Machine learning for variable selection and model smoothing for longitudinal causal inference

Mireille Schnitzer (Universite de Montreal)

Wednesday 10th December 14:00-15:00
Maths 311B

Abstract

Longitudinal causal inference answers the question of how time-varying exposures affect outcomes. One way of representing causal effects in this context is through marginal structural models. In observational studies, any baseline and time-dependent confounding must be adjusted for, typically through propensity scores or outcome models, or both. However, such adjustments dramatically reduce statistical power, sometimes needlessly as when the causal effect is estimable under more limited adjustment. Ad hoc statistical model smoothing is typically recommended in order to make the estimation problem tractable or to reduce estimation variance. However, if the ad hoc smoothing is incorrect, this can induce bias in the causal estimation. In this talk, I will first explain how to identify which covariates are important to adjust for and which only increase variance. I will then introduce our recent adaptive LASSO variable selection procedure for time-specific propensity scores that targets the optimal covariate set. Finally, moving to a nonparametric context, I will present a nonparametric extension of this method which allows for smoothing of propensity score functions and asymptotically linear causal estimation with theoretically justified inferential procedures.

Add to your calendar

Download event information as iCalendar file (only this event)