Abstract:
|
We propose a Bayesian nonparametric approach to estimate causal effects in the longitudinal treatment setting. We use Enriched Dirichlet process mixtures to model the longitudinal treatments, time-varying confounders, and the response and use G-computation for compute the causal effects. This flexible modeling avoids typical parametric G-computation approaches. In addition, most of the existing literature requires separate sets of models to estimate different causal effects (e.g., mean causal effect vs quantile causal effects). Contrary to such procedures, in this Bayesian nonparametric approach, we estimate the distribution of the potential outcomes, allowing us to calculate any causal effect. The proposed method can address ignorable missing covariates automatically through data augmentation; this method does not depend on a separate imputation model. We provide simulation studies and applications to real-world data to demonstrate the efficacy of the proposed method.
|