Diagnosing Multicollinearity in Exponential Random Graph Models

• In Event: 535 - Regular Session. Social Networks: New Methods and Measures

Abstract

Exponential random graph models (ERGM) have been widely applied in the social sciences since the development of Bayesian techniques that account for dependencies in network data. However, diagnostics for ERGM have lagged behind their application. Particularly, collinearity-type problems can still emerge without detection when fitting ERGM, skewing coefficients and biasing standard errors. This leads to a unique paradox in statistical models of social networks: as more endogenous effects are modeled, the likelihood of encountering poor model estimates also increases. This paper provides a method to detect multicollinearity when using ERGM. First, it outlines the problem and provides a method to estimate shared variance between ERGM parameters. Second, it provides formulas to estimate variance inflation factors for ERGM under certain specifications. Third, it shares the results from a series of simulation experiments examining when collinearity between endogenous terms bias ERGM estimates. The method is robust to a variety of network specifications, including density, size, topography, and dyadic dependencies.

Author

• Scott W. Duxbury