Nowcasting by Bayesian Smoothing: A flexible, generalizable model for real-time epidemic tracking

Sarah F. McGough; Michael A. Johansson; Marc Lipsitch; Nicolas A. Menzies

While NobBS mitigates the effects of a time-varying delay distribution on case estimation, i.e. that the history of cases is leveraged to anchor case estimates to recent values, it does not explicitly model temporal changes in that delay; in other words, the estimated probability of a case occurring with delay = d is assumed to apply to all reporting weeks in the moving window. Shorter moving windows can improve estimation of the delay in the presence of changes, but explicitly estimating changes in that distribution may be explored for additional robustness in the presence of systemic changes in reporting. For example, authors in (21) propose a smooth estimate of the time-varying reporting delay distribution using p-spline smoothing. Specifying a time-specific change has also been proposed (9), but empirical identification of a change point in real-time may be challenging or impossible in the context of nowcasting. The challenge that remains in all described approaches is the ability of the model to pick up on changes in the delay distribution that occur quickly, in other words that may otherwise be smoothed out by splines and long moving windows.