Two Likelihoods for Success Counts
Suppose each group contributes a count of successes from a known number of binary trials. A binomial model is appropriate when trials in a group share the same success probability and are independent. A beta-binomial model allows that probability to differ across groups: each group has its own latent success probability, drawn from a beta distribution. This extra layer represents group-to-group probability variation and, after averaging over it, produces more variable counts than the binomial model with the same mean.
The choice is about the response-generating process, not just which likelihood can accommodate a larger sample variance. The beta-binomial is especially relevant when several trials share a group, such as repeated outcomes within a subject or items within a batch, and groups may have different underlying success probabilities. It does not automatically explain every form of dependence or every source of extra variation.
For the beta-binomial model, one useful parameterization sets \(P_i\sim\operatorname{Beta}(\mu_i\phi,(1-\mu_i)\phi)\), with \(0<\mu_i<1\) and \(\phi>0\). Here \(\mu_i\) is the mean group probability and \(\phi\) controls how concentrated the distribution of group probabilities is. The corresponding intragroup correlation is \(\rho=1/(\phi+1)\): small \(\rho\) means little probability variation, while larger \(\rho\) means more.
Likelihood and Variance
For one group, the binomial probability of observing \(y_i\) successes is
For the beta-binomial, average the conditional binomial probability over the beta distribution of \(P_i\). Writing \(a_i=\mu_i\phi\) and \(b_i=(1-\mu_i)\phi\), the resulting probability is
where \(B(a,b)=\int_0^1 p^{a-1}(1-p)^{b-1}\,dp\) is the beta function. If groups are independent, the likelihood for all observed counts is the product of their group probabilities. The two likelihoods have the same response support, from zero through \(m_i\); their key distinction is how much probability they assign to counts far from the mean.
The variance result follows from conditioning on the latent probability. For \(P_i\sim\operatorname{Beta}(\mu_i\phi,(1-\mu_i)\phi)\), its mean is \(\mu_i\) and its variance is \(\mu_i(1-\mu_i)/(\phi+1)\). The law of total variance then gives
The added variance is a consequence of variation in the group probabilities, not a change in the number of trials. In the limiting case \(\phi\to\infty\), the beta distribution concentrates at \(\mu_i\), \(\rho\to0\), and the beta-binomial approaches the binomial model.
Worked Example: Comparing Variances at the Same Mean
Hypothetical setup. Each of many groups has \(m=10\) trials and mean success probability \(\mu=0.30\). Compare a binomial model with a beta-binomial model having \(\phi=4\). These assumptions specify the models; the calculation does not establish which one generated any particular data set.
For the binomial model, the expected count is \(10(0.30)=3\), and the variance is \(10(0.30)(0.70)=2.10\) successes squared.
For the beta-binomial model, \(\rho=1/(4+1)=0.20\). Its mean remains \(10(0.30)=3\), but its variance is \(10(0.30)(0.70)\{1+(10-1)(0.20)\}=2.10(2.80)=5.88\) successes squared.
Both models have the same mean count, but the beta-binomial predicts much greater variation among group counts. In this parameterization, each group’s probability is drawn from \(\operatorname{Beta}(1.2,2.8)\); groups with relatively high or low latent probabilities contribute more extreme counts.
What Extra Variation Looks Like
A practical signal for considering the beta-binomial is that observed group counts are more dispersed than a binomial model with the relevant trial counts and mean probabilities predicts. In particular, repeated groups may show more very low and very high counts, and fewer counts near their common mean, than expected under a fixed-probability model. A sample variance larger than the binomial variance can motivate further checking, but it does not by itself prove that the beta-binomial is appropriate.
Worked Example: Comparing Likelihood Probabilities for an Extreme Count
Hypothetical setup. A group has \(m=4\) trials and an expected success probability of \(0.50\). Compare a binomial model with a beta-binomial model using \(P\sim\operatorname{Beta}(2,2)\). This beta distribution has mean \(0.50\) and concentration \(\phi=4\).
Under the binomial model, the probability of four successes is \(\binom{4}{4}(0.50)^4(0.50)^0=0.0625\).
Under the beta-binomial model, it is \(\binom{4}{4}B(4+2,0+2)/B(2,2)\). Here \(B(6,2)=1/42\) and \(B(2,2)=1/6\), so the probability is \(1(1/42)/(1/6)=1/7\approx0.1429\).
Thus four successes are more than twice as probable under this beta-binomial model. This does not mean that four successes are always evidence for beta mixing: it shows what the two specified models predict for this count. Across groups, the beta-binomial reallocates probability toward both low and high counts. For comparison, the probability of exactly two successes is \(6/16=0.3750\) under the binomial, but \(9/35\approx0.2571\) under this beta-binomial model.
A useful moment diagnostic, when groups share a common \(m\) and mean probability, compares their observed count variance \(s_Y^2\) with the binomial variance \(m\bar{Y}(1-\bar{Y}/m)\), where \(\bar{Y}\) is the sample mean count. A rough moment estimate of the intragroup correlation is
This estimate is a diagnostic, not a model-selection rule. It can be unstable with few groups and may fall outside the allowable range \(0\leq\rho<1\). Variation in group sizes or mean probabilities also calls for care: raw count variances are not directly comparable when \(m_i\) or \(\mu_i\) differ. In those settings, compare observed counts with predictions from a fitted model rather than relying on one pooled variance.
