Is the dispersion parameter always estimated?

No. It is fixed in standard Poisson and Bernoulli models.

Does the framework itself specify predictors?

No. A regression structure, such as a GLM, adds them.

AI Glossary / Statistics and probability / Advanced

Exponential dispersion model

Definition

An exponential dispersion model is a family of probability distributions whose density or probability mass function separates a natural parameter from a dispersion parameter in a particular exponential form. In a common reproductive parameterization, the mean is μ and the variance is φV(μ), where φ controls dispersion and V is a family-specific variance function. These models provide a common mathematical framework for many generalized linear models.

What it means

Different outcomes have different relationships between their mean and variability. A normal model can have constant variance as the mean changes. A Poisson count has variance equal to its mean. A gamma model can have variance proportional to the square of its mean.
The exponential dispersion framework organizes such patterns. “Exponential” refers to the mathematical representation, not necessarily exponential growth or the exponential lifetime distribution. “Dispersion” does not automatically mean the variance itself.

How it works

A common form is f(y;θ,φ) = exp{[yθ − b(θ)]/φ + c(y,φ)}. Here θ is the natural parameter, b is the cumulant function, and c provides the remaining normalization and support information. The allowed parameter values must define an actual probability distribution.
Under the usual differentiability conditions, E[Y] = b′(θ) and Var(Y) = φb″(θ). Re-expressing the second derivative in terms of the mean gives the variance function V(μ). Consequently the mean and dispersion determine the variance within the chosen family.
For a normal model, θ = μ, b(θ) = θ²/2, V(μ) = 1, and φ is the variance. For an ordinary Poisson count, θ = log μ, b(θ) = exp θ, and φ = 1, so the variance is μ. A gamma model has V(μ) = μ², with φ equal to the reciprocal of its shape parameter in this convention.
Do not assume every family permits an arbitrary positive dispersion while retaining the same outcome support. The usual Bernoulli and Poisson models have fixed dispersion. A binomial proportion with m trials has dispersion 1/m and a support that depends on m. Parameterization and admissible values matter.
A generalized linear model adds a relationship between the conditional mean and predictors, usually g(μ_i) = x_iᵀβ. The distribution family and link function are separate choices. A log link does not, by itself, imply a gamma or Poisson distribution.
Estimate the regression and dispersion components as appropriate, then examine residuals, the mean - variance relationship, and whether the outcome support matches the data. Quasi-likelihood methods can specify a mean - variance relationship without specifying a full probability distribution; that is related to, but distinct from, a full exponential dispersion model.
The framework does not automatically handle clustering, temporal dependence, structural zeros, or contamination. Such features may require additional model structure and corresponding uncertainty calculations.

Three practical examples

An analyst models positive, right-skewed processing durations with a gamma response distribution and a suitable mean relationship.
A team models independent event counts with a Poisson model when its conditional variance assumptions are reasonable.
A researcher compares candidate distribution families by examining how observed variability changes with the fitted mean.

Think of it this way

Think of a family of containers sharing the same organizing design but holding different kinds of outcomes. The common design relates center and spread; the chosen container still determines what values are possible.

Worked example

Suppose an invented gamma model has mean μ = 40 and dispersion φ = 0.25. Because V(μ) = μ², its variance is 0.25 × 40² = 400 and its standard deviation is 20.
The corresponding gamma shape is 1/φ = 4 and scale is φμ = 10. Checking the familiar gamma formulas gives mean 4 × 10 = 40 and variance 4 × 10² = 400.
If the mean rises to 80 while φ stays 0.25, variance rises to 1,600 and standard deviation to 40. The coefficient of variation remains 0.5.
For four independent observations from the original model, the sample average has variance 400/4 = 100. The gamma average has the same mean and dispersion 0.25/4 = 0.0625. Independence and a common distribution are essential to this averaging calculation.

Illustrative case study

Fictional case: Brook Processing models task duration using a normal distribution with constant variance. Residual plots show much larger spread for longer expected tasks. Analysts consider a gamma mean - variance relationship, verify that durations are positive, and compare predictive checks. They investigate a separate group of interrupted tasks rather than forcing those records into the same model. This invented case shows why the family must fit both the support and variability.

Common misconceptions and mistakes

Do not confuse the dispersion parameter with standard deviation or assume it always equals variance. In the gamma example, variance also depends on the mean.
Do not conclude that choosing a flexible mean model fixes a wrong response distribution. Nor does a plausible variance function prove that the full distribution is correct.
A family name alone is insufficient documentation: state whether a variable is a count, proportion, average, or other quantity and identify the parameterization.

Business use

Exponential dispersion models give analysts a coherent way to match models to counts, proportions, and positive continuous outcomes. Business users benefit when uncertainty changes realistically with expected volume or duration. The practical goal is reliable predictions and intervals, not mathematical complexity for its own sake. Model checks and clear outcome definitions remain essential.

Check your understanding

Why does doubling the gamma mean quadruple variance here?

Answer: Its variance function is μ² while dispersion stays fixed.

What is φ in the example?

Answer: 0.25, the reciprocal of gamma shape.

Why does averaging four observations reduce variance?

Answer: Independent variances add, and dividing their sum by four scales variance by sixteen.

Related terms

Learn next

Learn exponential-family parameterizations and derivatives of cumulant functions. Then study how GLM link functions and response distributions play different roles.

Exponential family

Sources

Eugene Katsevich, STAT 9610, “Exponential dispersion model distributions”: https://katsevich-teaching.github.io/stat-9610-notes/exponential-dispersion-models.html Jørgensen, “Exponential Dispersion Models” (1987): https://academic.oup.com/jrsssb/article/49/2/127/6905366

Previous: Exponential dispersion family Next: Exponential family