i (3)
Columns (1) and (2) look at sex differences, grouping workers of each race and ethnicity, of the same sex, together, and including race and ethnicity dummy variables in the regression. For both the linear and log specifications, the ratio aoLS/aIV is considerably higher for men (.12-. 13) than for women (.02-.03). The lower estimate of a0LS for women implies that women’s starting wages are much more weakly related to their performance rating—which is measured after they have accumulated some time with the employer—than are men’s. On the other hand, the estimates of aIV are if anything higher for women, implying that their starting wages are at least as strongly related to expected productivity as are men’s. Together, this evidence suggests that employers have considerably worse information about new female employees than about new male employees. However, the standard errors of the estimated ratios of a0LS/aIV are relatively large, so the t-statistics for testing the null hypothesis of equality of these ratios for men and women are in the 1.4-1.5 range, implying that the evidence of a lower ratio for women is not statistically strong.

Turning to the results for whites versus non-whites, there is no evidence that employers have better information on white workers. In particular, in all of the specifications the estimate of the ratio a0LS/aIV is a bit lower for white than for minority workers, indicating if anything slightly worse information about white workers, although the differences are nowhere near significant.