Dimensionality Reduction Techniques for Improving Propensity Score Specification: An Application to a Cohort Study Using Claims Data.

PURPOSE: To apply various dimensionality reduction techniques for propensity score (PS) estimation in high-dimensional data and evaluate their performance against more conventional approaches in achieving covariate balance and confounding control in pharmacoepidemiological studies using claims data.

METHODS: We undertook a cohort study investigating the association between dialysis and mortality in older patients with heart failure and advanced chronic kidney disease using Optum's de-identified Clinformatics Data Mart Database. We compared PS estimated using investigator-specified covariates, high-dimensional propensity score (hdPS) algorithm, and three dimensionality reduction techniques: principal component analysis (PCA), logistic PCA, and autoencoders. Covariate balance was assessed using standardized mean differences (SMD) and propensity score overlap plots. Hazard ratios for in-hospital mortality were estimated using Cox proportional hazards models.

RESULTS: The analysis included 485 dialysis-exposed and 1455 unexposed individuals after matching. Autoencoder-based PS achieved the best covariate balance (8 covariates with SMD > 0.1), followed by PCA (20 covariates), logistic PCA (25 covariates), hdPS (37 covariates), and investigator-specified (83 covariates). Hazard ratios for in-hospital mortality were similar across PS methods.

CONCLUSIONS: Dimensionality reduction techniques, particularly autoencoders, outperformed traditional methods in achieving covariate balance when estimating PS in high-dimensional claims data. These methods may offer improved covariate balance in pharmacoepidemiological studies PS-matched designs in large healthcare databases.