Common Interpretation Errors in Overdispersed GLMs

Why Parameter Interpretation Matters

Overdispersion models add parameters to account for variability beyond a Poisson or binomial model. Those extra parameters are easy to misread: a negative-binomial size parameter is not itself a variance, beta-binomial correlation does not raise or lower the mean success probability, and a quasi-likelihood scale adjustment does not supply a complete probability distribution. Each parameter has a specific role, and that role should be checked against the model’s assumptions and moment formulas.

The distinctions matter in practice. Confusing size with variance can lead to nonsensical comparisons across fitted means. Treating within-group correlation as a mean effect changes the scientific interpretation of a grouped response. Treating a quasi-likelihood as a probability model invites unsupported predictions and likelihood-based comparisons. The calculations below make these limits explicit.

Negative-Binomial Size Is Not the Variance

Use the negative-binomial parameterization in which a count $Y$ has mean $\mu>0$, size $r>0$, and variance $\mu+\mu^2/r$. Its probability mass function is

P(Y=y)=\frac{\Gamma(y+r)}{\Gamma(r)y!} \left(\frac{r}{r+\mu}\right)^r \left(\frac{\mu}{r+\mu}\right)^y, \qquad y=0,1,2,\ldots .

In this parameterization, $r$ controls the amount of extra-Poisson variation: at a fixed mean, a larger $r$ makes the variance closer to the Poisson variance $\mu$. The size can be non-integer and is not measured in squared count units. In particular, it is not a direct estimate of $\operatorname{Var}(Y)$.

Definition: Under the negative-binomial mean-size parameterization, $\operatorname{Var}(Y)=\mu+\mu^2/r$. Equivalently, the excess variance over the Poisson variance is $\mu^2/r$. The size $r$ is a dispersion parameter; it is neither the mean nor the variance.

One way to verify the variance formula is through the gamma-Poisson representation. Suppose $\Lambda$ has a gamma distribution with shape $r$ and rate $r/\mu$, and $Y\mid\Lambda\sim\operatorname{Poisson}(\Lambda)$. Then $E[\Lambda]=\mu$ and $\operatorname{Var}(\Lambda)=\mu^2/r$. By conditional expectation and the law of total variance,

E[Y]=E\{E[Y\mid\Lambda]\}=E[\Lambda]=\mu,

\operatorname{Var}(Y) =E\{\operatorname{Var}(Y\mid\Lambda)\} +\operatorname{Var}\{E[Y\mid\Lambda]\} =E[\Lambda]+\operatorname{Var}(\Lambda) =\mu+\frac{\mu^2}{r}.

The two variance components have different sources: the conditional Poisson variation contributes $\mu$, and variation in the latent rate contributes $\mu^2/r$. This also shows why $r$ cannot be read as the total variance.

Worked Example: A Size of Four Does Not Mean a Variance of Four

Suppose a fitted negative-binomial model gives mean $\mu=8$ and size $r=4$. Calculate the variance and compare it with the size.

The variance is

\operatorname{Var}(Y)=8+\frac{8^2}{4}=8+16=24.

Thus $r=4$ is not the variance: the variance is 24. The excess above the Poisson variance is $24-8=16$, which equals $\mu^2/r=64/4$. If the mean were instead 4 with the same size, the variance would be $4+16/4=8$. The same size therefore does not imply the same variance across different means.

The standard deviation for the first model is $\sqrt{24}$, not 4. The size parameter describes how the variance grows with the mean; it does not give a count-scale spread on its own.

A related pitfall is to compare size estimates without considering the parameterization and the fitted means. In this model, $1/r$ is the coefficient of $\mu^2$ in the variance function. Some software or texts use other negative-binomial parameterizations, so the label “dispersion” alone is not enough to identify a parameter’s meaning. Check the stated variance function before interpreting an estimate.

Beta-Binomial Correlation Is Not a Mean Effect

For a group with $m$ trials, the beta-binomial model can be written as a mixture: a group-specific success probability $P$ is drawn from a beta distribution, then the success count $Y$, conditional on $P$, has a binomial distribution. Write $p=E[P]$ for the marginal success probability and $\rho$ for the intraclass correlation between two distinct trial outcomes in the same group. In the usual beta-mixture parameterization, if $P\sim\operatorname{Beta}(\alpha,\beta)$, then $p=\alpha/(\alpha+\beta)$ and $\rho=1/(\alpha+\beta+1)$.

The mean probability $p$ and the correlation $\rho$ answer different questions. The first describes the marginal chance of success on a trial. The second describes how strongly outcomes in the same group tend to move together. Increasing $\rho$ at fixed $p$ changes the spread of group counts, not their expected value.

