Why a Single Binomial Probability May Be Too Restrictive
A binomial count assumes that, within a group of \(m\) trials, every trial has the same success probability and trials are independent. This structure can be too restrictive when groups differ in their underlying propensity for success. For example, groups with higher latent risk may tend to have more successes across several trials, while lower-risk groups tend to have fewer. A beta-binomial model represents this group-to-group variation in probability and, as a consequence, permits more variability in counts than a binomial model with one fixed probability.
The count remains bounded between zero and the number of trials. What changes is the probability structure: each group has its own latent probability, drawn from a shared distribution. Conditional on that latent probability, the familiar binomial model applies; marginally, after accounting for variation in the latent probability, the count is beta-binomial.
Here \(Y_i\) is the number of successes among \(m_i\) trials in group \(i\). The beta distribution is a distribution for the group-specific success probability, not for the count itself. The beta-binomial count therefore has support \(0,1,\ldots,m_i\), just as the binomial count does.
Probability Mass Function and Parameters
Integrating out the latent probability gives the beta-binomial probability mass function. Define the beta function by \(B(u,v)=\int_0^1 t^{u-1}(1-t)^{v-1}\,dt\) for \(u,v>0\). For an observed count \(y_i\), the conditional binomial probability is multiplied by the beta density and integrated over all possible values of \(\Pi_i\):
This is a full probability likelihood for the count. The binomial coefficient accounts for the different arrangements of \(y_i\) successes among \(m_i\) trials. The beta-function ratio averages the conditional binomial probability over the group-probability distribution.
A useful parameterization separates the mean probability from the extent of heterogeneity. Write \(\kappa_i=a_i+b_i\), and define \(p_i=a_i/(a_i+b_i)\). Then \(a_i=p_i\kappa_i\), \(b_i=(1-p_i)\kappa_i\), and the latent probability has mean \(p_i\). The parameter \(\kappa_i\) is often called the concentration or precision: larger \(\kappa_i\), at a fixed \(p_i\), means less variation among the latent probabilities.
In a regression setting, the mean probability can depend on covariates, for example \(p_i=\operatorname{logit}^{-1}(x_i^\top\beta)\), while \(\kappa_i\) controls the amount of probability heterogeneity. This gives a beta-binomial response likelihood with a modeled mean. It is not necessary to use this specific link to define the distribution.
Worked Example: Calculating a Beta-Binomial Probability
Suppose a hypothetical group has \(m=4\) trials and a latent probability \(\Pi\sim\operatorname{Beta}(2,3)\). Find the probability of exactly two successes.
The beta-binomial probability is \({4\choose 2}B(2+2,3+4-2)/B(2,3)\). Thus \(B(4,5)=3!4!/8!=1/280\), and \(B(2,3)=1!2!/4!=1/12\). Substitution gives \[ P(Y=2)={4\choose 2}\frac{B(4,5)}{B(2,3)} =6\left(\frac{1/280}{1/12}\right) =6\left(\frac{12}{280}\right) =\frac{9}{35}\approx 0.2571. \]
The mean latent probability is \(p=2/(2+3)=0.4\), and \(\kappa=5\). The corresponding binomial probability at the same mean is \({4\choose2}(0.4)^2(0.6)^2=0.3456\). These are probabilities for one particular count, not measures of overall dispersion by themselves; the variance comparison below describes dispersion across the full distribution.
Mean, Variance, and Dependence
The expected count is \(m_i p_i\), the same mean as a binomial count with probability \(p_i\). The variance, however, is inflated when the latent probability varies. The derivation uses the law of total variance: the marginal variance equals the expected conditional variance plus the variance of the conditional mean.
\(E(Y_i\mid\Pi_i)=m_i\Pi_i\) and \(\operatorname{Var}(Y_i\mid\Pi_i)=m_i\Pi_i(1-\Pi_i)\).
\(E(Y_i)=E\{E(Y_i\mid\Pi_i)\}=m_iE(\Pi_i)=m_ip_i\).
\(\operatorname{Var}(Y_i)=E\{\operatorname{Var}(Y_i\mid\Pi_i)\}+\operatorname{Var}\{E(Y_i\mid\Pi_i)\}\), which yields the expression below.
To complete the variance calculation, use \(E\{\Pi_i(1-\Pi_i)\}=p_i(1-p_i)-\operatorname{Var}(\Pi_i)\), and substitute \(\operatorname{Var}(\Pi_i)=p_i(1-p_i)/(\kappa_i+1)\):
The first factor, \(m_i p_i(1-p_i)\), is the binomial variance. The multiplier is at least one, and is greater than one when \(m_i>1\) and \(\rho_i>0\). The beta-binomial therefore adds extra-binomial variation while preserving the count’s bounded support.
The same \(\rho_i\) has a dependence interpretation. Let \(X_{ij}\) be the success indicator for trial \(j\) in group \(i\), so \(Y_i=\sum_{j=1}^{m_i}X_{ij}\). For two distinct trials in the same group, conditional independence gives \(\operatorname{Cov}(X_{ij},X_{ik})=\operatorname{Var}(\Pi_i)\) marginally. Each indicator has marginal variance \(p_i(1-p_i)\), so their correlation is \(\operatorname{Var}(\Pi_i)/[p_i(1-p_i)]=1/(\kappa_i+1)=\rho_i\). Thus overdispersion in the count arises from positive within-group association induced by the shared latent probability, not from conditional dependence after \(\Pi_i\) is known.
Worked Example: Translating Mean and Correlation Into Beta Parameters
Suppose a hypothetical set of groups has \(m=10\) trials per group, mean success probability \(p=0.30\), and within-group correlation \(\rho=0.10\). Find the beta parameters and compare the two count variances.
