Abstract:
|
Traditional causal inference for vaccine efficacy against post-infection outcomes like severe illness or death employs principal stratification (PS). PS requires that infection status in control and treatment arms be identified without error, which does not hold in practice. For example, in COVID-19 vaccine trials polymerase chain reaction tests were used with sensitivities between 0.5 to 0.8 and specificity nearly 1. We design a principal stratification estimator that adjusts for imperfect diagnostic tests with unknown sensitivity and specificity. We show that for fixed values of specificity, the sensitivity and principal strata proportions are identifiable and we derive asymptotic bounds for the treatment effect. Alternatively, we develop a finite-sample Bayesian estimator for the treatment effect by placing a prior on the unknown specificity. We explore the model’s finite sample performance using a simulation study, and we apply our method to an influenza vaccine trial.
|