Gamma-Poisson Models With Exposure

Why Exposure Must Be Part of the Count Model

A claim count is not interpretable without its exposure. Four claims observed over 10 policy-years and four claims observed over 1,000 policy-years provide very different evidence about a risk’s underlying frequency. In a Poisson model, exposure scales the expected count; it is not itself the claim rate.

This tutorial sets out the Gamma-Poisson model for one risk with a fixed but unknown claim rate. The central tasks are to specify the count likelihood with the correct exposure, state the prior parameterization, and combine observations without losing track of units.

Definition: Let \(\theta\) be a risk’s unknown claim rate per unit of exposure. For an observation with known exposure \(e\), let \(N\) be the number of claims. The exposure-adjusted Poisson model is \(N\mid\theta\sim\operatorname{Poisson}(e\theta)\), so \(E[N\mid\theta]=e\theta\). A Gamma prior in the shape-rate parameterization is \[ \theta\sim\operatorname{Gamma}(\alpha,\beta),\qquad \pi(\theta)=\frac{\beta^\alpha}{\Gamma(\alpha)}\theta^{\alpha-1}e^{-\beta\theta},\quad \theta>0, \] where \(\alpha>0\) is the shape and \(\beta>0\) is the rate parameter.

Under this convention, \(E[\theta]=\alpha/\beta\). The parameter \(\theta\) has units of claims per exposure unit, while \(\beta\) has units of exposure; the product \(\beta\theta\) is dimensionless. Some references use a Gamma shape-scale parameterization instead. Here, \(\beta\) is a rate, so it appears in the density as \(e^{-\beta\theta}\), not as a denominator in the exponent.

Model and Conditional Assumptions

For one risk observed in periods \(j=1,\ldots,m\), write \(N_j=n_j\) for the observed claim count and \(e_j\) for the known exposure in period \(j\). The model assumes a single underlying rate \(\theta\) applies throughout the observations. Conditional on that rate, the counts in disjoint periods are independent:

N_j\mid\theta\ \text{independent},\qquad N_j\mid\theta\sim\operatorname{Poisson}(e_j\theta),\quad j=1,\ldots,m.

The likelihood is therefore the product of the period-specific Poisson probabilities:

L(\theta\mid n_1,\ldots,n_m) =\prod_{j=1}^{m}\frac{e^{-e_j\theta}(e_j\theta)^{n_j}}{n_j!} \ \propto\ \theta^{\sum_j n_j}\exp\left(-\theta\sum_j e_j\right).

The proportionality omits factors that depend on the observed counts and exposures but not on \(\theta\). In particular, those factors are part of the likelihood’s numerical value, but they do not affect the posterior distribution’s shape or rate.

The model is appropriate when the exposure amounts are known and the claim rate is constant for the risk over the modeled periods. The conditional Poisson assumption represents counts with conditional mean and variance both equal to \(e_j\theta\); conditional independence means that, once \(\theta\) is specified, one period’s count provides no additional information about another period’s count. A changing underlying rate, dependence between periods not explained by a shared \(\theta\), or exposure determined by the claim outcomes may require a different model. The Gamma prior describes uncertainty or heterogeneity in \(\theta\) across risks before observing this risk’s experience; it does not make the period counts independent without conditioning on \(\theta\).

Deriving the Gamma Update

The Gamma prior and the Poisson likelihood have matching powers and exponential terms. Multiplying them gives a posterior kernel of the same Gamma form.

\begin{aligned} \pi(\theta\mid \mathbf{n}) &\propto \theta^{\alpha-1}e^{-\beta\theta} \cdot \theta^{\sum_j n_j}e^{-\theta\sum_j e_j}\\ &= \theta^{\alpha+\sum_j n_j-1} e^{-\theta(\beta+\sum_j e_j)}. \end{aligned}

A Gamma density with shape \(\alpha+\sum_j n_j\) and rate \(\beta+\sum_j e_j\) has exactly this kernel. Normalizing it to integrate to one gives the posterior distribution:

Theorem: Under the conditional Poisson model with conditionally independent counts and the Gamma shape-rate prior above, \[ \theta\mid \mathbf{n}\sim \operatorname{Gamma}\left(\alpha+\sum_{j=1}^{m}n_j,\ \beta+\sum_{j=1}^{m}e_j\right). \] The observed total count adds to the prior shape, and the total exposure adds to the prior rate parameter.

