How Important Are Study Designs? A Simulation Assessment of Vaccine Effectiveness Estimation Bias with Time-Varying Vaccine Coverage, and Heterogeneous Testing and Baseline Attack Rates

ObjectiveWe studied how commonly used vaccine effectiveness (VE) study designs (variations of cohorts, and test-negative designs) perform under epidemiological nuances more prominent in the COVID-19 era, specifically time-varying vaccine coverage, and heterogeneous testing behaviour and baseline attack rates with selection on willingness to vaccinate. MethodologyWe simulated data from a multi-parameter conceptual model of the epidemiological environment using 888125 parameter sets. Four configurations of cohorts, and two test-negative designs, were conducted on the simulated data, from which estimation bias is computed. Finally, stratified and fixed effects linear regressions were estimated to quantify the sensitivity of estimation bias to model parameters. FindingsIrrespective of study designs, dynamic vaccine coverage, and heterogeneous testing behaviour and baseline attack rates are important determinants of bias. Study design choices have non-trivial effects on VE estimation bias even if these factors are absent. The importance of these sources of bias differ across study designs. ConclusionA re-benchmarking of methodology, especially for studying COVID-19 VE, and implementation of vaccine-preventable disease surveillance systems that minimise these sources of bias, are warranted. HighlightsO_LIThis paper simulated a theoretical model with frictions in vaccination, testing, baseline disease risks, and heterogeneous vaccine effectiveness to evaluate estimation bias across four cohort and two test-negative designs. C_LIO_LIIn theory, bias depends on behavioural asymmetries (in testing, and baseline risk) between the vax-willing and vax-unwilling, and the speed of vaccination rollout. C_LIO_LIThere is intrinsic estimation bias across all study designs, with the direction and magnitude contingent on specific conditions. C_LIO_LIIn scenarios that may be reflective of past SARS-CoV-2 waves, the degree of bias can be substantial, attributable to variation in assumed testing and baseline risk frictions. C_LIO_LIA regression-based decomposition indicates that study designs have visibly different primary sources of estimation bias, and degree of robustness in general. C_LIO_LIThis study warrants a re-benchmarking of methodology and reporting checklists for VE research, and informs the design of cost-effective surveillance by quantifying part of the bias-implementation cost trade-off. C_LI