Use of High-dimensional Propensity Scores (hdPS) in a Japanese National Claims Database
Background: There is much evidence demonstrating the utility of hdPS algorithms for confounding control in comparative safety and effectiveness studies using real-world data (RWD) from the US and UK. However, its performance using Japanese RWD remains unclear given the potential differences in healthcare practice and how the data are captured.
Objectives: Evaluate the performance of hdPS to improve confounding control in a national claims-based Japanese RWD source using the known moderate protective effect of COX-2 inhibitor (Cox-2i) on severe gastrointestinal (GI) complications as an example.
Methods: Using the Japanese Medical Data Center (JMDC) national RWD source with claims for >10 million lives, we selected a cohort of patients who had a new prescription claim for a COX-2i or a nonselective “traditional” NSAID (tNSAID) Jan 2007 - Dec 2011 with continuous enrollment and no evidence of either treatment during the 12 months prior (baseline). We employed 1:1 nearest neighbor matching on the propensity score (PS) for initiating a COX-2i given baseline covariates using 3 separate PS models: 1) prespecified covariates; 2) hdPS-selected covariates + age and sex, and 3) prespecified + hdPS-selected covariates. Candidate covariates for the hdPS model were autoselected from the following: 3-digit ICD-10 diagnosis codes, EphMRA ATC codes, and procedure category names. Covariate balance was evaluated using absolute standardized differences (ASD) between the treatment arms of pre-specified covariates, defining imbalance as an ASD > 0.1. Patients were followed from treatment initiation until the first occurrence of the outcome, severe GI complication (diagnostic claim for GI hemorrhage or peptic ulcer disease complications), or disenrollment, end of data, or a maximum of 180 days.
Results: We identified 4,731 initiators of COX-2i and 322,490 initiators of tNSAIDs before matching. COX-2i initiators had more comorbidities and healthcare resource utilization expected. While 7 out of 19 covariates were imbalanced in the unmatched cohort, covariate balance was achieved for all covariates in all PS models, with ASDs for the hdPS models generally lower than the prespecified. The unmatched HR for COX-2 versus tNSAIDs initiators was 1.11 ( 0.84 - 1.45), and matched HRs were 0.87 (0.61 - 1.26), 1.02 (0.80 - 1.30), and 0.72 (0.50 - 1.02) for the prespecified, hdPS alone, and prespecified + hdPS models, respectively.
Conclusions: Compared to the model with prespecified covariates alone, the addition of hdPS achieved greater balance and yielded an estimate closest to the expected effect for COX-2 inhibitors versus tNSAIDs, demonstrating its utility in Japanese RWD.