Why The Exposure Coefficient Is Not Usually Estimated

Why Exposure Usually Enters With a Fixed Coefficient

A frequency model needs to distinguish the amount of time or risk observed from the underlying claim rate. If two otherwise identical risks are observed for different lengths of time, the longer observation is ordinarily expected to produce proportionally more claims. The usual log-link generalized linear model (GLM) builds that assumption directly into the model by including the logarithm of exposure as an offset, with its coefficient fixed at one.

This is a modeling assumption, not an algebraic necessity. A fitted exposure coefficient can be useful when the data suggest that expected counts do not scale proportionally. But it changes the interpretation: the model no longer predicts a constant expected claim count per unit of exposure. It also asks the data to estimate a relationship that is often fixed for sound exposure and process reasons.

Definition: Let $N_i$ be the claim count for observation $i$, $e_i>0$ its exposure measured in a consistent unit, and $x_i$ its rating variables. Under a Poisson frequency model with a log link and proportional exposure, $$ N_i\mid e_i,x_i\sim\operatorname{Poisson}(\mu_i),\qquad \log(\mu_i)=\log(e_i)+x_i^\mathsf{T}\beta. $$ Equivalently, $\mu_i=e_i\exp(x_i^\mathsf{T}\beta)$. The term $\log(e_i)$ is an offset: its coefficient is fixed at one, rather than estimated. Conditional on the model inputs, the Poisson assumption also specifies $\operatorname{Var}(N_i\mid e_i,x_i)=\mu_i$; observations are assumed conditionally independent for the likelihood calculations below.

The resulting expected count per unit of exposure is $\mu_i/e_i=\exp(x_i^\mathsf{T}\beta)$, which does not depend on $e_i$. The rating variables can change the expected frequency per exposure unit, while exposure scales the total expected count. Earlier material on offsets distinguishes this use of exposure from using exposure as a weight; here the key issue is the proportionality assumption itself.

What a Free Exposure Coefficient Means

A more general model replaces the fixed offset coefficient with an estimated parameter $\alpha$:

\log(\mu_i)=\alpha\log(e_i)+x_i^\mathsf{T}\beta, \qquad \mu_i=e_i^\alpha\exp(x_i^\mathsf{T}\beta).

In this model, multiplying exposure by a factor $c$ multiplies expected claim count by $c^\alpha$, holding the rating variables fixed. Thus, $\alpha=1$ gives proportional scaling; $\alpha<1$ gives less-than-proportional scaling; and $\alpha>1$ gives more-than-proportional scaling. If $\alpha\ne1$, the expected count divided by exposure is $\mu_i/e_i=e_i^{\alpha-1}\exp(x_i^\mathsf{T}\beta)$, so the implied frequency per exposure unit varies with exposure.

The coefficient is not itself a claim rate. It is an elasticity: the percentage change in expected count for a small percentage change in exposure, holding other model terms fixed. Specifically, differentiating $\log(\mu_i)$ with respect to $\log(e_i)$ gives $\alpha$. A doubling of exposure therefore changes the predicted mean by $2^\alpha$, not by $2\alpha$.

Why Fixing the Coefficient Is Often Appropriate

The fixed coefficient encodes the idea that exposure accumulates over time or another defined risk unit. Under a constant underlying claim intensity, with comparable coverage and risk characteristics throughout the observed period, expected claims accumulate in proportion to the amount of exposure. For example, an observation covering twice as much time at the same risk level has twice the expected count. This is the actuarially meaningful basis for expressing a frequency as expected claims per exposure unit.

Estimating $\alpha$ can detect departures from proportionality, but a departure in the fitted model does not automatically establish that the exposure mechanism itself is nonlinear. Exposure can be associated with omitted characteristics: risks with long observation periods may differ systematically from short-period risks, or the mix of policyholders may change with term length. A free coefficient can absorb such differences. Data covering only a narrow range of exposures may also provide little information to separate the exposure pattern from the intercept and other predictors.

The practical choice is therefore not “estimate every coefficient that can be estimated.” It is whether the data and application support a reason to relax the proportionality assumption. A free coefficient can be considered as a diagnostic or alternative specification, but the fitted relationship should be checked for confounding, stability, and sensible interpretation. The usual offset is not a claim that all real-world data fit perfectly; it is a principled model specification when exposure is intended to represent accumulated time at risk.

A Short Likelihood Derivation

The simplest derivation shows how the fixed-offset model estimates a common frequency when there are no rating variables. Suppose $N_i$ are independent Poisson counts with mean $\mu_i=e_i\lambda$, where $\lambda>0$ is the expected count per exposure unit. The log-likelihood, omitting terms that do not involve $\lambda$, is

\ell(\lambda)=\sum_i\{N_i\log(e_i\lambda)-e_i\lambda\}.

