Log-Linear Models in Comparative Research: A Paradigm Lost?

Abstract

Major research questions in the field of social stratification and mobility deal with similarities and differences in the patterns of social mobility in space and time. Answers are typically given by means of regression or log-linear models. Kelley (1990) argues for the superiority of the former in detecting differences in bivariate correlations between two populations, and characterizes log-linear modelling of social mobility tables for contextual comparisons as a paradigm that has failed. In this paper we show that, other things being equal, log-linear models and OLS regression models have about the same statistical power provided that the statistical tests involved are based on the same number of degrees of freedom. This is true for detecting differences in linear as well as in non-linear relationships between two populations. Thus, there are no statistical grounds for choosing between OLS regression analysis and log-linear analysis. Kelley's substantive arguments against log-linear modelling in social stratification research can be matched by relatively new developments on multinomial logistic regression analysis.