Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 7-Day Trial for You or Your Team.

Learn More →

Computer Use and Earnings in Britain

Computer Use and Earnings in Britain Abstract This paper estimates various models of the effect of computer use on earnings using recent NCDS data. The cross‐section estimates are large and significant while the standard fixed effects estimates are small or insignificant. The panel estimates change considerably once we allow the coefficients to differ across individuals. Conditional on assumptions about when individuals use computers, conventional panel estimates may not identify the crucial parameters and alternative methods are needed. We conclude that there was a substantial premium associated with computer use for some individuals in the UK. The UK experienced an enormous expansion in the use of Information and Computing Technology (ICT) during the closing decades of the twentieth century. Rates of growth for investment were very high and its relative importance in total investment increased substantially. ICT became a major driver of growth in its own right and contributed directly to the growth in labour productivity. Oulton (2001) estimated that the average annual growth rates in capital services provided by computers and software in UK from 1980 to 1999 were 30% for computers and 32% for software (Table D7 p.76). Colecchia and Schreyer (2001) produce similarly large figures for the average annual percentage growth of volume investment,1 estimating that the shares of ICT equipment and software in total non‐residential investment for the UK doubled from 1980 to 1990 and tripled from 1980 to 2000.2 Oulton argues that ‘…despite its small share in GDP, ICT accounted for 13% of output growth in 1979–89 and 21% in 1989–99’ (p.31). Some of the economic issues associated with these large changes are addressed in the literature dealing with the implications of technological advance for wage inequality. These matters have been discussed recently in Acemoglu (2002), Card and DiNardo (2002) and Machin (2001). The present paper uses panel estimation to examine whether the large change in technology impacted on the wage structure in the most basic way by raising earnings. It estimates a variety of models using recently released data from the National Child Development Study (NCDS). Our cross‐section estimates are large and significant while the standard fixed effects estimates are small and insignificant for men. We show in our data that these estimates are consistent once we allow the coefficients to differ across individuals. Using simple fixed effects estimates different parameters to those estimated by our more general panel methods and does not identify the crucial parameters. We also obtain similar results from a ‘value‐added’ specification. We conclude that there has been a significant premium associated with computer use for some individuals in the UK in the last 15 years. The current consensus, post DiNardo and Pischke (1997), is that the return to using a computer must be very small, if not zero, and that the large estimates presented in the early empirical literature merely reflected the unobserved effects of ability or occupation. In our data, the earnings premium from using a computer does indeed fall if you condition on more variables but, properly measured, it still remains large. It is statistically robust and, therefore, economically important. We now move to a description of our data and then consider some of the problems that arise when estimating the impact of computing using panel data. First, we apply previous approaches to our data. These implicitly assume that the coefficient of interest is constant over time and individuals. If we examine data at the start and the end of a period of rapid growth in computer use, we might expect to see a fall in the average return to computer use if firms that make the largest gains use computers first or if the return is eroded as more workers acquire computing skills. Subsequent Sections therefore consider models where the impact of computers changes over time, first, in the same way for everyone who uses computers and, second, in different ways according to when individuals use them. 1. Data The data used in this paper are taken from NCDS. This is a study of all individuals born in Great Britain during one week in 1958. Information has been collected regularly over time and we use data from the latest two sweeps in 1991 when the respondents were 33 and 2000 when they were 42. The early sweeps give information on family background and the survey has comprehensive information on education. Later sweeps contain extensive labour market and other socio‐economic data. These data are particularly useful for our purposes as they contain information on computer use over a decade when the use of computers and their power accelerated particularly rapidly. Since the cohort are all the same age, everyone in the survey has experienced the same aggregate labour market conditions and been exposed to the same technological advances at the same point in their life cycle. Moreover, this particular cohort did not acquire computer skills at school and entered the labour market before the main micro‐computer revolution started in the early 1980s. By the same token, the effect of computers on the earnings of this cohort may be higher than in other cohorts, simply because they matured at the time when computers were becoming widely used. The variables used in our analysis are defined in the Appendix. In simple terms, we include variables corresponding to nearly all those that have been used in previous studies plus some extra ones specific to NCDS. Estimation is always constrained by the data available but we have one of the most comprehensive lists of controls available to any study in this area. The variables augment a set of standard human capital variables (schooling, work experience, tenure) with measures of attainment (highest qualification), ability (as measured by test scores at a young age), skills, occupation, industry, region, socio‐demographic characteristics (marital status, race, health) and other job characteristics. The sample comprises individuals who were full‐time employees in both 1991 and 2000. The dependent variable is the natural logarithm of real hourly wages. The present paper uses information on the use of computers at work in 1991 and 2000. We measure the impact of computer use by a dummy for ‘uses a computer at work’.3 We should note that the NCDS questions were asked explicitly at the time when the corresponding survey was undertaken. Some panel studies in the literature have difficulties with tracing computer use over time because questions on computer use only appear after the panel has been running for some time. The NCDS data are especially interesting in the present context because of the timing of the revolution in the use of computers. Computers would have had little practical impact as far as the schooling of the NCDS cohort was concerned. Some cohort members would have studied computer science in advanced courses in higher education but most would have completed higher education by the age of 23 in 1981. We could argue that the normal academic and vocational training routes for young people had a minor impact on the use of computers by the NCDS cohort members. Table 1 shows how computer use varies across our sample. The incidence of computer use increased over time for both genders and in aggregate from 60% to 75% of the sample.4 The key observations in panel estimation are the ‘changers’ (those who change from not using to using a computer or vice versa). 19% of the sample changed from not using a computer in 1991 to using one in 2000 while the converse applied to 4% of the sample. Women were more likely to use computers at each point in time, partly because of the occupations and sectors in which they work. There are, for example, higher fractions of women in secretarial and administrative jobs. 64% of women used a computer in their work in both 1991 and 2000 whereas the corresponding figure for men was only 53%. The proportion of men who did not use a computer in either time period is around 10 points higher than women at 24%. Table 1 Percentage Using a Computer at Work . Men . Women . All . Computer used in 1991 and 2000 53 64 56 Computer used in 2000 only 19 18 19 Computer used in 1991 only 4 4 4 Computer not used in 1991 or 2000 24 14 21 Number 2,707 987 3,694 . Men . Women . All . Computer used in 1991 and 2000 53 64 56 Computer used in 2000 only 19 18 19 Computer used in 1991 only 4 4 4 Computer not used in 1991 or 2000 24 14 21 Number 2,707 987 3,694 Open in new tab Table 1 Percentage Using a Computer at Work . Men . Women . All . Computer used in 1991 and 2000 53 64 56 Computer used in 2000 only 19 18 19 Computer used in 1991 only 4 4 4 Computer not used in 1991 or 2000 24 14 21 Number 2,707 987 3,694 . Men . Women . All . Computer used in 1991 and 2000 53 64 56 Computer used in 2000 only 19 18 19 Computer used in 1991 only 4 4 4 Computer not used in 1991 or 2000 24 14 21 Number 2,707 987 3,694 Open in new tab 2. Models With Constant Coefficients We have panel data showing the logarithm of earnings (Y) and whether the individual used a computer at work in each of two periods. The n individuals are indexed by i and the time periods by t. The variable Cit = 1 if individual i uses a computer in period t and 0 otherwise. The underlying model is: (1) where αi is an individual specific effect, λt and βt are parameters and uit is an error term with the familiar properties. Most previous studies have estimated the impact of computer use by applying OLS to a single cross‐section while panel studies typically report OLS results for pooled data assuming that β1 = β2. In either case, there is omitted variable bias if Cit is correlated with αi. Researchers have repeatedly sought to reduce the extent of this potential bias by adding proxies for the unobserved heterogeneity. These have included variables for occupation, industry, and region; see inter aliaKrueger (1993), DiNardo and Pischke (1997) and Oosterbeek (1997). DiNardo and Pischke (1997) and Dickerson and Green (2002) have included other job attributes such as use of tools and other skills. We employ highest qualification and ‘early test scores’, showing the separate scores on reading and mathematics tests taken at age 11.5 Economists have often interpreted these scores as measures of ability. This is debatable but they are certainly indicators of early attainment obtained largely independently of the normal system of education and public examinations. Bell (1996) was the first to use these test scores in the present context and Arabsheibani and Marin (2001) use the scores at age 7, although both studies restrict their attention solely to the 1991 data. Table 2 illustrates this methodology by starting from a basic human capital form and including successive groups of variables. The estimated impact of computing falls as more controls are added to the equation. This is exactly what we would expect if the use of computers was positively correlated with the previously omitted variables. However, the estimates for the broadest specification (labelled ‘Full’) are the same order of magnitude to those obtained in other UK studies. They indicate a premium of 13½% from the pooled data with a t‐value of over 12.6 This estimate is over twice those of Anger and Schwarz (2002) for Germany7 and Entorf and Kramarz (1997) for France8 but similar to that of Oosterbeek (1997) for The Netherlands. 9 The ‘Full’ specification arguably contains the most comprehensive list of control variables for this kind of exercise yet the effect of computer use remains large and robust. Nonetheless, this kind of argument is always open to the criticism that there may be some other omitted factor that should be included. Table 2 Estimates of the Impact of Computer Use for Different Specifications Control Variables . Basic Human Capital . Scores & Quals . Skills . SOC & SIC . Full . 1991 sample 0.221*** (0.012) 0.171*** (0.012) 0.159*** (0.012) 0.144*** (0.012) 0.118*** (0.012) R2 0.315 0.365 0.377 0.436 0.488 2000 sample 0.355*** (0.017) 0.274*** (0.018) 0.230*** (0.018) 0.167*** (0.019) 0.137*** (0.018) R2 0.307 0.350 0.378 0.462 0.504 Pooled sample 0.278*** (0.010) 0.214*** (0.011) 0.187*** (0.010) 0.155*** (0.011) 0.127*** (0.011) R2 0.334 0.377 0.402 0.465 0.509 Control Variables . Basic Human Capital . Scores & Quals . Skills . SOC & SIC . Full . 1991 sample 0.221*** (0.012) 0.171*** (0.012) 0.159*** (0.012) 0.144*** (0.012) 0.118*** (0.012) R2 0.315 0.365 0.377 0.436 0.488 2000 sample 0.355*** (0.017) 0.274*** (0.018) 0.230*** (0.018) 0.167*** (0.019) 0.137*** (0.018) R2 0.307 0.350 0.378 0.462 0.504 Pooled sample 0.278*** (0.010) 0.214*** (0.011) 0.187*** (0.010) 0.155*** (0.011) 0.127*** (0.011) R2 0.334 0.377 0.402 0.465 0.509 Notes: The table shows the estimates and the standard errors (in parenthesis). * means that the t‐value is greater than 1.64, ** 1.96 and *** 2.57. The results above refer to equations using the following sets of controls. These sets are defined in the Appendix. Basic  Basic Human Capital. Scores & Quals Basic Human Capital, Early Test Scores and Qualifications Skills  Scores & Quals and Measures of Skill SOC & SIC Skills, SOC and SIC Full  SOC & SIC, Region, Socio‐demographic and other variables All equations include a gender dummy. The pooled regressions include a cohort dummy. Open in new tab Table 2 Estimates of the Impact of Computer Use for Different Specifications Control Variables . Basic Human Capital . Scores & Quals . Skills . SOC & SIC . Full . 1991 sample 0.221*** (0.012) 0.171*** (0.012) 0.159*** (0.012) 0.144*** (0.012) 0.118*** (0.012) R2 0.315 0.365 0.377 0.436 0.488 2000 sample 0.355*** (0.017) 0.274*** (0.018) 0.230*** (0.018) 0.167*** (0.019) 0.137*** (0.018) R2 0.307 0.350 0.378 0.462 0.504 Pooled sample 0.278*** (0.010) 0.214*** (0.011) 0.187*** (0.010) 0.155*** (0.011) 0.127*** (0.011) R2 0.334 0.377 0.402 0.465 0.509 Control Variables . Basic Human Capital . Scores & Quals . Skills . SOC & SIC . Full . 1991 sample 0.221*** (0.012) 0.171*** (0.012) 0.159*** (0.012) 0.144*** (0.012) 0.118*** (0.012) R2 0.315 0.365 0.377 0.436 0.488 2000 sample 0.355*** (0.017) 0.274*** (0.018) 0.230*** (0.018) 0.167*** (0.019) 0.137*** (0.018) R2 0.307 0.350 0.378 0.462 0.504 Pooled sample 0.278*** (0.010) 0.214*** (0.011) 0.187*** (0.010) 0.155*** (0.011) 0.127*** (0.011) R2 0.334 0.377 0.402 0.465 0.509 Notes: The table shows the estimates and the standard errors (in parenthesis). * means that the t‐value is greater than 1.64, ** 1.96 and *** 2.57. The results above refer to equations using the following sets of controls. These sets are defined in the Appendix. Basic  Basic Human Capital. Scores & Quals Basic Human Capital, Early Test Scores and Qualifications Skills  Scores & Quals and Measures of Skill SOC & SIC Skills, SOC and SIC Full  SOC & SIC, Region, Socio‐demographic and other variables All equations include a gender dummy. The pooled regressions include a cohort dummy. Open in new tab Entorf and Kramarz (1997) and Anger and Schwarz (2002) have used fixed effects models to eliminate the effects of the unobservable individual characteristics on the assumption that β1 = β2. Their OLS estimates are under 6½% but have large t‐values. By contrast, the fixed effects estimates are insignificant and close to zero,10 pointing to the conclusion that the return to computing merely proxies unobserved ability. Table 3 compares cross‐section and panel estimates of the computing coefficient by gender. First differences give the same results as fixed effects in a 2‐period model and we use this model for ease of notation. The OLS and random effects results suggest that computer use increases earnings by 12% to 16% for men and by 10% to 12% for women. The panel estimate is small and insignificant for men although it remains large and significant for women. This supports the view that there is no return to computing, at least for men. This could be because the panel estimator is removing the fixed effect but we shall argue that it is because it ignores variation in the parameter values over time. We therefore consider what happens if the coefficients are the same for each individual at each point in time but differ over time, thus implicitly examining whether the returns are falling over time. Later we explore a situation where different groups are defined by their computer use over time. Table 3 OLS and Panel Estimates of the Impact of Using a Computer at Work . OLS . RE 

