Modelling and short-term forecasting of seasonal mortality

TL;DR

This study compares different Serfling-Poisson models with P-splines for short-term seasonal mortality forecasting, finding that the non-parametric trend with parametric seasonality model provides the most accurate predictions across European countries.

Contribution

It introduces and evaluates three variants of Serfling-Poisson models with P-splines, identifying the most effective specification for short-term mortality forecasting.

Findings

SP-STFS model yields the most accurate historical forecasts.

Non-parametric trend with parametric seasonality improves prediction accuracy.

Application to COVID-19 demonstrates utility in public health planning.

Abstract

Excess mortality, i.e. the difference between expected and observed mortality, is used to quantify the death toll of mortality shocks, such as infectious disease-related epidemics and pandemics. However, predictions of expected mortality are sensitive to model assumptions. Among three specifications of a Serfling-Poisson regression for seasonal mortality, we analyse which one yields the most accurate predictions. We compare the Serfling-Poisson models with: 1) parametric effect for the trend and seasonality (SP), 2) non-parametric effect for the trend and seasonality (SP-STSS), also known as modulation model, and 3) non-parametric effect for the trend and parametric effect for the seasonality (SP-STFS). Forecasting is achieved with P-splines smoothing. The SP-STFS model resulted in more accurate historical forecasts of monthly rates from national statistical offices in 25 European countries. An application to the COVID-19 pandemic years illustrates how excess mortality can be used to evaluate the vulnerability of populations and aid public health planning.