BGM-IV: AI-Powered Bayesian Generative Modeling for Instrumental Variable Regression with High-Dimensional Covariates
BGM-IV uses a latent Bayesian generative model to estimate causal effects when treatment is endogenous and covariates are high-dimensional. It partitions latent variables into four components representing confounding, outcome mechanism, treatment mechanism, and covariate variation. The method integrates instrumental variable assumptions by using an IV-integrated pseudo-likelihood to provide a principled way to estimate structural functions.