Modelling weekly mortality incorporating annual trends

Joanne Ellison, University of Southampton
Jason Hilton, University of Southampton
Erengul Dodd, University of Southampton
Jonathan J. Forster, University of Warwick
Peter W. F. Smith, University of Southampton
Jakub Bijak, University of Southampton

The global COVID-19 pandemic has focussed attention on methods for calculating excess deaths. Because of the difficulty of diagnosing novel viruses in the early stages of a pandemic, and the difficulty in determining cause of death in the presence of comorbidity, the effect of a pandemic on mortality is not clear-cut. Statistics based entirely on cause-of-death data have underestimated the true toll of the pandemic. This paper contributes to the growing literature on excess deaths by incorporating information about long-term trends in mortality from an annual mortality forecasting model. This provides a more realistic baseline for excess mortality estimation, as well as providing a method for decomposing weekly deaths by single year of age according to established mortality age patterns. The focus of our methodology is the modelling of a weekly period term, which captures weekly variation in mortality relative to the long-term trend captured in the annual model. Within a coherent Bayesian framework that fully integrates all sources of uncertainty, we investigate the use of dynamic harmonic regression models for capturing weekly mortality variation. Obtaining estimates of the weekly deaths expected in the absence of the pandemic, we estimate age-specific and overall excess deaths and compare with alternative estimates.

Keywords: COVID-19, Mortality, Population projections, forecasts, and estimations, Bayesian methods / estimation

See paper.

  Presented in Session 118. Estimating and Modelling Mortality