High-Dimensional Confounder Reduction for the Estimation of a Causal Quantity

• Activity Number: 110
• Type: Topic Contributed
• Date/Time: Monday, August 4, 2014 : 8:30 AM to 10:20 AM
• Sponsor: ENAR
• Author(s): Mireille Schnitzer*+ and Judith J. Lok and Susan Gruber
• Companies: University of Montreal and Harvard School of Public Health and Harvard School of Public Health
• Keywords: causal inference ; confounder ; variable selection ; propensity

Abstract

In observational cohorts, estimation of the effectiveness of a non-randomized treatment generally requires the implementation of causal modelling. Often this will involve the estimation of a propensity model, which predicts the probability of taking treatment conditional on all measured suspected confounders. However, in domains like pharmacoepidemiology, this set of hypothetical confounders might be very high-dimensional, while a sufficient adjustment set (that would properly control for confounding bias) might actually be much smaller. In this talk, I will give an overview of the recent literature on causal variable selection, highlight mistakes in common practice, and evaluate different strategies through simulation study and example.