Now-casting the COVID-19 epidemic: The use case of Japan, March 2020

Stephan Gloeckner; Gerard Krause; Michael Hoehle

For the purpose of this research, we define the reporting delay to be the time between the onset of disease symptoms of a case and the time the case report is available at the public health authority that generated the epidemic curve. 2 We used the line-list dataset provided by the nCoV-2019 Data Working Group 11 and WHO situation reports, 12 to allow for a comparison between the two sources. On March 6 th 2020, nCoV-2019 Data Working Group contained 204 cases (excluding all cases from the Diamond Princess and with the last case reported on March 2 nd 2020), of which 131 (64%) contained date of symptom onset and date of case confirmation. For the remaining 73 (36%) cases without information on symptom onset, we assume that the case still had a symptom onset date and it is just missing, because it could not be readily identified. We imputed the respective date using a generalized additive Weibull regression model for both the scale and shape parameter with week of confirmation as a covariate. 13 The now-casting approach consists of a Bayesian approach for modelling delay by which case reports appear in the continuously updated epidemic curve (also known as reporting triangle) and provides full predictive distributions for the daily delay-adjusted number of symptom onsets. 10 We used a sliding window of the past 14 days to estimate the time-varying delay distribution, in order to account for the assumption that the reporting delay decreases as the outbreak progresses, due to aforementioned changes in the health care seeking and reporting procedures. Uncertainty of the delay imputation was transported by repeating the imputation 100 times and averaging the resulting now-casting quantities.