COMPUTATIONAL SOCIAL SCIENCE

Friday, February 18 - 3:00 p.m.

TITLE: Exploring the Identification Question for Parameterized Social Simulation Models

Steven Wilcox

CSS PhD Student

George Mason University

Progress in bridging the gap between computational social science and conventional quantitative social science methodology will facilitate the application of simulation to practical issues involving a societal dimension and improve the level of rigor in social science methodology. Specifically, we examine the identification question as applied to simulation models rather than regression-type models. Thus, rather than plugging parameter values into a simulation model to get output measures, we consider the inverse problem in which the objective is to find parameter values that result in the observed data correlations. The identification question is whether such an inverse solution is uniquely available. Using linear regression to estimate metamodels of these correlations, estimates of the inverse problem can be found. Moreover, the resulting computations allow one to analyze the identification issues for a given simulation model in terms of its ability to generate a match to the calibration data.

To address the gap we utilize at an agent-based model of neighborhood crime as an example. In it, seemingly small differences in the simulation logic can make a big difference in the output measures. Merely having many parameters in the simulation model is not sufficient to permit a good match, as can be seen using data visualization techniques. Once the simulation model yields a range of output measures that bracket the data, we can then consider the calibration step. However, a great deal of work needs to be done in terms of demonstrating the computational effectiveness of candidate calibration algorithms and validating their results statistically.