Pharmacoepidemiology
Choices and Consequences of Alternative Estimands in Time-to-Event Safety Analyses Joy Zora Nakato* Joy Zora Nakato Nakato Nakato Nakato Nakato Nakato University of California, Berkeley
In pharmacoepidemiology, some post-authorization studies aim to assess the effects of treatment exposures (where multiple are available) on the risk of adverse events using electronic health records or insurance claims retrospectively comparing patients who initiated the treatment of interest versus an active comparator. Common challenges in these analyses include treatment switching, incomplete outcome ascertainment, and irregular monitoring (e.g., due to changes in insurance or healthcare provider). These challenges are exacerbated for time-to-event endpoints, where treatment switching and other intercurrent events affect risk sets, censoring mechanisms, and the interpretation of hazard-based estimands. Traditional approaches for time-to-event safety analyses allow the estimator (e.g., Cox proportional hazards) to implicitly determine the causal estimand and rely on strong causal and statistical assumptions. In contrast, alternative causal estimands address different scientific questions in the presence of treatment switching and other intercurrent events. We present several such estimands, including the treatment policy estimand (contrasting initial treatments regardless of subsequent changes), a hypothetical estimand under no treatment switching, a composite that incorporates switching into the outcome, a while-on-treatment estimand that censors at switching, and a principal stratum estimand defined among individuals unlikely to switch under either treatment. Following the Causal Roadmap, we discuss how each estimand fits into a wider research framework. Using simulations, we demonstrate the practical consequences of alternative choices for estimands and their corresponding estimators by comparing traditional parametric and modern machine-learning approaches (e.g., TMLE). We conclude that following the Causal Roadmap when designing safety analyses can help define a set of potentially interesting estimands that can be robustly estimated with modern methods.