This result follows because the posterior density is proportional to the displayed kernel, and the Gamma density with the stated parameters is the normalized density having that kernel. It is essential that the exposure and the prior parameter use compatible units. If exposure is measured in hundreds of vehicle-years, for example, the rate \(\theta\) must be defined per hundred vehicle-years.

Worked Applications

Worked Example: Combining Claim Counts Across Periods

A hypothetical commercial account is observed over three non-overlapping periods. It has 40, 25, and 35 vehicle-years of exposure, with respective claim counts 8, 3, and 7. Assume the conditional Poisson model with one constant rate \(\theta\) per vehicle-year, and a prior \(\theta\sim\operatorname{Gamma}(6,30)\) in shape-rate form. Find the likelihood kernel and posterior distribution.

Method and assumptions. The period exposures are known, the same \(\theta\) applies in all three periods, and counts are independent conditional on \(\theta\). The prior has shape \(\alpha=6\) and rate \(\beta=30\) vehicle-years.

First add the counts: \(n_1+n_2+n_3=8+3+7=18\). Then add exposure: \(e_1+e_2+e_3=40+25+35=100\) vehicle-years. The likelihood is

L(\theta)\propto \theta^{8+3+7}e^{-(40+25+35)\theta} =\theta^{18}e^{-100\theta}.

Multiply this kernel by the prior kernel, \(\theta^{6-1}e^{-30\theta}\):

\theta^{5}e^{-30\theta}\cdot\theta^{18}e^{-100\theta} =\theta^{23}e^{-130\theta}.

The exponent on \(\theta\) is \(23\), so the posterior shape is \(23+1=24\); the posterior rate is \(30+100=130\). Thus \(\theta\mid\mathbf n\sim\operatorname{Gamma}(24,130)\), with \(\theta\) still measured in claims per vehicle-year. The posterior rate parameter’s units are vehicle-years.

Check and common mistake. The data contribute 18 to shape and 100 to rate, matching the total observed count and exposure. A common error is to use 18 as the Poisson mean. The expected total count conditional on \(\theta\) is \(100\theta\), not \(\theta\) alone; 18 is the observed count.

Worked Example: Writing the Full Likelihood Before Updating

For a hypothetical insured risk, the observed exposures in three periods are 12, 18, and 30 policy-years, and the corresponding claim counts are 2, 5, and 4. Let \(\theta\) be the constant claim rate per policy-year and assume conditional independence given \(\theta\). The prior is \(\operatorname{Gamma}(3,20)\), with rate parameter in policy-years. Write the full likelihood, then identify the posterior.

Method and assumptions. Each period count has Poisson mean \(e_j\theta\). Retain the factorial and exposure factors when writing the full likelihood, then separate the terms involving \(\theta\).

For the first period, the probability is \(e^{-12\theta}(12\theta)^2/2!\). The next two factors are \(e^{-18\theta}(18\theta)^5/5!\) and \(e^{-30\theta}(30\theta)^4/4!\). Therefore,

L(\theta)= \frac{e^{-12\theta}(12\theta)^2}{2!} \frac{e^{-18\theta}(18\theta)^5}{5!} \frac{e^{-30\theta}(30\theta)^4}{4!} = \frac{12^2\,18^5\,30^4}{2!\,5!\,4!}\, \theta^{11}e^{-60\theta}.

The count total is \(2+5+4=11\); the exposure total is \(12+18+30=60\) policy-years. The likelihood kernel is \(\theta^{11}e^{-60\theta}\). Multiplying by the prior kernel \(\theta^{3-1}e^{-20\theta}\) gives \(\theta^{13}e^{-80\theta}\). The posterior is therefore \(\operatorname{Gamma}(14,80)\).

Check and common mistake. The three exponential terms combine to \(e^{-60\theta}\), not \(e^{-11\theta}\): the exponent uses exposure, while the power of \(\theta\) uses claim count. Omitting the factors \(12^2,18^5,30^4\) would not change the posterior, but they belong in the full likelihood.

Worked Example: Converting Exposure Units Consistently

A hypothetical fleet has 20 car-years of exposure in one period with 2 claims and 50 car-years in another period with 12 claims. A prior analyst specifies \(\theta\) as claims per 100 car-years and gives \(\theta\sim\operatorname{Gamma}(4,16)\), where the rate parameter is in units of 100 car-years. Find the posterior distribution.

