Statistical Classification Procedures for Polygraph Tests of Guilty Knowledge

Abstract

This paper presents statistical procedures for classifying records of psychophysiological responses obtained in the guilty knowledge polygraph tests. The procedures to be presented are divided into a within subject approach and a training-data approach. In the former approach, the classification is made using only the record to be classified, whereas in the latter the rule for the classification is derived using trainingdata. The procedures were evaluated with records obtained in an experiment. The major result was that both the approaches yielded almost 80% correct classifications for the records containing a sufficient number of observations, but the within subject approach became inaccurate for the records with few observations in contrast to the training-data approach showing stable outcomes. It was also shown that the statistical procedures were superior to the diagnostic procedure relying on human interpretations of the records.