Worked Example: A Moment Diagnostic Across Groups
Hypothetical data. Eight independent groups each have five trials, with observed success counts \(0,0,1,1,2,2,5,5\). The total is \(16\) successes in \(40\) trials, so \(\bar{Y}=16/8=2\) and \(\hat{\mu}=2/5=0.40\). The binomial variance at this estimated mean is \(5(0.40)(0.60)=1.20\) successes squared.
The sample variance of the eight counts is calculated around their mean of 2: \[ s_Y^2=\frac{(0-2)^2+(0-2)^2+(1-2)^2+(1-2)^2+(2-2)^2+(2-2)^2+(5-2)^2+(5-2)^2}{8-1} =\frac{28}{7}=4.00. \]
The moment estimate is \(\hat{\rho}_{\mathrm{mom}}=(4.00/1.20-1)/(5-1)=0.5833\). The corresponding concentration estimate is \(\hat{\phi}=1/\hat{\rho}_{\mathrm{mom}}-1\approx0.7143\). These values suggest substantial extra-binomial variation in this small example. They are not conclusive: only eight groups were observed, and the two counts of 0 and two counts of 5 could reflect sampling fluctuation, omitted group differences, or another process as well as beta-distributed latent probabilities.
Choosing Between the Models
The binomial likelihood is the more direct choice when the trials within a group can reasonably be treated as independent with a common probability, and there is no persuasive evidence of extra group-to-group probability variation. The beta-binomial is a candidate when the design supplies repeated trials within groups and the data show more between-group count variation than the binomial predicts. The model’s beta distribution is an assumption about how probabilities vary across groups; it is not a general-purpose correction for any mismatch.
| Question | Binomial model | Beta-binomial model |
|---|---|---|
| What varies among trials in a group? | Trials have a common fixed probability \(\mu_i\). | Trials share a group-specific probability \(P_i\), which varies across groups. |
| How is within-group dependence treated? | Trials are independent under the model. | Trials are independent conditional on \(P_i\), but positively associated marginally. |
| What count variability is expected? | Variance \(m_i\mu_i(1-\mu_i)\). | Variance inflated by \(1+(m_i-1)\rho\). |
| When are they hard to distinguish? | When groups have only one trial, or when the beta probabilities are highly concentrated near their mean. | |
If there is just one trial per group, both models give a Bernoulli response with probability \(\mu_i\): the count data contain no information about \(\phi\) or \(\rho\). Likewise, with very few groups, estimating probability variation can be imprecise even if each group has many trials. These are design limitations, not problems that a more elaborate fitting procedure can eliminate.
The beta-binomial also has structural limits. It represents positive, exchangeable association among trials in a group: every pair is linked through the same latent probability. It does not represent a specified time pattern, negative association, or arbitrary correlation structure. Nor does it substitute for accounting for known group differences that change the mean probability. When group covariates or other structure explain the variation, omitting them and attributing all dispersion to beta mixing can give a misleading account of the response process.
Diagnostics and Common Pitfalls
After fitting either model, check whether its predicted count distribution is compatible with the data, especially at zero and at the maximum \(m_i\), and whether variability is adequately represented across groups with comparable trial counts and covariates. Large, systematic discrepancies in such comparisons can signal a poor mean model, an unsuitable dependence structure, or a distribution other than either candidate. A beta-binomial can reduce extra-binomial dispersion, but it does not guarantee good fit.
- Mistake: “The sample variance exceeds the binomial variance, so the beta-binomial must be correct.” A noisy variance estimate or omitted group structure can also produce that pattern. Check the design, the fitted mean, and count-level predictions; treat the variance comparison as evidence to investigate, not proof.
- Mistake: “The beta-binomial makes every trial probability independent.” Conditional on a group’s latent probability, trials are independent. After that probability is averaged out, trials in the same group are positively associated.
- Mistake: “A beta-binomial is needed whenever counts come from clusters.” Clustering alone does not specify the response distribution. The beta-binomial is suitable only when its shared latent probability and exchangeable positive association are plausible for the grouped trials.
The main practical comparison is therefore between a simpler fixed-probability account and a specific model of probability heterogeneity. Prefer the binomial when its assumptions and predictive checks are adequate. Consider the beta-binomial when the grouping, extra variation, and predicted pattern of counts all support the additional latent-probability layer; avoid treating a better fit alone as proof of that mechanism.
Check Your Understanding
Answer each question, then check the solutions below.
- For \(m=6\), \(\mu=0.25\), and \(\phi=3\), calculate the beta-binomial variance and compare it with the binomial variance.
- Why can’t one trial per group distinguish a binomial model from a beta-binomial model?
- What does a positive \(\rho\) mean for two trials from the same group?
- A data set shows extra variation, but group sizes vary widely. Why is a single raw sample variance an inadequate diagnostic?
Solutions. (1) The binomial variance is \(6(0.25)(0.75)=1.125\). Here \(\rho=1/(3+1)=0.25\), so the beta-binomial variance is \(1.125\{1+(6-1)(0.25)\}=1.125(2.25)=2.531\) successes squared. (2) With \(m=1\), the count is a single Bernoulli outcome; mixing its probability leaves the marginal success probability equal to \(\mu\), so \(\phi\) is not identified from these counts. (3) The shared latent probability makes two trials in a group positively associated after averaging over that probability; \(\rho\) measures this intragroup correlation under the model. (4) Counts with different denominators have different means and variances even under a binomial model, so their raw dispersion is not directly comparable. A diagnostic should account for group sizes and fitted mean probabilities.