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

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

We observed greater volatility in the nowcasts when the initial number of cases reported increases suddenly from low values. Two weeks in the dengue time series highlight this: August 3, 1998 and August 16, 2010. In those weeks, the number of cases initially increased by 16 and 17, respectively, from the previous week after 10 week with an average absolute change of 2.6 and 1.8 cases, respectively. Because this increase is an outlier in the distribution of reporting delays, in particular for delay d=0, the model substantially overestimated the true number of cases before correcting the following week. We observed that shorter moving windows either exacerbated this issue (e.g. in 2010) or produced a similar overestimate (e.g. in 1998) (Fig. S6 ), which appears to be a consequence of the volatility in estimating the variance of the random walk process, despite more accurate estimation of the reporting delay probability (Fig. S7) . While the smooth, autocorrelated relationship fit in the NobBS model helps reduce the effect of week-to-week variability in early reporting, it remains a challenge. Users should keep in mind these trade-offs when seeking to apply NobBS to their data.