i (1)
Finally, the performance ratings pose some pure measurement problems, because they may vary for reasons other than the worker’s actual performance. In particular, the ratings that particular respondents provide may vary for random reasons, with some tending to give higher and some lower ratings for equally-productive workers. This case may be interpreted as one of pure measurement error in the performance rating. Unfortunately, in this case the instrumental variables procedure may be correcting for this pure measurement error, rather than that which arises in the imperfect information story because of discrepancies between current and expected productivity. Because the IV procedure may simply correct for standard measurement error bias, it could lead to spurious evidence in favor of the statistical discrimination model, with a,v larger in absolute value than aoLS, and bIV falling in absolute value relative to b0LS.
Nonetheless, if there is taste discrimination, blv should still be significantly different from zero, as opposed to the case of pure statistical discrimination. While this implies that a test of whether bIV differs from zero provides a test for taste discrimination, a significant difference between bIV and b0LS would not necessarily imply that there is statistical discrimination. As a result, considerable effort is devoted to distinguishing between the statistical discrimination and pure measurement error interpretations of the findings, an issue FR did not consider. so

Aside from these potential problems with the productivity proxy, the MCSUI data offer some advantages for this study. First, the wage measure is a starting wage, which, according to the statistical discrimination model, is the wage that should be set equal to expected productivity. On the other hand, the performance rating is current, and workers in this data set have average tenure of about two to three months, so that we would expect actual productivity to differ from expected productivity. Thus, the time frame to which the data refer are precisely what is required for the tests this paper considers.