1230Should multiple imputation be stratified by exposure group when estimating causal effects via outcome regression

Abstract Background Outcome regression remains widely applied for estimating causal effects in observational studies, in which causal inference is conceptualised as emulating a randomized controlled trial (RCT). Multiple imputation (MI) is a commonly used method for handling missing data, but while in RCTs it has been shown that MI should be conducted by treatment group to reduce bias, whether imputation should be conducted by exposure group in observational studies has not been studied. Methods We conducted a simulation study to evaluate the performance of seven methods for handling missing data: Complete-case analysis (CCA), MI of main effect, MI with interactions (between exposure and: outcome, a strong confounder, outcome and a strong confounder, all incomplete), and MI conducted by exposure group. We simulated data based on an example from the Victorian Adolescent Health Cohort Study. Three exposure prevalences and seven outcome generation models were considered, the latter ranging from no interaction to strong-positive or negative exposure-confounder interaction. Various missingness scenarios were examined: with incomplete outcome only or also incomplete confounders, and three levels of complexity regarding the missingness mechanism. Results For all scenarios, MI by exposure led to the least bias, followed by MI approaches that included exposure-confounder interactions. Conclusions If MI is adopted in outcome regression, we recommend conducting MI by exposure group and, when not feasible, including exposure-confounder interactions in the imputation model. Key messages Similar to RCTs, MI should be conducted by exposure group when estimating average causal effects using outcome regression in observational studies.