Bayesian nowcasting with Laplacian-P-splines

Abstract

AbstractDuring an epidemic, the daily number of reported (infected/death) cases is often lower than the actual number of cases due to underreporting. Nowcasting aims to estimate the cases that have not yet been reported and combine it with the already reported cases to obtain an estimate of the daily cases. In this paper, we present a fast and flexible Bayesian approach to nowcasting combining P-splines and Laplace approximations. The main benefit of Laplacian-P-splines (LPS) is the flexibility and faster computation time compared to Markov chain Monte Carlo (MCMC) algorithms that are often used for Bayesian inference. In addition, it is natural to quantify the prediction uncertainty with LPS in the Bayesian framework, and hence prediction intervals are easily obtained. Model performance is assessed through simulations, and the method is applied to the Belgian COVID-19 mortality cases for the year 2021. Simulation results show that our model has good predictive performance except when the nowcast date is near the peak date, where it has lower prediction interval coverage.