Testes para a seleção de provadores com base na proporção de acertos ao longo de ensaios triangulares

Abstract

Consumers have become increasingly critical and demanding when it comes to food choice. Seeking to improve the quality of these products, many studies involving sensory research become relevant. Sensory analysis is an experimental procedure that allows to measure, analyze and interpret reactions stimulated by food. This analysis is performed through sensory panels, formed by people responsible for sensorially analyzing the product. To select tasters and subject them to training, triangular tests are commonly used in order to assess the discriminative ability of judges. It’s desired to make the smallest possible error during selection, considering that these evaluators may have a constant hit proportion or not. In the latter situation, judges hit ratio decrease (due to fatigue), or increase (due to learning). Thus, the aim of this paper was to study interval estimators, rewriting them as tests to be applied for tasters selection, in order to recommend the best in terms of the lowest type I error rate and the highest power. For this, Monte Carlo simulations were performed considering tasters who hit a constant proportion (p) throughout the trials (n) and also those tasters who developed fatigue or learn in a linear and nonlinear way, evaluating test behavior without assumption (p constant). Additionally, real data were modeled for candidates submitted to olfactory coffee analysis, using segmented regression, which allows to verify changing points, helping in test decision and recommending the number of trials to be applied. Of the six tests analyzed, obtained from Normal estimators (TN1, TN2 and TN3), F-Distribution (TF), Sequential Test (TS) and Poisson (TP), TS presented the lowest type I error and the highest power for constant ratio and n < 20, while TN3 showed these characteristics for n 20. Without p constant assumption, type I error and power rates were very affected. For fatigue, there was a decrease in power, due to the small value of n when p increases, while with learning there was power increase simultaneous to n and p. Note that nonlinear functions affects the speed range of final power. Also, segmented regression is a great tool to be used to make a decision when the ratio of hits converges or reduces its variability. The recommendation was 15 trials for coffee olfactory analysis.