1G-01 - Performance of Different Methods to Handle Multicollinearity for Propensity Score Matching: A Monte Carlo Simulation Study

Background: Pharmacoepidemiologists often include comorbidities and co-medications in propensity score (PS) models, but these variables may be mathematically correlated (e.g., a drug and its target disease), leading to multicollinearity. Multicollinearity can be classified as null, concurrent, or reverse based on types of correlation between covariates (null, positive, or negative). While it affects parameter estimation in PS models, its impact on intervention effect estimation via PS matching remains unclear. This study used Monte Carlo simulation to compare traditional PS modeling (ignoring multicollinearity) with stepwise regression, ridge regression, and principal component analysis (PCA).

Objectives: To assess the impact of multicollinearity-handling techniques on outcome estimation for PS matching using Monte Carlo simulation.

Methods: A Monte Carlo simulation was conducted using 200 datasets with different types of multicollinearity (null, concurrent, reverse). Each dataset contained 10,000 individuals (5,000 treated, 5,000 controls) and 4 continuous covariates for each subjuct. Stepwise regression, ridge regression, and PCA were applied with PS matching. True effect sizes were hazard ratios (HRs) of 0.7, 1, and 1.3. Performance was measured using true effect coverage rate, defined as the proportion of 95% confidence intervals (CIs) containing the true HR.

Results: PS modeling without handling multicollinearity yielded high coverage rates (89.7%, 90.0%, and 86.8% for null, concurrent, and reverse multicollinearity, respectively). PCA slightly improved coverage in null (90%) and concurrent cases (91.2%) but was similar (85.5%) in reverse cases. Stepwise regression reduced coverage, particularly under null multicollinearity (87.3%), indicating potential information loss. Ridge regression showed stable coverage (88.3%, 89.7%, 87.0%).

Conclusions: PS modeling without handling multicollinearity provided robust results regardless of multicollinearity. Stepwise regression, ridge regression, and PCA yielded similar estimates. Multicollinearity had minimal impact on research findings.