the growth accounting method. In Table 4 we show output growth decomposition for one industry: Business services excluding real estate (ISIC 71 to 74) for the period 1995–2005. The decomposition is shown for a number of large European countries, Japan and the US by way of example. In the lowest panel we provide the share of each input in the value of output, which under the growth accounting assumptions, equals total costs, averaged over 1995 and 2005. In the middle panel one can find the growth rate of each input and in the upper panel the contribution of each input to gross output growth (which is derived by multiplying its share by its growth rate). It can be seen that output in this industry showed significant growth during the past decade in all countries, and that its sources of growth were highly varied. Labour is the most important input taking up about 40% to 50% of total costs in this industry. Labour’s contribution to output growth is high as hours worked increased rapidly in all countries and there was a concomitant shift of hours towards higher skilled workers as indicated by the positive and high contribution of the change in labour composition. Also the use of intermediate inputs grew rapidly for all types of intermediates but, due to its large share in total costs, growth in purchased services contributed most to output growth. As to be expected, the fastest growing input in all countries was ICT capital with annual average growth rates of 9% or more and, although its share in costs is still modest, its contribution to output was sometimes higher than for the traditional non‐ICT assets. Finally, MFP growth appeared to be small and often negative. It indicates that the overall efficiency with which intermediate, capital and labour inputs have been used was not increasing (see below for an interpretation of MFP figures).18 Table 4 
Decomposition of Gross Output Growth in Business Services, 1995–2005 . Gross output . Intermediate inputs . Labour input . Capital input . MFP . Total . Energy . Materials . Services . Total . Hours worked . Labour composition . Total . ICT . Non‐ICT . Contribution to gross output volume grcwth Spain 6.7 3.7 0.1 1.1 2.4 2.6 2.2 0.4 1.2 0.3 0.8 −0.8 France 3.9 2.2 0.0 0.4 1.9 1.5 1.3 0.2 0.8 0.4 0.5 −0.6 Germany 2.8 1.4 0.0 0.1 1.2 1.4 1.4 −0.1 2.7 1.1 1.5 −2.7 Italy 4.1 2.1 0.1 0.3 1.7 2.3 2.2 0.1 0.5 0.2 0.3 −0.9 Japan 4.1 1.3 0.0 0.1 1.3 1.5 1.2 0.3 1.0 0.7 0.3 0.3 UK 7.2 3.1 0.1 0.1 2.9 2.0 1.7 0.3 1.6 0.9 0.7 0.5 US 6.0 2.9 0.1 0.9 2.0 1.6 1.1 0.2 2.1 1.4 0.6 −0.5 Volume growth Spain 6.7 8.2 9.7 7.8 8.3 6.1 5.2 0.9 10.0 12.1 9.0 −0.8 France 3.9 5.0 1.4 5.1 5.1 3.4 3.0 0.4 6.7 9.7 5.4 −0.6 Germany 2.8 4.1 2.6 3.1 4.3 4.0 4.2 −0.2 8.5 17.4 6.0 −2.7 Italy 4.1 5.0 2.5 3.4 5.7 6.2 5.8 0.4 2.6 16.2 1.6 −0.9 Japan 4.1 3.0 3.4 1.1 4.4 3.4 2.7 0.8 7.3 11.0 4.1 0.3 UK 7.2 7.6 5.3 3.4 8.2 4.5 3.8 0.7 11.7 19.0 7.3 0.5 US 6.0 8.5 5.9 13.3 7.3 3.1 2.1 0.5 14.5 22.3 7.8 −0.5 Average share in nominal gross output of 1995 and 2005 Spain 100.0 45.3 1.5 14.4 29.4 42.6 12.1 2.7 9.4 France 100.0 45.2 1.2 7.2 36.7 42.5 12.3 3.9 8.4 Germany 100.0 33.5 0.7 4.0 28.9 34.6 32.0 6.4 25.6 Italy 100.0 41.9 2.4 9.4 30.1 37.7 20.4 1.5 18.9 Japan 100.0 42.7 0.8 12.6 29.2 44.0 13.3 6.2 7.1 UK 100.0 40.6 1.1 4.2 35.3 45.7 13.7 4.6 9.1 US 100.0 34.4 0.9 6.5 27.0 51.4 14.3 6.2 8.1 . Gross output . Intermediate inputs . Labour input . Capital input . MFP . Total . Energy . Materials . Services . Total . Hours worked . Labour composition . Total . ICT . Non‐ICT . Contribution to gross output volume grcwth Spain 6.7 3.7 0.1 1.1 2.4 2.6 2.2 0.4 1.2 0.3 0.8 −0.8 France 3.9 2.2 0.0 0.4 1.9 1.5 1.3 0.2 0.8 0.4 0.5 −0.6 Germany 2.8 1.4 0.0 0.1 1.2 1.4 1.4 −0.1 2.7 1.1 1.5 −2.7 Italy 4.1 2.1 0.1 0.3 1.7 2.3 2.2 0.1 0.5 0.2 0.3 −0.9 Japan 4.1 1.3 0.0 0.1 1.3 1.5 1.2 0.3 1.0 0.7 0.3 0.3 UK 7.2 3.1 0.1 0.1 2.9 2.0 1.7 0.3 1.6 0.9 0.7 0.5 US 6.0 2.9 0.1 0.9 2.0 1.6 1.1 0.2 2.1 1.4 0.6 −0.5 Volume growth Spain 6.7 8.2 9.7 7.8 8.3 6.1 5.2 0.9 10.0 12.1 9.0 −0.8 France 3.9 5.0 1.4 5.1 5.1 3.4 3.0 0.4 6.7 9.7 5.4 −0.6 Germany 2.8 4.1 2.6 3.1 4.3 4.0 4.2 −0.2 8.5 17.4 6.0 −2.7 Italy 4.1 5.0 2.5 3.4 5.7 6.2 5.8 0.4 2.6 16.2 1.6 −0.9 Japan 4.1 3.0 3.4 1.1 4.4 3.4 2.7 0.8 7.3 11.0 4.1 0.3 UK 7.2 7.6 5.3 3.4 8.2 4.5 3.8 0.7 11.7 19.0 7.3 0.5 US 6.0 8.5 5.9 13.3 7.3 3.1 2.1 0.5 14.5 22.3 7.8 −0.5 Average share in nominal gross output of 1995 and 2005 Spain 100.0 45.3 1.5 14.4 29.4 42.6 12.1 2.7 9.4 France 100.0 45.2 1.2 7.2 36.7 42.5 12.3 3.9 8.4 Germany 100.0 33.5 0.7 4.0 28.9 34.6 32.0 6.4 25.6 Italy 100.0 41.9 2.4 9.4 30.1 37.7 20.4 1.5 18.9 Japan 100.0 42.7 0.8 12.6 29.2 44.0 13.3 6.2 7.1 UK 100.0 40.6 1.1 4.2 35.3 45.7 13.7 4.6 9.1 US 100.0 34.4 0.9 6.5 27.0 51.4 14.3 6.2 8.1 Notes. Business services refer to NACE industries 71 to 74, thus excluding real estate. Contribution of inputs calculated as the share of input times the volume growth rate. Figures might not add due to rounding. Calculations based on EUKLEMS database, release March 2008, see Timmer et al. (2007). Open in new tab Table 4 
Decomposition of Gross Output Growth in Business Services, 1995–2005 . Gross output . Intermediate inputs . Labour input . Capital input . MFP . Total . Energy . Materials . Services . Total . Hours worked . Labour composition . Total . ICT . Non‐ICT . Contribution to gross output volume grcwth Spain 6.7 3.7 0.1 1.1 2.4 2.6 2.2 0.4 1.2 0.3 0.8 −0.8 France 3.9 2.2 0.0 0.4 1.9 1.5 1.3 0.2 0.8 0.4 0.5 −0.6 Germany 2.8 1.4 0.0 0.1 1.2 1.4 1.4 −0.1 2.7 1.1 1.5 −2.7 Italy 4.1 2.1 0.1 0.3 1.7 2.3 2.2 0.1 0.5 0.2 0.3 −0.9 Japan 4.1 1.3 0.0 0.1 1.3 1.5 1.2 0.3 1.0 0.7 0.3 0.3 UK 7.2 3.1 0.1 0.1 2.9 2.0 1.7 0.3 1.6 0.9 0.7 0.5 US 6.0 2.9 0.1 0.9 2.0 1.6 1.1 0.2 2.1 1.4 0.6 −0.5 Volume growth Spain 6.7 8.2 9.7 7.8 8.3 6.1 5.2 0.9 10.0 12.1 9.0 −0.8 France 3.9 5.0 1.4 5.1 5.1 3.4 3.0 0.4 6.7 9.7 5.4 −0.6 Germany 2.8 4.1 2.6 3.1 4.3 4.0 4.2 −0.2 8.5 17.4 6.0 −2.7 Italy 4.1 5.0 2.5 3.4 5.7 6.2 5.8 0.4 2.6 16.2 1.6 −0.9 Japan 4.1 3.0 3.4 1.1 4.4 3.4 2.7 0.8 7.3 11.0 4.1 0.3 UK 7.2 7.6 5.3 3.4 8.2 4.5 3.8 0.7 11.7 19.0 7.3 0.5 US 6.0 8.5 5.9 13.3 7.3 3.1 2.1 0.5 14.5 22.3 7.8 −0.5 Average share in nominal gross output of 1995 and 2005 Spain 100.0 45.3 1.5 14.4 29.4 42.6 12.1 2.7 9.4 France 100.0 45.2 1.2 7.2 36.7 42.5 12.3 3.9 8.4 Germany 100.0 33.5 0.7 4.0 28.9 34.6 32.0 6.4 25.6 Italy 100.0 41.9 2.4 9.4 30.1 37.7 20.4 1.5 18.9 Japan 100.0 42.7 0.8 12.6 29.2 44.0 13.3 6.2 7.1 UK 100.0 40.6 1.1 4.2 35.3 45.7 13.7 4.6 9.1 US 100.0 34.4 0.9 6.5 27.0 51.4 14.3 6.2 8.1 . Gross output . Intermediate inputs . Labour input . Capital input . MFP . Total . Energy . Materials . Services . Total . Hours worked . Labour composition . Total . ICT . Non‐ICT . Contribution to gross output volume grcwth Spain 6.7 3.7 0.1 1.1 2.4 2.6 2.2 0.4 1.2 0.3 0.8 −0.8 France 3.9 2.2 0.0 0.4 1.9 1.5 1.3 0.2 0.8 0.4 0.5 −0.6 Germany 2.8 1.4 0.0 0.1 1.2 1.4 1.4 −0.1 2.7 1.1 1.5 −2.7 Italy 4.1 2.1 0.1 0.3 1.7 2.3 2.2 0.1 0.5 0.2 0.3 −0.9 Japan 4.1 1.3 0.0 0.1 1.3 1.5 1.2 0.3 1.0 0.7 0.3 0.3 UK 7.2 3.1 0.1 0.1 2.9 2.0 1.7 0.3 1.6 0.9 0.7 0.5 US 6.0 2.9 0.1 0.9 2.0 1.6 1.1 0.2 2.1 1.4 0.6 −0.5 Volume growth Spain 6.7 8.2 9.7 7.8 8.3 6.1 5.2 0.9 10.0 12.1 9.0 −0.8 France 3.9 5.0 1.4 5.1 5.1 3.4 3.0 0.4 6.7 9.7 5.4 −0.6 Germany 2.8 4.1 2.6 3.1 4.3 4.0 4.2 −0.2 8.5 17.4 6.0 −2.7 Italy 4.1 5.0 2.5 3.4 5.7 6.2 5.8 0.4 2.6 16.2 1.6 −0.9 Japan 4.1 3.0 3.4 1.1 4.4 3.4 2.7 0.8 7.3 11.0 4.1 0.3 UK 7.2 7.6 5.3 3.4 8.2 4.5 3.8 0.7 11.7 19.0 7.3 0.5 US 6.0 8.5 5.9 13.3 7.3 3.1 2.1 0.5 14.5 22.3 7.8 −0.5 Average share in nominal gross output of 1995 and 2005 Spain 100.0 45.3 1.5 14.4 29.4 42.6 12.1 2.7 9.4 France 100.0 45.2 1.2 7.2 36.7 42.5 12.3 3.9 8.4 Germany 100.0 33.5 0.7 4.0 28.9 34.6 32.0 6.4 25.6 Italy 100.0 41.9 2.4 9.4 30.1 37.7 20.4 1.5 18.9 Japan 100.0 42.7 0.8 12.6 29.2 44.0 13.3 6.2 7.1 UK 100.0 40.6 1.1 4.2 35.3 45.7 13.7 4.6 9.1 US 100.0 34.4 0.9 6.5 27.0 51.4 14.3 6.2 8.1 Notes. Business services refer to NACE industries 71 to 74, thus excluding real estate. Contribution of inputs calculated as the share of input times the volume growth rate. Figures might not add due to rounding. Calculations based on EUKLEMS database, release March 2008, see Timmer et al. (2007). Open in new tab 3.2. Measurement in EU KLEMS: Some Health Warnings Some general remarks on usage of EU KLEMS data are also warranted. The data are suitable for both growth accounting and econometric exercises but the issues touched on below caution that the user should also be aware of their limitations. As with all data series there are some unresolved measurement issues. As a general rule the reliability of the data is likely to be lower the finer the industry detail, i.e. the more we move from the industry level identified in the published National Accounts, and often lower for services industries than for manufacturing. This is because to break down the national accounts series, we often had to rely on additional data sources which are more abundant and complete for manufacturing than for services. To this could be added that the further back in time the series, the greater the likelihood of error. Thus whereas growth accounting exercises that quantify the contribution of ICT to output growth in transport equipment manufacturing over the period 1995 to 2005 might be reasonable, a precise number for the change of energy input use in business services between 1970 and 1971 might not be. These issues may be less important in econometric analysis with judicious use of methods. In addition it should be emphasised that growth accounting is useful as a descriptive tool but that it is merely accounting and says nothing about causality. For example, MFP growth in computer manufacturing may lead to a price decline in ICT assets, which induces investment in ICT and growth in capital services. Therefore improved technology partly has its effect through the capital contribution. In addition complementarities between various types of inputs are not taken into account, e.g. between skills and ICT capital. More fundamentally, proximate sources of growth such as input growth are endogenous to deeper causes of growth such as technical change, institutions, geography or macro‐economic policies (Maddison, 1995). But growth accounting provides a useful starting point to the identification of the contributions of the proximate sources of growth. It also provides a consistent structure in which data on output and inputs can be collected, both across industries and between variables, and as such it is a powerful organising principle. Nevertheless the method is constrained by its assumptions and so researchers may prefer to work with the underlying data. We believe that by also providing the basic input‐data of the growth accounts, EU KLEMS can support a much wider variety of approaches to the study of economic growth, alongside growth accounting. Below we discuss some general issues which are important for potential users, on a variable‐by‐variable basis. At the same time, it must be stressed that the limitations of the EU KLEMS series vary widely by country, period and variable and prudent users of the data should familiarise themselves with the methods of construction as discussed on a country‐by‐country basis in Timmer et al. (2007). 