Having introduced the idea of dependent samples, where two responses are matched by subject or "cluster" (e.g., two ratings provided for each movie in a sample), and measuring ways these responses can be in agreement (e.g., marginal homogeneity or symmetry) for two-way tables, we now extend this concept to three or more responses clustered together. Thus, we have the categorical version of repeated measures and the additional challenge of deciding an appropriate structure for modeling the association among them.
The GEE approach is an extension of GLMs that provides a semi-parametric approach to repeated measures of categorical response; it can be also used for continuous measurements.
- Objective 12.1
Recognize when categorical responses are clustered within subjects and likely to be correlated.
- Objective 12.2
Interpret the parameters in a marginal model for clustered categorical data, including the covariance structure, and use software to estimate these parameters.
- Objective 12.3
Distinguish between sampling and structural zeros in categorical data and what the effect will be on the degrees of freedom when fitting a log-linear model to such data.