Tweedie Exponential-Dispersion Model

From a Mean–Variance Rule to a Distributional Model

Many response models begin by describing how variability changes with the mean. The Tweedie family uses a power variance function: for a specified power $p$, the variance is proportional to the mean raised to $p$. That relationship is useful, but it is not the whole model. A Tweedie model is an exponential-dispersion model (EDM): it specifies a family of distributions through an exponential form, a canonical parameter, a cumulant function, and a dispersion parameter.

This distinction matters in applications. A mean and variance do not, in general, determine a unique probability distribution. The EDM structure supplies additional information: it links the parameterization to a cumulant-generating function, which encodes the distribution. The power variance function then describes how the variance changes as the mean changes within that family.

Definition of the Tweedie Exponential-Dispersion Model

Definition: An exponential-dispersion model for a response $Y$ has a distribution of the form $$ f(y;\theta,\phi) = a(y,\phi)\exp\left\{\frac{y\theta-\kappa(\theta)}{\phi}\right\}, \qquad \phi>0. $$ Here $f$ is a density or mass function with respect to a specified base measure, $a(y,\phi)$ is the carrier factor, $\theta$ is the natural (canonical) parameter, $\kappa$ is the cumulant function, and $\phi$ is the dispersion parameter. The natural parameter must lie in a domain where the distribution is normalizable. A Tweedie model is an EDM whose variance function has the power form $V_p(\mu)=\mu^p$, on parameter values for which the specified EDM exists.

The mean parameter is $\mu=E[Y]$. For a regular EDM, the natural parameter and mean satisfy $\mu=\kappa'(\theta)$, and the variance is $\operatorname{Var}(Y)=\phi\kappa''(\theta)$. Thus the variance function is defined by $$ V(\mu)=\kappa''(\theta), \qquad \mu=\kappa'(\theta), $$ so that $\operatorname{Var}(Y)=\phi V(\mu)$. In the Tweedie family, $V(\mu)=V_p(\mu)=\mu^p$. The index $p$ controls the way the variance scales with the mean; $\phi$ multiplies that variance at every mean.

This definition separates quantities that are sometimes conflated. The mean $\mu$ locates the response distribution, the power $p$ specifies the shape of the mean–variance relationship, and the dispersion $\phi$ scales variability. The carrier factor and natural-parameter domain are also part of the distributional specification; knowing only $p$, $\mu$, and $\phi$ as numerical labels does not replace the EDM form.

Worked Example: Calculating Variance From the Power Function

Suppose a Tweedie EDM has mean $\mu=4$, dispersion $\phi=0.25$, and power $p=1.5$. Calculate its variance.

The variance function is $V_p(\mu)=\mu^p$, so $$ V_{1.5}(4)=4^{1.5}=4^{3/2}=(\sqrt{4})^3=2^3=8. $$ The EDM variance is dispersion times the variance function. Therefore, $$ \operatorname{Var}(Y)=\phi V_p(\mu)=0.25(8)=2. $$ The variance function value is $8$, while the response variance is $2$. Confusing these quantities would omit the dispersion multiplier.

Why the Cumulant Function Encodes the Distribution

The cumulant function is not merely a device for calculating moments. Its exponential form yields the moment-generating function, whenever that function exists near zero. To see the relationship precisely, let $\nu$ denote the base measure and write the normalizing integral as $$ Z_\phi(\theta)=\int a(y,\phi)\exp(y\theta/\phi)\,\nu(dy). $$ For a normalizable EDM, $\exp\{\kappa(\theta)/\phi\}=Z_\phi(\theta)$. The integral is finite for $\theta$ in the natural-parameter domain. The following result gives the mean and variance identities used in the definition.

Theorem: Suppose the EDM is normalizable on an open interval containing $\theta$, and its normalizer can be differentiated twice under the integral sign there. Then $$ E[Y]=\kappa'(\theta), \qquad \operatorname{Var}(Y)=\phi\kappa''(\theta). $$ A sufficient regularity condition for the differentiations is that the exponential integral be finite on an open neighborhood of $\theta$; this ensures the required local exponential moments.

Differentiate the normalizer.
Under the stated condition, differentiation under the integral gives $Z_\phi'(\theta)=\int (y/\phi)a(y,\phi)\exp(y\theta/\phi)\,\nu(dy)$.

Recognize the expectation.
Dividing the derivative by $Z_\phi(\theta)$ yields $Z_\phi'(\theta)/Z_\phi(\theta)=E[Y]/\phi$.

Use the cumulant normalization.
Since $\kappa(\theta)=\phi\log Z_\phi(\theta)$, differentiation gives $\kappa'(\theta)=\phi Z_\phi'(\theta)/Z_\phi(\theta)=E[Y]$.

Differentiate once more.
The second derivative of the log normalizer is $Z_\phi''/Z_\phi-(Z_\phi'/Z_\phi)^2=\operatorname{Var}(Y)/\phi^2$. Hence $\kappa''(\theta)=\operatorname{Var}(Y)/\phi$, proving the variance identity.

The same structure gives the moment-generating function. For any $t$ such that $\theta+\phi t$ is in the natural-parameter domain, $$ E[\exp(tY)] = \exp\left\{\frac{\kappa(\theta+\phi t)-\kappa(\theta)}{\phi}\right\}. $$ Indeed, inserting $\exp(tY)$ into the expectation changes the factor $\exp(Y\theta/\phi)$ to $\exp\{Y(\theta+\phi t)/\phi\}$; the ratio of the two normalizers is the displayed expression. Where a moment-generating function exists in a neighborhood of zero, it determines the probability distribution. This is why specifying the EDM structure and its cumulant—not just the first two moments—matters.