Random Effects . First differences . 1991 sample . 2000 sample . Pooled sample . β1 = β2 . β1≠β2 . Men (n = 2,707) Impact of computer use 0.126*** (0.014) 0.145*** (0.022) 0.135*** (0.013) 0.111*** (0.012) 0.015 (0.016) 0.044** (0.022) Change in impact 0.044* (0.025) R2 0.464 0.491 0.491 0.489 0.094 0.159 Women (n = 987) Impact of computer use 0.093*** (0.024) 0.115*** (0.033) 0.098*** (0.020) 0.097*** (0.019) 0.077*** (0.024) 0.105*** (0.034) Change in impact 0.037 (0.040) R2 0.558 0.578 0.574 0.569 0.144 0.257 . OLS . RE 

Random Effects . First differences . 1991 sample . 2000 sample . Pooled sample . β1 = β2 . β1≠β2 . Men (n = 2,707) Impact of computer use 0.126*** (0.014) 0.145*** (0.022) 0.135*** (0.013) 0.111*** (0.012) 0.015 (0.016) 0.044** (0.022) Change in impact 0.044* (0.025) R2 0.464 0.491 0.491 0.489 0.094 0.159 Women (n = 987) Impact of computer use 0.093*** (0.024) 0.115*** (0.033) 0.098*** (0.020) 0.097*** (0.019) 0.077*** (0.024) 0.105*** (0.034) Change in impact 0.037 (0.040) R2 0.558 0.578 0.574 0.569 0.144 0.257 All the estimations in Tables 3–5 use the full specification defined in Table 2 and, where appropriate, include levels and differences in the control variables. Open in new tab Table 3 OLS and Panel Estimates of the Impact of Using a Computer at Work . OLS . RE 

Random Effects . First differences . 1991 sample . 2000 sample . Pooled sample . β1 = β2 . β1≠β2 . Men (n = 2,707) Impact of computer use 0.126*** (0.014) 0.145*** (0.022) 0.135*** (0.013) 0.111*** (0.012) 0.015 (0.016) 0.044** (0.022) Change in impact 0.044* (0.025) R2 0.464 0.491 0.491 0.489 0.094 0.159 Women (n = 987) Impact of computer use 0.093*** (0.024) 0.115*** (0.033) 0.098*** (0.020) 0.097*** (0.019) 0.077*** (0.024) 0.105*** (0.034) Change in impact 0.037 (0.040) R2 0.558 0.578 0.574 0.569 0.144 0.257 . OLS . RE 

