ABSTRACT
As in many other settings, peak excess mortality preceded the officially reported `first wave' peak of the COVID-19 epidemic in Manaus, Brazil, reflecting delayed case recognition and limited initial access to diagnostic testing. To avoid early information bias, we used detailed age and gender stratified death certificate and hospitalisation data to evaluate the epidemic's trajectory and infer the cause of its decline using a stochastic model. Our results are consistent with heterogenous transmission reducing over time due to the development of herd immunity. Relative to a baseline model that assumed homogenous mixing across Manaus, a model that permitted a small, self-isolated population fraction raised the estimated herd-immunity threshold from 28% to 30% and reduced the final attack rate from 86% to 65%. In the latter scenario, a substantial proportion of vulnerable, older individuals remained susceptible to infection. Given uncertainties regarding the distancing behaviours of population subgroups with different social and economic characteristics, and the duration of sterilising or transmission-modifying immunity in exposed individuals, we conclude that the potential for epidemic outbreaks remains, but that future waves of infection are likely to be much less pronounced than that already experienced.