Abstract
We evaluated and compared two methods of data analysis for studies employing general recognition theory (GRT). GRT is a multidimensional signal detection theory that provides a rigorous framework for assessing and defining configural processing, and has been used to disambiguate between perceptual and decisional effects when perceptual dimensions interact (e.g., Cornes et al., 2011; Richler et al., 2008). The first method (Kadlec & Townsend, 1995; 1999) uses a set of marginal and non-parametric methods to indirectly estimate the configuration of underlying probability models, and the second uses probit regression models to estimate fully-parameterized probability models. The latter have previously been applied to GRT data (DeCarlo, 2003), but have not been evaluated in terms of relative strengths and weaknesses, either alone or in combination with marginal methods. Here, simulated data from known GRT configurations were used to determine the relative frequency of correct and incorrect inferences made by the two types of analysis. The two approaches were largely in agreement, but there were some important and striking differences. The results show that is the marginal methods are very conservative with respect to detecting correlations within perceptual distributions (violations of perceptual independence), while the probit analysis is conservative for detecting differences in sensitivity between pairs of perceptual distributions (violations of perceptual separability). The study suggests ways in which the two approaches may be combined (as sources of converging evidence) to support inferences regarding multidimensional signal detection models.
Meeting abstract presented at VSS 2013