Endogenous Stratification in Randomized Experiments (LIDS-SSRC Seminar)

Researchers and policy makers are often interested in estimating how policy interventions affect the outcomes of those individuals most in need of help. For example, a researcher may be interested in estimating the effect of financial aid on college enrollment for those students who are likely to drop out of college in the absence of financial aid. This concern has motivated the widespread practice of estimating policy effects separately by brackets of the predicted outcomes that the individuals in a sample would attain in the absence of a policy intervention. This presentation will discuss how substantial biases may arise in practice if predicted outcomes in the absence of the intervention are estimated, as is often the case, using the full sample of individuals not exposed to the intervention. We analyze the behavior of leave-one-out and repeated split sample estimators and show that they behave well in realistic scenarios, correcting the large bias problem of the full sample estimator. We use data from the National JTPA Study and the Tennessee STAR experiment to demonstrate the performance of alternative estimators and the magnitude of their biases.