A Two-stage Bayesian Causal Modeling with Instrumental Variables for Handling Nonnormal and Missing Data

Abstract: 

Confounding errors introduce selection bias. In this talk, I will discuss a Bayesian two-stage causal modeling with instrumental variables to mitigate selection bias, while simultaneously addressing nonnormal and missing data (including ignorable and non-ignorable missingness). The study discusses the treatment of both continuous and categorical treatment. Monte Carlo simulation studies show that the proposed Bayesian approach well address the contaminated data issues, with unbiased LATE (Local Average Treatment Effect) estimates and efficient standard error estimates. The proposed framework can be conveniently implemented using the R software package our team developed.