Method and assumptions. The two counts are independent conditional on the same rate. Because \(\theta\) is defined per 100 car-years, convert both exposures to hundreds of car-years before using them in the Poisson means.

The first exposure is \(20/100=0.2\) units and the second is \(50/100=0.5\) units. Their conditional Poisson means are \(0.2\theta\) and \(0.5\theta\). The total exposure in the model’s units is \(0.2+0.5=0.7\), and the total claim count is \(2+12=14\). Thus,

L(\theta)\propto\theta^{14}e^{-0.7\theta}.

Multiplying by the prior kernel \(\theta^{4-1}e^{-16\theta}\) gives \(\theta^{17}e^{-16.7\theta}\). Hence the posterior is \(\operatorname{Gamma}(18,16.7)\), with \(\theta\) measured in claims per 100 car-years and the posterior rate parameter in hundreds of car-years.

Check and common mistake. The data add \(0.7\), not 70, to the prior rate parameter because the prior rate and \(\theta\) use hundred-car-year units. Using 70 without also redefining \(\theta\) would mix units and produce an invalid update.

Common Traps and Scope of the Result

A reliable setup keeps three quantities distinct: the unknown rate \(\theta\), the exposure \(e\), and the expected count \(e\theta\). The observed count \(n\) is data; it is not the rate and is not generally equal to the expected count.

Using the wrong Gamma convention. Under the shape-rate form used here, the density contains \(e^{-\beta\theta}\) and the prior mean is \(\alpha/\beta\). If a source instead gives a scale parameter, convert it before applying this update.
Adding the wrong quantity to the shape. Add the number of claims, \(\sum_j n_j\), not the exposure or the observed rate \(n/e\).
Adding the wrong quantity to the rate. Add total exposure measured in units compatible with \(\theta\), not the number of periods.
Assuming a shared rate without justification. If the risk’s rate changes materially across periods, a single \(\theta\) does not describe the stated process. The displayed update then does not apply as written.
Ignoring conditional dependence. The product likelihood requires conditional independence given \(\theta\). Dependence that remains after conditioning can make that product inappropriate.

The Gamma-Poisson pairing is useful because its posterior stays within the Gamma family and can be calculated from total counts and total exposure. The formula is specific to the stated likelihood and prior: it does not by itself establish that the Poisson model is a good description of a particular portfolio or that the chosen prior is appropriate.

Key takeaway: With a Gamma shape-rate prior and conditionally independent Poisson counts sharing one rate, posterior shape equals prior shape plus total claims, and posterior rate equals prior rate plus total exposure. Check the rate and exposure units before combining them.

Check Your Understanding

For each problem, state the assumptions needed for the Gamma-Poisson update and track the exposure units.

A risk has exposure 15 policy-years and 3 claims. Its prior is \(\operatorname{Gamma}(2,10)\), with rate parameter in policy-years. What is the posterior distribution? (Answer: \(\operatorname{Gamma}(5,25)\).)
Over two conditionally independent periods for one risk, exposures are 10 and 20 vehicle-years and counts are 1 and 2. The prior is \(\operatorname{Gamma}(3,24)\) with rate parameter in vehicle-years. What is the posterior distribution? (Answer: \(\operatorname{Gamma}(6,54)\).)
A hypothetical account has exposures of 10, 20, and 30 property-years, with counts 1, 2, and 4. With a prior \(\operatorname{Gamma}(3,24)\) in shape-rate form, find the posterior distribution for a constant rate per property-year. (Answer: \(\operatorname{Gamma}(10,84)\).)
A risk has 40 car-years of exposure and 11 claims. Exposure is recorded in hundreds of car-years, the rate is claims per 100 car-years, and the prior is \(\operatorname{Gamma}(2,12)\) with rate parameter in hundreds of car-years. Find the posterior distribution. (Answer: \(\operatorname{Gamma}(13,12.4)\).)
For a single risk with one constant rate, counts are 2 and 9 over exposures of 5 and 25 years. State the likelihood kernel in \(\theta\) and the posterior if the prior is \(\operatorname{Gamma}(4,10)\) in shape-rate form. (Answer: Likelihood kernel \(\theta^{11}e^{-30\theta}\); posterior \(\operatorname{Gamma}(15,40)\).)

What You'll Learn