Background: Self-controlled designs, such as the case-crossover (CCO), case-time-control (CTC), self-controlled case series (SCCS) and sequence symmetry design (SSA) use the individuals’ own experience as reference, thereby being robust towards confounders that are stable over time. Self-controlled designs only include individuals with discordant exposure status in the analysis. With minor exposure misclassification, chronic drug users may falsely contribute to the analysis and potentially confer a strong bias towards the null as well as a false narrowing of the confidence interval.
Objectives: Our aim was to demonstrate both a hypersensitivity towards exposure misclassification and too narrow confidence intervals when chronic users are falsely included as discordant in self-controlled analyses.
Methods: In this simulation study, we used citalopram prescriptions from 41,835 users in Denmark. We established formally true treatment episodes by using the recorded days’ supply and assigning a grace period of 90 days to each prescription. Outcomes were simulated 1000 times using the true exposure classification with two different levels of baseline incidence rate (IR) and five levels of incidence rate ratios (IRR), applied while the subjects were truly exposed. In addition, we simulated five proportions of chronic users and five levels of exposure misclassification. We introduced exposure misclassification by shortening the grace period such that artificial treatment gaps were created. Using misclassified data sets, we analyzed the data according to a CCO, CTC, SCCS and SSA design as well as a conventional cohort design with time-dependent exposure, analyzed as treated.
Results: The CCO, CTC and SCCS showed considerably more bias by exposure misclassification than a cohort design. For example, under a scenario with 80% true chronic users, a baseline IR of 0.02 per person-year, and a true IRR of 10, the CCO yielded ORs of 10.3, 5.2, and 2.7 with exposure misclassifications of 0%, 2%, and 5%, respectively. The corresponding IRRs were 10.0, 9.7, and 9.1 in a cohort design. With 10% misclassification in the same scenario, we found relative measures of 1.45, 1.66, 2.05, 9.40 and 8.11 in the CCO, CTC, SCCS, SSA, and cohort design. In addition, confidence intervals narrowed with increasing exposure misclassification in the CCO, CTC and SCCS. Bias was most strongly dependent on the degree of misclassification, proportion of chronic users, IRR and IR, in that order.
Conclusions: Some self-controlled designs show strong bias due to exposure misclassification when the proportion of true chronic users is high
