Thurstonian models for sensory discrimination tests as generalized linear models

Sensory discrimination tests such as the triangle, duo-trio, 2-AFC and 3-AFC tests produce binary data and the Thurstonian decision rule links the underlying sensory difference δδ to the observed number of correct responses. In this paper it is shown how each of these four situations can be viewed as a so-called generalized linear model. The underlying sensory difference δδ becomes directly a parameter of the statistical model and the estimate d′d′ and it’s standard error becomes the “usual” output of the statistical analysis. The d′d′ for the monadic A-NOT A method is shown to appear as a standard linear contrast in a generalized linear model using the probit link function. All methods developed in the paper are implemented in our free R-package sensR (http://www.cran.r-project.org/package=sensR/). This includes the basic power and sample size calculations for these four discrimination tests. Examples using data from the literature and illustrational data will be given throughout.