Abstract:
|
Cholesterol testing and treatment events in children are rare, mutually dependent, and are affected by guideline promulgation in uneven ways. Observational data is essential to understand existing lipid management, but is difficult to disentangle personal, physician, and system level effects using standard approaches. We proposed a general Bayesian multilevel propensity score analysis in a test case studying the impact of prior lipid screening on the LDL levels of over 1 million children in 3 healthcare systems. This analysis utilized Bayesian multilevel models to estimate propensity score and treatment effect. In order to reduce bias in outcome model treatment estimation due to measured covariates, we stratified subjects from all clusters on estimated propensity score. Treatment effect was estimated with subclass indicators as covariates. To establish the model's comparative benefit, we contrasted the propensity analyses with multilevel versus non-multilevel models. Random intercept propensity score model emerged as the preferred model in term of absolute standardized bias. Posterior results were robust to outcome models with and without multilevel structure.
|