SOME HIERARCHICAL SCALING METHODS FOR CONFUSION MATRIX ANALYSIS

It is argued that current scaling models when they are applied to perceptual and mnemonic processes are out of touch with psychological theory. As an alternative, a psychological model is proposed in which stimuli are analysed by a series of feature tests, each feature test being dependent on the results of earlier feature tests in the series and on the other features present in the stimulus. Computer programs are described which fit such models to data, the output of such programs being the best-fitting hierarchy of feature tests. Detailed analyses are produced for data drawn from studies of speech perception and memory for speech; differences between normal, deaf and aphasic subjects are also discussed. It is claimed that, in comparison with normal scaling methods or with descriptions based simply on raw confusions, this approach offers a more insightful analysis of such data.