While there is an extensive literature on assessing racial bias in sentencing, no convincing quantitative study has been done for parole release. It is well documented that blacks are less likely to be paroled than white prisoners. Judges have answered that this is the consequence of statistical discrimination and not racism. We develop a model where the judge (or a parole board) is facing a population of prisoners eligible for parole. We assume that this information is summarized by a single-dimensional index about the likelihood that the prisoner will violate parole. Our main assumption is that the judge's objective is to minimize the number of violations subject to the constraint imposed by overcrowding in the prison system. The optimal rule is then to pick a common threshold that depends on this over-crowding constraint. We design an outcome test to see whether parole success data are consistent with this rule. The idea is similar to the strategy that has been used for the determination of racial discrimination in police car searches for drugs. In our context, the key variable is the parole violation rate which is constructed from the National Corrections Reporting Program data. A finding that paroles have systematically lower success for whites may raise concerns that the judge is using a race-contingent parole threshold. Yet, merely observing that the average violation rates vary across groups does not prove that the threshold was different. It is a common issue raised in this literature. We use data on prison capacity expansion to solve this so-called infra-marginality problem. Under the null hypothesis of no discrimination, the ratio of the change in parole violations over the change in the number of paroles granted should be equalized across groups.