Use of Aggregated Covariates In Propensity Score Analysis of Clustered Data

Abstract

Propensity score methods can be used to reduce selection bias and improve causal inferencing with nonrandomized data. However, there is little guidance for implementing a propensity score analysis when treatment exposure is a property of clusters rather than subjects. For example, education policies and practices are often implemented by school or district rather than by individual student. The three studies in this dissertation strive to clarify procedural quandaries for a propensity score analysis with cluster-level treatment exposure and subject-level outcomes. Additionally, omission of a true confounder from a propensity score analysis can bias treatment effect estimation. My dissertation also explores the utility of aggregated covariates as replacements for missing true cluster-level confounders. The first simulation study compared four procedures for generating aggregated covariates. The results highlight that: 1) researchers need to verify the comparability of generated samples to real world contexts; 2) a propensity score analysis with cluster-level treatment exposure requires at least 60 clusters. The second simulation compared covariate balance and treatment effect estimation when appraising treatment exposure by subjects or by clusters and including aggregated covariates of varying quality. Treatment appraisal by subjects outperformed appraisal by clusters under certain conditions. When highly correlated (r = .92 - .98) with the missing true confounders, aggregated covariates were viable replacements. The last study applied the guidance from the simulations to statewide survey data. The investigation found no association between the presence of a school resource officer and students’ social-emotional well-being and academic performance. A critical caveat is the results may not generalize to student populations that have historically been targeted by discrimination and school violence.