Using Paired Agent-Based Simulations To Test Strategies For Limiting The Effects Of Epidemics

Agent-based simulations are widely used nowadays in public health research for comparing different strategies for mitigating epidemics and for planning appropriate responses in the aftermath of crises in large urban areas because they can capture fine scale heterogeneities that may have important non-linear effects on the results. Given the costs of implementing strategies, decision-makers have to be convinced that the proposed treatment/procedure leads to a statistically significant improvement. This paper presents an innovative method for constructing paired agent-based simulations where exactly the same set of random effects is applied to simulations with and without the treatment/procedure. Statistical Analysis of Variance distinguishes the sum of squares between groups (BSS) from the sum of squares within groups (WSS). Our aim was to filter out the within sum of squares (WSS) leaving only the sum of squares between the control group and the treatment group (BSS). We propose to filter out the WSS by constructing paired simulations because as is well known, when paired t-tests can be used, they are much more powerful than ordinary t-tests. Pearsons Chi-squared goodness of fit, the Kolmogorov-Smirnov statistic and the Kullback-Leibler Divergence are then used to test whether the effect is statistically significant. This procedure has been tested on a case-study on the propagation of the Zika epidemic in Rio de Janeiro in 2015. Author summaryAgent-based simulations are emerging as a powerful tool in computational biology because they can capture fine scale heterogeneities that can have important effects on the propagation of epidemics. In silico experiments can be used to test different strategies for mitigating epidemics quickly and inexpensively. Given the inherent variability from one simulation to another, it is difficult to statistically prove their effectiveness. We have developed a powerful method rather like paired t-tests, for testing whether a given treatment is statistically better than the control. We do this by generating paired simulations with exactly the same random variables in the control simulation and the one with a treatment. Using the terminology of analysis of variance, we want to filter out the sum of squares within the group, leaving only the sum of squares between the control and the treatment. This procedure has been applied to a case-study to see whether enclosing and air-conditioning the transport hub in Rio de Janeiro would have slowed down the propagation of Zika.