3.2.1. Output and intermediate inputs As mentioned above, output series are taken primarily from National Accounts sources. However this does not mean that these series are by any means perfect. In fact there are significant unresolved measurement issues in the National Accounts, in particular for services. It is well‐known that the problem of measuring output is in general much more challenging in services than in goods‐producing industries. Most measurement problems boil down to the fact that service activities are intangible, more heterogeneous than goods and often dependent on the actions of the consumer as well as the producer. A distinction should be made for services which are traded in a market (market services) and non‐market services for which no prices exist. The measurement of nominal output in market services is generally less problematic, being mostly a matter of accurately registering total revenue. But the main bottleneck is the measurement of output volumes, which requires accurate price measurement adjusted for changes in the quality of services output. There is no doubt that problems in measuring market services output still exist, especially in finance and business services but many statistical offices have made great strides forward in the measurement of the nominal value and prices. Output measures in the National Accounts should give a fairly accurate – albeit not perfect – internationally comparable picture of developments in market services.19 If there are unresolved measurement issues in market sectors, these are magnified in the case of output in sectors where a large part of the services is provided by the public sector, namely public administration, education, health and social services.20 The main problems in measuring output in non‐market sectors relate to the lack of market prices that allow aggregation across diverse outputs, in addition to the need to incorporate quality improvements.21 Typically, in the past, nominal output was measured by wages, sometimes including an imputation for capital costs. If output is measured by inputs, productivity growth should be zero by definition. More recently there has been a move to employ quantity indicators to measure volumes of output, with EU countries facing a Eurostat target of removing the dependence on input measures. Until this process is complete, productivity measures for these sectors should therefore be interpreted with care, if at all. The data cannot be used as evidence that, say, health services in one country are more efficient or better than in another country in some overall sense. But EU KLEMS data may well be useful in considering the use of ICT or skilled labour in the health sector across countries. Finally on output measurement it is important to note that for the most part the output of the real estate sector (NACE 70) is imputed rent on owner‐occupied dwellings, so again productivity measures for this industry need to be interpreted with care. Given the measurement problems in regard to non‐market sector and real estate, EU KLEMS presents aggregates for the total market economy which excludes both. For an analysis of the use of intermediate inputs in production it is important to note that series of energy, materials and services are derived by using their shares in intermediate inputs from supply and use tables (SUTs) applied to series of intermediate inputs from the National Accounts. SUTs are generally available on a frequent basis from 1995 onwards for many countries but not in the period before. Earlier estimates in EU KLEMS are sometimes based on historical input–output tables which were not integrated with the National Accounts and only available for benchmark years, necessitating interpolation and on occasion assuming EMS shares constant over time or across a sub‐set of industries. 3.2.2. Labour input Series on number of workers and hours worked by industry present relatively few problems, although there are still some unresolved issues regarding differences in sources and methods for annual average hours worked, which mainly affects levels comparisons (OECD, 2008, Annex 1). Incorporating adjustments for composition is more contentious. In EU KLEMS, skill levels are divided into high, medium and low categories22– this division is dictated by the need to keep the number of categories relatively low given sample sizes in the underlying surveys. This fairly aggregate division can lead to biases in the aggregate composition adjustment if employment trends and wage shares differ within categories. The extent of these biases also relate to the comparability of educational attainment and qualifications across countries, since some sub‐categories with relatively high wages may be classified to high skill in one country and medium skill in another. Therefore, comparisons of skill shares across countries should be interpreted with care. In addition labour composition measures tend to be somewhat volatile over time since the underlying surveys are not designed to generate time series. For some uses, period averages might be preferred to a focus on year‐on‐year changes. It is also important to note that the level of independent industry detail is much lower for labour composition than other variables, again dictated by the survey samples. In many cases the detail is restricted to 15 industries, largely one‐digit sectors but with manufacturing divided into three groups: intermediate goods, investment goods and consumer goods. As growth accounts are provided at a more detailed industry‐level, there is an implicit assumption that hours and wage shares in sub‐industries are equal to those for aggregate industries. Researchers estimating labour demand equations should be aware that an attempt to do so at too fine an industry level will just reproduce this assumption. In addition, it should be noted that much of the information on self‐employed workers is not based on survey data but imputed from employees, as self‐employed are often not (sufficiently) covered in the labour force surveys. Similarly, for most countries, labour type characteristics are only available for the number of employees, rather than hours worked, with the implicit assumption that hours do not vary by characteristic. While employment and earnings are consistently measured so that growth accounting and wage share equations are not affected, this would affect, say, an analysis of female participation rates, as women typically work (many) fewer hours than men. The growth accounting section of EU KLEMS presents estimates of volume of labour input and labour services. Implicit in the construction of these series is the assumption that each type of labour is paid its marginal product. In some circumstances this assumption is not appropriate, e.g., if there is widespread monopsony power within an industry (Manning, 2003) or an industry approximates a bilateral monopoly. These problems might be addressed by inclusion in regression equations of variables that proxy for collective bargaining. An alternative might be that users include different types of labour directly in an estimating equation. The additional variables section of EU KLEMS contains data on hours worked and wage shares by skill type, and for some countries the underlying data cross‐classified by gender, age and skill are also available. 3.2.3. Capital input Industry‐level estimates of capital input require detailed asset‐by‐industry investment matrices. Aggregate investment by industry and aggregate investment by asset type are normally available from the National Accounts. However, the allocation of assets to using industries in the so‐called capital‐flow matrix is generally not made public by the NSIs. The main reason for this is that the construction of this matrix is much less reliable than the aggregate series and depends on a wide variety of assumptions.23 Also within EU KLEMS various assumptions have been used to generate the capital‐flow matrix, in particular for the breakdown of computing equipment (IT) and communications equipment (CT) by industry. In most cases, EU countries provide estimates of software by industry for recent years, although the extent of backdating and industry coverage varies, and sufficient survey information to allow separate identification of computing and communications equipment. However in some cases it was necessary to use assumptions about the hardware–software ratios from other countries, so that IT and CT could be distributed across industries. Hence there is more likelihood of error and non‐comparability in these series than for other assets, especially in earlier periods. Another particular problem concerns the issue of ownership versus use of capital assets. In general, assets are allocated to the industry of ownership, i.e. in the case of leasing, the assets are accounted for in the capital stock of the leasing industry and the using industry pays a rental fee which is recorded in its use of intermediate services. A particular example is infrastructure: public infrastructure is not allocated to the using industries but rather appears as part of the capital stock of public administration. This is an important asset in the transport industries and hence MFP growth in this industry includes the contribution of infrastructure to output growth. The assets covered by the EU KLEMS capital account are fixed assets as defined in the ESA 95, with the exception of inventories, land and natural resources due to a lack of data. Inventories can be especially important in trade and transportation industries, while the lack of land and natural resources data will mainly affect MFP estimates for agriculture and mining. It has little effect on input and productivity measures of most other industries, especially since land is often included with structures investment. Depreciation rates in EU KLEMS vary by asset and industry but are held constant over time and across countries. Most likely these assumptions do not hold, as depreciation also depends on the degree of turbulence and innovation within an industry which induces premature scrapping because of obsolescence. However, there is little empirical evidence to buttress this argument and so it is difficult to measure. As a second‐best solution constant rates are assumed. One of the more stringent assumptions in capital service measurement is the assumption of constant returns to scale. Capital services are constructed employing user costs of capital as weights assuming an ex post rate of return (see Appendix B for details). However ex post rates of return can only be derived under constant returns to scale as in the KLEMS accounting system nominal input costs equal nominal output revenue. Alternatively, user costs can be based on ex ante measures in which an exogenous rate of return is derived outside the accounting system. This enables one to estimate costs alongside revenues and allows for non‐zero profits. The use of ex ante rates of return in capital services has first been suggested by Diewert (1980) and is gaining stronger support, in particular in situations where not all assets are covered; see e.g. Oulton (2007). Further analysis has shown that that the impact of alternative methods on capital services growth is small for most industries (Erumban, 2008). 