Why the Tweedie Power Matters
A Tweedie model connects a response’s mean to its variance through a power parameter. Changing that power changes more than the amount of variability: at particular powers, the model corresponds to familiar distributions, while between powers one and two it gives a useful model for aggregate losses that can be zero or positive and continuous. Recognizing these forms helps an actuary interpret a fitted model rather than treating the power as just another tuning value.
The distributional interpretation requires the Tweedie exponential dispersion model assumption. In that model, the conditional mean is \(\mu\), the dispersion is \(\phi\), and the variance function is \(\phi\mu^p\). The power alone does not specify the model: its mean and dispersion are also needed. Nor does a matching mean-variance relationship, by itself, prove that observed data follow the corresponding distribution.
The Poisson case has a special restriction: an ordinary Poisson random variable has variance equal to its mean, so its Tweedie dispersion is \(\phi=1\). A variance proportional to the mean with another multiplier may be used as a quasi-Poisson variance specification, but it is not an ordinary Poisson distribution with that arbitrary dispersion.
Reading the Standard Power Ranges
| Power | Distribution form | Key variance or outcome feature |
|---|---|---|
| \(p=0\) | Normal | \(\operatorname{Var}(X)=\phi\); support is the real line |
| \(p=1\) | Poisson, with \(\phi=1\) | \(\operatorname{Var}(X)=\mu\); nonnegative integer outcomes |
| \(1<p<2\) | Compound Poisson-Gamma | Aggregate of a Poisson number of Gamma amounts |
| \(p=2\) | Gamma | Positive continuous outcomes; \(\operatorname{Var}(X)=\phi\mu^2\) |
| \(p=3\) | Inverse Gaussian | Positive continuous outcomes; \(\operatorname{Var}(X)=\phi\mu^3\) |
These are the principal named cases relevant to insurance modeling. The standard Tweedie exponential dispersion family does not provide a distribution for every real power: powers strictly between zero and one are not in its usual existence range. Powers above two are not all inverse Gaussian; \(p=3\) is the inverse Gaussian case. For aggregate insurance loss, the range \(1<p<2\) is especially useful because the model represents a random count of positive amounts.
Why Powers Between One and Two Give Compound Poisson-Gamma Losses
Let \(N\) be a Poisson count with mean \(\lambda\), and let \(Z_1,Z_2,\ldots\) be independent Gamma amounts, independent of \(N\), each with shape \(\alpha\) and scale \(\theta\). Define aggregate loss \(X=\sum_{j=1}^{N}Z_j\), with \(X=0\) if \(N=0\). A Gamma distribution with shape \(\alpha\) and scale \(\theta\) has mean \(\alpha\theta\) and variance \(\alpha\theta^2\).
The variance expression follows by conditioning on \(N\). Given \(N=n\), the sum has mean \(nE[Z_1]\) and variance \(n\operatorname{Var}(Z_1)\). Applying total variance gives \(\operatorname{Var}(X)=\lambda\operatorname{Var}(Z_1)+\lambda(E[Z_1])^2=\lambda E[Z_1^2]\).
Matching these moments to \(E[X]=\mu\) and \(\operatorname{Var}(X)=\phi\mu^p\) gives the compound model parameters below. These formulas apply for \(1<p<2\), \(\mu>0\), and \(\phi>0\).
For a short check of the match, the mean is \(\lambda\alpha\theta=\mu\). Substituting the parameters into the variance gives \(\lambda\alpha(\alpha+1)\theta^2=\phi\mu^p\). In this range, \(N=0\) produces a zero aggregate, while one or more Gamma amounts produce a positive continuous aggregate.
If the model is a Tweedie model with \(1<p<2\), use the compound Poisson-Gamma representation.
Use \(\mu,\phi,p\) in the formulas for \(\lambda,\alpha,\theta\), keeping the loss units consistent.
Check that \(\lambda\alpha\theta=\mu\) and \(\lambda\alpha(\alpha+1)\theta^2=\phi\mu^p\).
Worked Example: Inferring a Gamma Power From Cell Moments
Method and assumptions. Suppose a hypothetical portfolio has fitted conditional means of 2 and 8 thousand dollars in two rating cells, with corresponding conditional variances of 1 and 16 squared-thousand dollars. Assume both cells follow the same Tweedie variance function with a common \(\phi\), and that outcomes are positive and continuous. Interest and mortality are not part of this annual loss model.
Step 1: Solve for the power. Under a common dispersion, the ratio of variances equals the ratio of means raised to \(p\): $$ \frac{16}{1}=\left(\frac{8}{2}\right)^p,\qquad 16=4^p,\qquad p=2. $$
Step 2: Find the dispersion. Using the first cell, $$ \phi=\frac{\operatorname{Var}(X)}{\mu^2} =\frac{1}{2^2}=0.25. $$ The second cell confirms this value: \(0.25(8^2)=16\).
Step 3: Give the distribution parameters for the second cell. At \(p=2\), the model is Gamma. In the shape-scale parameterization, shape \(=1/\phi=4\), and scale \(=\phi\mu=0.25(8)=2\) thousand dollars. Its mean is \(4(2)=8\) thousand dollars and its variance is \(4(2^2)=16\) squared-thousand dollars.
Check and common mistake avoided. Both cells have the same implied shape, 4, as required by the common-dispersion Gamma model. It would be a mistake to infer \(p\) from only one cell: one mean-variance pair cannot determine both the power and dispersion.
Worked Example: Calculating a Compound Poisson-Gamma Model
Method and assumptions. A hypothetical insurer models annual aggregate covered loss in thousands of dollars with a Tweedie model having \(\mu=4\), \(\phi=2\), and \(p=1.5\). Assume independent Gamma severities and a Poisson claim count, as in the compound representation. Interest and mortality do not enter this one-year aggregate loss calculation.
