Categorical Data Analysis Using The SAS® System, 2nd Edition

10.5. General Conditional Logistic Regression

When you have paired data or responses that comprise profiles considered to come from a trinomial distribution, you can easily write down the possible response profiles and identify the informative and noninformative observations. In the case of the paired response, you can construct a likelihood that is similar to the unconditional likelihood for a dichotomous response; in the case of a trinomial outcome, you can create a model and likelihood that is based on generalized logits and is an extension of the loglinear model, which is discussed in Chapter 16, “Loglinear Models.”

However, for more complicated data situations, equivalent unconditional strategies are not conveniently available. Consider ...