Comparison of the accuracy of beta-binomial, multinomial, dirichlet-multinomial, and ordinal regression in modelling quality of life data

• Main Authors: Ali Ghanbari, Mir Saeed Yekaninejad, Amir Pakpour, Sobhan Sanjari, Keramat Nourijelyani
• Format: Article
• Language: English
• Published: Tehran University of Medical Sciences 2018-10-01
• Series: Journal of Biostatistics and Epidemiology
• Subjects: Regression Analysis; Surveys and Questionnaires; Quality of Life; Women; Statistical Model.
• Online Access: https://jbe.tums.ac.ir/index.php/jbe/article/view/210
• ISSN: 2383-4196; 2383-420X

Background & Aim: Questionnaires are used mostly as a tool in medical research. Due to the different varieties of questionnaires, we may face different score distributions. In many cases multiple linear regression assumptions are violated. Beta-binomial regression model has the high flexibility and compatibility with this situation. In previous studies there were no comparison between beta-binomial accuracy and other models to fitting quality of life data. So in this study, our aim is to compare the accuracy of models to prediction. Methods & Materials: In this cross-sectional study we collected the quality of life data from 511 healthy women in Qazvin, Iran. The data were used to compare accuracy of betabinomial model and with some other models. Since beta-binomial considers the discrete response variable, so it should be compared with other similar models which are mostly used such as multinomial, dirichlet-multinomial and ordinal regression models. The main method that we used in our study was cross-validation to determine the accuracy of different models. To compare the different aspects, vast variety of situations were made and considered. Results: Regarding to the accuracy of models that were obtained by cross-validation in different situations, beta-binomial model had better accuracy among all models. Conclusion: According to the results, we have concluded that beta-binomial model is more accurate in prediction and fitting to the quality of data than the other models. The main advantages of this model are its simplicity, more efficacy and accuracy than the similar models.