Step 1: Calculate the Poisson mean. $$ \lambda=\frac{4^{2-1.5}}{2(2-1.5)} =\frac{\sqrt{4}}{1}=2. $$
Step 2: Calculate the Gamma shape and scale. $$ \alpha=\frac{2-1.5}{1.5-1}=1,\qquad \theta=2(1.5-1)4^{1.5-1} =2(0.5)(2)=2. $$ Thus each claim amount is Gamma with shape 1 and scale 2 thousand dollars, with mean \(1(2)=2\) thousand dollars and variance \(1(2^2)=4\) squared-thousand dollars.
Step 3: Verify the aggregate mean and variance. $$ E[X]=\lambda\alpha\theta=2(1)(2)=4 $$ and $$ \operatorname{Var}(X)=\lambda\alpha(\alpha+1)\theta^2 =2(1)(2)(2^2)=16. $$ The Tweedie variance gives the same result: \(\phi\mu^p=2(4^{1.5})=2(8)=16\).
Step 4: Calculate the zero-claim aggregate probability. The aggregate is zero exactly when \(N=0\), so \(P(X=0)=e^{-\lambda}=e^{-2}=0.1353\), to four decimal places.
Check and common mistake avoided. The component mean and variance reproduce the specified Tweedie moments, and the zero probability is between zero and one. Do not use the Gamma severity variance alone as the aggregate variance: randomness in the Poisson count also contributes.
Gamma and Inverse Gaussian Parameters
The endpoint powers two and three have positive continuous outcome distributions, but their variance changes differently as the mean changes. This distinction can matter when comparing risks with different expected loss levels. For \(p=2\), the Gamma shape-scale parameters are \(1/\phi\) and \(\phi\mu\). For \(p=3\), use the inverse Gaussian parameterization with mean \(\mu\) and shape parameter \(\lambda_{\mathrm{IG}}\); its variance is \(\mu^3/\lambda_{\mathrm{IG}}\), so \(\lambda_{\mathrm{IG}}=1/\phi\).
Worked Example: Comparing Gamma and Inverse Gaussian Variances
Method and assumptions. Consider a hypothetical positive continuous claim severity with mean \(\mu=10\) thousand dollars and dispersion \(\phi=0.2\). Calculate the implications of a Gamma power \(p=2\) and an inverse Gaussian power \(p=3\). The annual severity model has no interest or mortality assumption.
Step 1: Calculate the Gamma case. $$ \operatorname{Var}(X)=0.2(10^2)=20 $$ in squared-thousand dollars. The shape is \(1/0.2=5\), and the scale is \(0.2(10)=2\) thousand dollars. Checking gives mean \(5(2)=10\) and variance \(5(2^2)=20\).
Step 2: Calculate the inverse Gaussian case. $$ \operatorname{Var}(X)=0.2(10^3)=200 $$ in squared-thousand dollars. The inverse Gaussian shape parameter is \(\lambda_{\mathrm{IG}}=\mu^3/\operatorname{Var}(X)=1000/200=5\). Checking its variance gives \(10^3/5=200\), as required.
Step 3: Compare. The mean is 10 thousand dollars in both cases, but the inverse Gaussian variance is ten times the Gamma variance: \(200/20=10\). The difference follows from the extra power of the mean in the variance function.
Check and common mistake avoided. Each model reproduces its specified mean and variance. Do not call every positive continuous Tweedie outcome Gamma: the power selects the named form, and \(p=3\) is inverse Gaussian, not Gamma.
Common Exam Traps
- Treating the power as the whole specification. The power identifies a variance-function form; \(\mu\) and \(\phi\) are still needed to calculate the variance and distribution parameters.
- Calling any \(p=1\) model Poisson. The ordinary Poisson case requires \(\phi=1\), giving variance \(\mu\). A free multiplier on \(\mu\) is not an ordinary Poisson variance.
- Using the wrong Gamma parameterization. With shape and scale, Gamma mean is shape times scale and variance is shape times scale squared. Do not treat scale as a rate.
- Confusing claim count with aggregate loss. In the compound model, \(N\) is the Poisson count; \(X\) is the sum of the severities. The variance of \(X\) includes variation in both count and amounts.
- Assuming all powers are valid distribution cases. In particular, powers strictly between zero and one are not standard Tweedie exponential dispersion distributions.
Check Your Understanding
For each problem, use the stated Tweedie assumptions and keep units consistent.
- A hypothetical loss response has Tweedie power \(p=0\), mean 6, and dispersion 3. Identify its distribution form and variance. (Answer: Normal; variance \(=3\).)
- A count response is specified as an ordinary Poisson distribution with mean 7. State its Tweedie power, dispersion, and variance. (Answer: \(p=1\), \(\phi=1\), variance \(=7\).)
- A compound Poisson-Gamma model has \(\mu=9\), \(\phi=1\), and \(p=1.5\). Find \(\lambda,\alpha,\theta\), and verify the aggregate mean. (Answer: \(\lambda=6,\ \alpha=1,\ \theta=1.5;\ E[X]=6(1)(1.5)=9\).)
- A hypothetical Gamma Tweedie model has \(\mu=5\), \(\phi=0.4\), and \(p=2\). Find its variance and its Gamma shape and scale. (Answer: variance \(=10\); shape \(=2.5\), scale \(=2\).)
- A hypothetical inverse Gaussian Tweedie model has \(\mu=4\), \(\phi=0.5\), and \(p=3\). Find its variance and inverse Gaussian shape parameter. (Answer: variance \(=32\); shape parameter \(\lambda_{\mathrm{IG}}=2\).)