Modelling approaches for estimating vaccine effectiveness of consecutive SARS-CoV-2 variant sublineages in the absence of study-specific genetic sequencing data, VEBIS hospital network, Europe, 2023/24

Abstract

Introduction Genetic changes in COVID-19 variants/sublineages (VSLs) can reduce vaccine effectiveness (VE). Timely VSL-specific VE estimates are essential, but study-specific VSL identification by whole genome sequencing (the “gold standard”) is expensive and time-consuming. Alternatively, VSL-specific VE has been estimated from external sequencing data (VSL predominance period by proxy: PP). We propose two novel approaches for use in test-negative design (TND) studies to estimate VSL-specific VE when study-specific VSL identification is not possible. Methods We demonstrate the variant category model (VCM) and the variant proportion model (VPM) approaches. Using data from a hospital-based TND study among adults ≥65 years, during the period of sequential predominance of XBB and BA.2.86 in 2023/24, we estimated the VE as (1-OR) x 100%. For the VCM, we used a binary variable categorising “most likely underlying sublineage” based on publicly available sequencing data. For the VPM, we used a continuous variable with values from 0 to 1 representing the weekly proportion of BA.2.86. We validated results using study-specific VSL identification from sequenced study data (SD) and the standard PP approach. Results Overall, at 14–59 days post vaccination, VE point estimates against XBB were within ±3% absolute for the VE estimated using both models, with an equivalent standard PP validation. We could not validate using SD, as there were no vaccinated XBB cases. Against BA.2.86, VE was lower than against XBB, and the VCM and VPM results were within ±7% absolute of each other, with lowest validation results from SD but equivalent results from the PP. Conclusions Both proposed approaches produced similar VE estimates to those from well-known methods. The VPM could also provide VE estimates when the validation techniques were limited by low sample size.