PRM72 - Review Of Methodological Approaches To Generate Propensity Scores In Multilevel Data

Abstract

Propensity score matching (PSM) techniques are frequently used in analyses of retrospective or observational data. Several approaches have been developed to account for the hierarchical structure of data in PSM analyses. The aim of this study is to identify and review existing multi-level PSM methodologies. Medline and PubMed databases were used to perform a targeted literature review to identify studies that use PSM methodologies in multi-level data. The following search terms were used: ‘Propensity score’, ‘Multi-level data’, ‘Hierarchical model’ and ‘Propensity score matching’. Methodologies that specifically considered the challenges of performing PSM with hierarchical data were included in the final review. Six strategies were identified in the literature to perform PSM in multi-level data. These included 1) Complete pooling (CP); 2) Partial pooling (PP); 3) No pooling (NP); 4) Simple single-level modeling (SSLM); 5) “Two stage” modeling (TSM); and 6) “Dummy” modeling (DM). CP ignores potential clustering in the data and is the most commonly used approach. SSLM differs from CP in that it matches patients only within a given cluster. In contrast, the NP method generates separate propensity scores (PS) for each cluster and matches prior to pooling. The PP method uses random intercept models to generate PS and patients are matched across all clusters. The TSM approach first estimates random errors separately and applies them in a subsequent PS model that account for clustering, after which patients are matched as in the PP method. The DM method simply includes the cluster identifier as a fixed effect in the PS model. Performance of each approach is dependent on the number of clusters and the sample size in each cluster. A thorough investigation of data should be undertaken before selecting an approach to use PSM in studies with multi-level data.