Propensity Score Matching

A previous EHR study of CCBs and neuropsychiatric outcomes had design flaws including potential reverse causality and protopathic bias because exposure qualification and outcome risk windows overlapped. More than 34,000 covariates on average were characterised and balanced within each data source. The propensity score model used LASSO (L1-regularised logistic regression) to select covariates predictive of treatment assignment. In the primary analysis, all covariates were required to achieve a standardised difference of mean of 0.20 or less after matching, with nearly all achieving 0.10 or less. Cohorts were matched 1-to-1 using a caliper of 0.2 of the standardised logit propensity score. Logistic regression was used to generate propensity scores. The study created comparable cohorts using nearest-neighbour propensity score matching at a 1:3 ratio. Matching variables were deliberately limited to age and sex. Residual confounding was acknowledged because the matching model did not include comorbidity indices, utilisation indicators, or socioeconomic proxies.