Assessing Treatment Effect Heterogeneity: Predictive Covariate Selection and Subgroup Identification

Lead Research Organisation: University of Manchester
Department Name: Computer Science


A common objective in a study that involves interventions is the evaluation of heterogeneity of the causal (or treatment) effect. Heterogeneity occurs when the effect varies in subsets of the data which may be defined by a few covariates/characteristics. The presence of heterogeneity can allow us to identify potentially promising interventions for different subsets. This project reviews methods for assessing the causal effect and focuses particularly on two closely related tasks: covariate selection and subgroup identification. The first task allows us to select a few covariates from a potentially large pool of candidates and focus on those that may explain the presence of treatment effect heterogeneity. The second allows us to identify subsets of the data with desirable characteristics, such as enhanced effects. In this project, we will study the properties of different algorithms for covariate selection, treatment effect estimation and subgroup identification. Based on the identified properties, we will explore possible extensions or suggest new approaches.

Studentship Projects

Project Reference Relationship Related To Start End Student Name
EP/N509565/1 01/10/2016 30/09/2021
1965487 Studentship EP/N509565/1 03/09/2016 31/03/2021 Konstantinos Papangelou
Description Machine learning methods for assessing the effect of an intervention are proposed and evaluated. The problem of identifying a few covariates from a potentially high dimensional dataset that can explain the presence of heterogeneity in the effect is discussed. The properties of a covariate selection algorithm, tailored for this task, are studied along with extensions of the algorithm. We additionally study the problem of identifying subgroups (subsets of the data) that exhibit heterogeneity of the effect using weighting methods as well as approaches for evaluating subgroup identification algorithms.
Exploitation Route Institutions working on the evaluation of the effectiveness of an intervention using data-driven approaches may benefit from the methods described in the project. The discussed properties of the algorithms can help to better understand existing methodologies and could be useful in identifying new data-driven methods.
Sectors Digital/Communication/Information Technologies (including Software)