The lay abstract featured today (for Estimation and Hypothesis Testing of Strain-Specific Vaccine Efficacy With Missing Strain Types With Application to a COVID-19 Vaccine Trial by Fei Heng, Yanqing Sun, Li Li and Peter B. Gilbert) is from Statistics in Medicine with the full article now available to read here.

How to cite

Heng, F., Sun, Y., Li, L. and B. Gilbert, P. (2025), Estimation and Hypothesis Testing of Strain-Specific Vaccine Efficacy With Missing Strain Types With Application to a COVID-19 Vaccine Trial. Statistics in Medicine, 44: e10345. https://doi.org/10.1002/sim.10345

Lay Abstract

Many pathogens that cause high-burden infectious diseases are Plastic Man escape artists, rapidly evolving resistance to immunity from infections and vaccinations while sustaining their capacity to efficiently replicate and keep wreaking havoc on public health. The SARS-CoV-2 virus is a great example. One of the first vaccines to be authorized – Moderna’s mRNA-1273 vaccine in December 2020 based on the COVE phase 3 trial – initially provided almost perfect protection to prevent the disease COVID-19, yet with many successive waves of viral evolution, 4 years later it would provide weak or even zero protection against current SARS-CoV-2 variants that have acquired more than 30 amino acid mutations in their Spike protein away from the original strain of the virus on which the vaccine is based. Hence today COVID-19 boosters are updated with optimally-selected strains that best match contemporary circulating strains. But how do we know that the mRNA-1273 vaccine’s protection weakened against new viral genotypes?  A gold-standard approach learns from a randomized, placebo-controlled, vaccine efficacy trial conducted in a geographic region with a diversity of circulating viral genotypes, enabling estimation of vaccine efficacy (VE) against each genotype and hypothesis testing for whether VE differs across the genotypes, an investigation called sieve analysis. Heng et al.’s sieve analysis methods measure genotype-specific VE by one minus the genotype-specific hazard ratio (vaccine / placebo) of COVID-19, where the randomization assures this quantity approximately reflects a causal effect of how well the vaccine worked to prevent genotype-specific COVID-19. Even though the gold-standard COVE trial was conducted at the beginning of the pandemic when the virus had only begun to evolve, several variants of SARS-CoV-2 did pop into existence during the trial, including Epsilon, Gamma, and Zeta, which acquired up to 10 amino acid sequence Spike mutations away from the vaccine strain, most with less than 4 mutations. Heng et al. tackled technical challenges presented by COVE including that several genotypes circulated instead of just two for which most available statistical methods had been developed; 53% of vaccine recipients and 25% of placebo recipients that acquired the COVID-19 primary endpoint had missing SARS-CoV-2 genotype; missing genotypes can be partially predicted from calendar time and geography; and the incidence of genotype-specific COVID-19 can rapidly change in calendar time and geography in unpredictable ways. The second challenge was especially motivating given that the probability of missing the viral genotype strongly depended on the amount of SARS-CoV-2 virus in nasal swabs, where swabs with lower virus levels had lower chance of sequencing success, to the point that swabs with very low values had zero chance of success. Heng et al.’s sieve analysis methods addressed these challenges through a genotype-specific proportional hazards model allowing K separate unspecified placebo-arm hazard functions for each of the J circulating genotypes to flexibly model COVID-19 incidence in an unpredictable pandemic, employing augmented inverse probability weighting to address the missing genotypes with a twist to accommodate the special problem of zero chance of sequencing success for some COVID-19 endpoints that generally thwarts inverse probability weighting based methods. The methods were applied to COVE for several ways of defining SARS-CoV-2 genotypes, with results affirming that indeed mRNA-1273’s efficacy had already started to wear off against evolving variants over the approximately 6-month follow-up period, where VE against a genotype defined by more than 7 amino acid mutations away from the vaccine strain (estimated at 84%) was significantly lower than VE against genotypes closer to the vaccine strain (estimated at 94%): this is a large difference representing an approximately 62% reduced degree of protection [1 – (1-0.94)/(1-0.84) = 0.625]. The Heng et al. methods are most appropriately applied to efficacy trial follow-up periods with a limited amount of waning efficacy, as genotype- and time-variations in VE can be conflated – under substantial waning additional flexibility in the methods is needed to assess how VE depends jointly on genotype and time since vaccination. Considering the vast number of ways to define genotypes and phenotypes of genetically diverse pathogens, an exciting frontier of sieve analysis deepens its pursuit to determine exactly which viral features are most critical for impacting VE, with deep learning computational algorithms including sequence-based predictors of pathogen protein structures coming to the fore.