Adjusting for Overdispersion in an Analysis of Comparative Social Mobility

Abstract

The authors discuss the problem of overdispersion in large-scale data sets and its potential impact on standard model selection strategies. Overdispersion is considered to be present when the data display more variability than is predicted by the assumed sampling model. In a recent cross-national analysis of social mobility, data were combined from nine national studies that employed somewhat different sampling schemes and related data collection procedures. Ignoring these features of the data is quite likely to introduce excess dispersion. Typically, the presence of overdispersion can be due to design effects, hidden clusters, or the absence of relevant explanatory variables in the model. When there is overdispersion, model selection based on the standard likelihood ratio test, the Akaike information criterion, or the Bayesian information criterion generally would be expected to perform poorly. A very simple adjustment to these model selection criteria, to account for overdispersion, is proposed.