Abstract
Vaccines are a crucial tool for controlling infectious diseases, yet rarely offer perfect protection. 'Vaccine efficacy' describes a population-level effect measured in clinical trials, but mathematical models used to evaluate the impact of vaccination campaigns require specifying how vaccines fail at the individual level, which is often impossible to measure. Does 90% efficacy imply perfect protection in 90% of people and no protection in 10% ('all-or-nothing') or that the per-exposure risk is reduced by 90% in all vaccinated individuals ('leaky') or somewhere in between? Here, we systematically investigate the role of vaccine failure mode in controlling ongoing epidemics. We find that the difference in population-level impact between all-or-nothing and leaky vaccines can be substantial when R(0) is higher, vaccines efficacy is intermediate, and vaccines slow but cannot curtail an outbreak. Comparing COVID-19 pandemic phases, we show times when model predictions would have been most sensitive to assumptions about vaccine failure mode. When determining the optimal risk group to prioritize for limited vaccines, we find that modelling a leaky vaccine as all-or-nothing (or vice versa) can change the recommended target group. Overall, we conclude that models of vaccination campaigns should include uncertainty about vaccine failure mode in their design and interpretation.