COMPUTERS HAVE CHANGED THE WAGE STRUCTURE 37 A linear probability regression of a computer-use dummy on experience and its square,education,and demographic variables indicates that the likelihood of using a computer increases with experience in the first fifteen years of experience,and declines thereafter. Tabulations of the 1989 CPS show that relatively few employ- ees (less than 5 percent of employees)use computers in the agriculture,construction,textile,lumber,and personal services industries,whereas computer use is widespread (exceeding 60 percent of employees)in the banking,insurance,real estate, communications,and public administration industries.The Octo- ber CPS does not contain information on employer size,but a 1989 establishment survey by the Gartner Group found that computer use is not strongly related to establishment size for establishments with more than twenty employees [Statistical Abstract of the United States,1990,p.951].And the growth in personal computers per worker between 1984 and 1989 was not strongly related to establishment size for establishments with more than twenty employees. II.COMPUTER USE AND WAGES I have estimated a variety of statistical models to try to answer the question:Do employees who use computers at work receive a higher wage rate as a result of their computer skills?I begin by summarizing some simple ordinary least squares(OLS)estimates. The analysis is based on data from the October 1984 and 1989 CPS. The sample consists of workers age 18-65.(See Appendix A for further details of the sample.) My initial approach is to augment a standard cross-sectional earnings function to include a dummy variable indicating whether an individual uses a computer at work.Let C:represent a dummy variable that equals one if the ith individual uses a computer at work,and zero otherwise.Observation i's wage rate W:is assumed to depend on Ci,a vector of observed characteristics Xi,and an error e.Adopting a log-linear specification, (1) lnW=X:β+C:a+ei, where B and a are parameters to be estimated.Section III considers the effect of bias because of possible correlation between C:and e. Table II reports results of fitting equation(1)by OLS,with varying sets of covariates(X).In columns(1)and(4)a computer- use dummy variable is the only right-hand-side variable.In these
COMPUTERS HAVE CHANGED THE WAGE STRUCTURE 37 A linear probability regression of a computer-use dummy on experience and its square, education, and demographic variables indicates that the likelihood of using a computer increases with experience in the first fifteen years of experience, and declines thereafter. Tabulations of the 1989 CPS show that relatively few employees (less than 5 percent of employees) use computers in the agriculture, construction, textile, lumber, and personal services industries, whereas computer use is widespread (exceeding 60 percent of employees) in the banking, insurance, real estate, communications, and public administration industries. The October CPS does not contain information on employer size, but a 1989 establishment survey by the Gartner Group found that computer use is not strongly related to establishment size for establishments with more than twenty employees [Statistical Abstract of the United States, 1990,p. 9511. And the growth in personal computers per worker between 1984 and 1989 was not strongly related to establishment size for establishments with more than twenty employees. I have estimated a variety of statistical models to try to answer the question: Do employees who use computers at work receive a higher wage rate as a result of their computer skills? I begin by summarizing some simple ordinary least squares (OLS) estimates. The analysis is based on data from the October 1984 and 1989 CPS. The sample consists of workers age 18-65. (See Appendix A for further details of the sample.) My initial approach is to augment a standard cross-sectional earnings function to include a dummy variable indicating whether an individual uses a computer at work. Let Ci represent a dummy variable that equals one if the ith individual uses a computer at work, and zero otherwise. Observation i's wage rate W, is assumed to depend on Ci, a vector of observed characteristics Xi, and an error E~. Adopting a log-linear specification, (1) lnWi =Xi@ +Cia + ei, where f3 and a are parameters to be estimated. Section I11 considers the effect of bias because of possible correlation between Ci and E,. Table I1 reports results of fitting equation (1)by OLS, with varying sets of covariates (X). In columns (1)and (4) a computeruse dummy variable is the only right-hand-side variable. In these
