5CONDUCTING A RATING SCALE EXPERIMENT
From the preceding chapters it should now be clear that the safest way of gaining information about an observer’s sensitivity is by obtaining from him a number of values of P(S|s) and P(S|n) which will allow the path of the ROC curve to be determined. In Chapter 2 it was found that although a single pair of hit and false alarm rates could be
used to make an approximate estimate of the area under the curve, the measure was not independent of response bias and would underestimate sensitivity to a degree dependent on the observer’s tendency to prefer S to N responses. In Chapter 4 it was seen
that a single pair of hit and false alarm rates could give different values of for points on the same ROC curve if the slope of the curve was not equal to 1. In a 2AFC task,
interval bias can affect the obtained value of P(c) and hence, . Two methods have been described for obtaining a set of hit and false alarm rates which
represent different degrees of response bias but equivalent degrees of sensitivity. The first required conducting a series of yes-no tasks with signal and noise strengths held constant but with the observer varying his bias from task to task. The second involved the use of a rating scale on which the observer simultaneously held several criteria of differing degrees of strictness. As the rating scale method is the most efficient as far as the number of trials needed to obtain the ROC curve points is concerned it is likely to be the procedure most favoured by potential users of detection theory. In the following sections of this chapter, a rating scale task is worked through and a number of problems which an experimenter may encounter are discussed. The example used is a yes-no rating scale task but most of the steps described could be applied equally to a 2AFC rating task.