A causal-inference method for observational data that estimates the individual treatment effect (ITE) by learning, for each unit, a posterior distribution over a low-dimensional set of latent features (latent confounders) that jointly drive treatment and outcome. Combining deep generative modeling with Bayesian inference and fitted by an iterative algorithm that alternately updates model parameters and latent features, it produces individualized effect estimates with well-calibrated posterior intervals while mitigating confounding in high-dimensional covariate settings.
Estimating causal effects from observational data is hard when high-dimensional covariates confound the relationship between treatment and outcome. This paper introduces CausalBGM, a Bayesian generat…