Theorem: If $P\sim\operatorname{Beta}(\alpha,\beta)$, $Y\mid P\sim\operatorname{Binomial}(m,P)$, $p=\alpha/(\alpha+\beta)$, and $\rho=1/(\alpha+\beta+1)$, then $$ E[Y]=mp,\qquad \operatorname{Var}(Y)=mp(1-p)\{1+(m-1)\rho\}. $$ For two distinct Bernoulli outcomes in the group, their correlation is $\rho$.

Proof. Let $X_j$ be the success indicator on trial $j$, for $j=1,\ldots,m$, so $Y=\sum_{j=1}^m X_j$. Conditional on $P$, each $X_j$ has expectation $P$, and $Y$ has a binomial distribution. The beta moments give $E[P]=p$ and $\operatorname{Var}(P)=p(1-p)\rho$. Therefore, $E[Y]=E\{E[Y\mid P]\}=E[mP]=mp$.

For the variance, use the law of total variance and $E[P(1-P)]=p(1-p)-\operatorname{Var}(P)=p(1-p)(1-\rho)$:

\begin{aligned} \operatorname{Var}(Y) &=E\{mP(1-P)\}+\operatorname{Var}(mP)\\ &=mp(1-p)(1-\rho)+m^2p(1-p)\rho\\ &=mp(1-p)\{1+(m-1)\rho\}. \end{aligned}

For distinct trials $j\ne k$, conditional independence gives $E[X_jX_k\mid P]=P^2$, so $\operatorname{Cov}(X_j,X_k)=E[P^2]-E[P]^2=\operatorname{Var}(P)$. Since each marginal $X_j$ is Bernoulli with variance $p(1-p)$, their correlation is $\operatorname{Var}(P)/\{p(1-p)\}=\rho$. This proves both claims. $\square$

Worked Example: Correlation Changes the Spread, Not the Expected Count

A group has $m=10$ trials, marginal success probability $p=0.30$, and beta-binomial intraclass correlation $\rho=0.20$. Find the mean and variance of the success count. Then compare the mean and variance with a binomial model having the same $m$ and $p$.

The beta-binomial mean is

$$ E[Y]=mp=10(0.30)=3. $$

Its variance is

\operatorname{Var}(Y) =10(0.30)(0.70)\{1+9(0.20)\} =2.1(2.8)=5.88.

The binomial model has the same mean, 3, but variance $mp(1-p)=2.1$. Thus the positive within-group correlation increases the count variance from 2.1 to 5.88 without changing the marginal success probability or the expected count. Calling $\rho=0.20$ a “20% increase in the mean” is incorrect.

Quasi-Likelihood Scale Is Not a Probability Model

A quasi-likelihood approach specifies a relationship between the mean and variance and uses it to form estimating equations. For example, a quasi-Poisson mean model uses $E[Y_i]=\mu_i$ and $\operatorname{Var}(Y_i)=\phi\mu_i$, where $\phi>0$ is a scale parameter. This variance assumption can support estimation and uncertainty calculations without specifying a complete probability mass function for each $Y_i$.

In general, for a variance function $V(\mu)$, a quasi-likelihood contribution $Q(y;\mu)$ is defined through its derivative with respect to the mean:

\frac{\partial Q(y;\mu)}{\partial\mu} =\frac{y-\mu}{V(\mu)}.

A multiplicative scale can be included in the variance function. For quasi-Poisson, $V(\mu)=\phi\mu$, so integrating with respect to $\mu$ yields

Q(y;\mu)=\frac{1}{\phi}\{y\log(\mu)-\mu\}+C(y),

where $C(y)$ does not depend on $\mu$. This expression resembles a scaled Poisson log-likelihood, but resemblance is not the same as being the log of a normalized probability mass function. A quasi-likelihood specifies the mean-variance structure and estimating procedure; by itself, it does not determine $P(Y=y)$ for every possible response value.

Worked Example: Matching a Variance Does Not Specify the Distribution

At a particular fitted mean, let $\mu=4$ and use a quasi-Poisson scale $\phi=2$. The specified variance is $\phi\mu=8$. A negative-binomial model with mean 4 and size 4 also has variance $4+4^2/4=8$.

These matching first two moments do not make the models the same. Under the negative-binomial model, the probability of zero is determined by its probability mass function:

P(Y=0)=\left(\frac{4}{4+4}\right)^4 =\left(\frac12\right)^4=\frac{1}{16}.

The quasi-Poisson mean and variance conditions alone do not give a corresponding value for $P(Y=0)$. Other distributions can share the same mean and variance while assigning different probabilities to zero and to other counts. Therefore the quasi-Poisson scale adjustment cannot, on its own, be used to calculate a full predictive distribution or a probability for a future count.

For a common scale multiplier, the quasi-likelihood score for regression coefficients is often multiplied by the same factor $1/\phi$; that factor does not change the score equation’s root. The scale nevertheless affects estimated uncertainty. This does not turn the fitted mean model into a fully specified likelihood model. In particular, likelihood-based quantities that require normalized response probabilities cannot be obtained merely by inserting a scale adjustment into a Poisson variance.