�����������������������������������������������������������������
als , heterogeneity in both individuals and the environment in which they live , change over time , and a desire to address counterfactuals of both theoretical and policy importance that are not easily amenable to experimental manipulation . An ABM is an ��� ������ instantiation of a theory of a complex phenomenon with a model based on that theory that is sufficiently well-specified that it can be translated to computer code and its behavior observed under various conditions and at various times . The model can be seen as an explicit theory of the phenomena under inquiry , and used to test various theoretical propositions ( Epstein , 1996 ; Epstein , 2006 ; Miller & Page , 2007 ). One advantage of an ABM approach is the explicit demands it makes of the modeler to specify anticipated relations explicitly ; absent that the model will not run . Potentially , this requires specific documentation of every assumption , every parameter and rule ( sometimes in the 100s ), and means a level of transparency seldom available otherwise , although documenting and making available the elaborate code for inspection by others brings with it its own logistical problems ( Bankes , 2002 ).
The ABM is an abstraction that permits us to simulate the impact of changes at any level of the model and see the evolution of the system in response to those changes . Because it explicitly allows an examination of change over time , it allows us to document different time signatures associated with these changes . Like any abstraction , the simulation is only valuable to the extent that it provides useful insights , provides explanation for observed phenomena , and stimulates new inquiry . Epstein provides a list of 16 benefits of such simulation modeling ( Epstein , 2008 ). Of course , many analytic techniques engage in an analog of simulation , for example , asking what would happen if the distribution of some covariate was the same in different groups . What is different about agent-based computational techniques is that they allow us to ask “ what-if ” questions in the context of far more complex circumstances where heterogeneity , nonlinearity , feedback , and connectivity are the rule , something not easily , or at all , accomplished with conventional analytic techniques ( Berry , Kiel , & Elliott , 2002 ). Furthermore , the goal of ABMs is generally not to accept or reject specific hypotheses or indicate our confidence in value of certain parameters , although they are not out of the reach of such simulations . There is another major difference . ABMs cannot practically be reduced to a set of equations ( Epstein , 2006 ). They are dynamic simulations of an abstract representation of an artificial society . As such , they constitute a “ generativist ’ s experiment ” allowing ��������� demonstrations of whether or not that abstract specification is sufficient to generate macro-phenomena of interest ( Epstein , 2006 ; Miller & Page , 2007 ).
As we have argued elsewhere ( Galea et al ., 2010 ), complex systems dynamic models , and ABMs in particular , may provide a new pathway to asking questions that are occasioned by our growing unease with simple models of disease causation . This seems particularly true for work on SDOH where as shown in Figures 2 – 3 , interaction , feedback , change , connection , and multiple-levels characterize our conceptual models . Above we have made the argument that the complexity portrayed in these figures means that simple interventions are likely to fail or give misleading answers . For the same reasons , the answers that come from applying our standard analytic tools to these systems of SDOH and their downstream pathways to health outcomes may also mislead us .
94