��������������������������
Grimm , 2012 ). In the case of this model , we choose school attendance by age , income distribution , smoking rates by age , and BMI by age as calibration targets . After assuring that the model closely conforms to observed patterns for these targets , we allow the model to run over several generations of agents and examine the impact of each policy on the average difference in BMI between Black and White agents over generations , and how that difference varies by policy strength over a hundred model runs . As can be seen in Figure 5 , more intensive policy changes reduce the BMI disparities , with interventions on physical activity infrastructure having the greatest effect , access to good food stores the smallest effect , with improvements in school quality in between . There is some evidence of nonlinearity of the effects , particularly for access to good food stores and school quality interventions .
However , the figure shows the average effects over several generations and holds two policy levers constant while the third is being changed . While a useful summary , it masks the reality that changes in one policy may , over various time scales , have an impact on risks and resources indexed by the other policy dimensions . For example , better education may through a variety of pathways lead to higher incomes that , in turn , lead to changes in residential location that may allow for better food and physical activity opportunities . For these reasons and because making policy decisions in complex systems often requires examination of a dense computational space representing the outcome of many simulations ( Lempert , 2002 ), feedback between different policy options , possible combinations of various policy alternatives and multiple time perspectives , we then consider various combinations of policies and policy strengths . Figure 6 shows the results when we consider all 125 combinations of the five strengths of the policy dimensions applied simultaneously . After a short run-in period , shown by the flat lines on the left , the three policies are introduced and BMI is assessed multiple times . As can be seen , there is considerable variation in the effects of policy combinations .
In general , there are much larger effects of the policies on BMI for Blacks compared to Whites , and while most of these changes are reductions in BMI , there are some combinations that result in increases in BMI as well . There is also considerable variation in BMI trajectories over time , and for some combinations of policies there is an elimination of BMI disparities . In those cases , the elimination of BMI disparities takes time , and occurs over approximately 20 years for some combinations of policies . Our purpose in presenting these results is not to present a detailed analysis of the results , but rather to show the potential value of considering such a wide policy space in a simulation .
The results of this simulation exercise are encouraging , but should be cautiously interpreted . First , a number of the simplifications in the model make it difficult to draw firm conclusions . For example , the agents have no gender , yet we know the dynamics and patterning of BMI is different for males and females . The model is an abstraction , and it is difficult to know without further elaboration of the model if the results would be different , for example , for males and females . While care has been taken to use the best available information , it should be acknowledged that there are over a hundred parameter estimates in this simulation model , the sources for the parameters vary in quality , and the complexity of the model makes it difficult to systematically conduct the full spectrum of sensitivity analyses , an issue to which we will return .
97