Mixed Beta Regression: A Bayesian Perspective

This paper builds on recent research that focuses on regression modeling of continuous bounded data, such as proportions measured on a continuous scale. Specifically, it deals with beta regression models with mixed effects from a Bayesian approach. We use a suitable parameterization of the beta law in terms of its mean and a precision parameter, and allow both parameters to be modeled through regression structures that may involve fixed and random effects. Specification of prior distributions is discussed, computational implementation via Gibbs sampling is provided, and illustrative examples are presented.

💡 Research Summary

The paper presents a comprehensive Bayesian framework for mixed‑effects beta regression, targeting continuous bounded outcomes that lie in the (0, 1) interval such as proportions, rates, and fractions. Traditional beta regression models usually parameterize the distribution by a mean μ and a precision (or dispersion) φ, but most applications only link μ to covariates while treating φ as a constant or a simple function of a few predictors. This work extends both parameters simultaneously: μ is modeled through a logistic link as a linear predictor that includes fixed effects (Xβ) and random effects (Zb), while φ is modeled on the log scale as another linear predictor (Wγ + Uc) that can also contain fixed and random components. By allowing random effects in the precision sub‑model, the approach captures heterogeneity not only in the central tendency but also in the variability of the data, which is particularly valuable when observations exhibit differing levels of dispersion across groups or clusters.

The authors adopt a fully Bayesian perspective, specifying prior distributions that balance non‑informativeness with computational tractability. Fixed‑effect coefficients β and γ receive diffuse normal priors (e.g., N(0, 10⁶)). Random effects b and c are assigned multivariate normal priors with zero mean and covariance matrices Σ_b and Σ_c, respectively; these covariance matrices are given inverse‑Wishart or LKJ‑correlation priors to ensure proper posterior behavior. For the precision parameter φ, a log‑normal prior is used, which leads to conjugate‑like conditional posteriors after transformation, simplifying Gibbs sampling. The paper devotes considerable space to prior‑sensitivity analysis, demonstrating that reasonable variations in hyper‑parameters do not materially alter posterior inferences, thereby providing practical guidance for applied researchers.

Computational implementation relies on a Gibbs sampler that cycles through closed‑form conditional distributions. After transforming μ and φ, the conditional posterior for each β and γ is normal, while the latent random effects follow multivariate normal conditionals. The precision φ, after log transformation, yields a gamma‑type conditional that can be sampled directly. Covariance matrices Σ_b and Σ_c are updated via inverse‑Wishart draws. Convergence diagnostics—including trace plots, Gelman‑Rubin R̂ statistics, and effective sample size calculations—are presented, and the authors supply R code (leveraging rjags and coda) that reproduces all analyses, ensuring reproducibility.

Two empirical illustrations showcase the method’s utility. The first example analyzes adverse‑event rates from a clinical trial, incorporating patient‑level random intercepts in both the mean and precision sub‑models, along with fixed covariates such as dose, age, and gender. The second example examines soil moisture fractions across multiple geographic sites and seasons, modeling site‑specific random effects and seasonal random slopes. In both cases, model comparison via Deviance Information Criterion (DIC) and Widely Applicable Information Criterion (WAIC) demonstrates substantial improvements over standard generalized linear models (GLMs) and over beta regressions that omit random effects in φ. Posterior predictive checks reveal tighter predictive intervals and better calibration, confirming that the mixed‑effects structure captures unobserved heterogeneity effectively.

The discussion acknowledges strengths and limitations. Strengths include the dual modeling of mean and dispersion, the natural incorporation of prior knowledge, and the relatively straightforward Gibbs sampling scheme that can be implemented with off‑the‑shelf software. Limitations involve potential slow mixing when the random‑effects dimension is large, and sensitivity of results to prior choices for variance components. The authors propose future extensions such as Hamiltonian Monte Carlo (e.g., via Stan) or variational Bayes to improve scalability, and hierarchical prior constructions that automate hyper‑parameter selection. Overall, the paper makes a significant methodological contribution by delivering a flexible, Bayesian mixed‑effects beta regression framework that can be readily applied to a wide range of proportion data across disciplines.