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

Area of Science:

• Pharmacoepidemiology
• Biostatistics
• Health Informatics

Background:

• Propensity score (PS) estimation is crucial for confounding control in observational studies.
• High-dimensional data presents challenges for traditional PS estimation methods.
• Effective covariate balance is essential for valid causal inference.

Purpose of the Study:

• To evaluate dimensionality reduction techniques for propensity score estimation in high-dimensional claims data.
• To compare the performance of these techniques against conventional methods for covariate balance and confounding control.
• To assess the utility of autoencoders, principal component analysis (PCA), and logistic PCA in pharmacoepidemiological research.

Main Methods:

• A cohort study used claims data to investigate dialysis and mortality in older patients with heart failure and chronic kidney disease.
• Propensity scores were estimated using investigator-specified covariates, high-dimensional propensity score (hdPS), principal component analysis (PCA), logistic PCA, and autoencoders.
• Covariate balance was assessed using standardized mean differences (SMD) and propensity score overlap plots.

Main Results:

• Autoencoder-based PS achieved the best covariate balance, with only 8 covariates exceeding an SMD of 0.1.
• Principal component analysis (PCA) and logistic PCA also showed improved balance compared to hdPS and investigator-specified covariates.
• Hazard ratios for in-hospital mortality were comparable across all propensity score estimation methods.

Conclusions:

• Dimensionality reduction techniques, particularly autoencoders, demonstrate superior performance in achieving covariate balance for propensity score estimation in high-dimensional claims data.
• These advanced methods hold promise for enhancing the validity of propensity score-matched designs in large-scale pharmacoepidemiological studies.
• Autoencoders offer a powerful approach to improve confounding control and covariate balance in real-world evidence research.