3.2.4. Multi‐factor productivity In our approach to growth accounting, MFP growth measures disembodied technological change. Technical change embodied in new capital goods is captured by our measure of capital input through the use of quality‐adjusted prices and user costs as weights in asset aggregation. In addition, one might also be interested in a proxy for embodied technological change. One way to address this is by measuring capital input as the capital stock deflated at real acquisition prices and aggregated with nominal asset shares (Greenwood et al., 1997). The difference between the EU KLEMS capital input series and this new series would be a proxy for embodied technological change. The EU KLEMS database provides the basic investment and capital stocks series to construct alternative measures of capital input. MFP growth rates in EU KLEMS are occasionally negative, especially for some services industries. This might seem improbable as, under strict neo‐classical assumptions, MFP growth measures disembodied technological change and negative MFP would indicate technological regress. However, in practice measured MFP includes a range of other effects.24 First, in addition to technical innovation it also includes the effects from organisational and institutional change. For example, the successful reorganisation of a business to streamline the production process will generally lead to higher MFP growth in the long run but in the short run might decrease measured MFP as resources are diverted to the reorganisation process (for a discussion see Basu et al., 2004). Second, MFP measures pick up any deviations from the neo‐classical assumption that marginal costs reflects marginal revenues. If, for example, ICT investments have been driven more by herd behaviour than by economic fundamentals, as may have occurred in the run up to the dot.com bubble, marginal costs might be higher than marginal revenues. Consequently, MFP is underestimated and the contributions of ICT investment to growth are overestimated as growth accounting assumes that marginal cost reflects marginal revenue. Conversely if there were above normal returns to ICT its contribution would be underestimated (O’Mahony and Vecchi, 2005). Or, in the case of imperfect competition, an increase in mark‐ups will be picked up by a decline in measured MFP, keeping the capital–labour ratio constant. One way to relax the underlying market‐clearing assumptions and allow for mark‐ups and varying returns to scale is to use cost shares rather than output value shares (Hall, 1988; Crafts and Mills, 2005). However this requires independent estimates of the cost of capital through ex ante rates of return as discussed above. Third, being a residual measure, MFP growth also includes the effects from changes in unmeasured inputs, such as research and development and other intangible investments (Corrado et al., 2006). Finally, MFP includes measurement errors in inputs and outputs, such as underestimation of the quality change of new services products, which might be proceeding faster today than in the past, although there is little hard evidence available so far. MFP measures can be derived at various levels of aggregation. Gross output decompositions are most meaningful at the lowest level of aggregation, viz., establishments. As soon as aggregates of gross output are decomposed, one runs into problems of comparability over time and across countries, depending on differences in vertical integration of firms. Ideally, decomposing gross output should be done on a sectoral output measure which excludes intra‐sectoral deliveries of intermediates (Gollop, 1979). Measures of sectoral output require detailed symmetric domestic input–output tables, which are not available on a sufficiently large scale for all European countries. Also, a coherent framework for aggregation in an open economy has not yet been developed, as the standard methods ignore the role of imports. Therefore, we present gross output decompositions only at the lowest possible industry level, depending on the level of detail of output and inputs, and do not show any industry aggregates. In the current database we also present the decomposition of value added growth, which is insensitive to the intra‐industry delivery problem. The decomposition results for the latter are shown for all aggregation levels, up to total economy. In summary, this Section has identified a number of issues that can affect the uses of EU KLEMS data. Some are unavoidable since the database relies heavily on National Accounts data and so need to await further developments in NSIs. In this respect, by confronting various data‐sources within and across countries, the EU KLEMS database is useful in indicating priority areas for further improvement in basic series including volume measures of services output, capital formation matrices and more generally consistency between output, labour and capital inputs at the industry level. Other caveats suggest prudence by the users, depending on the context in which the data are employed. But, as with all data analysis, a judicious use of econometric methods and sensible approaches to the use of the numbers should enable the database to be useful in a wide range of applications. 4. Future Developments This article describes the March 2008 release of the EU KLEMS database. The database will be revised and updated each year and gradually expanded in terms of country coverage. In the near future, extensions are planned to include Canada, China, India, South Korea and Taiwan. While the EU KLEMS data can provide descriptive analysis of growth and its contributors, potentially its greatest benefit will be in future research where it is linked to additional databases. The following extensions seem to be particularly promising: inclusion of innovation indicators and intangible investment; international trade and environmental pollution indicators and measures of firm‐level dynamics. The EU KLEMS consortium is engaged in research that should begin populating some of these variables, although the coverage across countries, industries and time will be less comprehensive than the variables in the current dataset. To explain differences in productivity growth, additional information on innovation inputs and outputs will be needed. Investment in intangibles, such as innovative property through research and development and firm‐specific economic competences such as organisational capital and brand equity, has become increasingly important for growth. Although some of these concepts are intuitive at the firm‐level, the development of industry‐aggregates provides particular challenges, in particular with respect to rates of depreciation and prices series to derive volume measures.25 In addition innovation output measures such as patent counts will be added to the database. For studies of outsourcing and international trade, further integration of trade statistics is highly desirable. Through the use of a supply‐and‐use framework, as in EU KLEMS, bilateral product‐level trade statistics can be mapped into an industry classification, and inter‐ and intra‐industry trade flows can be traced. A particular challenging extension would be the inclusion of environmental pollution indicators. This database can be instrumental in studying the relationships between economic growth, socio‐economic development and environmental quality within an international framework. Another promising avenue for further research is in the linking of firm‐level‐based variables that might affect industry productivity trends. Candidate variables are those related to market structure such as concentration rates and share of foreign firms, and dynamics of the industry such as entry and exit rates or average age of firms. An obvious link will be to firm level databases such as the Amadeus company accounts database or data on entry and exit at the plant level from national production surveys (Bartelsman et al., 2005). An additional potentially useful avenue of research is to match data from labour market databases to EU KLEMS. For example data from the UK Workplace Employee Relations Survey (WERS) is currently being aggregated to industry level and matched to EU KLEMS. Labour Force Surveys offer a potentially rich source of data, for example, on the use of migrant labour, the extent of workplace training and flexible working arrangements. Finally, while an industry database has its own uses, it may also facilitate comparative research based on data at the firm level that are unaffected by restrictions imposed by aggregation. In this way, the EU KLEMS database can provide industry‐country measures of variables such as productivity, ICT and skilled labour that can be used as control variables or interactions in conjunction with firm level data. Industry measures can also be used to benchmark firm‐level distributions, which are typically not comparable across countries due to different coverage of firms (Bartelsman et al., 2005). From the outset, the consortium and its European Commission sponsors were committed to ensuring the EU KLEMS database was a public good, with free access to both the research and policy community. It is hoped that it provides a stepping‐stone and encouragement for others to further develop the database through a process of addition or links to compatible data. Although open‐source databases are not particularly abundant in our profession, there is little doubt they generate external benefits for both producers and users. Footnotes 1 " To be more precise, A reflects Hicks‐neutral technical change. Because of our approach to capital measurement it only includes disembodied technical change, see also the discussion in Section 3.2.4. 