38 QUARTERLY JOURNAL OF ECONOMICS TABLE II OLS REGRESSION ESTIMATES OF THE EFFECT OF COMPUTER USE ON PAY (DEPENDENT VARIABLE:In (HOURLY WAGE)) October 1984 October 1989 Independent variable (1) (2) (3) (4) (5) (6) Intercept 1.937 0.750 0.928 2.086 0.905 1.094 (0.005)(0.023)(0.026)(0.006) (0.024)(0.026) Uses computer at work(1 yes)0.276 0.170 0.140 0.325 0.188 0.162 (0.010)(0.008) (0.008)(0.009)0.008 (0.008) Years of education 0.069 0.048 0.075 0.055 (0.001)(0.002) (0.002) (0.002) Experience 0.027 0.025 0.027 0.025 (0.001)(0.001) (0.001)(0.001) Experience-squared +100 -0.041-0.040 -0.041 -0.040 (0.002)(0.002) (0.002)(0.002) Black (1 yes) -0.098 -0.066 -0.121-0.092 (0.013)(0.012) (0.013)(0.012) Other race(1 yes) -0.105 -0.079 -0.029-0.015 (0.020)(0.019 (0.020) (0.020) Part-time(1 yes) -0.256 -0.216 -0.221 -0.183 (0.010)(0.010) (0.010)(0.010) Lives in SMSA(1 yes) 0.111 0.105 0.138 0.130 (0.007) (0.007) (0.007)(0.007) Veteran(1 yes) 0.038 0.041 0.025 0.031 (0.011) (0.011) (0.012)(0.011) Female(1 yes) -0.162-0.135 -0.172-0.151 (0.012)(0.012) (0.012)(0.012) Married(1 yes) 0.156 0.129 0.159 0.143 (0.011)0.011) (0.011)(0.011) Married*Female -0.168-0.151 -0.141-0.131 (0.015)(0.015) (0.015)(0.015) Union member(1 yes) 0.181 0.194 0.182 0.189 (0.009)(0.009) (0.010)(0.010) 8 Occupation dummies No No Yes No No Yes R2 0.051 0.446 0.4910.082 0.451 0.486 Notes.Standard errors are shown in parentheses.Sample size is 13,335 for 1984 and 13,379 for 1989. Columns (2).(3).(5),and (6)also include three region dummy variables. models the(raw)differential in hourly pay between workers who use computers on the job and those who do not is 31.8 percent (exp(0.276)-1)in1984,and38.4 percent(exp(0.325)-1)in1989.In columns(2)and(5)several covariates are added to the regression equation,including education,potential experience and its square, gender,and union status.Including these variables reduces the computer premium to 18.5 percent in 1984 and to 20.6 percent in
QUARTERLY JOURNAL OF ECONOMICS TABLE I1 OLS REGRESSION ESTIMATES OF THE EFFECT OF COMPUTER USEON PAY (DEPENDENT VARIABLE: In (HOURLY WAGE)) October 1984 October 1989 Independent variable (1) (2) (3) (4) (5) (6) Intercept 1.937 0.750 0.928 2.086 0.905 1.094 (0.005) (0.023) (0.026) (0.006) (0.024) (0.026) Uses computer at work (1= yes) 0.276 0.170 0.140 0.325 0.188 0.162 (0.010) (0.008) (0.008) (0.009) (0.008) (0.008) Years of education - 0.069 0.048 - 0.075 0.055 (0.001) (0.002) (0.002) (0.002) Experience - 0.027 0.025 - 0.027 0.025 (0.001) (0.001) (0.001) (0.001) Experience-squared i 100 - -0.041 -0.040 - -0.041 -0.040 (0.002) (0.002) (0.002) (0.002) Black (1 = yes) - -0.098 -0.066 - -0.121 -0.092 (0.013) (0.012) (0.013) (0.012) Other race (1= yes) - -0.105 -0.079 - -0.029 -0.015 (0.020) (0.019) (0.020) (0.020) Part-time (1= yes) - -0.256 -0.216 - -0.221 -0.183 (0.010) (0.010) (0.010) (0.010) Lives in SMSA (1 = yes) - 0.111 0.105 - 0.138 0.130 (0.007) (0.007) (0.007) (0.007) Veteran (1 = yes) - 0.038 0.041 - 0.025 0.031 (0.011) (0.011) (0.012) (0.011) Female (1= yes) - -0.162 -0.135 - -0.172 -0.151 (0.012) (0.012) (0.012) (0.012) Married (1 = yes) - 0.156 0.129 - 0.159 0.143 (0.011) (0.011) (0.011) (0.011) Married*Female - -0.168 -0.151 - -0.141 -0.131 (0.015) (0.015) (0.015) (0.015) Union member (1= yes) - 0.181 0.194 - 0.182 0.189 (0.009) (0.009) (0.010) (0.010) 8 Occupation dummies No No Yes No No Yes R 0.051 0.446 0.491 0.082 0.451 0.486 Notes. Standard errors are shown m parentheses. Sample srze is 13,335 for 1984 and 13,379 for 1989. Columns (21, (3), (51, and (6)also include three region dummy variables. models the (raw) differential in hourly pay between workers who use computers on the job and those who do not is 31.8 percent (exp(0.276)-1) in 1984, and 38.4 percent (exp(0.325)-1) in 1989. In columns (2) and (5) several covariates are added to the regression equation, including education, potential experience and its square, gender, and union status. Including these variables reduces the computer premium to 18.5 percent in 1984 and to 20.6 percent in
COMPUTERS HAVE CHANGED THE WAGE STRUCTURE 39 1989.5 Even after including these covariates,however,the com- puter dummy variable continues to have a sizable and statistically significant effect on wages,with t-ratios of 21.3 in 1984 and 23.1 in 1989. It is not clear whether occupation dummies are appropriate variables to include in these wage regressions because computer skills may enable workers to qualify for jobs in higher paying occupations and industries.For example,one would probably not want to control for whether a worker is in the computer program- ming occupation while estimating the effect of computer use on earnings.Nevertheless,columns (3)and(6)include a set of eight one-digit occupation dummies.These models still show a sizable pay differential for using a computer at work.In 1989,for example, employees