Random Effects . First differences . 1991 sample . 2000 sample . Pooled sample . β1 = β2 . β1≠β2 . Men (n = 2,707) Impact of computer use 0.126*** (0.014) 0.145*** (0.022) 0.135*** (0.013) 0.111*** (0.012) 0.015 (0.016) 0.044** (0.022) Change in impact 0.044* (0.025) R2 0.464 0.491 0.491 0.489 0.094 0.159 Women (n = 987) Impact of computer use 0.093*** (0.024) 0.115*** (0.033) 0.098*** (0.020) 0.097*** (0.019) 0.077*** (0.024) 0.105*** (0.034) Change in impact 0.037 (0.040) R2 0.558 0.578 0.574 0.569 0.144 0.257 All the estimations in Tables 3–5 use the full specification defined in Table 2 and, where appropriate, include levels and differences in the control variables. Open in new tab 3. Panel Estimates With Time Varying Coefficients 3.1. Ols and Fixed Effects The expected value of the fixed effect estimator is a weighted average of β1 and β2. It would give unbiased estimates if the impact of computing is constant over time (β1 = β2) and would be preferred to the OLS estimators if computer use is correlated with the omitted individual specific effect. The argument is less clear cut when β1≠β2. OLS applied to a single cross‐section overestimates the impact if computer use and the omitted specific effect are positively correlated. Even if this bias is small, there could be a large difference between the OLS and fixed effects estimates but it is not clear what the implications are because they are estimating two different things. The cross section estimate focuses on the coefficient in one period and the fixed effect estimate on a weighted average of the coefficients. A similar vein, the pooled and fixed effects estimators are estimating different weighted averages of the parameters. 3.2. Fixed Effects With Coefficients That Vary Over Time The differences in expected values raise the possibility that the two sets of estimates differ because the coefficients are not stable over time. This does not seem immediately plausible in the present case since the OLS estimates have similar magnitudes. Nonetheless there is a view that any advantage to using a computer will be competed away over time indicating that the value of β might fall over time. We can test for changes in the coefficients over time by estimating: (2) where ρ = λ2 − λ1. The last column of Table 3 shows the estimates for this specification and includes levels and changes for all the variables in the ‘Full’ specification. The row labelled ‘impact of computer use’ gives the estimate of β2. The women's results are consistent with the OLS results; the impact of computing is well defined at 11% and has not changed from 1991 to 2000. The men's results suggest that computing had a small impact in 2000 although the estimate is much less than the values produced by cross‐section techniques. The evidence for a change in the coefficient is weak and the resulting estimates appear implausible so we reject the hypothesis of a change in the impact of computing over time. By implication, this suggests that the large difference in the OLS and panel estimates for men is due to fixed effects rather than parameter instability. 4. Heterogeneity Across Individuals 4.1. Panel Estimation With Heterogeneity Across Individuals Computer use has often been viewed as an indicator of unobserved individual productivity or job characteristics. The main motivation for the panel model was that it removed these effects (assuming they do not change over time). If there are genuine differences across different computer users, then the panel model estimates an average of the effects for different individuals derived from those that change their computer use over time. To distinguish between the different types of computer user, we define dummy variables to identify individuals who used computers in both periods (Stay), only the first period (Leave), only the second period (Enter) and consider the ‘varying coefficients’ specification: (3) where Eliminating the fixed effect, we obtain (4) Many policy makers assume that computing skills are productive and vary across individuals. If the more skilful individuals enter the market first, we might expect . The ranking of is not clear cut. If leavers stopped using computers because they were not very good at it, but, if leavers moved up promotion ladders, may be relatively large. The expected value of the normal fixed effects estimator is a weighted average of the parameters for the two sets of movers ( and ) minus a term reflecting the change in the parameters for the stayers. Panel estimation will provide unbiased estimates under the maintained hypothesis that . If there are more subtle effects present, standard panel estimators may, by chance, produce a close estimate of the impact of computer use across the whole population but they are unlikely to be good indicators of the premium for stayers. If there is no omitted variable bias, the OLS estimator for period t measures the impact of computing for users in that period. (It is an average of the returns for the two types of computer user in that period, weighted by the relative proportions of each type.) The average premium estimated, albeit with error, by OLS is a parameter of considerable interest in contrast to the parameter estimated by the panel estimator. Table 4 presents the panel estimates for the varying coefficients model in (3). There has been no significant change over time for either male or female stayers. The null hypothesis that is formally rejected for both genders. There was no significant impact on earnings for male enterers in 2000 but there was a large positive impact for women. Male leavers received a large significant premium in contrast to the women. These results are consistent with the view that male stayers earned a stable return, leavers received a return of 9% from using computers but these returns are not available to enterers (i.e. , βL = 0.085 and βE = 0). A different argument may apply to women. It may be that female enterers received a return of 14% or more but these returns are not available to leavers.11 This interpretation highlights our main point that we want to know the individual values of and . We contend that our cross‐section results are high because and are large. Table 4 Estimates of the Impact of Using a Computer at Work with Heterogeneity . First differences: Varying coefficients . Value‐Added . Men . Women . Men . Women . Computer used in 1991 and 2000 (Stay) 0.039 (0.025) 0.049 (0.041) 0.134*** (0.022) 0.089*** (0.032) Computer used in 2000 only (Enter) (βE) 0.015 (0.025) 0.130*** (0.041) 0.058*** (0.022) 0.128*** (0.033) Computer used in 1991 only (Leave) (βL) 0.085** (0.039) 0.009 (0.062) – – R2 0.161 0.258 0.575 0.671 . First differences: Varying coefficients . Value‐Added . Men . Women . Men . Women . Computer used in 1991 and 2000 (Stay) 0.039 (0.025) 0.049 (0.041) 0.134*** (0.022) 0.089*** (0.032) Computer used in 2000 only (Enter) (βE) 0.015 (0.025) 0.130*** (0.041) 0.058*** (0.022) 0.128*** (0.033) Computer used in 1991 only (Leave) (βL) 0.085** (0.039) 0.009 (0.062) – – R2 0.161 0.258 0.575 0.671 Open in new tab Table 4 Estimates of the Impact of Using a Computer at Work with Heterogeneity . First differences: Varying coefficients . Value‐Added . Men . Women . Men . Women . Computer used in 1991 and 2000 (Stay) 0.039 (0.025) 0.049 (0.041) 0.134*** (0.022) 0.089*** (0.032) Computer used in 2000 only (Enter) (βE) 0.015 (0.025) 0.130*** (0.041) 0.058*** (0.022) 0.128*** (0.033) Computer used in 1991 only (Leave) (βL) 0.085** (0.039) 0.009 (0.062) – – R2 0.161 0.258 0.575 0.671 . First differences: Varying coefficients . Value‐Added . Men . Women . Men . Women . Computer used in 1991 and 2000 (Stay) 0.039 (0.025) 0.049 (0.041) 0.134*** (0.022) 0.089*** (0.032) Computer used in 2000 only (Enter) (βE) 0.015 (0.025) 0.130*** (0.041) 0.058*** (0.022) 0.128*** (0.033) Computer used in 1991 only (Leave) (βL) 0.085** (0.039) 0.009 (0.062) – – R2 0.161 0.258 0.575 0.671 Open in new tab 4.2. Modelling Heterogeneity Across Individuals We finally consider two extensions of our heterogeneity model. As Anger and Schwarz (2002) have observed, computers per se cannot affect earnings before they were actually used. More generally, any effect that computer use in one period has on earnings in a different period might suggest the role of unobserved factors. This insight underlies the reduced form approach of Jakubson (1991) that we use below. Our second model is taken from the education production function literature where an outcome typically depends on current regressors and values of variables and unobservables that reflect an individual's history. A common approach is to use a value‐added specification that proxies the individual's history by a lagged dependent variable. 4.3. Reduced Form Estimation If the fixed effect is correlated with current computer use, then in general it is correlated with computer use in each period. Generalising Jakubson (1991), we can write (5) where vi is orthogonal to the regressors by construction. Substituting into (3), we obtain the reduced form equations: (6) where These equations can be estimated for each time period or both periods together. In either case, the key hypothesis that and is accepted for women and rejected for men. Using both periods, the following joint hypotheses are accepted at the 5% level: These tests support the previous interpretation of the first difference results with heterogeneity. The one exception is that the coefficient of Enter for men should be interpreted as the difference between the coefficients for each period. Since  > 0, Table 4 implies that γE > 0 and for men. These tests also show that there are no effects from unobservables for male leavers and for women. Table 5 reports the OLS estimates of our reduced form heterogeneity model for each time period. Male enterers have a significant earnings premium in 1991 but leavers do not in 2000. We explain the first result by the unobserved ability of enterers and argue that male leavers have lower ability than the others because they were unable to take advantage of the opportunities offered by the new technology. Since there is no ability premium for leavers in 2000, the extra earnings that leavers receive in 1991 is solely due their use of computers. Applying this interpretation, the estimates of the coefficients for Enter and Leave for 1991 suggest, respectively, returns of about 8% to ability for enterers and 11% to computer use for leavers. The return for male stayers is statistically equal to the sum of these two effects in 1991 and the estimates do not change when this restriction is imposed. This is compatible with a common ability effect for those who use computers in the second period (γS = γE) and a common return to computer use for users in the first period (). Table 5 Reduced Form Estimates with Heterogeneity . Men . Women . 1991 sample . 2000 sample . 1991 sample . 2000 sample . Computer used in 1991 and 2000 (Stay) 0.183*** (0.018) 0.214*** (0.025) 0.123*** (0.032) 0.141*** (0.041) Computer used in 2000 only (Enter) 0.079*** (0.019) 0.093*** (0.024) 0.046 (0.033) 0.141*** (0.042) Computer used in 1991 only (Leave) 0.109*** (0.030) 0.054 (0.039) 0.136*** (0.053) 0.088 (0.062) R2 0.469 0.497 0.559 0.579 . Men . Women . 1991 sample . 2000 sample . 1991 sample . 2000 sample . Computer used in 1991 and 2000 (Stay) 0.183*** (0.018) 0.214*** (0.025) 0.123*** (0.032) 0.141*** (0.041) Computer used in 2000 only (Enter) 0.079*** (0.019) 0.093*** (0.024) 0.046 (0.033) 0.141*** (0.042) Computer used in 1991 only (Leave) 0.109*** (0.030) 0.054 (0.039) 0.136*** (0.053) 0.088 (0.062) R2 0.469 0.497 0.559 0.579 Open in new tab Table 5 Reduced Form Estimates with Heterogeneity . Men . Women . 1991 sample . 2000 sample . 1991 sample . 2000 sample . Computer used in 1991 and 2000 (Stay) 0.183*** (0.018) 0.214*** (0.025) 0.123*** (0.032) 0.141*** (0.041) Computer used in 2000 only (Enter) 0.079*** (0.019) 0.093*** (0.024) 0.046 (0.033) 0.141*** (0.042) Computer used in 1991 only (Leave) 0.109*** (0.030) 0.054 (0.039) 0.136*** (0.053) 0.088 (0.062) R2 0.469 0.497 0.559 0.579 . Men . Women . 1991 sample . 2000 sample . 1991 sample . 2000 sample . Computer used in 1991 and 2000 (Stay) 0.183*** (0.018) 0.214*** (0.025) 0.123*** (0.032) 0.141*** (0.041) Computer used in 2000 only (Enter) 0.079*** (0.019) 0.093*** (0.024) 0.046 (0.033) 0.141*** (0.042) Computer used in 1991 only (Leave) 0.109*** (0.030) 0.054 (0.039) 0.136*** (0.053) 0.088 (0.062) R2 0.469 0.497 0.559 0.579 Open in new tab If we accept that the wage premium given to the enterers is purely a return to ability, then the ability effect is approximately the same in 2000 at 9%. If we then assume that the return for stayers is the sum of an ability effect equal to that of enterers (γS = γE) and an enduring computer effect following early use (), then the second period return to computer use for male stayers is about 13% using the 2000 estimates. (Using both time periods with the restrictions imposed gives a return to unobservables of 8½% and to computer use of 12%). The estimates for stayers and enterers are similar for the 1991 and 2000 samples in Table 5 so we would not expect to see any significant changes over time in Table 4. The estimates in Tables 4 and 5 now suggest a return to computer use for particular groups of men of at least 9% and possibly as much as 13%. These are only slightly lower than the magnitudes suggested by Table 3. By contrast, using a computer only has an impact for women in the period when computers are used. Female enterers receive no significant increase in earnings in 1991 and the estimates for the female stayers and leavers are statistically the same in 1991. The two implied restrictions are accepted at the 5% level. The corresponding results also hold for 2000 so Table 3 reports appropriate estimates for women. Alternatively, using both samples with these restrictions imposed gives a return of 11%. (This differs from the pooled estimate because the coefficients of the control variables are allowed to differ in each period.) Table 3 suggests a return to computer use for women of 10% to 12% compared with 9% to 13% for men in Table 5. Male stayers appear to earn far more than women. Our results suggest that men receive an additional premium for computer use. Previous writers on this subject would probably interpret this as a return to ability but the literature on gender differentials would regard this as an unexplained difference. 