Having discussed the various specifications and analyses for men in detail, the results for women can be discussed more succinctly. In the levels specification in column (3) of Table 3, the OLS estimate of a is actually negative (-.0001) but insignificant. The estimates of the race/ethnicity differentials are similar to those excluding the performance rating. Instrumenting for the performance rating with the education variables, in column (4), causes the estimate of a to rise to .027 and become statistically significant. A similar result holds for the log specification. Columns (5) and (6) instead use the age variables along with education as instruments, with similar results. The overidentifying restrictions are not rejected for this specification, with p-values of .75 in the levels specification, and .52 in logs. Thus, these estimates are preferable. Note also that the F-statistics for the instruments in the first-stage regression are high for all of the specifications discussed so far. As for men, the model in columns (5) and (6) may be misspecifled by including the job requirements variables, so columns (7) and (8) report specifications excluding these variables. The results are little changed. Finally, in column (9) the job requirements variables are also used as instruments. The overidentifying restrictions are not rejected, although the p-values are relatively low (.20 for the levels specification, and .12 for the log specification). Thus, the specifications in columns (7) and (8) are the preferred ones, and the remaining results are discussed in reference to them; nonetheless, the qualitative conclusions are the same for the other specifications.

The IV estimate of a is considerably higher than the OLS estimate in both the linear and log specifications, and the IV estimates are statistically significant. For both the levels and log specification the null hypothesis of no bias in the estimated coefficient of the performance rating is rejected, with p-values of .00. For the linear specification, a one-standard deviation in the performance rating is associated with an increase of .49 in the log wage. For the log specification, a one-standard deviation in the performance rating is associated with a .55 increase in the log wage.

Again the IV estimates appear to generate estimated coefficients of the productivity proxy that map into wage differentials relatively well.