who use computers on the job earn 17.6 percent higher pay than employees who do not use computers on the job,holding education,occupation,and other characteristics constant.If 44 two-digit occupation dummies are included in the model in column (6)instead of the 8 one-digit occupation dummies,the computer- use wage differential is 13.9 percent,with a t-ratio of 15.5. A.Employer Characteristics Although I am mainly concerned about bias because of omitted employee characteristics that are correlated with computer use at work,it is possible that characteristics of employers are correlated with the provision of computers and the generosity of compensa- tion.Such a relationship might exist in a rent-sharing model,in which employees are able to capture some of the return to the employer's capital stock.Unfortunately,there is only a limited amount of information about employer characteristics in the CPS. However,if 48 two-digit industry dummies are included in a model that includes two-digit occupation dummies and the covariates in column (6),the computer-use wage differential is 11.4 percent, with a t-ratio of 13.0.6 Information on employer size is not available in the October CPS,but two findings suggest that the computer differential is not merely reflecting the effect of (omitted)employer size.First, establishment-level surveys do not show a strong relationship 5.The computer differential is about the same for men and women.For example,in 1989 the coefficient (and standard error)for computer use is 0.197 (0.012)for men and 0.185 (0.011)for women. 6.Results for 1984 are similar:the wage differential falls to 11.3 percent if 44 occupation dummies are included,and to 9.0 percent if 48 two-digit industry dummies are included
COMPUTERS HAVE CHANGED THE WAGE STRUCTURE 39 1989.5 Even after including these covariates, however, the computer dummy variable continues to have a sizable and statistically significant effect on wages, with t-ratios of 21.3 in 1984 and 23.1 in 1989. It is not clear whether occupation dummies are appropriate variables to include in these wage regressions because computer skills may enable workers to qualify for jobs in higher paying occupations and industries. For example, one would probably not want to control for whether a worker is in the computer programming occupation while estimating the effect of computer use on earnings. Nevertheless, columns (3) and (6) include a set of eight one-digit occupation dummies. These models still show a sizable pay differential for using a computer at work. In 1989, for example, employees who use computers on the job earn 17.6 percent higher pay than employees who do not use computers on the job, holding education, occupation, and other characteristics constant. If 44 two-digit occupation dummies are included in the model in column (6) instead of the 8 one-digit occupation dummies, the computeruse wage differential is 13.9 percent, with a t-ratio of 15.5. A. Employer Characteristics Although I am mainly concerned about bias because of omitted employee characteristics that are correlated with computer use at work, it is possible that characteristics of employers are correlated with the provision of computers and the generosity of compensation. Such a relationship might exist in a rent-sharing model, in which employees are able to capture some of the return to the employer's capital stock. Unfortunately, there is only ;I limited amount of information about employer characteristics in the CPS. However, if 48 two-digit industry dummies are included in a model that includes two-digit occupation dummies and the covariates in column (6), the computer-use wage differential is 11.4 percent, with a t-ratio of 13.0.6 Information on employer size is not available in the October CPS, but two findings suggest that the computer differential is not merely reflecting the effect of (omitted) employer size. First, establishment-level surveys do not show a strong relationship 5. The computer differential is about the same for men and women. For example, in 1989 the coefficient (and standard error) for computer use is 0.197 (0.012)for men and 0.185 (0.011) for women. 6. Results for 1984 are similar: the wage differential falls to 11.3 percent if 44 occupation dummies are included, and to 9.0 percent if 48 two-digit industry dummies are included