4.4. Value Added Specification We conclude by estimating the value‐added (VA) model: (7) The value added model adds previous earnings to the list of regressors in the ‘contemporaneous’ specification. This addition is taken as a sufficient statistic for observed factors which have changed over time before period 2 and unobserved individual endowments like ability which may affect earnings outcomes. The VA specification is interesting in this context as it is precisely these kinds of omitted factors that researchers mention when discussing ‘over‐estimates’ of the impact of computers on earnings. Although the VA model is considered preferable to the contemporaneous specification (Hanushek, 1996, 2003), Todd and Wolpin (2003) show that it relies on strong assumptions. Any fixed effect is only removed if the coefficients of the regressors and the fixed effect decay geometrically over time at the same rate as δ. Further, the estimates of VA specification are sensitive to the omission of regressors although, for obvious reasons, we regard missing regressors as a less important problem in this study. Table 4 shows the results of this estimation. The impact of ICT for stayers is similar to the previous estimates. It is large (14% for men and 9% for women) and significant. Enterers received higher earnings, 5% more for men and 14% for women. If lagged earnings are merely another proxy for unobservables, then we might want to derive the actual return to computer use for men from the difference between the estimates for stayers and enterers. In which case, the return falls but is still 8% for men. 5. Conclusion Our paper presents firm evidence that there was a large earnings premium to computer use in the UK. Over time, there have been repeated discussions of how to interpret the impact of computer use. Our paper focuses on DiNardo and Pischke's (1997) argument that any estimate merely measures unobservable job or individual characteristics and is likely to disappear as other factors are considered. This argument does not seem plausible in our case because we have many controls for ability, occupation, industry and skills, yet our cross‐section estimates are still about 13% and 14% for men. There are a range of econometric problems that arise in cross‐sectional models but Dolton and Makepeace (2002) report estimates from matching and selection models that are of a similar order of magnitude to those presented in this paper. Nonetheless, the fixed effects estimates of Entorf and Kramarz (1997) are much lower than their cross‐section estimates and insignificant, adding considerable weight to the argument that impact of computer use proxies unmeasured ability. We replicate this finding but seek to reconcile the cross‐section and fixed effects estimates by considering more complex panel models that allow the impact of computing to vary over time. We find no conclusive evidence that there was a uniform change in the value of the computing coefficient over time and argue instead that the ‘return’ to computing varied across individuals. We can determine whether an individual used a computer in both periods (stayers), the first period only (leavers), the second period only (enterers) or not at all. If we allow the coefficients to vary across individuals, the panel estimates for male leavers give an earnings premium of 9%. The simplest interpretation is that this boost to earnings was only available while these men worked with computers. Given that fixed effects are eliminated and the wide range of controls employed, it is not plausible to attribute this to unobserved ability. Indeed, the growth in this group's average earnings over the sample period was less than fifth of that of other men so they were not individuals in high demand who were moving on to better jobs. Formally, the panel model only identifies the change in coefficients for each group. Our ‘reduced form’ estimates for men show that the impact of computer use remained constant over time for stayers and enterers but was zero in 2000 for leavers. They are consistent with a return to early computer use of 11% to 13% and return to unobservables of 8% to 9%. Computers only affect women's earnings when they are used and our preferred estimates suggest a return of 10% to 12% for women with no return to unobservables. Our estimates of the value added model are slightly lower for men but similar for women. We conclude that there is a well defined earnings premium associated with computer use. Appendix Appendix: Variable Definitions This Appendix defines the control variables used in the regressions. The regressions also include dummies for missing values on each regressor although some of these are omitted because there are no missing values for that variable in the sample. Dependent Variable. Log of real hourly earnings, which is gross pay before deductions divided by usual hours of work. The nominal data were recoded to January 2000 prices. Full details of the variable are included in Dearden et al. (2003). Computer Use Dummies. Two binary variables taking the value 1 in 1991 and 2000, respectively, if the respondent uses at work a computer or word processor with a TV type screen (usually known as a VDU) and if the respondent uses a computer at work. Sets of Explanatory Variables Basic Human Capital Years of schooling, Years of work experience, Tenure with current employer. Qualifications Highest qualification achieved whether vocational or academic; dummies for 5 levels: NVQ level 1, NVQ level 2, NVQ level 3, NVQ level 4 or 5; omitted group ‐ No qualifications. Early Test Scores Dummies for quintile scores on tests at age 11; 5 dummies for reading and 5 dummies for mathematics: omitted groups – Bottom quintiles for reading and mathematics. Skills in 1991 survey Dummies for ‘good’ at 6 types of skill: communication (speaking clearly), carrying out mathematics, giving advice and support, using tools, caring, finance and accounts. Skills in 2000 survey Dummies for ‘good’ at 8 types of skill: communication, numbers and calculation, team work, learning new skills, problem solving, using tools, caring, finance and accounts. SOC Occupation, Dummies for 9 Major SOC Groups: managers and administrators, professional, associate professional and technical, clerical and secretarial, craft and related, personal and protective services, sales, other; omitted group – plant and machine operatives. SIC Industry, Dummies for 13 Major SIC Groups: Farming, Manufacturing, Construction, Sales (wholesale, retail and repair), Transport and communications, Financial intermediation, Real estate, renting and business activities, Public administration and defence, Education, Health and social work, Other community, social and personal services, Other industries (other jobs, mining, electricity, gas and water supply); omitted group – Hotels and restaurants. Region Dummies for 12 regions: London, East Anglia, South East, South West, East Midlands, West Midlands, Yorkshire and Humberside, North West, North, Scotland, Other: omitted group – Wales. Socio‐demographic Dummies for: Married, Non‐white, Long standing illness limits daily activities. Other Dummies for other job characteristics: Firm size (number of employees) 5 categories – 10–24, 25–99, 100–499, 500 or more (2000 survey); omitted group 1–9 (1–10 employees 1991 survey); Temporary job, Union member. Footnotes 1 " These are 28% (1980–90) and 25% (1990–2000) for IT equipment although their rates for software growth are lower at 27% (1980–90) and 10% (1990–2000). 2 " Other countries reacted differently. The UK share tripled from 1980 to 2000 but started from a low base (5% of total investment). The US share doubled but from a much higher base (15%). The German share only grew by a quarter from 12% to end slightly above the UK figure. 3 " We can also measure computer competence and intensity of computer use but this paper concentrates on the basic impact of computing. We are working on another paper that examines the effect of frequency and sophistication of use across different cohorts and countries. 4 " The use of a panel sample may mean that computer use is over‐represented because respondents have to be in work at both points in time. Our samples also exclude older people who are less likely to use computers. 5 " We have also estimated the models using scores at age 7. The estimates for the computing coefficients only change at the 3rd decimal place and their significance is not affected. 6 " Although the incidence of computer use is high, there seems to be no evidence of a saturation effect. 7 " 2½% for men and 6½% for women using pooled data for 1985–99. 8 " 3% using pooled data for 1985–7 and the significant ‘high’ autonomy figures (see Table 2). The ‘OLS’ figures use a maximum likelihood procedure to allow for grouping in the wage data. The fixed effects estimates are comparable. 9 " 12% in 1993. 10 " Oosterbeek's (1997) fixed effect estimate (10%) is only slightly smaller with a t‐value of 2.7. He assumed that computers were not used in his previous period (1983) and did not include any control variables in the panel model. 11 " There were only 38 female leavers so our estimate of is, at best, imprecise. References Acemoglu , D. ( 2002 ). ‘ Technical change, inequality and the labour market ’, Journal of Economic Literature , vol. 40 ( 1 ), pp. 7 – 72 . Google Scholar Crossref Search ADS WorldCat Anger , S. and Schwarz , J. ( 2002 ). ‘ Does future PC use determine our wages today: evidence from the German panel data ’, IZA Discussion Paper No.429. Arabsheibani , G. R. and Marin , A. ( 2001 ). ‘ If not computers then what? Returns to computer use in the UK revisited ’, School of Management and Business , Aberystwyth, mimeo. Bell , B. ( 1996 ). ‘ Skill‐biased technical change and wages: evidence from a longitudinal data set ’, Nuffield College , Oxford, mimeo. Card , D. and DiNardo , J. E. ( 2002 ). ‘ Skill‐biased technological change and rising wage inequality: some problems and puzzles ’, Journal of Labor Economics , vol. 20 . OpenURL Placeholder Text WorldCat Colecchia , A. and Schreyer , P. ( 2001 ). ‘ ICT investment and economic growth in the 1990s: is the United States a unique case? A comparative study of nine OECD countries ’, OECD Directorate for Science, Technology and Industry working paper 2001/7. Dearden , L. , Goodman , A. and Saunders , P. ( 2003 ). ‘Income and living standards’, in ( E. Ferri, J. Bynner and M. Wadsworth), Changing Britain, Changing Lives , ch. 6, London: Institute of Education, University of London. Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Dickerson , A. and Green F. ( 2002 ). ‘ The growth and valuation of generic skills ’, University of Kent, mimeo. DiNardo , J. and Pischke , J. ( 1997 ). ‘ The returns to computer use revisited: have pencils changed the wage structure too? ’, Quarterly Journal of Economics , vol. 112 , pp. 291 – 303 . Google Scholar Crossref Search ADS WorldCat Dolton , P and Makepeace , G. ( 2002 ). ‘ Returns to computer use: an empirical analysis for the UK ’, University of Newcastle, mimeo. Entorf , H. and Kramarz , F. ( 1997 ). ‘ Does unmeasured ability explain the higher wages of new technology workers? ’, European Economic Review , vol. 41 , pp. 1489 – 510 . Google Scholar Crossref Search ADS WorldCat Hanushek , E. A. ( 1996 ). ‘School resources and student performance’, in ( G. Burtless, ed.), Does Money Matter? The Effect of School Resources on Student Achievement and Adult Success , Washington, DC: Brookings Institution . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Hanushek , E. A. ( 2003 ). ‘ The failure of input‐based schooling policies ’, Economic Journal , vol. 113 , pp. F64 – 98 . Google Scholar Crossref Search ADS WorldCat Jakubson , G. ( 1991 ). ‘ Estmation and testing of the union wage effect using panel data ’, The Review of Economic Studies , vol. 58 , pp. 971 – 91. Google Scholar Crossref Search ADS WorldCat Krueger , A. ( 1993 ), ‘ How computers have changed the wage structure: evidence from microdata, 1984–1989 ’, Quarterly Journal of Economics , vol. 108 , pp. 33 – 60 . Google Scholar Crossref Search ADS WorldCat Machin , S. ( 2001 ). ‘ The changing nature of labour demand in the new economy and skill‐biased technological change ’, Oxford Bulletin of Economics and Statistics , vol. 63 , pp. 753 – 76. Google Scholar Crossref Search ADS WorldCat Oosterbeek , H. ( 1997 ). ‘ The returns from computer use: a simple test on the productivity interpretation ’, Economics Letters , vol. 55 , pp. 273 – 7. Google Scholar Crossref Search ADS WorldCat Oulton , N. ( 2001 ). ‘ ICT and productivity growth in the United Kingdom ’, Bank of England Working Paper no.140. Todd , P. and Wolpin , K. ( 2003 ). ‘ On the specification and estimation of the production function for cognitive achievement ’, Economic Journal , vol. 113 , pp. F3 – 33 . Google Scholar Crossref Search ADS WorldCat Author notes " The authors would like to thank Martyn Andrews, Ian Preston, Karl Taylor, two anonymous referees and seminar participants at Cardiff, ESPE 2003, Leicester, Manchester, LSE and RES 2003 for their comments on earlier versions of this paper. © Royal Economic Society 2004 http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png The Economic Journal Oxford University Press

