������������������
Model Results
Both model scenarios ( that is , where
consumers do not have access to re-inspection priority scores , and where they do and display risk averse behavior for 15 time steps following healing ) were repeated 100 times at each setting of 60 %, 70 %, 80 %, and 90 % compliant restaurants . The Kruskal-Wallis test was conducted to check the statistical significance of each increase in the percentage of compliant restaurants within each scenario ( see Table 2 ), and this was followed by post-hoc analysis using pair-wise Mann-Whitney-Wilcoxon tests , using the Holm correction to account for multiple comparisons . Unless otherwise stated , the pairwise analysis results are statistically significant ( � < 0.05 ). For comparisons between scenarios , Mann-Whitney-Wilcoxon tests were used to check for statistical significance of consumer risk aversion at each setting of 60 %, 70 %, 80 %, and 90 % compliant restaurants ( see Table 4 ).
In the first scenario , consumers do not have access to inspection scores and do not behave in a risk averse manner following an illness . Interestingly , the mean number of sick consumers is slightly lower than in the second scenario , where consumers can see inspection scores and behave in a risk averse way by only going to restaurants with a low re-inspection priority for 15 time steps following an illness . However , note that this difference between scenarios is only significant ( � < 0.05 ) when there are 70 % and 90 % compliant restaurants in the model ( see Table 4 ). Also , even though the mean number of sick consumers is slightly higher in the second scenario , the overall variation in the number of sick consumers is reduced substantially .
Figure 2 shows the results for the number of consumers in the at-risk group who experience sickness : in the first scenario , the mean number of sick , at-risk consumers is slightly lower , but there is less variation in the second scenario . However , this difference between scenarios is only statistically significant when there are 90 % compliant restaurants ( � < 0.001 ).
Generally , the results from the first scenario are more skewed and often leptokurtic ( see Table 3 ), whereas the results from the second scenario were less skewed and closer to a mesokurtic , or normal , distribution . A leptokurtic distribution is more peaked than normal and has fat tails , meaning that there are higher densities of values at the extremes . This phenomenon was most apparent in the numbers of naïve consumers ( see Figure 3 ), that is , the number of consumers that never became ill throughout the course of the model run .
Figure 4 shows the numbers of inspected restaurants in each scenario . The post-hoc analysis showed that , for the first scenario , that the differences between 60 % and 70 % compliant restaurants , between 60 % and 80 % compliant restaurants , between 70 % and 80 % compliant restaurants , and between 80 % and 90 % compliant restaurants were not statistically significant ( � > 0.05 ). For the second scenario , the posthoc analysis showed that the differences between 60 % and 70 % compliant restaurants , between 70 % and 80 % compliant restaurants , and between 80 % and 90 % compliant restaurants were not statistically significant ( � > 0.05 ). When the two scenarios were compared ( see Table 4 ), none of the differences were statistically significant ( � > 0.05 ). As well , for this indicator , the variation was not greatly reduced in the second scenario . The number of inspected restaurants appears to decline as compliance increases ; this is in part due to the model ’ s construction , since compliant restaurants are less likely to become contaminated in the first place .
148