Beta regression mixed model applied to sensory analysis

João César Reis Alves,Gabriel Rodrigues Palma,Idemauro Antonio Rodrigues de Lara

2024-08-06

Abstract:Sensory analysis is an important area that the food industry can use to innovate and improve its products. This study involves a sample of individuals who can be trained or not to assess a product using a hedonic scale or notes, where the experimental design is a balanced incomplete block design. In this context, integrating sensory analysis with effective statistical methods, which consider the nature of the response variables, is essential to answer the aim of the experimental study. Some techniques are available to analyse sensory data, such as response surface models or categorical models. This article proposes using beta regression as an alternative to the proportional odds model, addressing some convergence problems, especially regarding the number of parameters. Moreover, the beta distribution is flexible for heteroscedasticity and asymmetry data. To this end, we conducted simulation studies that showed agreement rates in product selection using both models. Also, we presented a motivational study that was developed to select prebiotic drinks based on cashew nuts added to grape juice. In this application, the beta regression mixed model results corroborated with the selected formulations using the proportional mixed model.

Quantitative Methods

What problem does this paper attempt to address?

The paper primarily addresses the issue of data modeling in sensory analysis, particularly focusing on how the food industry can evaluate and improve products through effective statistical methods. Specifically, the paper tackles the following key issues:

1. **Problems with traditional methods**: Traditionally, data in sensory analysis (e.g., ratings based on a 9-point scale) are often treated as continuous variables for analysis of variance. However, this approach ignores the heteroscedasticity and skewness of the data, which can lead to incorrect interpretations.

2. **Over-parameterization issue**: When using certain models (such as mixed proportional odds models), there can be an over-parameterization issue, meaning that too many parameters in the model can lead to convergence difficulties or unstable estimates.

3. **Insufficient consideration of random effects**: Sensory analysis experimental designs often use incomplete block designs to reduce costs and fatigue, which requires models to effectively handle the random effects in such designs.

To address the above issues, the paper proposes the following solutions:

- **Using Beta regression models**: The paper proposes using Beta regression models with random effects to analyze sensory data. These models can directly handle heteroscedastic and skewed data and do not require assumptions of normal distribution and homoscedasticity, thereby reducing the over-parameterization issue.

- **Simulation study**: Through simulation studies, the performance of Beta regression models was compared with commonly used proportional odds cumulative logit models. The results show that in different simulation scenarios, the selection results of the two models are highly consistent, especially in cases with larger sample sizes.

- **Application example**: The paper also provides a practical case, using Beta regression models to select a prebiotic beverage formula based on adding cashew nuts to grape juice. The results show that the Beta regression model and the proportional mixed model selected the same formula, validating the effectiveness of the model.

In summary, the paper aims to provide a more flexible, accurate, and easy-to-implement data analysis method for sensory analysis to support product innovation and improvement processes in the food industry.