time and time from event as covariates and restricting participants to 2-year observation periods surrounding the index event in order to reduce heterogeneity in observation time. We also used bidirectional sampling in order to reduce overlap bias resulting from control period selection as a function of event time.25 Third, even though self-matching is used by the case-control design, residual confounding by unmeasured time-invariant variables cannot be ruled out. Fourth, the MarketScan data only encompass insured patients with observed alcohol-related acute events, which limits our study’s generalizability.
