Abstract:
Subjective ratings are often used in the evaluation of hearing aids, loudspeakers, speech processors, and other systems involving sensory stimuli. The analysis of rating data, however, presents problems since the usual parametric assumptions underlying standard statistical techniques (e.g., t tests, analysis of variance) are not applicable to subjective ratings. A nonparametric method of analyzing rating data is described using concepts derived from signal detection theory. In order to compare two sets of ratings, a nonparametric estimate of the area under the ROC curve is obtained by first deriving the correlation function relating the two sets of ratings and then summing over half the range of the correlation function. This statistic, the half summed correlation (HSC), is relatively easy to compute and can be used to test, with known statistical power, for various differences between two sets of ratings (mean differences, variance differences, mean plus variance differences). It is also possible to rank sets of ratings using this statistic and to apply various multidimensional techniques designed for ordinal data. A variation of the analysis of proximities using HSC's will also be described. [Work supported by NIDCD.]