Applying two approaches to detect unmeasured confounding due to time-varying variables in a self-controlled risk interval design evaluating COVID-19 vaccine safety signals, using myocarditis as a case example.

Abstract

We test the robustness of the self-controlled risk interval (SCRI) design in a setting where time between doses may introduce time-varying confounding, using both negative control outcomes (NCOs) and quantitative bias analysis (QBA). All vaccinated cases identified from 5 European databases between September 1, 2020, and end of data availability were included. Exposures were doses 1-3 of the Pfizer, Moderna, AstraZeneca, and Janssen COVID-19 vaccines; outcomes were myocarditis and, as the NCO, otitis externa. The SCRI used a 60-day control window and dose-specific 28-day risk windows, stratified by vaccine brand and adjusted for calendar time. The QBA included two scenarios: (1) baseline probability of the confounder was higher in the control window and (2) vice versa. The NCO was not associated with any of the COVID-19 vaccine types or doses except Moderna dose 1 (IRR = 1.09; 95% CI 1.01-1.09). The QBA suggested that even the strongest literature-reported confounder (COVID-19; RR for myocarditis = 18.3) could only explain away part of the observed effect, from IRR = 3 to IRR = 1.40. The SCRI seems robust to unmeasured confounding in the COVID-19 setting, although a strong unmeasured confounder could bias the observed effect upward. Replication of our findings for other safety signals would strengthen this conclusion. This article is part of a Special Collection on Pharmacoepidemiology.