Variable Selection in Causal Inference

• Activity Number: 167 - SPEED: Missing Data and Causal Inference Methods, Part 1
• Type: Contributed
• Date/Time: Monday, July 29, 2019 : 10:30 AM to 12:20 PM
• Sponsor: Health Policy Statistics Section
• Author(s): Tingting Zhou* and Michael Elliott and Roderick J Little
• Companies: University of Michigan School of Public Health and University of Michigan and University of Michigan School of Public Health
• Keywords: PENCOMP; Causal Inference; Doubly Robust; Propensity Score; Variable Selection; Spline

Abstract

Inference about causal effects from observational studies requires the measurement and control of confounding variables. A class of methods is based on estimating the propensity of treatment assignment and using the estimated propensity score as a weight, or as a predictor in regression models for the outcome. We recently proposed a robust multiple-imputation based method, penalized spline of propensity for treatment comparisons (PENCOMP), that included a spline of the propensity score as a predictor. These methods are vulnerable to bias and inefficiency when the distributions of estimated propensities in the treatment groups have limited overlap. This situation can arise when conditioning on covariates that are strong predictors of the treatment but not predictors of the outcome. Here we examine by simulation studies the role of confounders in misspecified models. Specifically, we examine the impact of using variable selection techniques that restrict predictors in the propensity model to true confounders on inference from PENCOMP and weighted estimators. We also compare alternative approaches to standard error estimation that incorporates the variability of model selection.