OBJECTIVES: Cross-sectional studies showed that, for modelling health-related quality of life (HRQoL), beta regression is superior in terms of fit and predictive accuracy to other commonly used methods based on normality distribution assumption. Although, longitudinal HRQoL measurements are widely used in clinical trials, not much is known about beta regression suitability in this context. This is mainly due to software unavailability with classical estimation methods. This study proposes to model the longitudinal HRQoL outcome using a mixed beta regression estimated by Bayesian Markov chain Monte Carlo (MCMC) methods implemented in WinBUGS. Compared to the classical approach, not only the Bayesian estimation is considerably easier to implement but has other advantages; for example, the possibility of including informative priors, enabling analysts to incorporate multiple sources of evidence in a single model. METHODS: We used a 16‑year longitudinal follow-up for modelling the relationship between SF-6D HRQoL and variables age, gender and mortality risk by means of a mixed beta regression. Besides modelling the mean parameter, we also modeled the precision parameter using a regression structure. Regression coefficients and predictive accuracy from this model estimated using the Bayesian approach with vague priors were compared to those from a linear mixed effects model estimated classically. RESULTS: Our results indicated that beta distribution fitted the SF-6D outcome better than normal distribution. Furthermore, compared to the linear mixed effects model, the mixed beta regression model was superior in terms of predictive model accuracy. We found that mortality risk had a significant effect on both mean and precision parameters: we observed lower mean HRQoL for higher mortality risk and higher variation of health utilities for higher mortality risk.     CONCLUSIONS: Mixed beta regression offers a superior approach for modelling the HRQoL outcome longitudinally. Furthermore, such a model can be easily implemented using freely available Bayesian software.