What Overdispersion Means in a Fitted GLM

What Overdispersion Is Relative To

A count response can vary more from case to case than a Poisson model allows, and a grouped binary response can vary more than a binomial model allows. That comparison is the starting point for overdispersion. It is not a claim that the observed responses are “too variable” in isolation: it is a claim about variation relative to the variance the chosen model assigns at its conditional mean.

Let \(x_i\) denote the predictors for observation \(i\), and let \(\mu_i\) be the conditional mean specified by a fitted mean model. The Poisson benchmark is \(V(\mu_i)=\mu_i\). For a grouped binomial count \(Y_i\) from \(m_i\) trials, with success probability \(p_i\) and mean \(\mu_i=m_i p_i\), the benchmark is \(V(\mu_i)=m_i p_i(1-p_i)\). The meaning of “extra” variation depends on both this benchmark and the conditional mean being used.

The mean and variance are separate parts of the model. A Poisson GLM, for example, specifies both a conditional mean and the Poisson variance at that mean. Overdispersion means the mean can still be correctly described while the conditional variance is greater than the Poisson benchmark. If the mean model is wrong, residuals can also look too spread out; that is a different problem, not overdispersion as defined relative to a correct conditional mean.

Two Reference Variances

For a Poisson response, if \(Y_i\mid x_i\sim\operatorname{Poisson}(\mu_i)\), then \(\operatorname{Var}(Y_i\mid x_i)=\mu_i\). For a grouped binomial response, if \(Y_i\mid x_i\sim\operatorname{Binomial}(m_i,p_i)\), then \(\operatorname{Var}(Y_i\mid x_i)=m_i p_i(1-p_i)\). These formulas are the benchmarks, not general laws that every count or success count must obey.

A dispersion multiplier is often useful as a compact description: \(\phi=1\) matches the benchmark, \(\phi>1\) indicates greater variance, and \(\phi<1\) indicates less. A single multiplier is an approximation if the amount of extra variation differs across observations or does not follow the reference variance’s pattern. Thus “overdispersed” should always be understood as relative to a stated mean and variance benchmark.

The binomial case has an important boundary. For one binary response \(Y_i\in\{0,1\}\), its conditional mean \(p_i\) determines its conditional variance exactly: \(\operatorname{Var}(Y_i\mid x_i)=p_i(1-p_i)\). A binary response cannot have a larger variance at the same conditional mean while remaining binary. Extra variation relative to a binomial benchmark becomes possible for grouped counts, or when the assumed conditional structure is inadequate.

Variation in Success Probabilities Can Inflate Binomial Variance

One mechanism for extra variation in grouped binomial data is that the success probability varies across groups, even when the groups share the same predictors in the fitted mean model. The following identity makes the distinction precise.

Proof. By iterated expectation, \(E[Y]=E[E[Y\mid P]]=E[mP]=m\bar p\). By the law of total variance, \[ \operatorname{Var}(Y)=E[\operatorname{Var}(Y\mid P)]+\operatorname{Var}(E[Y\mid P]). \] Given \(P\), the binomial variance is \(mP(1-P)\) and the conditional mean is \(mP\). Therefore \[ \operatorname{Var}(Y)=mE[P-P^2]+m^2\operatorname{Var}(P). \] Since \(E[P^2]=\bar p^2+\operatorname{Var}(P)\), this becomes \[ m\bar p-m\bar p^2-m\operatorname{Var}(P)+m^2\operatorname{Var}(P) =m\bar p(1-\bar p)+m(m-1)\operatorname{Var}(P). \] The second term is positive when \(m>1\) and \(\operatorname{Var}(P)>0\), proving the claim. \(\square\)

Why a Wrong Mean Can Look Like Extra Variation

Overdispersion is defined relative to a conditional mean model. If that model assigns the wrong mean to some predictor values, deviations around its fitted means include both ordinary response variation and systematic mean error. The following identity shows exactly how the two enter.

Proof. Write \(Y-a(x)=(Y-\mu(x))+(\mu(x)-a(x))\). Squaring and taking the conditional expectation gives \[ E[(Y-a(x))^2\mid X=x] =E[(Y-\mu(x))^2\mid X=x] +2(\mu(x)-a(x))E[Y-\mu(x)\mid X=x] +(\mu(x)-a(x))^2. \] The middle term is zero because \(E[Y-\mu(x)\mid X=x]=0\), and the first term is the conditional variance. This proves the identity. \(\square\)

The squared mean error is nonnegative. A large average squared residual around an inadequate mean function can therefore occur even if the conditional variance at each predictor value exactly matches the Poisson or binomial benchmark. This identity concerns deviations around a proposed mean value; it does not say that every residual-based diagnostic has this simple expectation after estimating model parameters.

Common Mistakes and Practical Meaning

A dispersion statement is meaningful only when its reference is explicit. In a fitted GLM, the intended comparison is usually between conditional variation and the binomial or Poisson variance evaluated at the fitted mean, under the assumption that this mean adequately describes the response given its predictors. Estimation uncertainty in the fitted coefficients is not itself the definition of overdispersion.

• Wrong reasoning: “The residuals are large, so the response is overdispersed.” Correction: Large residuals can reflect a wrong conditional mean, as the two-group example shows. Overdispersion concerns variation beyond the benchmark at an adequate conditional mean.
• Wrong reasoning: “A dispersion multiplier changes the fitted mean.” Correction: A multiplier \(\phi\) in \(\operatorname{Var}(Y_i\mid x_i)=\phi V(\mu_i)\) describes variance relative to the benchmark. It does not alter \(\mu_i\) by definition.
• Wrong reasoning: “Any binary response can have extra binomial variance at the same probability.” Correction: For a single binary outcome with conditional mean \(p\), the variance must be \(p(1-p)\). Extra binomial variation is relevant to grouped counts or to structures beyond a single Bernoulli model.

The distinction matters because a variance adjustment cannot repair a misspecified mean function. Conversely, changing the mean model is not automatically a solution when conditional variation truly exceeds its benchmark. A sound interpretation first states the mean being conditioned on, then names the binomial or Poisson reference variance, and only then describes variation relative to it.