2 " Aggregate input is unobservable and it is common to express it as a translog function of its individual components. Then the corresponding index is a Törnqvist volume index (see Jorgenson et al., 1987). 3 " The first term is also known as ‘labour quality’ in the growth accounting literature (see e.g. Jorgenson et al. 2005). However, this terminology has a normative connotation which easily leads to confusion. For example, lower female wages would suggest that hours worked by females have a lower ‘quality’ than hours worked by males. Instead we prefer to use the more positive concept of ‘labour composition’. 4 " The growth accounting calculations in EU KLEMS included this division into volume and composition for labour input to summarise all aspects of labour composition. Alternatively, in keeping with the divisions for intermediate and capital input one could subdivide the contribution of labour into groups, e.g. high‐skilled and low‐skilled labour. The data necessary for such a division is also available in the database – see below for further details. 5 " These assets are residential structures, non‐residential structures, transport equipment, information technology equipment, communication technology equipment, other machinery and equipment, software and other fixed capital assets. 6 " Taxes have not been included due to a lack of data. Also, the assets do not cover land and inventories. 7 " See Jorgenson et al. (2005, ch. 8) for an elaborate discussion. 8 " Several European NSIs are experimenting with growth accounting statistics, including Statistics Netherlands, Statistics Sweden, Statistics Finland and ISTAT (the Italian NSI). Outside Europe, the Australian Bureau of Statistics, Statistics Canada and the US Bureau of Labor Statistics (BLS) maintain a detailed productivity programme. 9 " For OECD productivity programme, see http://www.oecd.org. For GGDC series, see Total Economy Growth Accounting database at http://www.ggdc.nl, described in Timmer and van Ark (2005). Also see O’Mahony (1999) and Inklaar et al. (2005) for international comparisons at the industry‐level for a limited set of countries. 10 " Instead STAN was intended for tracking knowledge spillovers (in combination with other OECD databases such as ANBERD and the input‐output database) and general structural analyses. 11 " STAN includes a number of variables not included in EU KLEMS, most notably imports and exports by product group. 12 " Possible differences are due to differences in the vintage on the National Accounts series used and in the use of different index‐number formulae for industry‐aggregation. We use the theoretically based Törnqvist indices, whereas in most National Accounts and in STAN chained Laspeyres type indices are used. These differ only in cases of very high or low growth rates. 13 " See Inklaar et al. (2008) for a discussion of the sensitivity of results to alternative productivity measures in the context of a study on the impact of regulation and skilled labour on productivity catch up. 14 " Market economy in EU KLEMS excludes the real estate sector, public administration and education, health and social services, due to problems in measuring productivity in these sectors, see Section 4.2. 15 " See e.g. Cameron et al. (2005), Griffith et al. (2004), Nicoletti and Scarpetta (2003), Vandenbussche et al. (2006) and Inklaar et al. (2008). 16 " The basic labour and capital input data series are also publicly available at the EU KLEMS website, except for some countries where confidentiality had to be respected. 17 " Aggregate tables are provided for four institutional country groupings: EU‐25 (all member states of the EU as of 1 May 2004), EU‐15 (all member states of the EU as of 1 January 1995), EU‐10 (all states which joined the EU on 1 May 2004) and Euro (all countries in the euro zone as of 1 January 2001). We also provide an aggregation for those countries for which there is long‐run capital and labour composition data. These groups are called EU‐15ex and Euroex. 18 " Negative MFP growth in business services in the US has also been found by Jorgenson et al. (2005) and Triplett and Bosworth (2006). 19 " See Appendix A in Inklaar et al. (2008) for a survey of the current state of services output measurement practices. 20 " In EU KLEMS as elsewhere we refer to these sectors as ‘non‐market services’, recognising that some output of these sectors is provided by the private sector and the extent of this varies across countries. 21 " For general discussions of the issues involved see Atkinson (2005) and O’Mahony and Stevens (2006); the reader is referred to Castelli et al. (2007) for discussion and possible resolution in the particular example of health sector output. 22 " See EU KLEMS methodology document (Timmer et al., 2007) for definition of each educational group in each country. 23 " For example, to distribute parts of equipment, computers and software the BEA uses occupation‐by‐industry data, rather than investment survey data (Meade et al., 2003). 24 " See Hulten (2001) for an elaborate biography of the MFP concept. 25 " See Corrado et al. (2006) on the measurement of intangibles. References Aghion , P. , Bloom , N., Blundell , R., Griffith , R. and Howitt , P. ( 2005 ). ‘Competition and innovation: an inverted‐U relationship’ , Quarterly Journal of Economics , vol. 120 ( 2 ), pp. 701 – 28 . OpenURL Placeholder Text WorldCat Aghion , P. and Griffith , R. ( 2005 ). Competition and Growth. Reconciling Theory and Evidence , Cambridge, MA: MIT Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Antràs , P. and Helpman , E. ( 2004 ). ‘Global sourcing’ , Journal of Political Economy , vol. 112 ( 3 ), pp. 552 – 80 . Google Scholar Crossref Search ADS WorldCat Van Ark , B. , O’Mahony , M. and Timmer , M.P. ( 2008 ). ‘The productivity gap between Europe and the U.S.: trends and causes’ , Journal of Economic Perspectives , vol. 22 ( 1 ), vol. 25 – 44 . Google Scholar Crossref Search ADS WorldCat Atkinson , T. ( 2005 ) Measurement of Government Output and Productivity for the National Accounts, Atkinson Review: Final Report , London: HMSO . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Autor , D.H. , Katz L.F. and Krueger , A.B. ( 1998 ). ‘Computing inequality: have computers changed the labour market?’ , Quarterly Journal of Economics , vol. 113 ( 4 ), pp. 1169 – 210 . Google Scholar Crossref Search ADS WorldCat Bartelsman , E.J. , Haltiwanger , J. and Scarpetta , S. ( 2005 ). ‘ Micro–economic evidence of creative destruction in industrial and developing countries ’, World Bank Policy Research Working Paper Series, 3464. Basu , S. , Fernald , J., Oulton , N. and Srinivasan , S. ( 2004 ). ‘The case of the missing productivity growth: or, does information technology explain why productivity accelerated in the United States but not in the United Kingdom?’, in ( M. Gertler and K. Rogoff, eds.), NBER Macroeconomics Annual , pp. 9 – 63 , Cambridge MA: MIT Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Cameron , G. , Proudman , J. and Redding , S ( 2005 ), ‘Technological convergence, R&D, trade and productivity growth’ , European Economic Review , vol. 49 , pp. 775 – 807 . Google Scholar Crossref Search ADS WorldCat Castelli , A. , Dawson , D., Gravelle , H., Jacobs , R., Kind , P., Loveridge , P., Martin , S., O’Mahony , M., Stevens , P., Stokes , L., Street , A. and Weale , M. ( 2007 ). ‘A new approach to measuring health system output and productivity’ , National Institute Economic Review , vol. 200 (April), pp. 105 – 17 . Google Scholar Crossref Search ADS WorldCat Caves , D.W. , Christensen , L.R. and Diewert , W.E. ( 1982 ). ‘Multilateral comparisons of output, input and productivity using superlative index numbers’ , Economic Journal , vol. 92 ( 365 ), pp. 73 – 86 . Google Scholar Crossref Search ADS WorldCat Corrado , C.A. , Hulten , C.R. and Sichel , D.E. ( 2006 ). ‘ Intangible capital and economic growth ’, NBER Working Paper, no. 11948. Crafts , N. and Mills , T. ( 2005 ). ‘TFP growth in British and German manufacturing’ , 1950–1996 , Economic Journal , vol. 115 ( 505 ), pp. 649 – 70 . Google Scholar Crossref Search ADS WorldCat Diewert , W. ( 1980 ). ‘Aggregation problems in the measurement of capital’, in ( D. Usher, ed.), The Measurement of Capital , pp. 433 – 52 , Chicago: University of Chicago Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Erumban , A.A . ( 2008 ), ‘ Rental prices, rates of return, capital aggregation and productivity: evidence from EU and US ’, CESifo Economic Studies , vol. 54 ( 3 ), pp. 499 – 533 . Google Scholar Crossref Search ADS WorldCat European Commission ( 2007 ). Competitiveness and Economic Reforms: European Competitiveness Report, 2007 , Luxembourg: Office for Official Publications of the European Communities . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Gollop , F . ( 1979 ). ‘Accounting for intermediate input: the link between sectoral and aggregate measures of productivity’, in Measurement and Interpretation of Productivity , pp. 318 – 33 , Washington, DC: National Academy of Sciences . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Greenwood , J. , Hercowitz , Z, and Krusell , P. ( 1997 ). ‘Long‐run implications of investment‐specific technological change’ , American Economic Review , vol. 87 , pp. 342 – 62 . OpenURL Placeholder Text WorldCat Griffith , R. , Redding , S. and Van Reenen , J. ( 2004 ). ‘Mapping the two faces of R&D: productivity growth in a panel of OECD countries’ , Review of Economics and Statistics , vol. 84 ( 4 ), pp. 883 – 95 . Google Scholar Crossref Search ADS WorldCat Hall , R.E. ( 1988 ). ‘The relation between price and marginal cost in US industry’ , Journal of Political Economy , vol. 96 ( 5 ), pp. 921 – 47 . Google Scholar Crossref Search ADS WorldCat Harrigan , J. ( 1997 ). ‘Technology, factor supplies and international specialisation’ , American Economic Review , vol. 87 ( 4 ), pp. 475 – 94 . OpenURL Placeholder Text WorldCat Hulten , C.R. ( 2001 ). ‘Total factor productivity: a short biography’, in ( C.R. Hulten, E.R. Dean and M.J. Harper, eds.), New Developments in Productivity Analysis , Studies in Income and Wealth, vol. 63 , pp. 1 – 47 , Chicago: University of Chicago Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Inklaar , R. , O’Mahony , M. and Timmer , M.P. ( 2005 ). ‘ICT and Europe’s productivity performance; industry‐level growth account comparisons with the United States’ , Review of Income and Wealth , vol. 51 ( 4 ), pp. 505 – 36 . Google Scholar Crossref Search ADS WorldCat Inklaar , R. , Timmer , M. and Van Ark , B. ( 2008 ), ‘Market services productivity across Europe and the US’ , Economic Policy , vol. 23 , pp. 141 – 94 . Google Scholar Crossref Search ADS WorldCat Inklaar , R. and Timmer , M.P. ( 2008 ). ‘ GGDC productivity level database: international comparisons of output, inputs and productivity at the industry level ’, GGDC Research Memorandum GD‐104, University of Groningen . Jalava , J. and Pohjola , M. ( 2008 ). ‘The roles of electricity and ICT in economic growth: the case of Finland’ , Explorations in Economic History , vol. 45 ( 3 ), pp. 270 – 87 . Google Scholar Crossref Search ADS WorldCat Jorgenson , D.W. and Griliches , Z. ( 1967 ). ‘The explanation of productivity change’ , Review of Economic Studies , vol. 34 ( 3 ), pp. 249 – 83 . Google Scholar Crossref Search ADS WorldCat Jorgenson , D.W. , Gollop , F.M. and Fraumeni , B. ( 1987 ). Productivity and U.S. Economic Growth , Cambridge, MA: Harvard Economic Studies . