Impact of exposure time trends on estimators of self-controlled designs in drug safety assessment using electronic healthcare databases: a simulation study

In drug safety assessments, the secondary use of electronic health databases can be challenging because information on confounding factors for assessing the association between drugs and adverse drug reactions is often incomplete. Self-controlled designs can mitigate confounding by time-invariant factors. However, drug use frequency may change over time, and exposure time trends can affect the estimates obtained from these designs. We conducted a simulation study to evaluate the performance of five self-controlled designs: case-crossover, case-time-control, case-case-time-control, symmetric bidirectional case-crossover, and self-controlled case series under various exposure time trends and scenarios. The exposure time trends included four patterns: constant, linear increase, cosine function-pattern temporal change, and mountain-pattern temporal change. We investigated three scenarios: the impact of exposure time trends alone, events restricting the length of the observation periods, and events influencing subsequent exposures. Case-crossover exhibited the most pronounced positive bias in estimates when the exposure time trend demonstrated a linear increase of more than 20% per year or a mountain-pattern temporal change. Additionally, under the cosine function-pattern temporal change, symmetric bidirectional case-crossover exhibited a negative bias. By contrast, case-time-control, case-case-time-control, and self-controlled case series showed little bias from exposure time trends, although case-time-control and case-case-time-control were less precise than the other self-controlled designs. Furthermore, when adverse drug reaction occurrence affected the observation period or subsequent exposure, the estimation performance of symmetric bidirectional case-crossover and self-controlled case series was substantially affected. When applying self-controlled designs to drug safety assessment, exposure time trends affected bias and estimation precision differently depending on the pattern of the exposure time trend and self-controlled design. Adverse drug reaction-dependent changes in the observation period or subsequent exposure substantially affected the estimation performance of self-controlled case series and symmetric bidirectional case-crossover. These findings suggest that design selection should first consider whether adverse drug reaction occurrence may affect the observation period or subsequent exposure, followed by consideration of exposure time trends and the available data structure.