Inference and control of the nosocomial transmission of methicillin-resistant Staphylococcus aureus

Pei, Sen; Morone, Flaviano; Liljeros, Fredrik; Makse, Hernán; Shaman, Jeffrey L

We validated the IF-EAKF inference framework for different synthetic scenarios. Figure 2 presents the synthetic situation where nosocomial transmission accounts for the majority of incidence (see Figure 2D -E). To evaluate the goodness of fit for incidence number in Figure 2B , we performed the following statistical analysis. As the agent-based model is a highly stochastic system, the observed incidence in Figure 2B is only one possible outcome of the actual dynamics, whereas in our analysis, the stochasticity of incidence number needs to be considered. To this end, we compared several summary statistics quantifying the goodness of fit in Figure 2B with their distributions calculated from synthetic outbreaks (surrogate data) generated from the inferred dynamics. We first considered the log likelihood of observations. In particular, we generated 1000 synthetic outbreaks using the inferred parameters, and approximated the distribution of incidence number at each week. Then we calculated the log likelihood for the observed incidence in each synthetic outbreak, and estimated its distribution using these 1000 log likelihood values computed from the surrogate data. In Figure 2 -figure supplement 1A, we compared the log likelihood computed from Figure 2B (vertical red line) with this distribution (blue bars) and calculated the 2-sided p-value. The p-value is well above zero, indicating that, in terms of log likelihood, our inferred dynamics span and thus agree well with the observed incidence. In other words, the observed incidence in Figure 2B is a typical outcome from our inferred dynamics. The same analysis was also applied to root-mean-square error (RMSE), coefficient of determination (R 2 ) and Pearson correlation coefficient (Figure 2figure supplement 1B-D). The RMSE, R 2 and Pearson correlation coefficient were calculated using the incidence time series in each synthetic outbreak and the mean incidence time series averaged over 1000 simulations.