Power and sample size considerations for test-negative design with bias correction: a case study on the world first...

Test-negative design (TND) studies are increasingly common in evaluating vaccine effectiveness (VE) for various infectious diseases. TND studies are susceptible to bias due to disease outcome misclassification caused by imperfect test sensitivity and specificity. Several bias correction methods have been proposed. However, sample size or power considerations for TND studies incorporating bias correction for such misclassification have not yet been investigated.

We used Monte Carlo simulations to assess how bias correction affects the statistical power and sample size for VE estimation in TND studies. Simulations were conducted under varying levels of diagnostic test sensitivities (60%, 80%, and 95%). Bias correction was implemented using the multiple over-imputation method, which accounts for test misclassification through a parametric bootstrapping approach. Using a malaria vaccine as an example, we defined six vaccination status categories based on the time since receipt of the third or fourth vaccine dose. In the simulated target population, vaccination coverage was assumed to be low (< 10%) except for the group vaccinated more than 12 months after dose 4. We assumed relatively low VE (< 50%) against clinical malaria cases and a 30% malaria positivity rate among unvaccinated individuals presenting with malaria-related symptoms. Statistical power to detect VE was estimated for each vaccination status, both with and without bias correction.

Estimated VEs based on observed data were consistently underestimated across all vaccination status groups due to diagnostic misclassification. In contrast, bias-corrected estimates were approximately unbiased but displayed wider confidence intervals, with their precision decreasing at lower test sensitivities. Statistical power to detect VE declined substantially when diagnostic test sensitivity was low. For instance, at 80% sensitivity, only three vaccination status groups reached 80% power with a sample size of 10,000, whereas the same power was achieved with just 6,000 individuals under a perfect test.

Bias due to imperfect diagnostic testing can substantially reduce the power of TND studies. Power calculations should account for outcome misclassification and potential correction methods. Failure to do so may lead to underpowered studies and misleading VE estimates.