THE LOG-LINEAR TECHNIQUE PROVIDES A MEANS OF EXAMINING ALL POSSIBLE COMPLEXITIES OF A MULTIPLE CONTINGENCY TABLE AND ISOLATING THE MOST IMPORTANT ASSOCIATIONS AND INTERACTIONS. IN PAROLE PREDICTION, THE METHOD PROVIDES ESTIMATES OF THE EFFECTS OF VARIABLES ACTING ALONE OR IN CONJUNCTION WITH OTHER VARIABLES, AND A MEANS OF INDIRECTLY TESTING HYPOTHESIZED RELATIONSHIPS AMONG VARIABLES BY SETTING UP COMPETING MODELS AND COMPARING THEM WITH OBSERVED EVENTS. A SAMPLE OF 5,587 OHIO PAROLEES IS EMPLOYED IN A VALIDATION STUDY COMPARING THE LOG-LINEAR TECHNIQUE WITH THE BURGESS TECHNIQUE (SAID TO BE THE LEAST SOPHISTICATED STATISTICAL METHOD USED IN CONSTRUCTING PREDICTION INSTRUMENTS), AND IN A REPLICATION STUDY OF THE LOG-LINEAR MODEL. IN THE VALIDATION STUDY, THE LOG-LINEAR TECHNIQUE ACHIEVED ABOUT THE SAME PREDICTIVE EFFICIENCY AS THE BURGESS METHOD FOR BOTH CONSTRUCTION AND VALIDATION SAMPLES. THE HIERARCHICAL MODEL THAT EMERGED FROM THE LOG-LINEAR ANALYSIS OF THE CONSTRUCTION SAMPLE DID NOT REPLICATE ON THE VALIDATION SAMPLE. IN LIGHT OF OTHER STUDIES, THE FINDINGS INDICATE THAT IMPROVING STATISTICAL TECHNIQUES MAY NOT BE THE BEST MEANS OF IMPROVING PREDICTIVE EFFICIENCY. IMPLICATIONS OF THE RESULTS FROM PREDICTION RESEARCH AND THEORY TESTING INCLUDE THE INDICATION THAT REPLICATION OF HIERARCHICAL MODELS CANNOT BE ASSUMED BUT MUST BE STUDIED EMPIRICALLY. SUPPORTING DATA AND A LIST OF REFERENCES ARE INCLUDED. (AUTHOR ABSTRACT MODIFIED--LKM)