High-Dimensional Propensity Scores for Mitigating Confounding: Implementation Using Primary and Secondary Care Data...

Area of Science:

• Observational Health Data Science
• Pharmacoepidemiology
• Biostatistics

Background:

• Confounding is a major challenge in observational healthcare database studies.
• High-dimensional propensity score (HDPS) algorithms leverage comprehensive data to reduce residual confounding.
• Implementing novel methods like HDPS in new settings is crucial for robust research.

Purpose of the Study:

• To implement and evaluate High-dimensional propensity score (HDPS) methods in a Hong Kong (HK) healthcare database.
• To investigate the association between antihypertensive drug classes and incident dementia risk using HDPS.
• To assess the impact of HDPS on covariate balance and identify potential confounders.

Main Methods:

• A cohort study was conducted using HK data on new users of antihypertensives.
• High-dimensional propensity score (HDPS) was implemented, including the top 250 covariates, for inverse probability of treatment weighting.
• Covariate balance was assessed, and sensitivity analyses were performed.

Main Results:

• The study included 434,506 new users of antihypertensives.
• Traditional propensity score (PS) methods showed no association between antihypertensives and dementia risk.
• HDPS implementation revealed moderate evidence of reduced dementia hazard with beta-blockers compared to ACE inhibitors (HR: 0.90, 95% CI: 0.82-0.98), and improved covariate balance.

Conclusions:

• Successful implementation of HDPS in HK data demonstrated improved covariate balance.
• HDPS identified potential database-specific frailty markers, enhancing confounder adjustment strategies.
• The findings highlight the utility of HDPS in mitigating confounding in real-world health data.