Evaluating Wine-Tasting Results and Randomness with a Mixture of Rank Preference Models

Evaluating observed wine-tasting results as a mixture distribution, using linear regression on a transformation of observed results, has been described in the wine-tasting literature. This article advances the use of mixture models by considering that existing work, examining five analyses of ranking and mixture model applications to non-wine food tastings and then deriving a mixture model with specific application to observed wine-tasting results. The mixture model is specified with Plackett-Luce probability mass functions, solved with the expectation maximization algorithm that is standard in the literature, tested on a hypothetical set of wine ranks, tested with a random-ranking Monte Carlo simulation, and then employed to evaluate the results of a blind tasting of Pinot Gris by experienced tasters. The test on a hypothetical set of wine ranks shows that a mixture model is an accurate predictor of observed rank densities. The Monte Carlo simulation yields confirmatory results and an estimate of potential Type I errors (the probability that tasters appear to agree although ranks are actually random). Application of the mixture model to the tasting of Pinot Gris, with over a 95% level of confidence based on the likelihood ratio and t statistics, shows that agreement among tasters exceeds the random expectation of illusory agreement. (JEL Classifications: A10, C10, C00, C12, D12)