COVID-19 Pandemic
Vaccine effectiveness estimates from a test negative design without controlling for prior infection underestimate true vaccine effectiveness Ryan Wiegand* Ryan Wiegand Bruce Fireman Morgan Najdowski Mark Tenforde Ruth Link-Gelles Jill Ferdinands
Test negative designs (TNDs) are frequently used to assess vaccine effectiveness (VE). For COVID-19 vaccines, differences in infection-induced protection between vaccinated and unvaccinated individuals may confound the association between vaccination and outcomes but ascertaining infection history can be challenging, especially for capturing asymptomatic or paucisymptomatic infections. Failure to account for infection history may bias VE estimates. We simulated individual-level SARS-CoV-2 infection and COVID-19 mRNA vaccination histories from January 19, 2020, to May 13, 2023. Simulated vaccination rollouts and TNDs were used to calculate VE against infection and against severe disease. VE was calculated from logistic regression models as (1-OR)*100%. Models were either unadjusted for infection history (uncontrolled) or were adjusted for the months since last infection and number of prior infections (controlled). Bias was assessed in estimates of VE from controlled as well as uncontrolled models compared to VE calculated without infection-induced protection. Uncontrolled models typically underestimated VE more than controlled models (Figure). For vaccination in the last 3 months, mean bias for protection against infection was -4.8 percentage points (pp) (95% CI -5.7, -3.9) for uncontrolled analyses and -3.0 pp (CI -3.9, -2.0) for controlled analyses, and mean bias for protection against severe disease was -4.0 pp (CI -5.5, -2.4) for uncontrolled and -2.4 pp (CI -4.0, -0.7) for controlled. When VE approaches zero, such as when vaccine-induced protection has waned, failure to account for infection history may result in a VE estimate below zero, e.g., in uncontrolled analyses, VE against infection 5-11 months since vaccination was -2.4% (CI -2.7, -2.0). VE estimates from TNDs without prior infection information tended to underestimate VE, in most scenarios by less than 8 percentage points.