Since \(\rho=1/(\kappa+1)\), \(\kappa=1/\rho-1=10-1=9\). Therefore \(a=p\kappa=0.30(9)=2.7\) and \(b=(1-p)\kappa=0.70(9)=6.3\). The mean count is \(mp=10(0.30)=3.0\) successes.
The beta-binomial variance is \(mp(1-p)\{1+(m-1)\rho\} =10(0.30)(0.70)\{1+9(0.10)\} =2.1(1.9)=3.99\) successes squared.
A binomial count with the same \(m\) and \(p\) has variance \(mp(1-p)=10(0.30)(0.70)=2.10\) successes squared. The beta-binomial variance is \(3.99/2.10=1.90\) times as large. This increase reflects variation in the latent probability across groups; it does not change the mean count of 3.0.
Worked Example: Probability of Extreme Group Counts
Consider three hypothetical, independent groups, each with four trials. Under a beta-binomial model let \(a=b=1\), so the latent probability is uniform on \([0,1]\). Suppose the observed counts are \(0,2,4\). Calculate their joint likelihood under this specified model and compare it with the likelihood under a binomial model having \(m=4\) and \(p=0.5\).
For \(a=b=1\), the beta-binomial probability at any count \(y=0,\ldots,4\) is \({4\choose y}B(y+1,5-y)/B(1,1)=1/5\). For example, at \(y=2\) this is \(6B(3,3)=6(1/30)=1/5\); the same result holds at each count. Independence across groups gives the likelihood \((1/5)(1/5)(1/5)=1/125=0.008000\).
Under the specified binomial model, the probabilities of counts 0, 2 and 4 are respectively \({4\choose0}(0.5)^4=1/16\), \({4\choose2}(0.5)^4=6/16\), and \({4\choose4}(0.5)^4=1/16\). The joint likelihood is therefore \((1/16)(6/16)(1/16)=6/4096\approx0.001465\). The beta mixture assigns more probability to this pattern containing extreme group counts. This arithmetic illustrates the model’s probability structure; it is not a formal model-selection conclusion, because the example compares specified parameter values rather than fitted models.
Assumptions, Diagnostics, and Limits
The beta-binomial model is appropriate as a probability model only when its assumptions are plausible. In the basic formulation, each group has a known trial count \(m_i\); its latent probability lies in \((0,1)\) and follows a beta distribution; conditional on that probability, the trials are independent Bernoulli outcomes with the same success probability; and the latent probabilities are independent across groups, conditional on any modeled covariates. The beta distribution imposes a particular form of heterogeneity: it varies a group’s probability while keeping probabilities within the unit interval.
Diagnostics should examine both the mean pattern and the distribution of counts. If observed counts show more zeros or more upper-end counts than fitted probabilities predict, that may indicate remaining heterogeneity or a misspecified mean structure. Compare observed and fitted frequencies across the full support \(0,\ldots,m_i\), and check whether residual patterns or dispersion differ by covariate values or trial count. A single variance comparison is not sufficient: a model can match the variance while missing important features of the count distribution.
The main alternative at this stage is the binomial likelihood. It is recovered as \(\kappa_i\) grows without bound while \(p_i\) remains fixed: the beta distribution for \(\Pi_i\) concentrates around \(p_i\), \(\rho_i\) tends to zero, and the beta-binomial variance tends to the binomial variance. A beta-binomial model is not automatically preferable whenever counts vary. It assumes a shared beta mixing structure, and it can be inadequate if groups are dependent for other reasons, if covariates do not capture systematic mean differences, or if the observed response process is not a count of successes out of a known number of trials.
In a generalized linear mixed model, a random effect can also induce group-level variation and within-group association. Beta-binomial mixing is a specific probability-scale representation: it places a beta distribution directly on the group success probability. It is not identical in general to placing a normally distributed random intercept on the log-odds scale. The distinction matters because the two mixing distributions imply different marginal count probabilities and dependence patterns.
Common Mistakes
- Calling the conditional trials dependent. Given \(\Pi_i\), the trials are independent in this model. Their marginal positive association arises because they share the random probability.
- Changing the mean when describing overdispersion. At fixed \(p_i\), the beta-binomial and binomial means are both \(m_ip_i\). The beta-binomial variance changes because the latent probability varies.
- Confusing \(\rho_i\) with the count variance. \(\rho_i\) is the correlation between two distinct trial indicators in the same group. The count variance also depends on \(m_i\) and \(p_i\).
- Assuming every extra-variable pattern is beta-binomial. A larger variance alone does not verify the beta mixing assumption; fitted counts should also be checked across their support and covariate patterns.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- If \(a=3\) and \(b=7\), what are the mean latent probability and the within-group correlation?
- For \(m=5\), \(p=0.40\), and \(\rho=0.20\), calculate the beta-binomial mean and variance.
- Why can two trials in the same group be marginally correlated even though they are conditionally independent?
- What happens to the beta-binomial variance as \(\kappa\) increases with \(p\) fixed?
- Does a variance larger than the binomial variance alone establish that the beta-binomial model fits well? Explain.
Solutions. (1) \(p=a/(a+b)=3/10=0.30\), and \(\rho=1/(a+b+1)=1/11\approx0.09091\). (2) The mean is \(mp=5(0.40)=2.0\). The variance is \(5(0.40)(0.60)\{1+4(0.20)\}=1.2(1.8)=2.16\) successes squared. (3) Sharing \(\Pi_i\) makes outcomes tend to be high together when the group probability is high and low together when it is low; their marginal covariance is \(\operatorname{Var}(\Pi_i)>0\). (4) Since \(\rho=1/(\kappa+1)\), increasing \(\kappa\) reduces the variance multiplier toward one, so the count variance approaches the binomial variance. (5) No. Extra variance is consistent with beta-binomial mixing, but the full count distribution and covariate patterns must also be assessed.