40 QUARTERLY JOURNAL OF ECONOMICS between computer use and establishment size (e.g.,Hirschorn [1988]).Second,in a recent paper Reilly [1991]uses a sample of 607 employees who worked in 60 plants in Canada in 1979 to investigate the relationship between establishment size and wages. Reilly estimates wage regressions including a dummy variable indicating access to a computer.Without controlling for establish- ment size,he finds that employees who have access to a computer earn 15.5 percent(t=5.7)higher pay.When he includes the log of establishment size,the computer-wage differential is 13.4 percent (t=3.9). Finally,I have estimated the model in column(5)separately for union and nonunion workers.The premium for computer use is 20.4 percent(t =23)in the nonunion sector,and just 7.8 percent (t=4.3)in the union sector.Since unions have been found to compress skill differentials(see Lewis [1986]and Card [1991]),this finding should not be surprising.However,if one believes that the premium for work-related computer use is a result of employees capturing firms'capital rents rather than a return to a skill,it is difficult to explain why the premium is so much larger in the nonunion sector than in the union sector. B.Computer Premium over Time The results in Table II indicate that,if anything,the estimated reward for using a computer at work increased slightly between 1984 and 1989.For example,based on the models in columns(3) and(6),between 1984 and 1989 the computer (log)wage premium increased by 0.022.The standard error of this estimate is 0.011,so the increase is on the margin of statistical significance.There is certainly no evidence of a decline in the payoff for computer skills in this period. This finding is of interest for two reasons.First,given the substantial expansion in the supply of workers who have computer skills between 1984 and 1989,one might have expected a decline in the wage differential associated with computer use at work,ceteris paribus.The failure of the wage differential for computer use to decline suggests that the demand for workers with computer skills may have shifted out as fast as,or faster than,the outward shift in the supply of computer-literate workers.This hypothesis is plausi- ble given the remarkable decline in the price of computers and the expansion in uses of computers in the 1980s. A second reason why the slight increase in the wage differen- tial associated with computer use is of interest concerns the effect
40 QUARTERLY JOURNAL OF ECONOMICS between computer use and establishment size (e.g., Hirschorn [1988]). Second, in a recent paper Reilly [1991] uses a sample of 607 employees who worked in 60 plants in Canada in 1979 to investigate the relationship between establishment size and wages. Reilly estimates wage regressions including a dummy variable indicating access to a computer. Without controlling for establishment size, he finds that employees who have access to a computer earn 15.5 percent (t = 5.7) higher pay. When he includes the log of establishment size, the computer-wage differential is 13.4 percent (t = 3.9). Finally, I have estimated the model in column (5) separately for union and nonunion workers. The premium for computer use is 20.4 percent (t = 23) in the nonunion sector, and just 7.8 percent (t = 4.3) in the union sector. Since unions have been found to compress skill differentials (see Lewis [I9861 and Card [19911), this finding should not be surprising. However, if one believes that the premium for work-related computer use is a result of employees capturing firms7 capital rents rather than a return to a skill, it is difficult to explain why the premium is so much larger in the nonunion sector than in the union sector. B. Computer Premium over Time The results in Table I1 indicate that, if anything, the estimated reward for using a computer at work increased slightly between 1984 and 1989. For example, based on the models in columns (3) and (6), between 1984 and 1989 the computer (log) wage premium increased by 0.022. The standard error of this estimate is 0.011, so the increase is on the margin of statistical significance. There is certainly no evidence of a decline in the payoff for computer skills in this period. This finding is of interest for two reasons. First, given the substantial expansion in the supply of workers who have computer skills between 1984 and 1989, one might have expected a decline in the wage differential associated with computer use at work, ceteris paribus. The failure of the wage differential for computer use to decline suggests that the demand for workers with computer skills may have shifted out as fast as, or faster than, the outward shift in the supply of computer-literate workers. This hypothesis is plausible given the remarkable decline in the price of computers and the expansion in uses of computers in the 1980s. A second reason why the slight increase in the wage differential associated with computer use is of interest concerns the effect
