Why the Exponential-Family Form Matters
In ratemaking, a distribution is often described first using familiar parameters such as a mean or a scale. Generalized linear models use another representation that makes those models easier to connect to a systematic predictor. The exponential-family form provides a common structure for many count and severity distributions, including the Poisson and Gamma examples developed here.
The main task is to recognize the parts of the form, not to assume that every distribution has one universal parameterization. A single distribution can be written in different ways, and the natural parameter depends on the chosen representation. In each case, the algebra must reproduce the original probability mass or density, including its normalizing constant.
Here \(y\) is an observed response, \(\theta\) is the natural parameter, \(b(\theta)\) is the cumulant function, \(\phi\) is a dispersion parameter, and \(a(\phi)\) is a specified positive function of dispersion. The term \(c(y,\phi)\) is the normalizing term; it may depend on the observed value and dispersion, but not on \(\theta\). For a discrete response, \(f\) denotes a probability mass function; for a continuous response, it denotes a density.
The representation is useful only when its pieces together give a properly normalized distribution: the probabilities must sum to one, or the density must integrate to one over its support. The usual form assumes the support of \(Y\) does not change with the parameters. Some distributions require a different or more general treatment, so one should not force this form onto a model without checking its support and normalization.
Reading the Components
The natural parameter is the parameter that multiplies the response \(y\) in the exponent. The cumulant function is the accompanying function of that parameter; together they provide the parameter-dependent part of the log-probability or log-density. The dispersion parameter controls the scale of that part through \(a(\phi)\), while \(c(y,\phi)\) supplies the remaining terms needed for the exact probability or density.
A useful way to identify the components is to take the logarithm of a known probability mass function or density, then collect terms involving \(y\theta\), terms involving only \(\theta\), and all remaining terms. This is safer than memorizing a formula without checking what parameterization it uses. In particular, a distribution's mean is not generally its natural parameter.
Worked Example: Writing a Poisson Count in Exponential-Family Form
Suppose a hypothetical insurance segment has claim count \(Y\sim\operatorname{Poisson}(\mu)\), where \(\mu=2.5\). Write the mass function in exponential-family form and calculate the probability of observing \(Y=3\). Identify every component.
Method. The Poisson mass function for a nonnegative integer \(y\) is \(P(Y=y)=e^{-\mu}\mu^y/y!\), with \(\mu>0\). Set \(\theta=\log\mu\), so \(\mu=e^\theta\). Substitution gives
Thus \(a(\phi)=1\), the Poisson dispersion is fixed at \(\phi=1\) in this standard model, \(b(\theta)=e^\theta\), and \(c(y,\phi)=-\log(y!)\). The natural parameter is \(\theta=\log(2.5)\approx0.916291\).
For \(y=3\), the parameter-dependent term is
The normalizing term is \(c(3,1)=-\log(3!)=-\log 6\approx-1.791759\). Therefore the log-probability is \(0.248872-1.791759=-1.542887\), and the probability is \(\exp(-1.542887)\approx0.213763\). Direct substitution checks the result: \(2.5^3e^{-2.5}/6\approx0.213763\). The two calculations agree.
Common mistake avoided: using \(\mu\) itself as the natural parameter. In this representation, the coefficient of \(y\) is \(\theta=\log\mu\), not \(\mu\).
Why the Cumulant Function Normalizes the Model
The role of \(b(\theta)\) is not just to make the expression look tidy. It ensures that the probabilities or density integrate to one. This can be seen directly for the Poisson model.
For \(y=0,1,2,\ldots\), write \(P(Y=y)=\exp\{y\theta-e^\theta-\log(y!)\}\), where \(\theta=\log\mu\).
The total probability is \(e^{-e^\theta}\sum_{y=0}^{\infty}e^{y\theta}/y!\).
Since \(\sum_{y=0}^{\infty}z^y/y!=e^z\), setting \(z=e^\theta\) gives \(e^{-e^\theta}e^{e^\theta}=1\).
This verifies the normalization and shows why \(b(\theta)=e^\theta\) appears: its contribution in the exponent is precisely what cancels the sum over all possible counts. More generally, in a valid exponential-family representation, the parameter-only term \(b(\theta)\) is tied to normalization; it is not an arbitrary adjustment.
Gamma Distribution in the Same Form
For a positive continuous response, one standard Gamma parameterization uses mean \(\mu>0\) and dispersion \(\phi>0\). In this parameterization, the shape is \(1/\phi\) and the scale is \(\phi\mu\). Its density, for \(y>0\), is
To match the standard form, choose \(\theta=-1/\mu\), so \(\theta<0\), and \(b(\theta)=-\log(-\theta)\). Since \(-\theta=1/\mu\), this is equivalent to \(b(\theta)=\log\mu\). Set \(a(\phi)=\phi\). The remaining term is
The restriction \(\theta<0\) is necessary because \(\theta=-1/\mu\) for a positive mean. The support is \(y>0\), and the density is a continuous density, not a probability of observing one exact value.
Worked Example: Checking a Gamma Density Representation
A hypothetical positive claim severity is modeled with a Gamma distribution having mean \(\mu=2{,}000\) dollars and dispersion \(\phi=0.25\). Calculate its density at \(y=1{,}000\) dollars using the exponential-family form. Report the density to seven significant figures and check it against the original density.
