(341) Evaluating impact of time-varying confounders in self-controlled risk interval analysis in a vaccine-safety cohort event monitoring-based simulation study

Background: Self-controlled case series (SCCS) and self-controlled risk interval (SCRI) analyses are increasingly popular tools to study vaccine and medicine safety using real-world data (RWD), where multiple routine care data sources are combined. In these analyses, the incidence of adverse events (AEs) is compared between risk and control periods for the same individual, eliminating time-invariant confounding. However, little is known about the performance of these methods in cohort event monitoring studies, in which time-varying confounding and loss to follow-up are likely bigger problems than when using RWD.

Objectives: This study evaluates the performance of the SCRI design in the presence of time-varying confounding and loss to follow-up common in cohort event monitoring studies, and how design choices common in RWD settings (e.g. counting only the first record of an event when multiple recordings of the same events appear) impact this performance.

Methods: We simulate recurrent AEs per individual over 180 days from taking the vaccine or medicine, considering the first 42 days as risk period. We vary the cumulative incidence in the control window (1–10%), the incidence rate ratio (IRR: (1–5), and presence/absence of a time-varying confounder that is predominantly prevalent in the first 28 days and doubles the risk of experiencing the AE. Conditional Poisson regression models were fitted considering all AEs versus only an individual’s first AE and controlling versus ignoring a time-varying confounder. We calculated bias, root mean square error (RMSE) and coverage probabilities of the 95% confidence intervals.

Results: Preliminary results show that incorrectly counting only the first event results in biased estimates (bias 0.039–0.55 for a true IRR=5) and coverage probabilities below 95% (0.000–0.928 for true IRR=5). Counting all events, but ignoring a time-varying confounder in the analysis also led to biased estimates (bias of 0.039 to 0.150 for IRR = 5). After adjustment for the time-varying confounder, unbiased estimates (bias close to 0.001 for IRR = 5) and adequate coverage probability of 95% confidence intervals (close to 0.95 or higher) were obtained, except in settings where the risk of being lost to follow increased after the subject experienced a first event.

Conclusion: The simulation showed that SCRI analyses using only the first event can be prone to bias when subjects truly experience distinct events multiple times in an observation window, and that ignoring time-varying confounding can produce bias. Loss to follow-up that is dependent on having experienced an AE poses challenges for standard analyses. The next step will be to investigate alternative statistical methods that can deal with outcome-dependent loss to follow-up and unmeasured confounders.

Keywords: Bias, Conditional Poisson models, COVID-19 vaccines; Cohort Event Monitoring; Safety; Self-controlled risk interval, Simulation