JMIR Publications

ABSTRACT

Background:

Applying nowcasting methods to partially accrued reportable disease data can help policymakers interpret recent epidemic trends and quickly identify and remediate health inequities. During the 2022 mpox outbreak in New York City (NYC), we applied Nowcasting by Bayesian Smoothing (NobBS) to estimate recent cases, citywide and stratified by race or ethnicity. However, in real time, it was unclear if estimates were accurate.

Objective:

We evaluated the accuracy of estimated mpox case counts across a range of NobBS implementation options.

Methods:

We evaluated NobBS performance for NYC residents with confirmed or probable mpox diagnosis or illness onset from July 8 through September 30, 2022, as compared with fully accrued cases. We used the mean absolute error (MAE), relative root mean square error (rRMSE), and 95% prediction interval (PI) coverage to compare moving window lengths, stratifying or not by race or ethnicity, diagnosis and onset time elements, and daily and weekly time units.

Results:

During the study period, 3305 NYC residents were diagnosed with mpox (median 4 days from diagnosis to diagnosis report), and 2278 patients had known illness onset (median 10 days from onset to onset report). No single moving window length performed best. As window lengths increased from 14 to 49 days, generally, MAE improved (decreased), while rRMSE worsened (increased). For the diagnosis time element, for the 14-day moving window used in real time, MAE was 9, rRMSE was 0.23, and 95% PI coverage was 96%; ranges for longer moving windows were MAE: 3–9, rRMSE: 0.25–0.30, and 95% PI coverage: 93%–100%. For the onset time element, for the 21-day moving window used in real time, MAE was 12, rRMSE was 1.07, and 95% PI coverage was 84%; ranges for other moving windows were MAE: 7–11, rRMSE: 0.75–1.42, and 95% PI coverage: 75%–99%. For any given moving window length, rRMSE worsened (increased) for stratified compared with unstratified estimates. For stratified daily diagnosis hindcasts, for the 14-day moving window used in real time, rRMSE was 0.32, and 95% PI coverage was 95%; ranges for other moving window lengths were rRMSE: 0.35–0.50 and 95% PI coverage: 96%–100%. Performance generally worsened when using onset compared with diagnosis time elements and weekly compared with daily time units. Hindcasts underestimated diagnoses in early August after the epidemic peaked, then overestimated diagnoses in late August during epidemic waning. Estimates were most accurate during September, when cases were low and stable.

Conclusions:

For nowcasting this outbreak using NobBS, accuracy depended on the moving window length and whether cases were stratified. Health departments need additional nowcasting guidance, particularly to promote health equity by ensuring stratified estimates are accurate and to improve robustness, such as by incorporating multiple methods.