Keywords: latent variables analysis, treatment effect heterogeneity, health policy research
For making a causal inference in the presence of unmeasured confounders, IV analysis plays a crucial role. Most of the existing IV methods utilize a method of moments approach within a structural models framework. In this talk, we explore likelihood-based IV estimators and necessary assumptions to make a valid causal inference. We focus on a Bayesian approach to make posterior inference on the estimates of causal effects of interest, with particular emphasis when there is treatment effect heterogeneity. We extend existing parametric approach by accounting for an unobserved heterogeneity via a latent structure that leverages on Dirichlet process mixture priors. Our approach provides a flexible framework to model complex latent structures and to account for correlation that standard approach does not. We utilize simulations to characterize operating characteristics of the estimators. A novel application to determine causal effect of radial artery access on vascular outcomes compared to femoral artery access for patients undergoing cardiovascular procedures demonstrates the utility of the proposed IV analysis. Funded by R01-GM111339.