Nonparametric Causal Mediation Analysis for Stochastic Interventions


Mediation analysis in causal inference has traditionally centered on static interventions and binary exposures, with classical theory introducing the natural (in)direct effects through a decomposition of the average treatment effect. From a decomposition of the population intervention effect, defined via stochastic interventions on exposure and mediators, we outline a framework for defining a variety of interesting causal contrasts, including effects for continuous and categorical exposures. Our (in)direct effects have been shown to require weaker assumptions than their average treatment effect analogs, making them a suitable choice for settings in which the cross-world independencies of traditional mediation analysis are unverifiable; moreover, we discuss extensions of our proposed effects to settings featuring intermediate confounders. We construct and evaluate efficient estimators of our (in)direct effects, all of which accommodate state-of-the-art machine learning in the estimation of nuisance parameters. We discuss theoretical conditions for establishing the asymptotic linearity of our efficient estimators and investigate their practical performance in simulation studies. This is joint work with Iván Díaz and Mark van der Laan, much of it based on

Wed, Apr 29, 2020 1:30 PM
Causal Inference Research Seminar, Department of Statistics, University of California, Berkeley
Berkeley, California, United States (remote due to COVID-19)
Nima Hejazi
Nima Hejazi
NSF Postdoctoral Research Fellow in Biostatistics

My research interests lie at the intersection of causal inference and machine learning, especially as applied to the statistical analysis of complex data from observational studies and experiments in the biomedical and health sciences.