Philippe Boileau: Assumption-Lean Differential Variance Inference for Heterogeneous Treatment Effect Detection

Assumption-Lean Differential Variance Inference for Heterogeneous Treatment Effect Detection

The conditional average treatment effect (CATE) is frequently estimated to uncover heterogeneous treatment effects (HTEs). Indeed, CATE inference can be used to refute the homogeneous treatment effect assumption, which states that all units in a population experience identical benefit from a treatment. Uncovering HTEs using CATE inference procedures requires, however, that the pre-treatment covariates modifying the effect of treatment be included in the data. CATE-based techniques will necessarily fail to detect HTEs when the effect modifiers are missing. Such omissions occur when, for example, effect modifiers are not collected owing to limited mechanistic knowledge about the treatment, resource constraints prevent the measurement of hypothesized effect modifiers, or HTE inference only becomes of interest after data is collected. To address this limitation, we propose to investigate the homogeneous treatment effect assumption through inference about the contrasts of the potential outcomes' variances. We derive causal machine learning estimators of these contrasts and study their asymptotic properties. We establish that these estimators are doubly robust and asymptotically linear under mild conditions, permitting formal hypothesis testing about the homogeneous treatment effect assumption in settings where effect modifiers are missing from the data. Numerical experiments demonstrate that these estimators' asymptotic guarantees are approximately achieved in randomized and observational studies alike. These inference procedures are then used to uncover HTEs in the re-analysis of randomized controlled trials.

BIO: Philippe Boileau is an Assistant Professor of Biostatistics at McGill University with a joint appointment in the Department of Epidemiology, Biostatistics, and Occupational Health and the Department of Medicine. He is broadly interested in the development of assumption-lean statistical methods and their application to quantitative problems in the health and life sciences. Assumption-lean procedures combine causal inference and machine learning techniques to avoid unjustified assumptions about the data-generating process, encouraging dependable statistical inference.