Propensity Score Methods for Continuous Treatments in Multilevel Settings: A Deep Learning Approach

• In Event: Advances in Multilevel Modeling for Causal Inferences

Abstract

Unbiased estimation of average treatment effects in multilevel observational studies requires adjustment for pre-treatment individual-level and cluster-level characteristics. Although clusters such as classrooms and schools are typically imbalanced on pretreatment characteristics, much research in propensity score methods ignore cluster effects. This study presents the first deep learning model to successfully estimate generalized propensity scores (GPS) and estimate average dose response functions with the outcome model proposed by Hirano and Imbens (2004) for continuous treatments. In a Monte Carlo simulation study, we compare the estimation and inference of a random effects model, a fixed effects model, and a single-level model to a more flexible feedforward deep learning model.

Authors

• Zachary Kendall Collier, The University of Delaware

• Haobai Zhang, University of Delaware