Statistical Versus Taste Discrimination

The empirical approach to testing for employer taste discrimination versus statistical discrimination that is the starting point for this paper was originally developed by FR. The discussion here is geared more closely towards the data used in the present paper, as discussed fully in Section III. Suppose that data are available on starting wages (ws), race (R, a dummy variable defined as one for minorities and zero otherwise), and (marginal) productivity (P).4 It is assumed that P is constant over the time horizon considered by the employer (ruling out human capital investment), and that there are no incentive considerations (as in Lazear, 1979) that lead wages to diverge from productivity. According to the simple statistical discrimination model, ws is set equal to the expected value of P when the worker begins the job, denoted Ps\ and defined as
where Is is all information about the worker available to the employer when the starting wage is set.5 Under the null hypothesis of no race discrimination, in the regression

we should find that p = 0. However, we do not have data on expected productivity, but only on actual productivity P, where because of equation (1),

Thus, the estimated equation is website


where Var(P/)/Var(P) is the reliability of the information available on new hires.

Now suppose that P and R are negatively correlated, which will occur if minorities are on average less productive. In this case, the OLS estimate of p (bOLS) will also be biased downward.

As a result, OLS estimates of equation (4) may lead to evidence that taste discrimination generates race differences in wages, because controlling for productivity, minorities are paid lower starting wages. But the downward bias in the estimate of p implies that p could nonetheless equal zero, with starting wages conditional on expected productivity not reflecting any race differential.