Estimation and Hypothesis Testing of Strain-Specific Vaccine Efficacy with Missing Strain Types, with Applications to a COVID-19 Vaccine Trial

Statistical methods are developed for analysis of clinical and virus genetics data from phase 3 randomized, placebo-controlled trials of vaccines against novel coronavirus COVID-19. Vaccine efficacy (VE) of a vaccine to prevent COVID-19 caused by one of finitely many genetic strains of SARS-CoV-2 may vary by strain. The problem of assessing differential VE by viral genetics can be formulated under a competing risks model where the endpoint is virologically confirmed COVID-19 and the cause-of-failure is the infecting SARS-CoV-2 genotype. Strain-specific VE is defined as one minus the cause-specific hazard ratio (vaccine/placebo). For the COVID-19 VE trials, the time to COVID-19 is right-censored, and a substantial percentage of failure cases are missing the infecting virus genotype. We develop estimation and hypothesis testing procedures for strain-specific VE when the failure time is subject to right censoring and the cause-of-failure is subject to missingness, focusing on $J \ge 2$ discrete categorical unordered or ordered virus genotypes. The stratified Cox proportional hazards model is used to relate the cause-specific outcomes to explanatory variables. The inverse probability weighted complete-case (IPW) estimator and the augmented inverse probability weighted complete-case (AIPW) estimator are investigated. Hypothesis tests are developed to assess whether the vaccine provides at least a specified level of efficacy against some viral genotypes and whether VE varies across genotypes, adjusting for covariates. The finite-sample properties of the proposed tests are studied through simulations and are shown to have good performances. In preparation for the real data analyses, the developed methods are applied to a pseudo dataset mimicking the Moderna COVE trial.