Testes triangulares considerando a distribuição binomial: um exemplo com qualidade de café

Abstract

Discrimination tests can be used to detect adulteration in food products. Those tests allow to estimate the psychometric threshold to testers detect differences with adulteration. Triangular tests are a widely used version of discrimination experiments, as they are easy to understand by the testers and easy to design and implement. In a triangular test the tester proves three coded samples, being two equal and one different. He must pick up the different one. Usually this kind of experiments are analyzed under normal approximation. However, we found it better to model the number of correct calls using a Binomial distribution. It can be easily done representing the ability to spot differences as a latent variable in a Thurstonian model. In this paper we compare both ways to analyze a triangular test for coffee adulterants. Special attention is given to the precision of estimates by the different models. Treatment structure is a factorial scheme with two factors: type of adulterant and percentage of adulteration. Regression models were fitted to percentage of adulteration. Point and interval estimates for regression models and detection thresholds were obtained by Maximum Likelihood. Binomial fitting result in narrower confidence regions and changed some hypothesis on detection threshold.