Sensitivity analysis
In the context of a growing epidemic, sensitivity analysis using the method of Latin hypercube sampling and partial rank correlation (Marino et al., 2008) showed that the fraction of travellers detected was moderately sensitive to all parameters considered -- most coefficient estimates fell between 0.1 and 0.3 in absolute value (Figure 4). Sensitivity to R0 was somewhat higher than sensitivity to other parameters, but the difference was not statistically remarkable. R0 and the mean incubation period were negatively associated with the fraction of cases detected. An increase in either of these parameters implies an increase in the probability an infected traveller will be undetectable, either because they have been recently exposed (R0), or have not yet progressed to the symptomatic stage (mean incubation time). The positive association between the fraction of cases detected and the sensitivity of thermal scanners, sensitivity of risk questionnaires, or the fraction of travellers aware of exposure risk is intuitive. Finally, the duration from onset to isolation effectively describes the window of time in which we assume a symptomatic individual could initiate travel. Here, a wider window is associated with increased screening effectiveness, because it will lead to a higher proportion of infected travellers who are symptomatic. Figure 4 shows results from the middle case scenario, in which 25% of cases are subclinical. Considering scenarios where more or fewer cases are subclinical, we see increased influence of the factors based on exposure risk (questionnaire sensitivity and the fraction of cases aware of their exposure) as the proportion of cases with detectable symptoms declines (Figure 4—figures supplement 1).
Figure 4.  Sensitivity analysis showing partial rank correlation coefficient (PRCC) between each parameter and the fraction (per-simulation) of 30 infected travellers detected.
Outcomes were obtained from 1000 simulations, each using a candidate parameter sets drawn using Latin hypercube sampling. Text shows PRCC estimate, and * indicates statistical significance after Bonferroni correction (threshold = 9e-4 for 54 comparisons).
Figure 4—source data 1.  Source data for Figure 4, and Figure 4—figures supplement 1.
Source data for Figure 4—figures supplement 2 can be found as a. csv file in the supplementary code.
Figure 4—figure supplement 1.  PRCC analysis comparing cases where 5%, 25% or 50% of cases are subclinical.
(Middle panel is identical to Figure 4, but repeated for ease of comparison).
Figure 4—figure supplement 2.  PRCC analysis assuming the source epidemic is no longer growing.
By construction, R0 has no impact in a flat epidemic. Small PRCC estimates for R0 arise from stochasticity in simulated outcomes, but are never significantly different from zero. In the context of a stable epidemic, a greater proportion of infected travellers will have progressed to show detectable symptoms, and so screening effectiveness was more sensitive to parameters that impact symptom screening efficacy (thermal scanner sensitivity, and to the time from symptom onset to isolation). Note that by construction, model outcomes are insensitive to parameter R0 in the stable epidemic context. As a result, R0 coefficient estimates are very small (non-zero due to stochasticity in simulation outcomes), and never significant. (Figure 4—figures supplement 2).