B1173
Title: Optimal weighting for estimating treatment effects
Authors: Michele Santacatterina - New York University (United States) [presenting]
Nathan Kallus - Cornell University (United States)
Abstract: Weighted methods based on Inverse Probability Weights (IPW) have been widely used to estimate causal effects using observational data. Despite their wide use, IPW methods rely on the correct specification of the propensity score model, in which violations lead to biased estimates, and on the positivity assumption, in which practical violations lead to extreme weights and erroneous inferences. We will present novel approaches based on modern optimization techniques and machine learning methods that mitigate model misspecification while simultaneously controlling for precision. We will describe two methods that find weights that balance confounders to estimate the effect of binary and continuous treatments on continuous and time-to-event outcomes. We will also describe a method that finds weights that optimally balance time-dependent confounders for marginal structural models. We will present these approaches using HIV, spine surgery, and heart disease data.
