The increased impact of high-dimensional data on a number of scientific frontiers requires an understanding of treatment effect heterogeneity Even if those heterogeneous treatement effects can be estimated accurately, researchers might iteratively search for the treatments or subgroups with high treatment levels and then report the results with the most positive effects. Such practices naturally pave the way for spurious conclusions. To prevent false scientific discoveries, we propose a statistical framework to evaluate the identified most promising treatment or subgroup after a data-dependent search. Our proposed inference procedure is easy to compute, asymptotically sharp and can be generalized to a variety of settings. We demonstrate the merit of proposed method through extensive simulation studies and by analyzing the UK Biobank data.
Speaker: Jingshen Wang, PhD, Assistant Professor of Biostatistics, School of Public Health, University of California, Berkel.