Models for prediction analysis

This paper discusses models for prediction analysis. It describes prediction analysis as devised by Hildebrand, Laing, and Rosenthal (1977) and criticisms of the method. Paradoxical results are illustrated. The paradoxical characteristics are explained from differences between models underlying statistical testing and hypothesis formulation. The paper proposes nonstandard log‐linear modeling as an alternative to Hildebrand et al.’s (1977) approach. This new approach parameterizes deviations from independence between predictors and criteria. Implications are laid out. The discussion focuses on the type of hypotheses analyzed.

Key words:

• Prediction analysis
• log‐linear models
• nonstandard models
• quasi‐independence models

Notes

Alexander von Eye's work on this project was supported in part by NIA Grant #5T32 AG00110–07, “Training in Aging Research Methodology.”

Michael Rovine's work on this project was supported in part by NIMH Grant #MH43373.