Optimising Renewal Models for Real-Time Epidemic Prediction and Estimation

KV Parag; CA Donnelly

Removing these biases might be expected to lead to smoother APE-justified 302 reproduction numbers. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/835181 doi: bioRxiv preprint Figure 7 : Selection for empirical datasets. Left graphs of each panel com-pareR Ï„ (s) estimates (cyan with 95% confidence intervals) at APE window length k * to those when k = 7 (weekly sliding windows). Right graphs give corresponding one-step-ahead predicted incidence curves (cyan with 95% prediction intervals). Dashed lines are the R = 1 threshold (left) and dots are the true incidence counts (right). The panels examine epi-data from outbreaks of (a) pandemic influenza (1918) and (b) SARS (2003), which are available in EpiEstim [8] . Panels (c) and (d) repeat (a) and (b) respectively, but smooth outliers in the incidence data using 5-day moving averages as in [17] . Interestingly, while k * leads to less believable reproduction number fluctuations, these are justified by the available data in both smoothed and un-smoothed cases. This hints that a Poisson renewal model may not be the most appropriate for these data. The longer weekly windows, which are heuristic recommendations of previous analyses [8] , miss important characteristics of the epi-curves.