OBJECTIVES : In the era of big data and machine learning, researchers benefit from efficiency and improved quality in methods. High-dimensional propensity score (HDPS) methods are an extension of traditional propensity score (PS) methods taking full advantage of large data sources and attempting to control for factors not observed in data. This study compares covariate selection, match quality, and investigator time between matched cohorts produced using these methods.

METHODS : Data from a burden of illness study among patients with Wilson’s Disease (WD) using IQVIA’s PharMetrics Plus database were used. Four hundred binary covariates were identified and selected for the HDPS logistic model using the Pharamcoepi Toolbox. The PS logistic model included nine covariates associated with WD using a manual covariate selection process and modeling with SAS macros. Both models included baseline demographic covariates. Patients were matched using a Greedy 1:1 algorithm and absolute standardized difference was used to assess matched cohort balance.

RESULTS : Among the 753 WD patients and 340,256 non-WD patients in the study population, 648 and 689 WD patients were matched using the HDPS and PS methods, respectively. The HDPS covariate selection process identified one third of the study covariates of interest for inclusion in the HDPS model. Absolute standardized differences less than 10% were observed for 92% of covariates in the HDPS model and 84% of covariates in the PS model. The traditional PS process took longer and involved more manual steps than the HDPS process.

CONCLUSIONS : While both methods produced well balanced cohorts, the HDPS method had a greater percentage of balanced covariates and this balance was observed across hundreds of covariates. The HDPS method took the guesswork out of manually identifying and selecting covariates as done in the PS method, and the HDPS method was faster. When considering efficiency and quality, HDPS methods outperformed PS methods in this study.