The criminal justice system is a complex institution with a great deal of power over the lives of the individuals who enter that system as defendants. There is justifiably a great deal of concern both in policy circles and the academic literature about unwarranted disparity in general, and racial discrimination in particular. Often, research on this question focuses on differences in treatment across racial groups. However, differential treatment may in fact be warranted by differences in individuals that are observed by decisionmakers but unobserved but researchers. A common approach is to control for observables, even when there is only relatively sparse information available to researchers. An alternative approach, what we call Standard Outcome Analysis (SOA), recently has gained prominence in the economics literature in the context of racial profiling by police. This approach is founded on rational choice theory and starts with a clear statement of the objective function of the criminal justice actor. Proponents of SOA claim that observed (average) levels of the objective function should be constant across racial groups if the actor in question is unbiased. In this proposal, we consider this approach in the context of bail-setting by judges. In the older, control-for-observables literature, authors using regression-based models consistently find substantial racial disparities in bail levels. Ian Ayres has conjectured that in such a case—where the decision variable is continuous—SOA will provide a consistent test for racial discrimination. To use SOA in this context, one simply compares failure-to-appear (FTA) rates across race, among those offered positive bail levels. Using a simple model, however, we show that equality of FTA rates generally is neither necessary nor sufficient for unbiased bail-setting. In this proposal, we develop an alternative approach, which we call General Outcome Analysis. To implement this approach, we need exogenous variation in bail amounts. We propose to estimate the model using interesting new administrative data from the Washington DC Pretrial Services Agency; as part of the project, we will assemble and clean a dataset suitable for estimation. Exogenous variation in bail levels comes from the arguably exogenous assignment of defendants to judges. We will estimate both standard and generalized outcome models using these data. While these results will be of independent interest, they will also allow us to see whether our generalized approach can be practically useful in an important application.
