Bayesian Multilevel Propensity Score Approach for Causal Inference Using a Latent Modeling Strategy

• In Event: Propensity Scores and Causal Inferences

Abstract

Multilevel observational studies are common in educational research, where students
are nested in schools and treatment selection is not random. Extending propensity
score analysis into multilevel settings can account for both selection bias and
dependencies within clusters. Within-cluster and across-cluster matching are
commonly used, but they often suffer from lack of overlap between treatment and
control units and violation of the strong ignorability assumption, respectively. This
paper introduced a fully Bayesian multilevel propensity score approach that
incorporated a recently proposed latent modeling strategy and compared it with
within-cluster and across-cluster matching. Bayesian propensity score weighting was
utilized to match subjects across clusters but within latent classes. Preliminary
results of a simulation study indicated that the proposed approach was reasonable.

Author

• Jianshen Chen, Educational Testing Service