Discrimination tests (e.g. triangle test, m-AFC, duo-trio, and recently the tetrads) are useful methodologies to detect whether process or recipe changes will be detected by consumers. The statistical analysis of such data can be done in two-ways:
- (1) estimating the proportion PD of discriminators who can detect the differences (Guessing model), or
- (2) estimating the perceptual distance δ between pairs of products (Thurstonian approach).
In both cases, the results obtained (whether it is PD or d’ observed) is compared to a threshold value define by the user. If the value observed is larger than the threshold, one would conclude that the difference is large enough to be detected by consumers.
These two methods have their own advantages and inconveniences. Although the guessing model is easy to apply (easy calculation) and to interpret, it is “method-specific”. Indeed, different methods having different sensitivities yield different proportion of correct answers PC. Since the calculation of PD only depends on PC and the probability of guessing right PG, different conclusions might be drawn from data obtained with different methods involving the same PG (Gridgeman paradox). Such paradox does not happen with the Thurstonian model since the decision rule associated with each method is taken into consideration in the model. However, Thurstonian model requires higher statistical skills and the interpretation of δ is not so straightforward.
Since both Thurstonian and Guessing model involve the same parameter in PC, it is possible to link PD directly to δ via PC. By doing so, it is possible to take the best of both models and get adequate results for each test by adapting the threshold value of PD within each test using δ. Such association induces in more adequate (thanks to the Thurstonian approach) and more interpretable (thanks to the Guessing model) results.