The 1970s and 1980s were times when communication behavior was a primary interest of many communication scholars. The aim of this book is to reignite some interest in and passion about how human communication behavior should be studied. It presents the best advice, techniques, cautions, and controversies from the 1970s and 1980s and then updates them. Several chapters also introduce statistical methods and procedures to allow readers to analyze behavioral data.
This book is a useful resource for communication scholars and graduate students to guide their study of communication behavior.
11 Linear Models of Categorical Data
| 304 →
Linear Models of Categorical Data
The log-linear models presented in Chapter 10 were shown to be quite useful for addressing several types of questions that are common to the analysis of communication behavior. Despite this utility, the primary constraint in the use of log-linear models is that all models must be fit in the loge scale, even if the substantive research question would be better addressed by analyzing the data in a scale other than loge; e.g., the original probabilities.
Accordingly, the purpose of this chapter is to introduce a statistical approach designed to address questions of differences in response probabilities among and between several subpopulations of individuals. The original paper on this topic was written by Grizzle, Starmer, and Koch (1969), and the procedure has subsequently been referred to simply as the GSK approach. The GSK approach is grounded in the logic and basic calculations of the general linear model (GLM). Unlike the GLM which is designed to analyze a continuous outcome or criterion measure, however, the GSK approach applies the methods of weighted least squares to a set of response probabilities that have been obtained by classifying (into a vector) or cross-classifying (into a contingency table) one or more categorical variables.