(352) Quantifying Bias in COVID-19 Vaccine Effectiveness (VE) Estimates Due to Missing Information on Prior SARS-CoV-2 Infection: an Assessment Using Data from the European COVIDRIVE Study

Background: COVIDRIVE is an ongoing, hospital based, test-negative design case-control study of VE for COVID-19 vaccines in Europe. Prior SARS-CoV-2 infection is a frequently missing and important confounder of VE, as it provides protection against future infection and influences individuals’ choice to be vaccinated.

Objectives: To quantify the level of bias in AZD1222 VE estimates in COVIDRIVE resulting from uncontrolled confounding due to missing information on prior SARS-CoV-2 infection.

Methods: VE of an AZD1222 primary series against COVID-19 hospitalisation was derived from multivariable logistic regression models where VE=1-OR. Generalised estimating equations were used to obtain confounder-adjusted estimates, adjusting for symptom onset date, number of chronic conditions, sex and age. Two approaches were used to quantify bias. The E-value quantified the magnitude of uncontrolled confounding required to bring the VE estimate to a prespecified threshold (based on EMA guidance) of 50%. Probabilistic bias analysis quantified the magnitude and direction of the presumed bias by estimating a bias factor to be applied to the unadjusted VE estimate. A bias factor was derived with evidence-based assumptions on the association between prior SARS-CoV-2 infection and COVID-19 hospitalisation (protective effect 82%), prevalence of prior infection among unvaccinated (7-20%), and prevalence of prior infection among vaccinated (5-15% less than in unvaccinated) used in a Monte Carlo simulation. Results were summarised by calculating the median and percentile-based 95% uncertainty intervals.

Results: 1333 patients hospitalised for SARI between 01-MAY-21 and 01-MAR-23 were included in the analysis of AZD1222 VE; 332 test negative controls and 1001 test positive cases. Information on prior infection was missing in 42% of controls and 21% of cases. Over the 22mth study period, with mean time of 142d since vaccination, VE was 65% (95% CI 53-75). The amount of unmeasured confounding required to bring the AZD1222 VE estimate to the 50% threshold was E-value 2.2; meaning the OR of prior infection as a confounder between and vaccination and COVID-19 hospitalisation would have to be ≥2.2 to lower the VE to 50%. After applying bias corrections, AZD1222 VE was 68% (95% CI 67-70), indicating a small underestimation of VE.

Conclusions: Our quantitative bias analysis using evidence-based estimates shows that unknown prior SARS-CoV-2 infection status in COVIDRIVE resulted in non-meaningful underestimation of AZD1222 VE against COVID-19 hospitalisation. The analysis only considered prior infection as a confounder, and not additionally as an effect modifier.