COMPUTERS HAVE CHANGED THE WAGE STRUCTURE 41 of possible nonrandom selection of the workers who use computers. Companies are likely to provide computer training and equipment first to the workers whose productivity is expected to increase the most from using a computer.This would pose a problem for the interpretation of the OLS estimates if these workers would have earned higher wages in the absence of computer use.The large increase in the number of workers who used computers at work between 1984 and 1989 is likely to have reduced the average quality of workers who work with computers,which would be expected to drive down the average wage differential associated with computer use.However,the slight increase in the computer wage premium between 1984 and 1989 suggests that nonrandom selection of the workers who use computers is not the dominant factor behind the positive association between computer use and wages. The other variables in Table II generally have their typical effects on wages,and their coefficients are relatively stable between 1984 and 1989.One notable exception is the rate of return to education,which increased by 0.6 percentage points between 1984 and 1989,even after holding computer use constant.And the black-white wage gap increased,while the wage gap between whites and other races declined in these years. C.Specific Computer Tasks The 1989 CPS asked workers what tasks they use their computer for.Respondents were allowed to indicate multiple tasks. Table III presents estimates of the coefficients on the specific computer tasks for a wage regression that also includes the covariates listed in column (6)of Table II (including occupation dummies).Importantly,the regression includes a dummy that equals one if the individual used a computer for any task at all,as well as dummies for the specific tasks.Thus,the coefficients on the specific tasks should be interpreted as indicating the additional payoff associated with a specific task relative to any computer use at all. Interestingly,these results show that the most highly re- warded task computers are used for is electronic mail,probably reflecting the fact that high-ranking executives often use E-mail. On the other hand,the results indicate a negative premium for individuals who use a computer for playing computer games.In fact,the-0.11 coefficient on computer games virtually negates the 0.145 coefficient for using computers at all.This result is signifi-
COMPUTERS HAVE CHANGED THE WAGE STRUCTURE 41 of possible nonrandom selection of the workers who use computers. Companies are likely to provide computer training and equipment first to the workers whose productivity is expected to increase the most from using a computer. This would pose a problem for the interpretation of the OLS estimates if these workers would have earned higher wages in the absence of computer use. The large increase in the number of workers who used computers at work between 1984 and 1989 is likely to have reduced the average quality of workers who work with computers, which would be expected to drive down the average wage differential associated with computer use. However, the slight increase in the computer wage premium between 1984 and 1989 suggests that nonrandom selection of the workers who use computers is not the dominant factor behind the positive association between computer use and wages. The other variables in Table I1 generally have their typical effects on wages, and their coefficients are relatively stable between 1984 and 1989. One notable exception is the rate of return to education, which increased by 0.6 percentage points between 1984 and 1989, even after holding computer use constant. And the black-white wage gap increased, while the wage gap between whites and other races declined in these years. C. Specific Computer Tasks The 1989 CPS asked workers what tasks they use their computer for. Respondents were allowed to indicate multiple tasks. Table I11 presents estimates of the coefficients on the specific computer tasks for a wage regression that also includes the covariates listed in column (6) of Table I1 (including occupation dummies). Importantly, the regression includes a dummy that equals one if the individual used a computer for any task at all, as well as dummies for the specific tasks. Thus, the coefficients on the specific tasks should be interpreted as indicating the additional payoff associated with a specific task relative to any computer use at all. Interestingly, these results show that the most highly rewarded task computers are used for is electronic mail, probably reflecting the fact that high-ranking executives often use E-mail. On the other hand, the results indicate a negative premium for individuals who use a computer for playing computer games. In fact, the -0.11 coefficient on computer games virtually negates the 0.145 coefficient for using computers at all. This result is signifi-