Computer Use and Earnings in Britain

Dolton,, Peter; Makepeace,, Gerry
The Economic Journal , Volume 114 (494) – Mar 1, 2004
Download PDF
13 pages

Loading next page...
 
/lp/oxford-university-press/computer-use-and-earnings-in-britain-15BzDJWoZA

References (33)

Publisher
Oxford University Press
Copyright
© Royal Economic Society 2004
ISSN
0013-0133
eISSN
1468-0297
DOI
10.1111/j.0013-0133.2004.00201.x
Publisher site
See Article on Publisher Site

Abstract

Abstract This paper estimates various models of the effect of computer use on earnings using recent NCDS data. The cross‐section estimates are large and significant while the standard fixed effects estimates are small or insignificant. The panel estimates change considerably once we allow the coefficients to differ across individuals. Conditional on assumptions about when individuals use computers, conventional panel estimates may not identify the crucial parameters and alternative methods are needed. We conclude that there was a substantial premium associated with computer use for some individuals in the UK. The UK experienced an enormous expansion in the use of Information and Computing Technology (ICT) during the closing decades of the twentieth century. Rates of growth for investment were very high and its relative importance in total investment increased substantially. ICT became a major driver of growth in its own right and contributed directly to the growth in labour productivity. Oulton (2001) estimated that the average annual growth rates in capital services provided by computers and software in UK from 1980 to 1999 were 30% for computers and 32% for software (Table D7 p.76). Colecchia and Schreyer (2001) produce similarly large figures for the average annual percentage growth of volume investment,1 estimating that the shares of ICT equipment and software in total non‐residential investment for the UK doubled from 1980 to 1990 and tripled from 1980 to 2000.2 Oulton argues that ‘…despite its small share in GDP, ICT accounted for 13% of output growth in 1979–89 and 21% in 1989–99’ (p.31). Some of the economic issues associated with these large changes are addressed in the literature dealing with the implications of technological advance for wage inequality. These matters have been discussed recently in Acemoglu (2002), Card and DiNardo (2002) and Machin (2001). The present paper uses panel estimation to examine whether the large change in technology impacted on the wage structure in the most basic way by raising earnings. It estimates a variety of models using recently released data from the National Child Development Study (NCDS). Our cross‐section estimates are large and significant while the standard fixed effects estimates are small and insignificant for men. We show in our data that these estimates are consistent once we allow the coefficients to differ across individuals. Using simple fixed effects estimates different parameters to those estimated by our more general panel methods and does not identify the crucial parameters. We also obtain similar results from a ‘value‐added’ specification. We conclude that there has been a significant premium associated with computer use for some individuals in the UK in the last 15 years. The current consensus, post DiNardo and Pischke (1997), is that the return to using a computer must be very small, if not zero, and that the large estimates presented in the early empirical literature merely reflected the unobserved effects of ability or occupation. In our data, the earnings premium from using a computer does indeed fall if you condition on more variables but, properly measured, it still remains large. It is statistically robust and, therefore, economically important. We now move to a description of our data and then consider some of the problems that arise when estimating the impact of computing using panel data. First, we apply previous approaches to our data. These implicitly assume that the coefficient of interest is constant over time and individuals. If we examine data at the start and the end of a period of rapid growth in computer use, we might expect to see a fall in the average return to computer use if firms that make the largest gains use computers first or if the return is eroded as more workers acquire computing skills. Subsequent Sections therefore consider models where the impact of computers changes over time, first, in the same way for everyone who uses computers and, second, in different ways according to when individuals use them. 1. Data The data used in this paper are taken from NCDS. This is a study of all individuals born in Great Britain during one week in 1958. Information has been collected regularly over time and we use data from the latest two sweeps in 1991 when the respondents were 33 and 2000 when they were 42. The early sweeps give information on family background and the survey has comprehensive information on education. Later sweeps contain extensive labour market and other socio‐economic data. These data are particularly useful for our purposes as they contain information on computer use over a decade when the use of computers and their power accelerated particularly rapidly. Since the cohort are all the same age, everyone in the survey has experienced the same aggregate labour market conditions and been exposed to the same technological advances at the same point in their life cycle. Moreover, this particular cohort did not acquire computer skills at school and entered the labour market before the main micro‐computer revolution started in the early 1980s. By the same token, the effect of computers on the earnings of this cohort may be higher than in other cohorts, simply because they matured at the time when computers were becoming widely used. The variables used in our analysis are defined in the Appendix. In simple terms, we include variables corresponding to nearly all those that have been used in previous studies plus some extra ones specific to NCDS. Estimation is always constrained by the data available but we have one of the most comprehensive lists of controls available to any study in this area. The variables augment a set of standard human capital variables (schooling, work experience, tenure) with measures of attainment (highest qualification), ability (as measured by test scores at a young age), skills, occupation, industry, region, socio‐demographic characteristics (marital status, race, health) and other job characteristics. The sample comprises individuals who were full‐time employees in both 1991 and 2000. The dependent variable is the natural logarithm of real hourly wages. The present paper uses information on the use of computers at work in 1991 and 2000. We measure the impact of computer use by a dummy for ‘uses a computer at work’.3 We should note that the NCDS questions were asked explicitly at the time when the corresponding survey was undertaken. Some panel studies in the literature have difficulties with tracing computer use over time because questions on computer use only appear after the panel has been running for some time. The NCDS data are especially interesting in the present context because of the timing of the revolution in the use of computers. Computers would have had little practical impact as far as the schooling of the NCDS cohort was concerned. Some cohort members would have studied computer science in advanced courses in higher education but most would have completed higher education by the age of 23 in 1981. We could argue that the normal academic and vocational training routes for young people had a minor impact on the use of computers by the NCDS cohort members. Table 1 shows how computer use varies across our sample. The incidence of computer use increased over time for both genders and in aggregate from 60% to 75% of the sample.4 The key observations in panel estimation are the ‘changers’ (those who change from not using to using a computer or vice versa). 19% of the sample changed from not using a computer in 1991 to using one in 2000 while the converse applied to 4% of the sample. Women were more likely to use computers at each point in time, partly because of the occupations and sectors in which they work. There are, for example, higher fractions of women in secretarial and administrative jobs. 64% of women used a computer in their work in both 1991 and 2000 whereas the corresponding figure for men was only 53%. The proportion of men who did not use a computer in either time period is around 10 points higher than women at 24%. Table 1 Percentage Using a Computer at Work . Men . Women . All . Computer used in 1991 and 2000 53 64 56 Computer used in 2000 only 19 18 19 Computer used in 1991 only 4 4 4 Computer not used in 1991 or 2000 24 14 21 Number 2,707 987 3,694 . Men . Women . All . Computer used in 1991 and 2000 53 64 56 Computer used in 2000 only 19 18 19 Computer used in 1991 only 4 4 4 Computer not used in 1991 or 2000 24 14 21 Number 2,707 987 3,694 Open in new tab Table 1 Percentage Using a Computer at Work . Men . Women . All . Computer used in 1991 and 2000 53 64 56 Computer used in 2000 only 19 18 19 Computer used in 1991 only 4 4 4 Computer not used in 1991 or 2000 24 14 21 Number 2,707 987 3,694 . Men . Women . All . Computer used in 1991 and 2000 53 64 56 Computer used in 2000 only 19 18 19 Computer used in 1991 only 4 4 4 Computer not used in 1991 or 2000 24 14 21 Number 2,707 987 3,694 Open in new tab 2. Models With Constant Coefficients We have panel data showing the logarithm of earnings (Y) and whether the individual used a computer at work in each of two periods. The n individuals are indexed by i and the time periods by t. The variable Cit = 1 if individual i uses a computer in period t and 0 otherwise. The underlying model is: (1) where αi is an individual specific effect, λt and βt are parameters and uit is an error term with the familiar properties. Most previous studies have estimated the impact of computer use by applying OLS to a single cross‐section while panel studies typically report OLS results for pooled data assuming that β1 = β2. In either case, there is omitted variable bias if Cit is correlated with αi. Researchers have repeatedly sought to reduce the extent of this potential bias by adding proxies for the unobserved heterogeneity. These have included variables for occupation, industry, and region; see inter aliaKrueger (1993), DiNardo and Pischke (1997) and Oosterbeek (1997). DiNardo and Pischke (1997) and Dickerson and Green (2002) have included other job attributes such as use of tools and other skills. We employ highest qualification and ‘early test scores’, showing the separate scores on reading and mathematics tests taken at age 11.5 Economists have often interpreted these scores as measures of ability. This is debatable but they are certainly indicators of early attainment obtained largely independently of the normal system of education and public examinations. Bell (1996) was the first to use these test scores in the present context and Arabsheibani and Marin (2001) use the scores at age 7, although both studies restrict their attention solely to the 1991 data. Table 2 illustrates this methodology by starting from a basic human capital form and including successive groups of variables. The estimated impact of computing falls as more controls are added to the equation. This is exactly what we would expect if the use of computers was positively correlated with the previously omitted variables. However, the estimates for the broadest specification (labelled ‘Full’) are the same order of magnitude to those obtained in other UK studies. They indicate a premium of 13½% from the pooled data with a t‐value of over 12.6 This estimate is over twice those of Anger and Schwarz (2002) for Germany7 and Entorf and Kramarz (1997) for France8 but similar to that of Oosterbeek (1997) for The Netherlands. 9 The ‘Full’ specification arguably contains the most comprehensive list of control variables for this kind of exercise yet the effect of computer use remains large and robust. Nonetheless, this kind of argument is always open to the criticism that there may be some other omitted factor that should be included. Table 2 Estimates of the Impact of Computer Use for Different Specifications Control Variables . Basic Human Capital . Scores & Quals . Skills . SOC & SIC . Full . 1991 sample 0.221*** (0.012) 0.171*** (0.012) 0.159*** (0.012) 0.144*** (0.012) 0.118*** (0.012) R2 0.315 0.365 0.377 0.436 0.488 2000 sample 0.355*** (0.017) 0.274*** (0.018) 0.230*** (0.018) 0.167*** (0.019) 0.137*** (0.018) R2 0.307 0.350 0.378 0.462 0.504 Pooled sample 0.278*** (0.010) 0.214*** (0.011) 0.187*** (0.010) 0.155*** (0.011) 0.127*** (0.011) R2 0.334 0.377 0.402 0.465 0.509 Control Variables . Basic Human Capital . Scores & Quals . Skills . SOC & SIC . Full . 1991 sample 0.221*** (0.012) 0.171*** (0.012) 0.159*** (0.012) 0.144*** (0.012) 0.118*** (0.012) R2 0.315 0.365 0.377 0.436 0.488 2000 sample 0.355*** (0.017) 0.274*** (0.018) 0.230*** (0.018) 0.167*** (0.019) 0.137*** (0.018) R2 0.307 0.350 0.378 0.462 0.504 Pooled sample 0.278*** (0.010) 0.214*** (0.011) 0.187*** (0.010) 0.155*** (0.011) 0.127*** (0.011) R2 0.334 0.377 0.402 0.465 0.509 Notes: The table shows the estimates and the standard errors (in parenthesis). * means that the t‐value is greater than 1.64, ** 1.96 and *** 2.57. The results above refer to equations using the following sets of controls. These sets are defined in the Appendix. Basic  Basic Human Capital. Scores & Quals Basic Human Capital, Early Test Scores and Qualifications Skills  Scores & Quals and Measures of Skill SOC & SIC Skills, SOC and SIC Full  SOC & SIC, Region, Socio‐demographic and other variables All equations include a gender dummy. The pooled regressions include a cohort dummy. Open in new tab Table 2 Estimates of the Impact of Computer Use for Different Specifications Control Variables . Basic Human Capital . Scores & Quals . Skills . SOC & SIC . Full . 1991 sample 0.221*** (0.012) 0.171*** (0.012) 0.159*** (0.012) 0.144*** (0.012) 0.118*** (0.012) R2 0.315 0.365 0.377 0.436 0.488 2000 sample 0.355*** (0.017) 0.274*** (0.018) 0.230*** (0.018) 0.167*** (0.019) 0.137*** (0.018) R2 0.307 0.350 0.378 0.462 0.504 Pooled sample 0.278*** (0.010) 0.214*** (0.011) 0.187*** (0.010) 0.155*** (0.011) 0.127*** (0.011) R2 0.334 0.377 0.402 0.465 0.509 Control Variables . Basic Human Capital . Scores & Quals . Skills . SOC & SIC . Full . 1991 sample 0.221*** (0.012) 0.171*** (0.012) 0.159*** (0.012) 0.144*** (0.012) 0.118*** (0.012) R2 0.315 0.365 0.377 0.436 0.488 2000 sample 0.355*** (0.017) 0.274*** (0.018) 0.230*** (0.018) 0.167*** (0.019) 0.137*** (0.018) R2 0.307 0.350 0.378 0.462 0.504 Pooled sample 0.278*** (0.010) 0.214*** (0.011) 0.187*** (0.010) 0.155*** (0.011) 0.127*** (0.011) R2 0.334 0.377 0.402 0.465 0.509 Notes: The table shows the estimates and the standard errors (in parenthesis). * means that the t‐value is greater than 1.64, ** 1.96 and *** 2.57. The results above refer to equations using the following sets of controls. These sets are defined in the Appendix. Basic  Basic Human Capital. Scores & Quals Basic Human Capital, Early Test Scores and Qualifications Skills  Scores & Quals and Measures of Skill SOC & SIC Skills, SOC and SIC Full  SOC & SIC, Region, Socio‐demographic and other variables All equations include a gender dummy. The pooled regressions include a cohort dummy. Open in new tab Entorf and Kramarz (1997) and Anger and Schwarz (2002) have used fixed effects models to eliminate the effects of the unobservable individual characteristics on the assumption that β1 = β2. Their OLS estimates are under 6½% but have large t‐values. By contrast, the fixed effects estimates are insignificant and close to zero,10 pointing to the conclusion that the return to computing merely proxies unobserved ability. Table 3 compares cross‐section and panel estimates of the computing coefficient by gender. First differences give the same results as fixed effects in a 2‐period model and we use this model for ease of notation. The OLS and random effects results suggest that computer use increases earnings by 12% to 16% for men and by 10% to 12% for women. The panel estimate is small and insignificant for men although it remains large and significant for women. This supports the view that there is no return to computing, at least for men. This could be because the panel estimator is removing the fixed effect but we shall argue that it is because it ignores variation in the parameter values over time. We therefore consider what happens if the coefficients are the same for each individual at each point in time but differ over time, thus implicitly examining whether the returns are falling over time. Later we explore a situation where different groups are defined by their computer use over time. Table 3 OLS and Panel Estimates of the Impact of Using a Computer at Work . OLS . RE 