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Jorgenson , D.W. , Ho , M.S. and Stiroh , K.J. ( 2005 ). Information Technology and the American Growth Resurgence , Cambridge, MA: MIT Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Koeniger , W. and Leonardi , M. ( 2007 ). ‘Capital deepening and wage differentials: Germany versus US’ , Economic Policy , vol. 22 ( 49 ), pp. 71 – 116 . Google Scholar Crossref Search ADS WorldCat Kratena , K. ( 2007 ). ‘Technical change, investment and energy intensity’ , Economic Systems Research , vol. 19 ( 3 ), pp. 295 – 314 . Google Scholar Crossref Search ADS WorldCat Krueger , A. B. ( 1999 ). ‘Measuring labor’s share’ , American Economic Review, Papers and Proceedings , vol. 89 ( 2 ), pp. 45 – 51 . Google Scholar Crossref Search ADS WorldCat Machin , S. and Van Reenen , J. ( 1998 ). ‘Technology and changes in skill structure: evidence from seven OECD countries’ , Quarterly Journal of Economics , vol. 113 ( 4 ), pp. 1215 – 44 . Google Scholar Crossref Search ADS WorldCat Maddison , A. ( 1995 ). Explaining the Economic Performance of Nations. Essays in Time and Space , Aldershot: Edward Elgar Publishing . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Manning , A. ( 2003 ). Monopsony in Motion: Imperfect Competition in Labour Markets , New Jersey: Princeton University Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Maudos , J. , Pastor , J.M. and Serrano , L. ( 2008 ). ‘Explaining the US‐EU productivity growth gap: structural change vs. intra sectoral effect’ , Economic Letters , vol. 100 ( 2 ), pp. 311 – 3 . Google Scholar Crossref Search ADS WorldCat Meade , D.S. , Stanislaw J. Rzeznik , S.J. and Robinson‐Smith , D.C ( 2003 ). ‘Business investment by industry in the U.S. economy for 1997’, Survey of Current Business , November, pp. 18 – 70 . Nicoletti , G. and Scarpetta , S. ( 2003 ). ‘Regulation, productivity and growth: OECD evidence’ , Economic Policy , vol. 36 , pp. 9 – 72 . Google Scholar Crossref Search ADS WorldCat O’Mahony , M. ( 1999 ). Britain’s Relative Productivity Performance, 1950‐1996: An International Perspective ,London: National Institute of Economic and Social Research . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC O’Mahony , M. and Stevens P. ( 2006 ). ‘International comparisons of output and productivity in public services provision: a review’, in ( G. A. Boyne, K. A. Meier, L. J. O’Toole and R. M. Walker, eds.), Public Service Performance: Perspectives on Measurement and Management , pp. 233 – 53 . Cambridge: Cambridge University Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC O’Mahony , M. and Vecchi , M. ( 2005 ). ‘Quantifying the impact of ICT capital on output growth: a heterogeneous dynamic panel approach’ , Economica , vol. 72 ( 288 ), pp. 615 – 34 . Google Scholar Crossref Search ADS WorldCat OECD ( 2008 ). OECD Compendium of Productivity Indicators 2008 , Paris: OECD . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Oulton , N. ( 2007 ). ‘Ex post versus ex ante measures of the user cost of capital’ , Review of Income and Wealth , vol. 53 ( 2 ), pp. 295 – 317 . Google Scholar Crossref Search ADS WorldCat Schreyer , P. ( 2001 ). Measuring Productivity – OECD Manual. Measurement of Aggregate and Industry‐Level Productivity Growth , Paris: OECD . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Schreyer , P. ( 2002 ). ‘Computer price indices and international growth and productivity comparisons’ , Review of Income and Wealth , vol. 48 ( 1 ), pp. 15 – 31 . Google Scholar Crossref Search ADS WorldCat Timmer , M.P. and Van Ark , B. ( 2005 ). ‘Does information and communication technology drive EU‐US productivity growth differentials?’ , Oxford Economic Papers , vol. 57 ( 4 ), pp. 693 – 716 . Google Scholar Crossref Search ADS WorldCat Timmer , M.P. , Van Moergastel , T., Stuivenwold , E., Ypma , G., O’Mahony , M. and Kangasniemi , M. ( 2007 ). EU KLEMS Growth And Productivity Accounts , March, available at http://www.euklems.net. Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Triplett , J.E and Bosworth , B.P. ( 2006 ). ‘Baumol’s disease has been cured: IT and multifactor productivity in U.S. service industries’, in ( D. W. Jansen, ed.), The New Economy and Beyond: Past, Present, and Future , pp. 34 – 71 , Cheltenham: Edward Elgar . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Vandenbussche , J. , Aghion , P. and Meghir , C. ( 2006 ). ‘Growth, distance to the frontier and composition of human capital’ , Journal of Economic Growth , vol. 11 , pp. 97 – 127 . Google Scholar Crossref Search ADS WorldCat Appendices Appendix A. EU KLEMS Institutes The following is a list of institutes participating in the EU KLEMS project: Rijksuniversiteit Groningen (RUG), Groningen; National Institute of Economic and Social Research (NIESR), London; Centre d’études prospectives et d’informations internationales (CEPII), Paris; Centre for Economic and Business Research (CEBR), Copenhagen; CPB Netherlands Bureau for Economic Policy Analysis, The Hague; Statistics Netherlands, The Hague; Deutsches Institut für Wirtschaftsforschung e.V. (DIW), Berlin; Federaal Planbureau, Brussels; Istituto di Studi e Analisi Economica (ISAE), Rome; ISTAT, Rome; Instituto Valenciano De Investigaciones Economicas (IVIE), Valencia; Helsingin kauppakorkeakoulu (Helsinki School of Economics), Helsinki; Statistics Finland, Helsinki; Austrian Institute of Economic Research (WIFO), Vienna; The Vienna Institute for International Economic Studies (WIIW); AMsterdam Business and Economic Research (AMBER); The Conference Board Europe (TCBE), Brussels; Dale Jorgenson and associates, Cambridge; University of Konstanz; University of Birmingham; Pellervo Economic Research Institute (PTT), Helsinki; Hitotsubashi University, Tokyo; Statistics Luxembourg; Statistics Sweden. See also http://www.euklems.com. Appendix B. Measurement of Capital Services Growth According to the perpetual inventory model (PIM), the capital stock (S) is defined as a weighted sum of past investments with weights given by the relative efficiencies of capital goods at different ages: (A.1) with Si,T the capital stock (for a particular asset type i) at time T, ∂i,t the efficiency of a capital good i of age t relative to the efficiency of a new capital good and Ii,T−t the investments in period T − t. An important implicit assumption made here is that the services by assets of different vintages are perfect substitutes for each other. As in most studies, a geometric depreciation pattern is applied. With a given constant rate of depreciation δi which is constant over time, but different for each asset type, we get ∂i,t = (1 − δi)t−1, so that: (A.2) If one assumes that the flow of capital services from each asset type i (Ki) is proportional to the average of the stock available at the end of the current and the prior period (Si,T and Si,T−1), one can aggregate capital service flows from these asset types as a translog quantity index to: (A.3) where weights are given by the average shares of each component in the value of capital compensation and . The estimation of the compensation share of each asset, vi, is related to the user cost of each asset. The user cost approach is crucial in any analysis of the contribution of ICT capital to growth, because the annual amount of capital services delivered per euro of investment in ICT is much higher than that of an euro invested in, say, buildings. While an ICT asset may typically be scrapped after 5 years, buildings may provide services for decades. In addition, ICT assets have a high user cost due to rapidly declining prices. For example, the user cost of IT‐machinery is typically 50% to 60% of the investment price, while that of buildings is less than 10%. Therefore one euro of IT capital stock gets a heavier weight in the growth decomposition than one euro of building stock. This different weight on capital services is picked up by using the rental price of capital services, , which reflects the price at which the investor is indifferent between buying or renting the capital good for a one‐year lease in the rental market. In the absence of taxation the equilibrium condition can be rearranged, yielding the familiar cost‐of‐capital equation: (A.4) with it representing the nominal rate of return, δk the depreciation rate of asset type k, and , the investment price of asset type k. This formula shows that the rental fee is determined by the nominal rate of return, the rate of economic depreciation and the asset specific capital gains. The nominal rate of return is determined ex‐post as in the endogenous approach (Jorgenson et al., 2005). It is assumed that the total value of capital services for each industry equals its compensation for all assets. This procedure yields an internal rate of return that exhausts capital income and is consistent with constant returns to scale. This nominal rate of return is the same for all assets in an industry but is allowed to vary across industries and derived as a residual as follows: ( A.5) where the first term is the capital compensation in industry j, which under constant returns to scale can be derived as value added minus the compensation of labour. Table A1 
