On discovery of subpopulations with differential treatment effects

N. T. Longford

Abstract

In a commonly adopted approach, the search for a subpopulation in which a treatment effect is different than in the complementary subpopulation entails complex modelling of interactions of the treatment indicator with background variables. We introduce a relatively simple graphical method for this problem. It is an adaptation of the general idea of post-observational design, which can be traced to the potential outcomes framework. Its essence is to construct a dataset comprising matched pairs, in which the influence of the background variables not involved in the search is eliminated. The method is equally suitable for data from randomised trials and observational studies. An application to transfers in neonatal care is presented.
 
  Submitted.
February 2017