Random Effects . First differences . 1991 sample . 2000 sample . Pooled sample . β1 = β2 . β1≠β2 . Men (n = 2,707) Impact of computer use 0.126*** (0.014) 0.145*** (0.022) 0.135*** (0.013) 0.111*** (0.012) 0.015 (0.016) 0.044** (0.022) Change in impact 0.044* (0.025) R2 0.464 0.491 0.491 0.489 0.094 0.159 Women (n = 987) Impact of computer use 0.093*** (0.024) 0.115*** (0.033) 0.098*** (0.020) 0.097*** (0.019) 0.077*** (0.024) 0.105*** (0.034) Change in impact 0.037 (0.040) R2 0.558 0.578 0.574 0.569 0.144 0.257 . OLS . RE 

Random Effects . First differences . 1991 sample . 2000 sample . Pooled sample . β1 = β2 . β1≠β2 . Men (n = 2,707) Impact of computer use 0.126*** (0.014) 0.145*** (0.022) 0.135*** (0.013) 0.111*** (0.012) 0.015 (0.016) 0.044** (0.022) Change in impact 0.044* (0.025) R2 0.464 0.491 0.491 0.489 0.094 0.159 Women (n = 987) Impact of computer use 0.093*** (0.024) 0.115*** (0.033) 0.098*** (0.020) 0.097*** (0.019) 0.077*** (0.024) 0.105*** (0.034) Change in impact 0.037 (0.040) R2 0.558 0.578 0.574 0.569 0.144 0.257 All the estimations in Tables 3–5 use the full specification defined in Table 2 and, where appropriate, include levels and differences in the control variables. Open in new tab Table 3 OLS and Panel Estimates of the Impact of Using a Computer at Work . OLS . RE 

Random Effects . First differences . 1991 sample . 2000 sample . Pooled sample . β1 = β2 . β1≠β2 . Men (n = 2,707) Impact of computer use 0.126*** (0.014) 0.145*** (0.022) 0.135*** (0.013) 0.111*** (0.012) 0.015 (0.016) 0.044** (0.022) Change in impact 0.044* (0.025) R2 0.464 0.491 0.491 0.489 0.094 0.159 Women (n = 987) Impact of computer use 0.093*** (0.024) 0.115*** (0.033) 0.098*** (0.020) 0.097*** (0.019) 0.077*** (0.024) 0.105*** (0.034) Change in impact 0.037 (0.040) R2 0.558 0.578 0.574 0.569 0.144 0.257 . OLS . RE 

