Bayesian propensity score analysis for clustered observational data

Observational data with clustered structure may have confounding at one or more levels which when combined critically undermine result validity. We propose using multilevel models in Bayesian propensity score analysis to account for cluster and individual level confounding in the estimation of both propensity score and in turn treatment effect. In addition, our approach includes confounders in the outcome model for more flexibility to model outcome-covariate surface, minimizing the influence of feedback effect in Bayesian joint modeling of propensity score model and outcome model. In an extensive simulation study, we compare several propensity score analysis approaches with varying complexity of multilevel modeling structures. With each of proposed propensity score model, random intercept outcome model augmented with covariates adjustment well maintains the property of propensity score as balancing score and outperforms single level outcome model. To illustrate the proposed models, a case study is considered, which investigates the impact of lipid screening on lipid management in youth from three different health care systems.