Using a Two-Stage Propensity Score Matching Strategy and Multilevel Modeling to Estimate Treatment Effects in a Multisite Observational Study
In this study I present, demonstrate, and test a method that extends the Stuart and Rubin (2008) multiple control group matching strategy to a multisite setting. Three primary phases define the proposed method: (1) a design phase, in which one uses a two-stage matching strategy to construct treatment and control groups that are well balanced along both unit- and site-level key pretreatment covariates; (2) an adjustment phase, in which the observed outcomes for non-local control group matches are adjusted to account for differences in the local and non-local matched control units; and (3) an analysis phase, in which one estimates average causal effects for the treated units and investigates heterogeneity in causal effects through multilevel modeling. The main novelty of the proposed method occurs in the design phase, where propensity score matching is executed in two stages. In the first stage, treatment units are matched to control units within the same site. In the second stage, treatment units without an acceptable within-site match are matched to control units in another site (between-site match). The two-stage matching method provides researchers with an alternative to strict within-site matching or matching that ignores the nested data structure (pooled matching). I employ an empirical illustration and a set of simulation studies to test the utility and feasibility of the proposed two-stage matching method. The results document the two-stage matching method's conceptual appeal, but indicate that effect estimation under the two-stage matching method does not, in general, outperform more traditional matching-based or regression-based methods. Alternative specifications within the proposed method can improve performance of two-stage matching. In addition to extending the work of Stuart and Rubin, this study complements the small set of studies that have examined propensity score matching in multisite settings and provides guidance for researchers looking to estimate treatment effects from a multisite observational study. The dissertation concludes with directions for future research and considerations for researchers conducting multisite observational studies.