Random Effects . First differences . 1991 sample . 2000 sample . Pooled sample . β1 = β2 . β1≠β2 . Men (n = 2,707) Impact of computer use 0.126*** (0.014) 0.145*** (0.022) 0.135*** (0.013) 0.111*** (0.012) 0.015 (0.016) 0.044** (0.022) Change in impact 0.044* (0.025) R2 0.464 0.491 0.491 0.489 0.094 0.159 Women (n = 987) Impact of computer use 0.093*** (0.024) 0.115*** (0.033) 0.098*** (0.020) 0.097*** (0.019) 0.077*** (0.024) 0.105*** (0.034) Change in impact 0.037 (0.040) R2 0.558 0.578 0.574 0.569 0.144 0.257 All the estimations in Tables 3–5 use the full specification defined in Table 2 and, where appropriate, include levels and differences in the control variables. Open in new tab 3. Panel Estimates With Time Varying Coefficients 3.1. Ols and Fixed Effects The expected value of the fixed effect estimator is a weighted average of β1 and β2. It would give unbiased estimates if the impact of computing is constant over time (β1 = β2) and would be preferred to the OLS estimators if computer use is correlated with the omitted individual specific effect. The argument is less clear cut when β1≠β2. OLS applied to a single cross‐section overestimates the impact if computer use and the omitted specific effect are positively correlated. Even if this bias is small, there could be a large difference between the OLS and fixed effects estimates but it is not clear what the implications are because they are estimating two different things. The cross section estimate focuses on the coefficient in one period and the fixed effect estimate on a weighted average of the coefficients. A similar vein, the pooled and fixed effects estimators are estimating different weighted averages of the parameters. 3.2. Fixed Effects With Coefficients That Vary Over Time The differences in expected values raise the possibility that the two sets of estimates differ because the coefficients are not stable over time. This does not seem immediately plausible in the present case since the OLS estimates have similar magnitudes. Nonetheless there is a view that any advantage to using a computer will be competed away over time indicating that the value of β might fall over time. We can test for changes in the coefficients over time by estimating: (2) where ρ = λ2 − λ1. The last column of Table 3 shows the estimates for this specification and includes levels and changes for all the variables in the ‘Full’ specification. The row labelled ‘impact of computer use’ gives the estimate of β2. The women's results are consistent with the OLS results; the impact of computing is well defined at 11% and has not changed from 1991 to 2000. The men's results suggest that computing had a small impact in 2000 although the estimate is much less than the values produced by cross‐section techniques. The evidence for a change in the coefficient is weak and the resulting estimates appear implausible so we reject the hypothesis of a change in the impact of computing over time. By implication, this suggests that the large difference in the OLS and panel estimates for men is due to fixed effects rather than parameter instability. 4. Heterogeneity Across Individuals 4.1. Panel Estimation With Heterogeneity Across Individuals Computer use has often been viewed as an indicator of unobserved individual productivity or job characteristics. The main motivation for the panel model was that it removed these effects (assuming they do not change over time). If there are genuine differences across different computer users, then the panel model estimates an average of the effects for different individuals derived from those that change their computer use over time. To distinguish between the different types of computer user, we define dummy variables to identify individuals who used computers in both periods (Stay), only the first period (Leave), only the second period (Enter) and consider the ‘varying coefficients’ specification: (3) where Eliminating the fixed effect, we obtain (4) Many policy makers assume that computing skills are productive and vary across individuals. If the more skilful individuals enter the market first, we might expect . The ranking of is not clear cut. If leavers stopped using computers because they were not very good at it, but, if leavers moved up promotion ladders, may be relatively large. The expected value of the normal fixed effects estimator is a weighted average of the parameters for the two sets of movers ( and ) minus a term reflecting the change in the parameters for the stayers. Panel estimation will provide unbiased estimates under the maintained hypothesis that . If there are more subtle effects present, standard panel estimators may, by chance, produce a close estimate of the impact of computer use across the whole population but they are unlikely to be good indicators of the premium for stayers. If there is no omitted variable bias, the OLS estimator for period t measures the impact of computing for users in that period. (It is an average of the returns for the two types of computer user in that period, weighted by the relative proportions of each type.) The average premium estimated, albeit with error, by OLS is a parameter of considerable interest in contrast to the parameter estimated by the panel estimator. Table 4 presents the panel estimates for the varying coefficients model in (3). There has been no significant change over time for either male or female stayers. The null hypothesis that is formally rejected for both genders. There was no significant impact on earnings for male enterers in 2000 but there was a large positive impact for women. Male leavers received a large significant premium in contrast to the women. These results are consistent with the view that male stayers earned a stable return, leavers received a return of 9% from using computers but these returns are not available to enterers (i.e. , βL = 0.085 and βE = 0). A different argument may apply to women. It may be that female enterers received a return of 14% or more but these returns are not available to leavers.11 This interpretation highlights our main point that we want to know the individual values of and . We contend that our cross‐section results are high because and are large. Table 4 Estimates of the Impact of Using a Computer at Work with Heterogeneity . First differences: Varying coefficients . Value‐Added . Men . Women . Men . Women . Computer used in 1991 and 2000 (Stay) 0.039 (0.025) 0.049 (0.041) 0.134*** (0.022) 0.089*** (0.032) Computer used in 2000 only (Enter) (βE) 0.015 (0.025) 0.130*** (0.041) 0.058*** (0.022) 0.128*** (0.033) Computer used in 1991 only (Leave) (βL) 0.085** (0.039) 0.009 (0.062) – – R2 0.161 0.258 0.575 0.671 . First differences: Varying coefficients . Value‐Added . Men . Women . Men . Women . Computer used in 1991 and 2000 (Stay) 0.039 (0.025) 0.049 (0.041) 0.134*** (0.022) 0.089*** (0.032) Computer used in 2000 only (Enter) (βE) 0.015 (0.025) 0.130*** (0.041) 0.058*** (0.022) 0.128*** (0.033) Computer used in 1991 only (Leave) (βL) 0.085** (0.039) 0.009 (0.062) – – R2 0.161 0.258 0.575 0.671 Open in new tab Table 4 Estimates of the Impact of Using a Computer at Work with Heterogeneity . First differences: Varying coefficients . Value‐Added . Men . Women . Men . Women . Computer used in 1991 and 2000 (Stay) 0.039 (0.025) 0.049 (0.041) 0.134*** (0.022) 0.089*** (0.032) Computer used in 2000 only (Enter) (βE) 0.015 (0.025) 0.130*** (0.041) 0.058*** (0.022) 0.128*** (0.033) Computer used in 1991 only (Leave) (βL) 0.085** (0.039) 0.009 (0.062) – – R2 0.161 0.258 0.575 0.671 . First differences: Varying coefficients . Value‐Added . Men . Women . Men . Women . Computer used in 1991 and 2000 (Stay) 0.039 (0.025) 0.049 (0.041) 0.134*** (0.022) 0.089*** (0.032) Computer used in 2000 only (Enter) (βE) 0.015 (0.025) 0.130*** (0.041) 0.058*** (0.022) 0.128*** (0.033) Computer used in 1991 only (Leave) (βL) 0.085** (0.039) 0.009 (0.062) – – R2 0.161 0.258 0.575 0.671 Open in new tab 4.2. Modelling Heterogeneity Across Individuals We finally consider two extensions of our heterogeneity model. As Anger and Schwarz (2002) have observed, computers per se cannot affect earnings before they were actually used. More generally, any effect that computer use in one period has on earnings in a different period might suggest the role of unobserved factors. This insight underlies the reduced form approach of Jakubson (1991) that we use below. Our second model is taken from the education production function literature where an outcome typically depends on current regressors and values of variables and unobservables that reflect an individual's history. A common approach is to use a value‐added specification that proxies the individual's history by a lagged dependent variable. 4.3. Reduced Form Estimation If the fixed effect is correlated with current computer use, then in general it is correlated with computer use in each period. Generalising Jakubson (1991), we can write (5) where vi is orthogonal to the regressors by construction. Substituting into (3), we obtain the reduced form equations: (6) where These equations can be estimated for each time period or both periods together. In either case, the key hypothesis that and is accepted for women and rejected for men. Using both periods, the following joint hypotheses are accepted at the 5% level: These tests support the previous interpretation of the first difference results with heterogeneity. The one exception is that the coefficient of Enter for men should be interpreted as the difference between the coefficients for each period. Since  > 0, Table 4 implies that γE > 0 and for men. These tests also show that there are no effects from unobservables for male leavers and for women. Table 5 reports the OLS estimates of our reduced form heterogeneity model for each time period. Male enterers have a significant earnings premium in 1991 but leavers do not in 2000. We explain the first result by the unobserved ability of enterers and argue that male leavers have lower ability than the others because they were unable to take advantage of the opportunities offered by the new technology. Since there is no ability premium for leavers in 2000, the extra earnings that leavers receive in 1991 is solely due their use of computers. Applying this interpretation, the estimates of the coefficients for Enter and Leave for 1991 suggest, respectively, returns of about 8% to ability for enterers and 11% to computer use for leavers. The return for male stayers is statistically equal to the sum of these two effects in 1991 and the estimates do not change when this restriction is imposed. This is compatible with a common ability effect for those who use computers in the second period (γS = γE) and a common return to computer use for users in the first period (). Table 5 Reduced Form Estimates with Heterogeneity . Men . Women . 1991 sample . 2000 sample . 1991 sample . 2000 sample . Computer used in 1991 and 2000 (Stay) 0.183*** (0.018) 0.214*** (0.025) 0.123*** (0.032) 0.141*** (0.041) Computer used in 2000 only (Enter) 0.079*** (0.019) 0.093*** (0.024) 0.046 (0.033) 0.141*** (0.042) Computer used in 1991 only (Leave) 0.109*** (0.030) 0.054 (0.039) 0.136*** (0.053) 0.088 (0.062) R2 0.469 0.497 0.559 0.579 . Men . Women . 1991 sample . 2000 sample . 1991 sample . 2000 sample . Computer used in 1991 and 2000 (Stay) 0.183*** (0.018) 0.214*** (0.025) 0.123*** (0.032) 0.141*** (0.041) Computer used in 2000 only (Enter) 0.079*** (0.019) 0.093*** (0.024) 0.046 (0.033) 0.141*** (0.042) Computer used in 1991 only (Leave) 0.109*** (0.030) 0.054 (0.039) 0.136*** (0.053) 0.088 (0.062) R2 0.469 0.497 0.559 0.579 Open in new tab Table 5 Reduced Form Estimates with Heterogeneity . Men . Women . 1991 sample . 2000 sample . 1991 sample . 2000 sample . Computer used in 1991 and 2000 (Stay) 0.183*** (0.018) 0.214*** (0.025) 0.123*** (0.032) 0.141*** (0.041) Computer used in 2000 only (Enter) 0.079*** (0.019) 0.093*** (0.024) 0.046 (0.033) 0.141*** (0.042) Computer used in 1991 only (Leave) 0.109*** (0.030) 0.054 (0.039) 0.136*** (0.053) 0.088 (0.062) R2 0.469 0.497 0.559 0.579 . Men . Women . 1991 sample . 2000 sample . 1991 sample . 2000 sample . Computer used in 1991 and 2000 (Stay) 0.183*** (0.018) 0.214*** (0.025) 0.123*** (0.032) 0.141*** (0.041) Computer used in 2000 only (Enter) 0.079*** (0.019) 0.093*** (0.024) 0.046 (0.033) 0.141*** (0.042) Computer used in 1991 only (Leave) 0.109*** (0.030) 0.054 (0.039) 0.136*** (0.053) 0.088 (0.062) R2 0.469 0.497 0.559 0.579 Open in new tab If we accept that the wage premium given to the enterers is purely a return to ability, then the ability effect is approximately the same in 2000 at 9%. If we then assume that the return for stayers is the sum of an ability effect equal to that of enterers (γS = γE) and an enduring computer effect following early use (), then the second period return to computer use for male stayers is about 13% using the 2000 estimates. (Using both time periods with the restrictions imposed gives a return to unobservables of 8½% and to computer use of 12%). The estimates for stayers and enterers are similar for the 1991 and 2000 samples in Table 5 so we would not expect to see any significant changes over time in Table 4. The estimates in Tables 4 and 5 now suggest a return to computer use for particular groups of men of at least 9% and possibly as much as 13%. These are only slightly lower than the magnitudes suggested by Table 3. By contrast, using a computer only has an impact for women in the period when computers are used. Female enterers receive no significant increase in earnings in 1991 and the estimates for the female stayers and leavers are statistically the same in 1991. The two implied restrictions are accepted at the 5% level. The corresponding results also hold for 2000 so Table 3 reports appropriate estimates for women. Alternatively, using both samples with these restrictions imposed gives a return of 11%. (This differs from the pooled estimate because the coefficients of the control variables are allowed to differ in each period.) Table 3 suggests a return to computer use for women of 10% to 12% compared with 9% to 13% for men in Table 5. Male stayers appear to earn far more than women. Our results suggest that men receive an additional premium for computer use. Previous writers on this subject would probably interpret this as a return to ability but the literature on gender differentials would regard this as an unexplained difference. 