Comparison of uncertainty in chikungunya case count estimates for the baseline nowcasting model and nowcasting model using Google searches.

We compare the mean prediction interval widths (MPIs) for nowcasts from the baseline model and model using Google searches. The heuristic model is omitted as this approach does not allow a prediction interval to be calculated. The first three columns show results across the sample as a whole. The first column is the MPI reported in number of cases; the second column is the MPI relative to the baseline nowcasting model; and the third column is the percentage of actual weekly case counts within the prediction interval. The nowcasting model using Google searches is more precise than the baseline nowcasting model, reducing MPI by 7%. It is also slightly more reliable, capturing 93% of actual weekly case counts relative to 91% for the baseline model. In Table C in S1 Appendix, we show that the slight overconfidence of the prediction intervals is largely driven by the first epidemic, where the model had little data to train on. The final three columns show similar results when considering epidemic periods only. While the intervals are much wider for both models, the model using Google searches reduces MPI by 8%. The intervals produced by the model using Google searches are also even more reliable relative to the baseline during epidemics, capturing 93% of weekly case counts relative to 88%.