A Comparison of Log-Linear and Regression Models for Systems of Dichotomous Variables

The relative abilities of dummy variable regression and log-linear models to locate significant relationships in systems of dichotomous variables are compared. On logical grounds log-linear models are superior to regression since the data more readily meet the assumptions of the former. Two illustrative examples suggest that the methods converge in their findings when the range in proportions of the dependent dichotomy is between .25 and. 75, but may differ on which effects are significant when proportions are more extreme. Substantive differences under the two methods are likely to be small, however.