Comments and Discussion

Harry J. Holzer: Andrew Warner has written a very interesting paper on patterns in wages and wage growth across developed and developing countries around the world. Indeed, he addresses one of the most important questions concerning global labor markets—namely, the extent to which different groups of workers in different countries share in the economic growth associated with international trade and are compensated for the higher risks and growing inequality that trade often generates.

Warner's results are based, to a large extent, on data that he has gathered in his own large-scale survey of companies around the world to make his case. This survey effort has been very ambitious and likely yields data that are more comparable across countries than has usually been true for other such studies (as the data are based on very specific occupations and wage definitions as well as on consistent timing).

There are essentially three sets of results in Warner's paper. The first set is from the simple regressions across countries of wages against per capita GDP that are highlighted in figures 1a to 1e. The findings that wages in less-skilled occupations tend to rise proportionately with real GDP, but that managerial wages rise less than proportionately, is quite interesting. The results also imply declining wage dispersion with rising GDP, both between and within countries.

The second set of results appears in tables 3 to 6, from regressions based on a Cobb-Douglas production function. These results show strong effects of capital-labor ratios on wages, as well as negative effects of average levels of skill accumulation on wages in highly-skilled occupations across countries. The third set of results then appears in table 7, where a variety of additional market and institutional characteristics—such as minimum wages, product market competition, and foreign language proficiency among managers—appear to have effects on wages as well, in ways that are broadly consistent with our expectations.

Before getting to the particular empirical findings, one might quibble a bit with the quality of the survey data that Warner has generated. For one thing, the data cover over 50 countries, and the wage measure for each country is based on only about 50 to 60 employers in each case. Warner argues fairly convincingly that his average wage measures are not heavily affected by sample size and are not too sensitive to observable characteristics of the firms in the survey. But there is little information here on how the samples of firms themselves were generated, how response rates varied with firm characteristics (both observable and unobservable), and therefore how representative the data really are within countries and across them.

Regarding the three sets of empirical results, the first set of simple regressions presents clean and straightforward evidence on how average wages and their dispersion varies with real GDP across countries. The tendency of wages to rise strongly with real GDP in each occupation considered cannot be disputed. But the fact that responsiveness to real GDP declines with the rise in skill level of the job is somewhat contradicted by the other studies that Warner cites (Freeman and Oostendorp as well as Ashenfelter and Jurajda), and it is unclear why his results differ from theirs and which are more plausible.

The second set of results, based on production function estimates, clearly indicates the importance of capital-labor ratios in explaining wages across countries within each occupation. It would have been nice to see a bit more about the partial contribution of this variable to the results and about the relationship between this instrument variable and real GDP. More important, there are some questions about the two variables, skill and employment share, that Warner has constructed from other data sources. The descriptions in the appendix of their construction lead me to believe they might be measured with some error, which should lead to downward biases in their coefficients (if the measurement error is "classical," that is, uncorrelated with other variables in the model). This, in turn, might lead the estimated coefficients on capital-labor ratios to be upward biased. Furthermore, whether the skill measures really account...