Containing Emerging Epidemics: a Quantitative Comparison of Quarantine and Symptom Monitoring

Corey M Peak; Lauren M Childs; Yonatan H Grad; Caroline O Buckee

As compared to characteristics related to the natural history of symptoms and illness, key aspects of the natural history of infectiousness tend to be harder to observe and measure (28). Therefore, we use a Sequential Monte Carlo particle filtering algorithm (29, 30) to create a joint probability space of the time offset between the latent period and incubation period ( !""#$% = !"# − !"# ), time of peak infectiousness ( ! ), and duration of infectiousness ( !"# ). From an uninformative prior distribution of each parameter bounded by published observations, we simulate five infection generations of 500 initial individuals and record the simulated serial interval (i.e., the time between symptom onset in infector-infectee pairs). Parameter sets are resampled with importance weights determined by the degree to which the distribution of simulated serial intervals match published serial interval distributions, using the Kolmogorov-Smirnov test of the difference between cumulative distribution functions ( Table 2 ) (31, 32). After perturbation, the process is repeated until convergence, which we define to be when the median Kolmogorov-Smirnov statistic was within 10% of the previous two iterations. This cutoff criterion balances the objectives of finding a stationary posterior set of particles while preserving some of the heterogeneity in input parameters.