1

Paper

An Embarrassment of Riches: Parameter Choice in Agent-Based Models Ragan, Robi Uploaded 07-19-2010 Keywords agent-based ABM generative complexity computational emergence Abstract In this paper, I develop a set of criteria that researchers and reviewers may use to determine whether the omission of a parameter in an agent-based computational model calls into question the model’s results. The use of agent-based modeling and other computational modeling techniques is growing within political science. The ability to model without the constraints and associated unrealistic assumptions of traditional analytical techniques has attracted a new generation of scholars. However, now that we are freed from having to make assumptions in order to find closed form solutions or game theoretic equilibria, a new problem arises. Which simplifying assumptions and parameter omissions are still useful, and which should be adjusted? The lack of clear guidelines for this type of question is a hindrance to the acceptance of agent-based models. Every computational modeler makes a fundamental choice with every model: which parameters to include and which to exclude. Most applied methodologists make similar choices: which control variables to include in an estimation in order to isolate the partial effect of an independent variable of interest on the dependent variable. In most cases, there are clear criteria to help an empirical researcher decide. However, when creating a theoretical computational model, the decision is rarely so cut and dry. Often, adding more parameters will do little to change the results generated by the model. In other cases, new parameters radically change the results and may render the model useless. In this paper, I propose a set of criteria for determining which parameters should be included or excluded. I use the analogy of an OLS regression to motivate the discussion. Omitting a parameter that “matters” is likened to omitted variable bias in OLS, and including parameters deemed “unnecessary” is likened to a loss of efficiency due to multi-colinearity. After developing and discussing the set of criteria, I illustrate the point using a set of canonical computational models, including Schelling’s “sorting and mixing” model of housing segregation (Schelling 1978) and Kollman, Miller and Page’s model of Tiebout sorting (1997).