Abstract
Processing of the result of so-called triangle preference testing is discussed from a statistical point of view. Although criticized by some people, the testing is commonly practiced as convenient means to investigate the preference of consumers regarding new products, for example, without designating the characteristics of the difference in products to subjects. By imposing a discrimination testing on the subjects in the first stage, the tester can obtain the preference information selectively from those subjects who could perceive the difference in the products in the second stage. However, there has been much argument about the treatment of the answers from those that failed in the first stage. Important point to consider is, as far as there are a number of subjects who fail to give a correct answer in the first stage, there should also be a correlated number of subjects who give right answer by chance, and that the latter can be statistically estimated from the former. On this basis, three methods are devised to estimate the number of subjects who prefer each product perceiving the difference in this paper. First method makes use of a test based on the likelihood ratio method while the second, a test of goodness of fit. In the last method, the distribution of the number of subjects who fail in the first stage is directly estimated.
