ABSTRACT
Log-linear methods are applicable when most variables are nominal (although a few covariates are allowed). These methods are more general and flexible than the ones presented in Chapters 4 and 5 for analyzing contingency tables. For example, although the Mantel-Haenszel (MH) test can be used with 2 × 2 × k tables to control for the confounding effect of a third variable, the test cannot be used with larger contingency tables (e.g., 3 × 3 × k). In contrast, log-linear analysis can be used for any size contingency table and, unlike the MH test, does not necessarily assume the absence of interaction effects. Logistic regression can also handle situations where all the variables are nominal, as can Poisson regression, and log-linear analysis can be considered a special case of these techniques when one variable is being predicted by others. However, it is preferable to use log-linear analysis when one does not have clear hypotheses about causal directionality and/or when all the variables in the model could be considered response variables.