Variables in EU KLEMS Database Labour productivity variables . GO Gross output at current basic prices (in millions of local currency) II Intermediate inputs at current purchasers’ prices (in millions of local currency) VA Gross value added at current basic prices (in millions of local currency) COMP Compensation of employees (in millions of local currency) GOS Gross operating surplus (in millions of local currency) TXSP Taxes minus subsidies on production (in millions of local currency) EMP Number of persons engaged (thousands) EMPE Number of employees (thousands) H_EMP Total hours worked by persons engaged (millions) H_EMPE Total hours worked by employees (millions) GO_P Gross output, price indices, 1995 = 100 II_P Intermediate inputs, price indices, 1995 = 100 VA_P Gross value added, price indices, 1995 = 100 GO_QI Gross output, volume indices, 1995 = 100 II_QI Intermediate inputs, volume indices, 1995 = 100 VA_QI Gross value added, volume indices, 1995 = 100 LP_I Gross value added per hour worked, volume indices, 1995=100 Growth accounting variables LAB Labour compensation (in millions of local currency) CAP Capital compensation (in millions of local currency) LAB_QI Labour services, volume indices, 1995 = 100 CAP_QI Capital services, volume indices, 1995 = 100 IIE Intermediate energy inputs at current purchasers’ prices (in millions of local currency) IIM Intermediate material inputs at current purchasers’ prices (in millions of local currency) IIS Intermediate service inputs at current purchasers’ prices (in millions of local currency) IIE_QI Intermediate energy inputs, volume indices, 1995 = 100 IIM_QI Intermediate material inputs, volume indices, 1995 = 100 IIS_QI Intermediate service inputs, volume indices, 1995 = 100 VA_Q Growth rate of value added volume (% per year) VAConH Contribution of hours worked to value added growth (percentage points) VAConLC Contribution of labour composition change to value added growth (percentage points) VAConKIT Contribution of ICT capital services to output growth (percentage points) VAConKNIT Contribution of non‐ICT capital services to output growth (percentage points) VAConTFP Contribution of TFP to value added growth (percentage points) TFPva_I TFP (value added based) growth, 1995=100 GO_Q Growth rate of gross output volume (% per year) GOConII Contribution of intermediate inputs to output growth (percentage points) GOConIIM Contribution of intermediate energy inputs to output growth (percentage points) GOConIIE Contribution of intermediate material inputs to output growth (percentage points) GOConIIS Contribution of intermediate services inputs to output growth (percentage points) GOConH Contribution of hours worked to output growth (percentage points) GOConLC Contribution of labour composition change to output growth (percentage points) GOConKIT Contribution of ICT capital services to output growth (percentage points) GOConKNIT Contribution of non‐ICT capital services to output growth (percentage points) GOConTFP Contribution of TFP to output growth (percentage points) TFPgo_I TFP (gross output based) growth, 1995=100 Additional labour and capital variables CAPIT ICT capital compensation (share in total capital compensation) CAPNIT Non‐ICT capital compensation (share in total capital compensation) CAPIT_QI ICT capital services, volume indices, 1995 = 100 CAPNIT_QI Non‐ICT capital services, volume indices, 1995 = 100 CAPIT_QPH ICT capital services per hour worked, 1995 reference CAPNIT_QPH Non‐ICT capital services per hour worked, 1995 reference LABHS High‐skilled labour compensation (share in total labour compensation) LABMS Medium‐skilled labour compensation (share in total labour compensation) LABLS Low‐skilled labour compensation (share in total labour compensation) LAB_QPH Labour services per hour worked, 1995 reference H_HS Hours worked by high‐skilled persons engaged (share in total hours) H_MS Hours worked by medium‐skilled persons engaged (share in total hours) H_LS Hours worked by low‐skilled persons engaged (share in total hours) H_M Hours worked by male persons engaged (share in total hours) H_F Hours worked by female persons engaged (share in total hours) H_29 Hours worked by persons engaged aged 15‐29 (share in total hours) H_49 Hours worked by persons engaged aged 30‐49 (share in total hours) H_50+ Hours worked by persons engaged aged 50 and over (share in total hours) Labour productivity variables . GO Gross output at current basic prices (in millions of local currency) II Intermediate inputs at current purchasers’ prices (in millions of local currency) VA Gross value added at current basic prices (in millions of local currency) COMP Compensation of employees (in millions of local currency) GOS Gross operating surplus (in millions of local currency) TXSP Taxes minus subsidies on production (in millions of local currency) EMP Number of persons engaged (thousands) EMPE Number of employees (thousands) H_EMP Total hours worked by persons engaged (millions) H_EMPE Total hours worked by employees (millions) GO_P Gross output, price indices, 1995 = 100 II_P Intermediate inputs, price indices, 1995 = 100 VA_P Gross value added, price indices, 1995 = 100 GO_QI Gross output, volume indices, 1995 = 100 II_QI Intermediate inputs, volume indices, 1995 = 100 VA_QI Gross value added, volume indices, 1995 = 100 LP_I Gross value added per hour worked, volume indices, 1995=100 Growth accounting variables LAB Labour compensation (in millions of local currency) CAP Capital compensation (in millions of local currency) LAB_QI Labour services, volume indices, 1995 = 100 CAP_QI Capital services, volume indices, 1995 = 100 IIE Intermediate energy inputs at current purchasers’ prices (in millions of local currency) IIM Intermediate material inputs at current purchasers’ prices (in millions of local currency) IIS Intermediate service inputs at current purchasers’ prices (in millions of local currency) IIE_QI Intermediate energy inputs, volume indices, 1995 = 100 IIM_QI Intermediate material inputs, volume indices, 1995 = 100 IIS_QI Intermediate service inputs, volume indices, 1995 = 100 VA_Q Growth rate of value added volume (% per year) VAConH Contribution of hours worked to value added growth (percentage points) VAConLC Contribution of labour composition change to value added growth (percentage points) VAConKIT Contribution of ICT capital services to output growth (percentage points) VAConKNIT Contribution of non‐ICT capital services to output growth (percentage points) VAConTFP Contribution of TFP to value added growth (percentage points) TFPva_I TFP (value added based) growth, 1995=100 GO_Q Growth rate of gross output volume (% per year) GOConII Contribution of intermediate inputs to output growth (percentage points) GOConIIM Contribution of intermediate energy inputs to output growth (percentage points) GOConIIE Contribution of intermediate material inputs to output growth (percentage points) GOConIIS Contribution of intermediate services inputs to output growth (percentage points) GOConH Contribution of hours worked to output growth (percentage points) GOConLC Contribution of labour composition change to output growth (percentage points) GOConKIT Contribution of ICT capital services to output growth (percentage points) GOConKNIT Contribution of non‐ICT capital services to output growth (percentage points) GOConTFP Contribution of TFP to output growth (percentage points) TFPgo_I TFP (gross output based) growth, 1995=100 Additional labour and capital variables CAPIT ICT capital compensation (share in total capital compensation) CAPNIT Non‐ICT capital compensation (share in total capital compensation) CAPIT_QI ICT capital services, volume indices, 1995 = 100 CAPNIT_QI Non‐ICT capital services, volume indices, 1995 = 100 CAPIT_QPH ICT capital services per hour worked, 1995 reference CAPNIT_QPH Non‐ICT capital services per hour worked, 1995 reference LABHS High‐skilled labour compensation (share in total labour compensation) LABMS Medium‐skilled labour compensation (share in total labour compensation) LABLS Low‐skilled labour compensation (share in total labour compensation) LAB_QPH Labour services per hour worked, 1995 reference H_HS Hours worked by high‐skilled persons engaged (share in total hours) H_MS Hours worked by medium‐skilled persons engaged (share in total hours) H_LS Hours worked by low‐skilled persons engaged (share in total hours) H_M Hours worked by male persons engaged (share in total hours) H_F Hours worked by female persons engaged (share in total hours) H_29 Hours worked by persons engaged aged 15‐29 (share in total hours) H_49 Hours worked by persons engaged aged 30‐49 (share in total hours) H_50+ Hours worked by persons engaged aged 50 and over (share in total hours) Open in new tab Table A1 
Variables in EU KLEMS Database Labour productivity variables . GO Gross output at current basic prices (in millions of local currency) II Intermediate inputs at current purchasers’ prices (in millions of local currency) VA Gross value added at current basic prices (in millions of local currency) COMP Compensation of employees (in millions of local currency) GOS Gross operating surplus (in millions of local currency) TXSP Taxes minus subsidies on production (in millions of local currency) EMP Number of persons engaged (thousands) EMPE Number of employees (thousands) H_EMP Total hours worked by persons engaged (millions) H_EMPE Total hours worked by employees (millions) GO_P Gross output, price indices, 1995 = 100 II_P Intermediate inputs, price indices, 1995 = 100 VA_P Gross value added, price indices, 1995 = 100 GO_QI Gross output, volume indices, 1995 = 100 II_QI Intermediate inputs, volume indices, 1995 = 100 VA_QI Gross value added, volume indices, 1995 = 100 LP_I Gross value added per hour worked, volume indices, 1995=100 Growth accounting variables LAB Labour compensation (in millions of local currency) CAP Capital compensation (in millions of local currency) LAB_QI Labour services, volume indices, 1995 = 100 CAP_QI Capital services, volume indices, 1995 = 100 IIE Intermediate energy inputs at current purchasers’ prices (in millions of local currency) IIM Intermediate material inputs at current purchasers’ prices (in millions of local currency) IIS Intermediate service inputs at current purchasers’ prices (in millions of local currency) IIE_QI Intermediate energy inputs, volume indices, 1995 = 100 IIM_QI Intermediate material inputs, volume indices, 1995 = 100 IIS_QI Intermediate service inputs, volume indices, 1995 = 100 VA_Q Growth rate of value added volume (% per year) VAConH Contribution of hours worked to value added growth (percentage points) VAConLC Contribution of labour composition change to value added growth (percentage points) VAConKIT Contribution of ICT capital services to output growth (percentage points) VAConKNIT Contribution of non‐ICT capital services to output growth (percentage points) VAConTFP Contribution of TFP to value added growth (percentage points) TFPva_I TFP (value added based) growth, 1995=100 GO_Q Growth rate of gross output volume (% per year) GOConII