Using recursive partitioning to find and estimate heterogenous treatment effects in randomized clinical trials

When for an RCT heterogeneous treatment effects are inductively obtained, significant complications are introduced. Special loss functions may be needed to find local, average treatment effects followed by techniques that properly address post-selection statistical inference. Reanalyzing a recidivism RCT, we use a new form of classification trees to seek heterogeneous treatment effects and then correct for “data snooping” with novel inferential procedures. There are perhaps increases in recidivism for a small subset of offenders whose risk factors place them toward the right tail of the risk distribution. A legitimate but partial account for uncertainty might well reject the null hypothesis of no heterogenous treatment effects. An equally legitimate but far more complete account of uncertainty for this study fails to reject the null hypothesis of no heterogeneous treatment effects.