Primary Submission Category: Bayesian Causal Inference
A New Bayesian Spike and Slab Approach for Causal Effect Estimation
Authors: Brandon Koch,
Presenting Author: Brandon Koch*
Two Bayesian approaches have recently been developed for causal effect estimation that use spike and slab priors. Rather than modeling the outcome only (e.g., using a lasso to model the outcome as a function of treatment and covariates), the approaches consider models for both outcome and treatment with priors that aim to control for important confounding variables weakly related to outcome. Both approaches have been shown to estimate the treatment effect with small bias and variability compared to alternative approaches across a variety of simulations, especially in settings when the sample size is small in relation to the number of total covariates under consideration. Coverage rates of confidence intervals are also generally higher when using the spike and slab approaches compared to alternative techniques. However, one approach cannot be directly extended to non-continuous outcomes or account for treatment effect heterogeneity, and the other approach requires separately fitting heterogeneous and non-heterogeneous models and choosing one based on an information criterion. In this talk, we discuss a new Bayesian spike and slab method that addresses these limitations by introducing a model and prior that is applicable to non-continuous outcomes and allows the treatment effect to be homogeneous or heterogeneous based on the data without fitting two separate models. Simulations demonstrate improved effect estimation with the new approach over the alternative approaches.