4.4. Value Added Specification We conclude by estimating the value‐added (VA) model: (7) The value added model adds previous earnings to the list of regressors in the ‘contemporaneous’ specification. This addition is taken as a sufficient statistic for observed factors which have changed over time before period 2 and unobserved individual endowments like ability which may affect earnings outcomes. The VA specification is interesting in this context as it is precisely these kinds of omitted factors that researchers mention when discussing ‘over‐estimates’ of the impact of computers on earnings. Although the VA model is considered preferable to the contemporaneous specification (Hanushek, 1996, 2003), Todd and Wolpin (2003) show that it relies on strong assumptions. Any fixed effect is only removed if the coefficients of the regressors and the fixed effect decay geometrically over time at the same rate as δ. Further, the estimates of VA specification are sensitive to the omission of regressors although, for obvious reasons, we regard missing regressors as a less important problem in this study. Table 4 shows the results of this estimation. The impact of ICT for stayers is similar to the previous estimates. It is large (14% for men and 9% for women) and significant. Enterers received higher earnings, 5% more for men and 14% for women. If lagged earnings are merely another proxy for unobservables, then we might want to derive the actual return to computer use for men from the difference between the estimates for stayers and enterers. In which case, the return falls but is still 8% for men. 5. Conclusion Our paper presents firm evidence that there was a large earnings premium to computer use in the UK. Over time, there have been repeated discussions of how to interpret the impact of computer use. Our paper focuses on DiNardo and Pischke's (1997) argument that any estimate merely measures unobservable job or individual characteristics and is likely to disappear as other factors are considered. This argument does not seem plausible in our case because we have many controls for ability, occupation, industry and skills, yet our cross‐section estimates are still about 13% and 14% for men. There are a range of econometric problems that arise in cross‐sectional models but Dolton and Makepeace (2002) report estimates from matching and selection models that are of a similar order of magnitude to those presented in this paper. Nonetheless, the fixed effects estimates of Entorf and Kramarz (1997) are much lower than their cross‐section estimates and insignificant, adding considerable weight to the argument that impact of computer use proxies unmeasured ability. We replicate this finding but seek to reconcile the cross‐section and fixed effects estimates by considering more complex panel models that allow the impact of computing to vary over time. We find no conclusive evidence that there was a uniform change in the value of the computing coefficient over time and argue instead that the ‘return’ to computing varied across individuals. We can determine whether an individual used a computer in both periods (stayers), the first period only (leavers), the second period only (enterers) or not at all. If we allow the coefficients to vary across individuals, the panel estimates for male leavers give an earnings premium of 9%. The simplest interpretation is that this boost to earnings was only available while these men worked with computers. Given that fixed effects are eliminated and the wide range of controls employed, it is not plausible to attribute this to unobserved ability. Indeed, the growth in this group's average earnings over the sample period was less than fifth of that of other men so they were not individuals in high demand who were moving on to better jobs. Formally, the panel model only identifies the change in coefficients for each group. Our ‘reduced form’ estimates for men show that the impact of computer use remained constant over time for stayers and enterers but was zero in 2000 for leavers. They are consistent with a return to early computer use of 11% to 13% and return to unobservables of 8% to 9%. Computers only affect women's earnings when they are used and our preferred estimates suggest a return of 10% to 12% for women with no return to unobservables. Our estimates of the value added model are slightly lower for men but similar for women. We conclude that there is a well defined earnings premium associated with computer use. Appendix Appendix: Variable Definitions This Appendix defines the control variables used in the regressions. The regressions also include dummies for missing values on each regressor although some of these are omitted because there are no missing values for that variable in the sample. Dependent Variable. Log of real hourly earnings, which is gross pay before deductions divided by usual hours of work. The nominal data were recoded to January 2000 prices. Full details of the variable are included in Dearden et al. (2003). Computer Use Dummies. Two binary variables taking the value 1 in 1991 and 2000, respectively, if the respondent uses at work a computer or word processor with a TV type screen (usually known as a VDU) and if the respondent uses a computer at work. Sets of Explanatory Variables Basic Human Capital Years of schooling, Years of work experience, Tenure with current employer. Qualifications Highest qualification achieved whether vocational or academic; dummies for 5 levels: NVQ level 1, NVQ level 2, NVQ level 3, NVQ level 4 or 5; omitted group ‐ No qualifications. Early Test Scores Dummies for quintile scores on tests at age 11; 5 dummies for reading and 5 dummies for mathematics: omitted groups – Bottom quintiles for reading and mathematics. Skills in 1991 survey Dummies for ‘good’ at 6 types of skill: communication (speaking clearly), carrying out mathematics, giving advice and support, using tools, caring, finance and accounts. Skills in 2000 survey Dummies for ‘good’ at 8 types of skill: communication, numbers and calculation, team work, learning new skills, problem solving, using tools, caring, finance and accounts. SOC Occupation, Dummies for 9 Major SOC Groups: managers and administrators, professional, associate professional and technical, clerical and secretarial, craft and related, personal and protective services, sales, other; omitted group – plant and machine operatives. SIC Industry, Dummies for 13 Major SIC Groups: Farming, Manufacturing, Construction, Sales (wholesale, retail and repair), Transport and communications, Financial intermediation, Real estate, renting and business activities, Public administration and defence, Education, Health and social work, Other community, social and personal services, Other industries (other jobs, mining, electricity, gas and water supply); omitted group – Hotels and restaurants. Region Dummies for 12 regions: London, East Anglia, South East, South West, East Midlands, West Midlands, Yorkshire and Humberside, North West, North, Scotland, Other: omitted group – Wales. Socio‐demographic Dummies for: Married, Non‐white, Long standing illness limits daily activities. Other Dummies for other job characteristics: Firm size (number of employees) 5 categories – 10–24, 25–99, 100–499, 500 or more (2000 survey); omitted group 1–9 (1–10 employees 1991 survey); Temporary job, Union member. Footnotes 1 " These are 28% (1980–90) and 25% (1990–2000) for IT equipment although their rates for software growth are lower at 27% (1980–90) and 10% (1990–2000). 2 " Other countries reacted differently. The UK share tripled from 1980 to 2000 but started from a low base (5% of total investment). The US share doubled but from a much higher base (15%). The German share only grew by a quarter from 12% to end slightly above the UK figure. 3 " We can also measure computer competence and intensity of computer use but this paper concentrates on the basic impact of computing. We are working on another paper that examines the effect of frequency and sophistication of use across different cohorts and countries. 4 " The use of a panel sample may mean that computer use is over‐represented because respondents have to be in work at both points in time. Our samples also exclude older people who are less likely to use computers. 5 " We have also estimated the models using scores at age 7. The estimates for the computing coefficients only change at the 3rd decimal place and their significance is not affected. 6 " Although the incidence of computer use is high, there seems to be no evidence of a saturation effect. 7 " 2½% for men and 6½% for women using pooled data for 1985–99. 8 " 3% using pooled data for 1985–7 and the significant ‘high’ autonomy figures (see Table 2). The ‘OLS’ figures use a maximum likelihood procedure to allow for grouping in the wage data. The fixed effects estimates are comparable. 9 " 12% in 1993. 10 " Oosterbeek's (1997) fixed effect estimate (10%) is only slightly smaller with a t‐value of 2.7. He assumed that computers were not used in his previous period (1983) and did not include any control variables in the panel model. 11 " There were only 38 female leavers so our estimate of is, at best, imprecise. References Acemoglu , D. ( 2002 ). ‘ Technical change, inequality and the labour market ’, Journal of Economic Literature , vol. 40 ( 1 ), pp. 7 – 72 . Google Scholar Crossref Search ADS WorldCat Anger , S. and Schwarz , J. ( 2002 ). ‘ Does future PC use determine our wages today: evidence from the German panel data ’, IZA Discussion Paper No.429. Arabsheibani , G. R. and Marin , A. ( 2001 ). ‘ If not computers then what? Returns to computer use in the UK revisited ’, School of Management and Business , Aberystwyth, mimeo. Bell , B. ( 1996 ). ‘ Skill‐biased technical change and wages: evidence from a longitudinal data set ’, Nuffield College , Oxford, mimeo. Card , D. and DiNardo , J. E. ( 2002 ). ‘ Skill‐biased technological change and rising wage inequality: some problems and puzzles ’, Journal of Labor Economics , vol. 20 . OpenURL Placeholder Text WorldCat Colecchia , A. and Schreyer , P. ( 2001 ). ‘ ICT investment and economic growth in the 1990s: is the United States a unique case? A comparative study of nine OECD countries ’, OECD Directorate for Science, Technology and Industry working paper 2001/7. Dearden , L. , Goodman , A. and Saunders , P. ( 2003 ). ‘Income and living standards’, in ( E. Ferri, J. Bynner and M. Wadsworth), Changing Britain, Changing Lives , ch. 6, London: Institute of Education, University of London. Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Dickerson , A. and Green F. ( 2002 ). ‘ The growth and valuation of generic skills ’, University of Kent, mimeo. DiNardo , J. and Pischke , J. ( 1997 ). ‘ The returns to computer use revisited: have pencils changed the wage structure too? ’, Quarterly Journal of Economics , vol. 112 , pp. 291 – 303 . Google Scholar Crossref Search ADS WorldCat Dolton , P and Makepeace , G. ( 2002 ). ‘ Returns to computer use: an empirical analysis for the UK ’, University of Newcastle, mimeo. Entorf , H. and Kramarz , F. ( 1997 ). ‘ Does unmeasured ability explain the higher wages of new technology workers? ’, European Economic Review , vol. 41 , pp. 1489 – 510 . Google Scholar Crossref Search ADS WorldCat Hanushek , E. A. ( 1996 ). ‘School resources and student performance’, in ( G. Burtless, ed.), Does Money Matter? The Effect of School Resources on Student Achievement and Adult Success , Washington, DC: Brookings Institution . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Hanushek , E. A. ( 2003 ). ‘ The failure of input‐based schooling policies ’, Economic Journal , vol. 113 , pp. F64 – 98 . Google Scholar Crossref Search ADS WorldCat Jakubson , G. ( 1991 ). ‘ Estmation and testing of the union wage effect using panel data ’, The Review of Economic Studies , vol. 58 , pp. 971 – 91. Google Scholar Crossref Search ADS WorldCat Krueger , A. ( 1993 ), ‘ How computers have changed the wage structure: evidence from microdata, 1984–1989 ’, Quarterly Journal of Economics , vol. 108 , pp. 33 – 60 . Google Scholar Crossref Search ADS WorldCat Machin , S. ( 2001 ). ‘ The changing nature of labour demand in the new economy and skill‐biased technological change ’, Oxford Bulletin of Economics and Statistics , vol. 63 , pp. 753 – 76. Google Scholar Crossref Search ADS WorldCat Oosterbeek , H. ( 1997 ). ‘ The returns from computer use: a simple test on the productivity interpretation ’, Economics Letters , vol. 55 , pp. 273 – 7. Google Scholar Crossref Search ADS WorldCat Oulton , N. ( 2001 ). ‘ ICT and productivity growth in the United Kingdom ’, Bank of England Working Paper no.140. Todd , P. and Wolpin , K. ( 2003 ). ‘ On the specification and estimation of the production function for cognitive achievement ’, Economic Journal , vol. 113 , pp. F3 – 33 . Google Scholar Crossref Search ADS WorldCat Author notes " The authors would like to thank Martyn Andrews, Ian Preston, Karl Taylor, two anonymous referees and seminar participants at Cardiff, ESPE 2003, Leicester, Manchester, LSE and RES 2003 for their comments on earlier versions of this paper. © Royal Economic Society 2004

Journal

The Economic Journal – Oxford University Press

Published: Mar 1, 2004

There are no references for this article.