Paul R. Yarnold & Fred B. Bryant
Optimal Data Analysis, LLC & Loyola University Chicago
Attributes measured on a categorical response scale are common in the literature. Categorical scales for attributes such as, for example, political affiliation, ethnic origin, marital status, state of residence, or diagnosis may consist of many qualitative response categories. Such disorganized variables rarely appear in multivariable models: some effects are missed in analysis due to inadequate statistical power for the many categories, and some findings are dismissed due to inability of the investigator to recognize the dimension(s) underlying segmented categories. This research note recommends that such multicategorical attributes are replaced by a new set of attributes created via content analysis. In this approach observations are scored on new dimensions all theoretically motivated to predict the class variable. The methodology is illustrated using a hypothetical example in the field of investment realty.