Contribution of intermediate inputs to output growth (percentage points) GOConIIM Contribution of intermediate energy inputs to output growth (percentage points) GOConIIE Contribution of intermediate material inputs to output growth (percentage points) GOConIIS Contribution of intermediate services inputs to output growth (percentage points) GOConH Contribution of hours worked to output growth (percentage points) GOConLC Contribution of labour composition change to output growth (percentage points) GOConKIT Contribution of ICT capital services to output growth (percentage points) GOConKNIT Contribution of non‐ICT capital services to output growth (percentage points) GOConTFP Contribution of TFP to output growth (percentage points) TFPgo_I TFP (gross output based) growth, 1995=100 Additional labour and capital variables CAPIT ICT capital compensation (share in total capital compensation) CAPNIT Non‐ICT capital compensation (share in total capital compensation) CAPIT_QI ICT capital services, volume indices, 1995 = 100 CAPNIT_QI Non‐ICT capital services, volume indices, 1995 = 100 CAPIT_QPH ICT capital services per hour worked, 1995 reference CAPNIT_QPH Non‐ICT capital services per hour worked, 1995 reference LABHS High‐skilled labour compensation (share in total labour compensation) LABMS Medium‐skilled labour compensation (share in total labour compensation) LABLS Low‐skilled labour compensation (share in total labour compensation) LAB_QPH Labour services per hour worked, 1995 reference H_HS Hours worked by high‐skilled persons engaged (share in total hours) H_MS Hours worked by medium‐skilled persons engaged (share in total hours) H_LS Hours worked by low‐skilled persons engaged (share in total hours) H_M Hours worked by male persons engaged (share in total hours) H_F Hours worked by female persons engaged (share in total hours) H_29 Hours worked by persons engaged aged 15‐29 (share in total hours) H_49 Hours worked by persons engaged aged 30‐49 (share in total hours) H_50+ Hours worked by persons engaged aged 50 and over (share in total hours) Labour productivity variables . GO Gross output at current basic prices (in millions of local currency) II Intermediate inputs at current purchasers’ prices (in millions of local currency) VA Gross value added at current basic prices (in millions of local currency) COMP Compensation of employees (in millions of local currency) GOS Gross operating surplus (in millions of local currency) TXSP Taxes minus subsidies on production (in millions of local currency) EMP Number of persons engaged (thousands) EMPE Number of employees (thousands) H_EMP Total hours worked by persons engaged (millions) H_EMPE Total hours worked by employees (millions) GO_P Gross output, price indices, 1995 = 100 II_P Intermediate inputs, price indices, 1995 = 100 VA_P Gross value added, price indices, 1995 = 100 GO_QI Gross output, volume indices, 1995 = 100 II_QI Intermediate inputs, volume indices, 1995 = 100 VA_QI Gross value added, volume indices, 1995 = 100 LP_I Gross value added per hour worked, volume indices, 1995=100 Growth accounting variables LAB Labour compensation (in millions of local currency) CAP Capital compensation (in millions of local currency) LAB_QI Labour services, volume indices, 1995 = 100 CAP_QI Capital services, volume indices, 1995 = 100 IIE Intermediate energy inputs at current purchasers’ prices (in millions of local currency) IIM Intermediate material inputs at current purchasers’ prices (in millions of local currency) IIS Intermediate service inputs at current purchasers’ prices (in millions of local currency) IIE_QI Intermediate energy inputs, volume indices, 1995 = 100 IIM_QI Intermediate material inputs, volume indices, 1995 = 100 IIS_QI Intermediate service inputs, volume indices, 1995 = 100 VA_Q Growth rate of value added volume (% per year) VAConH Contribution of hours worked to value added growth (percentage points) VAConLC Contribution of labour composition change to value added growth (percentage points) VAConKIT Contribution of ICT capital services to output growth (percentage points) VAConKNIT Contribution of non‐ICT capital services to output growth (percentage points) VAConTFP Contribution of TFP to value added growth (percentage points) TFPva_I TFP (value added based) growth, 1995=100 GO_Q Growth rate of gross output volume (% per year) GOConII Contribution of intermediate inputs to output growth (percentage points) GOConIIM Contribution of intermediate energy inputs to output growth (percentage points) GOConIIE Contribution of intermediate material inputs to output growth (percentage points) GOConIIS Contribution of intermediate services inputs to output growth (percentage points) GOConH Contribution of hours worked to output growth (percentage points) GOConLC Contribution of labour composition change to output growth (percentage points) GOConKIT Contribution of ICT capital services to output growth (percentage points) GOConKNIT Contribution of non‐ICT capital services to output growth (percentage points) GOConTFP Contribution of TFP to output growth (percentage points) TFPgo_I TFP (gross output based) growth, 1995=100 Additional labour and capital variables CAPIT ICT capital compensation (share in total capital compensation) CAPNIT Non‐ICT capital compensation (share in total capital compensation) CAPIT_QI ICT capital services, volume indices, 1995 = 100 CAPNIT_QI Non‐ICT capital services, volume indices, 1995 = 100 CAPIT_QPH ICT capital services per hour worked, 1995 reference CAPNIT_QPH Non‐ICT capital services per hour worked, 1995 reference LABHS High‐skilled labour compensation (share in total labour compensation) LABMS Medium‐skilled labour compensation (share in total labour compensation) LABLS Low‐skilled labour compensation (share in total labour compensation) LAB_QPH Labour services per hour worked, 1995 reference H_HS Hours worked by high‐skilled persons engaged (share in total hours) H_MS Hours worked by medium‐skilled persons engaged (share in total hours) H_LS Hours worked by low‐skilled persons engaged (share in total hours) H_M Hours worked by male persons engaged (share in total hours) H_F Hours worked by female persons engaged (share in total hours) H_29 Hours worked by persons engaged aged 15‐29 (share in total hours) H_49 Hours worked by persons engaged aged 30‐49 (share in total hours) H_50+ Hours worked by persons engaged aged 50 and over (share in total hours) Open in new tab Table A2 
Industry List for Growth Accounting Variables Description . EU KLEMS Code . TOTAL INDUSTRIES TOT MARKET ECONOMY MARKT ELECTRICAL MACHINERY, POST AND COMMUNICATION SERVICES ELECOM Electrical and optical equipment 30t33 Post and telecommunications 64 GOODS PRODUCING, EXCLUDING ELECTRICAL MACHINERY GOODS TOTAL MANUFACTURING, EXCLUDING ELECTRICAL MexElec Consumer manufacturing Mcons Food products, beverages and tobacco 15t16 Textiles, textile products, leather and footwear 17t19 Manufacturing nec; recycling 36t37 Intermediate manufacturing Minter Wood and products of wood and cork 20 Pulp, paper, paper products, printing and publishing 21t22 Coke, refined petroleum products and nuclear fuel 23 Chemicals and chemical products 24 Rubber and plastics products 25 Other non‐metallic mineral products 26 Basic metals and fabricated metal products 27t28 Investment goods, excluding hightech Minves Machinery, nec 29 Transport equipment 34t35 OTHER GOODS PRODUCTION OtherG Mining and quarrying C Electricity, gas and water supply E Construction F Agriculture, hunting, forestry and fishing AtB MARKET SERVICES, EXCLUDING POST AND TELECOMMUNICATIONS MSERV DISTRIBUTION DISTR Trade 50t52 Sale, maintenance and repair of motor vehicles and motorcycles; retail sale of fuel 50 Wholesale trade and commission trade, except of motor vehicles and motorcycles 51 Retail trade, except of motor vehicles and motorcycles; repair of household goods 52 Transport and storage 60t63 FINANCE AND BUSINESS, EXCEPT REAL ESTATE FINBU Financial intermediation J Renting of m&eq and other business activities 71t74 PERSONAL SERVICES PERS Hotels and restaurants H Other community, social and personal services O Private households with employed persons P NON‐MARKET SERVICES NONMAR Public admin, education and health LtN Public admin and defence; compulsory social security L Education M Health and social work N Real estate activities 70 Description . EU KLEMS Code . TOTAL INDUSTRIES TOT MARKET ECONOMY MARKT ELECTRICAL MACHINERY, POST AND COMMUNICATION SERVICES ELECOM Electrical and optical equipment 30t33 Post and telecommunications 64 GOODS PRODUCING, EXCLUDING ELECTRICAL MACHINERY GOODS TOTAL MANUFACTURING, EXCLUDING ELECTRICAL MexElec Consumer manufacturing Mcons Food products, beverages and tobacco 15t16 Textiles, textile products, leather and footwear 17t19 Manufacturing nec; recycling 36t37 Intermediate manufacturing Minter Wood and products of wood and cork 20 Pulp, paper, paper products, printing and publishing 21t22 Coke, refined petroleum products and nuclear fuel 23 Chemicals and chemical products 24 Rubber and plastics products 25 Other non‐metallic mineral products 26 Basic metals and fabricated metal products 27t28 Investment goods, excluding hightech Minves Machinery, nec 29 Transport equipment 34t35 OTHER GOODS PRODUCTION OtherG Mining and quarrying C Electricity, gas and water supply E Construction F Agriculture, hunting, forestry and fishing AtB MARKET SERVICES, EXCLUDING POST AND TELECOMMUNICATIONS MSERV DISTRIBUTION DISTR Trade 50t52 Sale, maintenance and repair of motor vehicles and motorcycles; retail sale of fuel 50 Wholesale trade and commission trade, except of motor vehicles and motorcycles 51 Retail trade, except of motor vehicles and motorcycles; repair of household goods 52 Transport and storage 60t63 FINANCE AND BUSINESS, EXCEPT REAL ESTATE FINBU Financial intermediation J Renting of m&eq and other business activities 71t74 PERSONAL SERVICES PERS Hotels and restaurants H Other community, social and personal services O Private households with employed persons P NON‐MARKET SERVICES NONMAR Public admin, education and health LtN Public admin and defence; compulsory social security L Education M Health and social work N Real estate activities 70 Notes. EU KLEMS code based on NACE rev 1 industrial classification. Open in new tab Table A2 
