Key points are not available for this paper at this time.
A statistical technique is proposed for comparing an empirically obtained confusion matrix against a set of distinctive features that supposedly characterize the stimuli on which the given confusion matrix is based. Each distinctive feature corresponds to a partition of the stimulus set, and the term “confusion matrix” refers to the measures of “closeness” collected on all pairs of stimuli. The suggested paradigm can be considered an analysis-of-variance generalization and is dependent on a randomization strategy for evaluating the size of a goodness-of-fit index calculated between the given confusion matrix and a single partition. An example of the inference scheme is carried out on a data set dealing with the 26 Roman capital letters; in addition, an exploratory strategy is illustrated that tries to locate “good” partitionings of a stimulus set in a post-hoc manner.
Hubert et al. (Wed,) studied this question.