12-10 |
Is the beta-binomial model always better? |
R. E. LIGGETT and J. F. Delwiche. Sensory Science Group, Ohio State Univ., Dept. of Food Science & Technology, 2015 Fyffe Ct., 110 Parker Food Science Bldg., Columbus, OH 43210-1007 Justification: Currently, whether conducting a discrimination test or a preference test, sensory evaluationists typically analyze the resultant data with binomial statistics. However, binomial statistics do not account for panelist differences. A newly introduced model, the beta-binomial, does account for panelist variability by the determination of gamma, a measure of overdispersion, or dissimilarity. When high, panelists are quite dissimilar. When very low, panelists are acting in an identical fashion. While this new model does account for panelist differences, it is more complicated and time consuming to calculate than the binomial model and it requires that a replicate be collected. It remains to be determined when these drawbacks are justified by the improvement in model fit. Objectives: The objective of the study was to compare panelist variability in terms of overdispersion, or gamma, when panelists are asked to complete preference tasks vs. discrimination tasks. Methods: Two fruit beverages varying only in sucrose level were presented to fifty panelists in a series of four comparisons. In the first two comparisons, panelists indicated which sample they preferred. In the subsequent two comparisons, panelists indicated which sample was sweeter. Data was then analyzed using the beta-binomial model and gamma was determined. Results: Preliminary results indicate that when performing a preference task, there is enough overdispersion in the responses (gamma is high enough) that the beta-binomial model provides a better fit to the data; however, when performing a discrimination task, panelists variability is reduced, overdispersion is low (gamma is very small) and the binomial model provides a better fit. Significance: These initial findings suggest that the binomial model is sufficient when analyzing data collected from a discrimination task but the beta-binomial model may be a more appropriate and useful tool when analyzing preference data.
Session 12, Sensory Evaluation: Methods and consumer testing
|