Assessing Treatment Effect Heterogeneity: Predictive Covariate Selection and Subgroup Identification

Abstract

A key objective in an interventional study, such as a randomised clinical trial, is the evaluation of heterogeneity of treatment effect in the population. This allows us to identify the most promising intervention for a given observation. In this thesis we approach this by targeting two tightly coupled sub-problems. The first concerns the identification of covariates and the second the identification of subgroups associated with treatment effect heterogeneity. Regarding the first problem we study an information theoretic approach. This can be motivated by phrasing the predictive covariate selection problem in log-likelihood terms. We study the properties of this approach in the case of randomised studies and evaluate low-dimensional approximations that are better suited for small-sample and/or high-dimensional studies. We identify some limitations and propose extensions based on propensity score weighting and stratification that extend this criterion in scenarios when the treatment assignment depends on the covariates. Regarding the second problem, we discuss recursive partitioning approaches coupled with weighting methods for treatment effect estimation. The purpose of these methods is to tackle the problem of subgroup identification in the presence of confounders in the data. Finally, studying the literature of subgroup identification we identify a significant number of approaches. Given such a large number of methods to choose from, an important question is how to select the best for a given task. We introduce a framework that uses the subgroup stability as a measure to capture the variations in the identified subgroups due to small changes in the data.

Date of Award	31 Dec 2021
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Gavin Brown (Supervisor) & Tingting Mu (Supervisor)