Abstract:
|
*Student Paper Award* Subgroup identification techniques are often used to distinguish patients who may benefit from a particular treatment from those who may not. While post hoc or unprincipled approaches to subgroup identification may lead to spurious results, theoretically justified approaches are essential to precision medicine. We propose a one-step value difference estimator to test for the existence of a subgroup that benefits from an active treatment. The test statistic is valid under the exceptional law and converges in distribution to a standard normal random variable as the sample size goes to infinity. If the null hypothesis is rejected, subgroups are identified using a readily available estimated treatment decision rule. We consider four versions of the test statistic. The four versions arise from allowing two forms of an estimator for the value difference, and two methods for how observations are partitioned into smaller data sets, which we call 'chunks.' Simulations were performed to study the type I error and power of the four test statistics. In certain simulation settings, the chunking method drastically increases the power of the test.
|