Method. For this parameterization, the shape is \(1/\phi=4\), and the scale is \(\phi\mu=0.25(2{,}000)=500\) dollars. The natural parameter is \(\theta=-1/\mu=-0.0005\), and \(b(\theta)=\log\mu=\log(2{,}000)\).
First calculate the parameter-dependent part of the log-density:
Next calculate the normalizing term. Since \(\Gamma(4)=3!=6\),
The log-density is therefore \(-32.403609838+24.476683812=-7.926926026\), or approximately \(-7.926926\). Exponentiating gives \(f(1{,}000)\approx0.0003608941\), reported to seven significant figures.
For a direct check, the original Gamma density is \(1{,}000^3e^{-1{,}000/500}/[6(500)^4]\). This is \(10^9e^{-2}/(6\cdot62{,}500{,}000{,}000)\approx0.0003608941\). The result agrees. Because the response is measured in dollars, this density is per dollar; it is not a probability.
Common mistake avoided: omitting the term \(-\frac{1}{\phi}\log\phi\) from \(c(y,\phi)\). Although it does not depend on \(y\) or \(\theta\), it is needed to reproduce the correct Gamma density.
Likelihoods and a Common Exam Trap
For independent observations, the likelihood is the product of their individual probability masses or densities. Taking logarithms turns that product into a sum. Each observation contributes its own response-dependent term and normalizing term; if the observations have different parameter values, the corresponding natural parameters and cumulant terms can differ as well.
Worked Example: Combining Two Poisson Contributions
In a hypothetical two-segment portfolio, independent claim counts have \(Y_1\sim\operatorname{Poisson}(1.5)\) and \(Y_2\sim\operatorname{Poisson}(2.5)\). The observed counts are \(y_1=0\) and \(y_2=3\). Write the joint likelihood in exponential-family form and calculate it.
Method. For the first observation, \(\theta_1=\log(1.5)\), \(b(\theta_1)=1.5\), and \(c(0,1)=-\log(0!)=0\). Its log-probability is \(0\theta_1-1.5-0=-1.5\).
For the second observation, \(\theta_2=\log(2.5)\), \(b(\theta_2)=2.5\), and \(c(3,1)=-\log 6\). Its log-probability is
Independence means the joint likelihood is the product of the two probabilities, so its log is the sum \(-1.5-1.542887=-3.042887\). Thus the joint likelihood is \(\exp(-3.042887)\approx0.047697\). Directly, the product is \(e^{-1.5}\cdot(2.5^3e^{-2.5}/6)\approx0.047697\). This check confirms that the two individual normalizing terms have both been included.
Common mistake avoided: treating the two segment means as if they were one common parameter. The observations are independent, but their means—and therefore their natural parameters—are different.
A frequent error is to report only the part of the log-density that contains the response and the natural parameter. For example, writing \(y\theta-b(\theta)\) for a Poisson log-probability omits \(-\log(y!)\). That omission may not affect some comparisons made with the same observed data, but it does change the actual probability and likelihood value. When asked for a full distribution or a numerical probability, retain every term.
Check Your Understanding
For each problem, identify the requested components or calculate the requested probability or density. Keep probability answers to four decimals.
- A hypothetical Poisson claim count has mean \(\mu=3\). For an observed count \(y=1\), identify \(\theta\), \(b(\theta)\), \(a(\phi)\), and \(c(y,\phi)\), then give the probability. (Answer: \(\theta=\log 3\approx1.098612\), \(b(\theta)=3\), \(a(\phi)=1\), \(c(1,1)=0\), and \(P(Y=1)=3e^{-3}\approx0.1494\).)
- A Poisson count has mean \(\mu=0.8\). For \(y=0\), write the log-probability using the exponential-family form and calculate the probability. (Answer: \(\log P(Y=0)=-0.8\), and \(P(Y=0)=e^{-0.8}\approx0.4493\).)
- A hypothetical severity has Gamma mean \(\mu=4\), dispersion \(\phi=0.5\), and observed value \(y=2\). Using shape \(1/\phi\) and scale \(\phi\mu\), calculate the density. (Answer: shape \(2\), scale \(2\), \(\theta=-0.25\), \(b(\theta)=\log4\), \(c(2,0.5)=\log2-2\log(0.5)=2.079442\), log-density \(-1.693147\), and density \(e^{-1.693147}\approx0.1839\).)
- For a Poisson model with mean \(\mu=2.4\) and observed count \(y=5\), identify the natural parameter, cumulant function value, and normalizing term. (Answer: \(\theta=\log(2.4)\approx0.875469\), \(b(\theta)=2.4\), and \(c(5,1)=-\log(5!)=-\log120\approx-4.787492\).)
- A positive response is modeled with the Gamma parameterization used here, with \(\mu=3\), \(\phi=1\), and \(y=3\). Calculate the density and state why it is not the probability of exactly \(Y=3\). (Answer: shape \(1\), scale \(3\), \(\theta=-1/3\), \(b(\theta)=\log3\), \(c(3,1)=0\), and density \(e^{-1-\log3}\approx0.1226\); it is a density value, not a point probability, because the response is continuous.)