What (if any) are the returns to computer use?

Abstract

Using North American data, we revisit the question first broached by Krueger (1993) and re-examined by DiNardo and Pischke (1997) of whether there exists a real wage differential associated with computer use. Employing a mixed effects model with matched employer–employee data to correct for the fact that workers and workplaces that use computers are self-selected, we find that computer users enjoy an almost 4% wage premium over nonusers. Failure to correct for worker and workplace selection effect leads to a more than twofold overestimate of this premium.

Acknowledgements

We thank Larry Blume, Richard Chaykovski, Rob Clark, Nicole Fortin, Avi Goldfard, David Green, Paul Lanoie, Pierre-Thomas Leger, Thomas Lemieux, Asaf Razin, Paul Ruud, Klaus Schmidt, Michael Smart, Pierre-Yves Steunou and participants in the 2004 CEA and SCSE Conferences for valuable comments. We greatfully acknowledge financial support from the Social Sciences and Humanities Research Council of Canada through the Initiatives on the New Economy programme, Human Resources and Social Development Canada and HEC Montreal. The usual caveat applies.

Notes

¹Several recent papers have examined the effect of technological change on individual wages using European data. Dolton and Makepeace (2004) find a computer wage premium of 10–13% in Britain, while Anger and Schwarze (2003) in Germany and Entorf and Kramarz (1997) and Entorf et al. (1999) in France find no significant premium. Pabilonia and Zoghi (2005) obtain a similar result for Canada. Of these only the last three utilize linked employer–employee data, potentially allowing one to correct for worker and workplace unobserved components and only Entorf et al. (1999) correct for these two simultaneously. We compare our results in Section IV.

²Acemoglu (2002) provides a comprehensive review of the literature on technological change and wage inequality and a lucid analysis of the main arguments.

³To use standard fixed effects, we would have to observe the same worker in different firms–a feature our data does not permit, therefore necessitating the use of this somewhat involved mixed model. We argue later, however, that fixed effects methods may not be appropriate with short panels characterized by little variation. It is also worth noting that this mixed model does not require the orthogonality conditions pertaining to the unobserved components typically demanded of random effects models. Moreover, it can be shown that fixed effects estimates are a special case of mixed model estimates (Abowd and Kramarz, 1999b).

⁴Details about the model, estimation procedure as well as properties of the estimators described in this can be found in Abowd and Kramarz (1999b).

⁵This data constraint also precludes the inclusion of worker–firm match effects.

⁶This is a restricted-access data set available in Statistics Canada Research Data Centers (RDC).

⁷Abowd and Kramarz (1999a) classify WES as a survey in which both the sample of workplaces and the sample of workers are cross-sectionally representative of the target population.

⁸According to the 2001 Current Population Survey (CPS)–the data source used by Krueger (1993)–computer use in the United States among the industries covered by WES was 54%.

⁹Krueger (1993, p. 42).

¹⁰The coefficients on computer use and computer experience are robust to alternative specifications which include, among other things, a large number of organizational practices.

¹¹Furthermore, the fixed effects model relies upon the assumption that staying, leaving and entering computer use each have the same effect on wages. A simple test for this along the lines of Jakubson (1991) and Dolton and Makepeace (2004) leads to the rejection of this assumption for our data, suggesting that fixed effects estimates would be biased in this context.

¹²In the workplace fixed-effects model, the implied bias is 9.7%−5.1%=4.6% compared to 9.7%−6.1%=3.6% in the workplace mixed model presented in column 4 of Table 7.