However, the presence of neuroinflammation could be misclassified, and the severity of neuroinflammation is not adjusted. Third, we use pharmacy claims data to define drug exposure, while we cannot confirm drug adherence and/or cumulative exposure. We largely assume pharmacy claims data are consistent to drug exposure in the overall population, but this assumption may not be hold for all study individuals. Fourth, our findings are subject to unmeasured confounding. The potential confounders include genetic risk factors, social determinants of health, cognitive status. Fifth, our findings are subject to potential model violations such as the proportional hazard assumption in Cox model, while the associations are consistent between covariate-stratified log-rank tests and Cox models (Table S1). Despite these limitations, we believe our findings deserve further investigations, especially for the differential drug responses in AD in a quasi-difference-in-difference setting.
