Turning to the question of substantive interest, the estimated wage shortfalls for blacks and Hispanics in the OLS estimates of the levels specification are -.178 and -.066 respectively, with the latter only significant at the ten-percent level. However, instrumenting for the performance rating causes the differential for blacks to fall to -.062, and that for Hispanics to change sign; both estimated differentials become statistically insignificant. Qualitatively, these reductions in the wage shortfalls for black and Hispanic men are consistent with a substantial part of these shortfalls being attributable to statistical discrimination rather than taste discrimination.

(1) (2) (3) (4) (5) (6) (7)
В. Using loe performance ratine
Performance rating .289(.068) .220(.057) .034(.657) .252(.068) 2.012(1.025) 1.899(.689)
Black -.143(.032) -.124(.033) -.136(.053) -.175(.039) -.059(.088) -.067(.070)
Hispanic -.039(.030) -.008(.030) -.020(.052) -.068(.036) .043(.082) .036(.065)
R2 .384 .027 .347 .056
P-value from F-test of instruments in first-stage regression:
Age variables only .09
Education variables only .32
Job requirement variables only .40
Instruments Age Age Age, education, job requirements
F-statistic on instruments in
first-stage regression: 2.47 2.92 1.51
Overidentifying restrictions, p-value: .00
Bias in OLS estimates, p-value from Hausman test:
Performance rating .78 .09 .02
Black .78 .14 .06
Hispanic .78 .13 .06

The results are similar for the log specification. Hausman tests to gauge the statistical significance of the differences are reported in the last two rows of each panel. In the levels specification, the p-values for the test of the null of pure taste discrimination—implying no bias in the OLS estimates of the race/ethnicity wage differentials—are .11 for blacks and .07 for Hispanics. In the log specification, the corresponding p-values are .14 and .13. Thus, there is some evidence against the lower wages paid to blacks and Hispanics reflecting solely taste discrimination, although it is not overwhelming.

Finally, column (7) repeats the IV estimation, but now using education and job requirements as instruments as well. The estimates are more precise as we would expect, leading to lower p-values from the tests of the null of taste discrimination (in the .06-.09 range). However, the overidentifying restrictions are rejected, and the full set of instruments has relatively weak predictive power for the performance rating, as reflected in the F-statistics of 1.50 or 1.51. Thus, this column probably does not provide very reliable evidence.