Beta-Binomial Model

Definition

The beta-binomial model (Crowder 1978) is a distribution for counted proportions that allows over-dispersion — more variability than a binomial permits. The number of successes $Y$ in $m$ trials is binomial given a success probability $p$, but $p$ itself varies from unit to unit as a Beta$(\alpha,\beta)$ random variable; marginalizing over $p$ gives the beta-binomial. It is the standard remedy for clustered or replicated binary data (litters, batches, replicates) whose proportions are "too variable" for a common binomial.

Key Ideas

• Binomial mixed over a Beta. $Y\mid p \sim \text{Bin}(m,p)$, $p \sim \text{Beta}(\alpha,\beta)$; integrating out $p$ yields $\Pr(Y=y)=\binom{m}{y}\frac{B(y+\alpha,\,m-y+\beta)}{B(\alpha,\beta)}$, a two-parameter over-dispersed alternative to the binomial.
• Over-dispersion as intraclass correlation. With mean $\pi=\alpha/(\alpha+\beta)$, the variance is $m\pi(1-\pi)\,[1+(m-1)\rho]$ where $\rho=1/(\alpha+\beta+1)$ is an intraclass correlation; $\rho\to0$ recovers the binomial, larger $\rho$ means more clustering.
• Regression / ANOVA structure. Covariate effects (treatments, groups) enter through the mean $\pi$ (and possibly the dispersion), giving an analysis-of-variance or regression model for proportions, estimated by maximum likelihood.
• Why not just binomial. Ignoring over-dispersion leaves standard errors too small and tests anti-conservative; the beta-binomial restores honest uncertainty with a single extra parameter.

Why It Matters

• Ubiquitous for clustered binary data. Toxicology (litter effects), agriculture (germination), survey and quality data — anywhere binary outcomes cluster within units — routinely show extra-binomial variation the beta-binomial handles.
• A building block. It is the over-dispersion device in the latent-class dependence-structure toolkit and a compact alternative to the logistic-normal GLMM and quasi-likelihood approaches to the same problem.

Open Questions

• Beta vs. logistic-normal mixing. The Beta mixing distribution is convenient but arbitrary; a logistic-normal random effect (GLMM) is often more natural for adding covariates and hierarchies, and the two can disagree in the tails.
• Dispersion modeling. Whether and how to let the over-dispersion parameter depend on covariates (double-GLM style) is model- and data-dependent.
• Correlation sign. The beta-binomial only accommodates non-negative intraclass correlation; under-dispersion (negative correlation) needs other models.

Related

• Latent Class Model
• Generalized Linear Mixed Model
• Mixture Model
• Crowder (1978) Beta-Binomial ANOVA for Proportions
• Martin J. Crowder