Connecting the Mean, Natural Parameter, and Variance Function

The variance function expresses the second cumulant derivative in terms of the mean, rather than in terms of $\theta$. Since $\mu=\kappa'(\theta)$, differentiating with respect to $\theta$ gives $$ \frac{d\mu}{d\theta}=\kappa''(\theta)=V(\mu). $$ Where $V(\mu)>0$, this can be written as $d\theta/d\mu=1/V(\mu)$. This identity shows how a variance function determines the relationship between the mean and canonical parameter. For a Tweedie model, the specified function is $V_p(\mu)=\mu^p$; the exact cumulant and its parameter domain must be compatible with a normalizable EDM.

The following calculation uses one power value to illustrate the connection. It does not assert a general existence result for every formal power specification: the normalizability and parameter-domain requirements remain part of the model.

Worked Example: Recovering Canonical Quantities for a Specified Power

Take $V(\mu)=\mu^{3/2}$ on positive means. Choose constants of integration to be zero at the level of the following antiderivatives. Find $\theta$ as a function of $\mu$, and then find a compatible $\kappa$.

From $d\theta/d\mu=1/V(\mu)$, $$ \frac{d\theta}{d\mu}=\mu^{-3/2}, \qquad \theta(\mu)=\int\mu^{-3/2}\,d\mu=-2\mu^{-1/2}. $$ Since $d\kappa/d\theta=\mu$, the chain rule gives $d\kappa/d\mu=\mu(d\theta/d\mu)$. Thus $$ \frac{d\kappa}{d\mu}=\mu\mu^{-3/2}=\mu^{-1/2}, \qquad \kappa(\mu)=\int\mu^{-1/2}\,d\mu=2\mu^{1/2}. $$ Check the relationship directly. The derivatives with respect to $\mu$ are $d\theta/d\mu=\mu^{-3/2}$ and $d\kappa/d\mu=\mu^{-1/2}$, so $$ \frac{d\kappa}{d\theta} = \frac{d\kappa/d\mu}{d\theta/d\mu} = \frac{\mu^{-1/2}}{\mu^{-3/2}} = \mu. $$ Differentiating this identity with respect to $\theta$ gives $d\mu/d\theta=\mu^{3/2}$, as required. Additive integration constants can change the parameterization without changing these derivative identities.

This example illustrates how the power variance function constrains the canonical parameter and cumulant. But the calculation of these functions is not, by itself, a complete density calculation: the carrier measure and normalizing domain still determine which response values are represented and whether the exponential expression integrates to one.

Worked Example: Same Mean, Different Variance Scaling

Compare two Tweedie variance specifications at mean $\mu=2$ and dispersion $\phi=0.5$, using $p=1.5$ and $p=1.8$.

For $p=1.5$, $$ V_{1.5}(2)=2^{1.5}=2\sqrt{2}\approx 2.8284, \qquad \operatorname{Var}(Y)=0.5(2.8284)\approx 1.4142. $$ For $p=1.8$, $$ V_{1.8}(2)=2^{1.8}\approx 3.4822, \qquad \operatorname{Var}(Y)=0.5(3.4822)\approx 1.7411. $$ The two specifications share the same mean and dispersion, but their variances differ because the power changes the variance function. This comparison establishes a difference in second moments; it does not, by itself, calculate or compare the full densities.

What the Mean–Variance Relationship Specifies

Within an EDM, the variance function is part of a structured distributional model. Given a specified Tweedie power, the function $V_p(\mu)$ fixes how variance relates to the mean; through $d\theta/d\mu=1/V_p(\mu)$, it also constrains the canonical parameterization and cumulant. Together with the carrier measure, dispersion, and natural-parameter domain, these ingredients define the probability law. In this sense, a Tweedie mean–variance relationship specifies the response distribution as part of the EDM construction.

It would be incorrect to generalize this claim to arbitrary distributions. Infinitely many distributions can share a given mean and variance, and even a mean–variance rule across several means does not automatically identify a unique family without additional structure. The exponential-dispersion form is the additional structure that connects the variance function to a cumulant-generating function and the corresponding law.

In a generalized linear model, the response distribution is specified conditionally for each observation. Its mean may vary with predictors, while the Tweedie power controls how the conditional variance changes with that mean and the dispersion scales it. This makes the family useful when a constant-variance model is not appropriate, while keeping the mean–variance relationship explicit.

Common Pitfall

Wrong reasoning: “I know the mean, dispersion, and power, so I know the distribution just by calculating $\phi\mu^p$.” This calculation gives the variance, not the whole probability law.
Correction: Treat $\phi\mu^p$ as the EDM variance relation. A complete distributional specification also uses the exponential form, cumulant, carrier measure, and natural-parameter domain. Do not infer an exact density or response support from the variance calculation alone.
Another common error: Calling $V_p(\mu)$ the response variance. The variance is $\phi V_p(\mu)$; the two agree only when $\phi=1$.

Key takeaway: A Tweedie model is an exponential-dispersion distribution with power variance function $V_p(\mu)=\mu^p$. Its response variance is $\phi\mu^p$, while its canonical parameter and cumulant link the mean–variance rule to the full distributional model.

Check Your Understanding

Answer each question using the distinction between the variance function and the full EDM specification.

For mean $\mu=3$, dispersion $\phi=0.4$, and power $p=1.5$, calculate $V_p(\mu)$ and $\operatorname{Var}(Y)$.
State the roles of the natural parameter, cumulant function, and dispersion in an exponential-dispersion model.
Why does a specified mean and variance not generally determine a unique probability distribution?
In an EDM, what condition relates the derivative of the mean with respect to the natural parameter to the variance function?
Besides the power variance function, what components are needed to specify the response distribution through the EDM form?