Industry List for Growth Accounting Variables Description . EU KLEMS Code . TOTAL INDUSTRIES TOT MARKET ECONOMY MARKT ELECTRICAL MACHINERY, POST AND COMMUNICATION SERVICES ELECOM Electrical and optical equipment 30t33 Post and telecommunications 64 GOODS PRODUCING, EXCLUDING ELECTRICAL MACHINERY GOODS TOTAL MANUFACTURING, EXCLUDING ELECTRICAL MexElec Consumer manufacturing Mcons Food products, beverages and tobacco 15t16 Textiles, textile products, leather and footwear 17t19 Manufacturing nec; recycling 36t37 Intermediate manufacturing Minter Wood and products of wood and cork 20 Pulp, paper, paper products, printing and publishing 21t22 Coke, refined petroleum products and nuclear fuel 23 Chemicals and chemical products 24 Rubber and plastics products 25 Other non‐metallic mineral products 26 Basic metals and fabricated metal products 27t28 Investment goods, excluding hightech Minves Machinery, nec 29 Transport equipment 34t35 OTHER GOODS PRODUCTION OtherG Mining and quarrying C Electricity, gas and water supply E Construction F Agriculture, hunting, forestry and fishing AtB MARKET SERVICES, EXCLUDING POST AND TELECOMMUNICATIONS MSERV DISTRIBUTION DISTR Trade 50t52 Sale, maintenance and repair of motor vehicles and motorcycles; retail sale of fuel 50 Wholesale trade and commission trade, except of motor vehicles and motorcycles 51 Retail trade, except of motor vehicles and motorcycles; repair of household goods 52 Transport and storage 60t63 FINANCE AND BUSINESS, EXCEPT REAL ESTATE FINBU Financial intermediation J Renting of m&eq and other business activities 71t74 PERSONAL SERVICES PERS Hotels and restaurants H Other community, social and personal services O Private households with employed persons P NON‐MARKET SERVICES NONMAR Public admin, education and health LtN Public admin and defence; compulsory social security L Education M Health and social work N Real estate activities 70 Description . EU KLEMS Code . TOTAL INDUSTRIES TOT MARKET ECONOMY MARKT ELECTRICAL MACHINERY, POST AND COMMUNICATION SERVICES ELECOM Electrical and optical equipment 30t33 Post and telecommunications 64 GOODS PRODUCING, EXCLUDING ELECTRICAL MACHINERY GOODS TOTAL MANUFACTURING, EXCLUDING ELECTRICAL MexElec Consumer manufacturing Mcons Food products, beverages and tobacco 15t16 Textiles, textile products, leather and footwear 17t19 Manufacturing nec; recycling 36t37 Intermediate manufacturing Minter Wood and products of wood and cork 20 Pulp, paper, paper products, printing and publishing 21t22 Coke, refined petroleum products and nuclear fuel 23 Chemicals and chemical products 24 Rubber and plastics products 25 Other non‐metallic mineral products 26 Basic metals and fabricated metal products 27t28 Investment goods, excluding hightech Minves Machinery, nec 29 Transport equipment 34t35 OTHER GOODS PRODUCTION OtherG Mining and quarrying C Electricity, gas and water supply E Construction F Agriculture, hunting, forestry and fishing AtB MARKET SERVICES, EXCLUDING POST AND TELECOMMUNICATIONS MSERV DISTRIBUTION DISTR Trade 50t52 Sale, maintenance and repair of motor vehicles and motorcycles; retail sale of fuel 50 Wholesale trade and commission trade, except of motor vehicles and motorcycles 51 Retail trade, except of motor vehicles and motorcycles; repair of household goods 52 Transport and storage 60t63 FINANCE AND BUSINESS, EXCEPT REAL ESTATE FINBU Financial intermediation J Renting of m&eq and other business activities 71t74 PERSONAL SERVICES PERS Hotels and restaurants H Other community, social and personal services O Private households with employed persons P NON‐MARKET SERVICES NONMAR Public admin, education and health LtN Public admin and defence; compulsory social security L Education M Health and social work N Real estate activities 70 Notes. EU KLEMS code based on NACE rev 1 industrial classification. Open in new tab Table A3 
Country, Period and Variable Coverage in EU KLEMS Database Country and regions . Abbreviation . Labour productivity variables . Growth accounting variables . MFP . Labour composition . Capital composition . Intermediate input composition . Australia aus 1970 1982 1982 1970 – Austria aut 1970 1980 1980 1976 1970 Belgium bel 1970 1980 1980 1970 1980 Cyprus cyp 1995 – – – – Czech Republic cze 1995 1995 1995 1995 1995 Denmark dnk 1970 1980 1980 1970 1970 Estonia est 1995 – – – – Finland fin 1970 1970 1970 1970 1970 France fra 1970 1980 1980 1970 1978 Germany ger 1970 1970 1970 1970 1978 Greece grc 1970 – 1992 – 1995 Hungary hun 1992 1995 1995 1995 1995 Ireland irl 1970 1995 1988 1970 Italy ita 1970 1970 1970 1970 1970 Japan jpn 1973 1973 1973 1970 1973 Latvia lva 1995 – – – – Lithuania ltu 1995 – – – – Luxembourg lux 1970 1992 1992 1970 1995 Malta mlt 1995 – – – – Netherlands nld 1970 1979 1979 1970 1987 Poland pol 1995 – 1995 – 1995 Portugal prt 1970 1995 1992 1970 1977 Slovak Republic svk 1995 – 1995 – 1995 Slovenia svn 1995 1995 1995 1995 1995 Spain esp 1970 1980 1980 1970 1980 Sweden swe 1970 1993 1981 1993 1993 United Kingdom uk 1970 1970 1970 1970 1970 United States (NAICS based) usa‐naics 1977 1977 – 1970 – United States (SIC based) usa‐sic 1970 1970 1970 1970 1970 West Germany dew 1970 1970 1970 1970 1978 EU‐25 EU‐25 1995 – – – – EU‐15 EU‐15 1970 – – – – EU‐10 EU‐10 1995 – – – – EU‐15ex EU‐15ex 1970 1980 1980 1980 1980 Eurozone Euro 1970 – – – – Eurozone ex Euro‐ex 1970 1980 1980 1980 1980 Country and regions . Abbreviation . Labour productivity variables . Growth accounting variables . MFP . Labour composition . Capital composition . Intermediate input composition . Australia aus 1970 1982 1982 1970 – Austria aut 1970 1980 1980 1976 1970 Belgium bel 1970 1980 1980 1970 1980 Cyprus cyp 1995 – – – – Czech Republic cze 1995 1995 1995 1995 1995 Denmark dnk 1970 1980 1980 1970 1970 Estonia est 1995 – – – – Finland fin 1970 1970 1970 1970 1970 France fra 1970 1980 1980 1970 1978 Germany ger 1970 1970 1970 1970 1978 Greece grc 1970 – 1992 – 1995 Hungary hun 1992 1995 1995 1995 1995 Ireland irl 1970 1995 1988 1970 Italy ita 1970 1970 1970 1970 1970 Japan jpn 1973 1973 1973 1970 1973 Latvia lva 1995 – – – – Lithuania ltu 1995 – – – – Luxembourg lux 1970 1992 1992 1970 1995 Malta mlt 1995 – – – – Netherlands nld 1970 1979 1979 1970 1987 Poland pol 1995 – 1995 – 1995 Portugal prt 1970 1995 1992 1970 1977 Slovak Republic svk 1995 – 1995 – 1995 Slovenia svn 1995 1995 1995 1995 1995 Spain esp 1970 1980 1980 1970 1980 Sweden swe 1970 1993 1981 1993 1993 United Kingdom uk 1970 1970 1970 1970 1970 United States (NAICS based) usa‐naics 1977 1977 – 1970 – United States (SIC based) usa‐sic 1970 1970 1970 1970 1970 West Germany dew 1970 1970 1970 1970 1978 EU‐25 EU‐25 1995 – – – – EU‐15 EU‐15 1970 – – – – EU‐10 EU‐10 1995 – – – – EU‐15ex EU‐15ex 1970 1980 1980 1980 1980 Eurozone Euro 1970 – – – – Eurozone ex Euro‐ex 1970 1980 1980 1980 1980 Notes: This Table indicates for each country and variable the first year for which data is available in the EU KLEMS database, March 2008. ‘–’ indicates not available. See Table A1 for sets of labour productivity and growth accounting variables. Open in new tab Table A3 
Country, Period and Variable Coverage in EU KLEMS Database Country and regions . Abbreviation . Labour productivity variables . Growth accounting variables . MFP . Labour composition . Capital composition . Intermediate input composition . Australia aus 1970 1982 1982 1970 – Austria aut 1970 1980 1980 1976 1970 Belgium bel 1970 1980 1980 1970 1980 Cyprus cyp 1995 – – – – Czech Republic cze 1995 1995 1995 1995 1995 Denmark dnk 1970 1980 1980 1970 1970 Estonia est 1995 – – – – Finland fin 1970 1970 1970 1970 1970 France fra 1970 1980 1980 1970 1978 Germany ger 1970 1970 1970 1970 1978 Greece grc 1970 – 1992 – 1995 Hungary hun 1992 1995 1995 1995 1995 Ireland irl 1970 1995 1988 1970 Italy ita 1970 1970 1970 1970 1970 Japan jpn 1973 1973 1973 1970 1973 Latvia lva 1995 – – – – Lithuania ltu 1995 – – – – Luxembourg lux 1970 1992 1992 1970 1995 Malta mlt 1995 – – – – Netherlands nld 1970 1979 1979 1970 1987 Poland pol 1995 – 1995 – 1995 Portugal prt 1970 1995 1992 1970 1977 Slovak Republic svk 1995 – 1995 – 1995 Slovenia svn 1995 1995 1995 1995 1995 Spain esp 1970 1980 1980 1970 1980 Sweden swe 1970 1993 1981 1993 1993 United Kingdom uk 1970 1970 1970 1970 1970 United States (NAICS based) usa‐naics 1977 1977 – 1970 – United States (SIC based) usa‐sic 1970 1970 1970 1970 1970 West Germany dew 1970 1970 1970 1970 1978 EU‐25 EU‐25 1995 – – – – EU‐15 EU‐15 1970 – – – – EU‐10 EU‐10 1995 – – – – EU‐15ex EU‐15ex 1970 1980 1980 1980 1980 Eurozone Euro 1970 – – – – Eurozone ex Euro‐ex 1970 1980 1980 1980 1980 Country and regions . Abbreviation . Labour productivity variables . Growth accounting variables . MFP . Labour composition . Capital composition . Intermediate input composition . Australia aus 1970 1982 1982 1970 – Austria aut 1970 1980 1980 1976 1970 Belgium bel 1970 1980 1980 1970 1980 Cyprus cyp 1995 – – – – Czech Republic cze 1995 1995 1995 1995 1995 Denmark dnk 1970 1980 1980 1970 1970 Estonia est 1995 – – – – Finland fin 1970 1970 1970 1970 1970 France fra 1970 1980 1980 1970 1978 Germany ger 1970 1970 1970 1970 1978 Greece grc 1970 – 1992 – 1995 Hungary hun 1992 1995 1995 1995 1995 Ireland irl 1970 1995 1988 1970 Italy ita 1970 1970 1970 1970 1970 Japan jpn 1973 1973 1973 1970 1973 Latvia lva 1995 – – – – Lithuania ltu 1995 – – – – Luxembourg lux 1970 1992 1992 1970 1995 Malta mlt 1995 – – – – Netherlands nld 1970 1979 1979 1970 1987 Poland pol 1995 – 1995 – 1995 Portugal prt 1970 1995 1992 1970 1977 Slovak Republic svk 1995 – 1995 – 1995 Slovenia svn 1995 1995 1995 1995 1995 Spain esp 1970 1980 1980 1970 1980 Sweden swe 1970 1993 1981 1993 1993 United Kingdom uk 1970 1970 1970 1970 1970 United States (NAICS based) usa‐naics 1977 1977 – 1970 – United States (SIC based) usa‐sic 1970 1970 1970 1970 1970 West Germany dew 1970 1970 1970 1970 1978 EU‐25 EU‐25 1995 – – – – EU‐15 EU‐15 1970 – – – – EU‐10 EU‐10 1995 – – – – EU‐15ex EU‐15ex 1970 1980 1980 1980 1980 Eurozone Euro 1970 – – – – Eurozone ex Euro‐ex 1970 1980 1980 1980 1980 Notes: This Table indicates for each country and variable the first year for which data is available in the EU KLEMS database, March 2008. ‘–’ indicates not available. See Table A1 for sets of labour productivity and growth accounting variables. Open in new tab Author notes " This research was supported by the European Commission, Research Directorate General as part of the 6th Framework Programme, Priority 8, ‘Policy Support and Anticipating Scientific and Technological Needs’ and is part of the ‘EU KLEMS project on Growth and Productivity in the European Union’. The project was carried out by a consortium of 24 research institutes and national statistical institutes (see Appendix 1 for list). We are grateful to all participants in the EU KLEMS consortium for their contribution without which the construction of this database would have been impossible. The article has also benefited from comments from two anonymous referees © The Author(s). Journal compilation © Royal Economic Society 2009