Differentiating and setting the score equal to zero gives

\frac{d\ell}{d\lambda} =\frac{\sum_iN_i}{\lambda}-\sum_ie_i=0 \quad\Longrightarrow\quad \widehat{\lambda}=\frac{\sum_iN_i}{\sum_ie_i}.

The estimate is total claims divided by total exposure, because the model says each exposure unit contributes the same expected count. With a free exposure coefficient, the mean is instead $\lambda e_i^\alpha$, and exposure contributes through a power that must be estimated. If all observations have the same exposure, $\alpha$ cannot be separated from the intercept using those observations alone: changing $\alpha$ can be offset by changing $\lambda$. Variation in exposure is needed to estimate the exposure pattern, and other systematic differences across exposure levels can still confound it.

Worked Calculations

Worked Example: Applying a Fixed Exposure Offset

A hypothetical insurer models annual claim counts for a specified coverage. Assume conditionally independent Poisson counts, a constant frequency within each stated rating cell, and proportional scaling with vehicle-years of exposure. The fitted annual frequency for the base territory and vehicle class is $0.08$ claims per vehicle-year. A territory relativity is $1.5$, and a class relativity is $0.8$. Find the mean count for a policy with $0.25$ vehicle-years and the mean count for a policy with $1.5$ vehicle-years, both in the stated territory and class.

Method. Use the log-link offset model, so the cell frequency per vehicle-year is the base frequency multiplied by the two relativities. First calculate that frequency:

0.08\times1.5\times0.8=0.096 \quad\text{claims per vehicle-year}.

For $0.25$ vehicle-years, the expected claim count is

\mu=0.25\times0.096=0.024.

For $1.5$ vehicle-years, it is

\mu=1.5\times0.096=0.144.

Check. The second policy has $1.5/0.25=6$ times the exposure, and $0.144/0.024=6$ times the mean count. The common mistake avoided is treating the relativity-adjusted annual frequency as the final mean count without multiplying by the policy’s exposure.

Worked Example: Estimating a Free Exposure Coefficient

Consider a hypothetical portfolio divided into two homogeneous cells. Their exposures are $e_1=1$ and $e_2=4$ exposure units, and their observed counts are $N_1=6$ and $N_2=12$. Assume the cell counts are independent Poisson variables with $\mu_j=\lambda e_j^\alpha$, where $\lambda>0$. Estimate $\lambda$ and $\alpha$, then compare the fit with the proportional model $\alpha=1$.

Method. There are two parameters and two observed cell counts. For any positive observed count, the Poisson likelihood for a cell is maximized when its fitted mean equals its observed count. The two cell means can be matched by solving $\lambda e_1^\alpha=6$ and $\lambda e_2^\alpha=12$. Since $e_1=1$, the first equation gives $\lambda=6$. Substitution into the second gives

6(4^\alpha)=12,\qquad 4^\alpha=2,\qquad \widehat{\alpha}=\frac{\log(2)}{\log(4)}=0.5.

Thus the estimated model is $\widehat{\mu}_j=6e_j^{0.5}$. It predicts $\widehat{\mu}_1=6(1^{0.5})=6$ and $\widehat{\mu}_2=6(4^{0.5})=6(2)=12$, exactly matching the observed counts.

Under the proportional model, $\mu_j=\lambda e_j$. The likelihood estimate of $\lambda$, from the total-count-over-total-exposure formula, is

\widehat{\lambda}=\frac{6+12}{1+4}=\frac{18}{5}=3.6.

The fitted means are $3.6(1)=3.6$ and $3.6(4)=14.4$. These are not the observed counts. The free model’s fitted mean doubles when exposure increases from 1 to 4, so its estimated coefficient is $0.5$, since $4^{0.5}=2$. Check. Both free-model means equal their respective counts, and the fixed-model means sum to $3.6+14.4=18$, the total observed count. The common mistake avoided is interpreting the fitted $\alpha=0.5$ as a frequency rate; it describes how the mean count scales with exposure. With only two cells and two fitted parameters, the exact fit is not strong evidence that the relationship will generalize.

Worked Example: Translating an Estimated Coefficient Into Rate Meaning

A hypothetical frequency analysis uses a free-exposure-coefficient Poisson model $\mu=0.10e^{0.7}$ for a fixed rating cell, with exposure measured in policy-years. Assume independent Poisson counts and that the coefficient $0.7$ is the fitted value for this illustration. Calculate the predicted mean counts at $e=0.5$ and $e=2$, the corresponding means per policy-year, and the effect of doubling exposure.

Method. At $e=0.5$, the predicted mean is

\mu=0.10(0.5^{0.7})\approx0.0616.

The mean per policy-year is $0.0616/0.5\approx0.1231$. At $e=2$, the predicted mean is

\mu=0.10(2^{0.7})\approx0.1625,

and the mean per policy-year is $0.1625/2\approx0.0812$. Doubling exposure from $0.5$ to $1$, or from $1$ to $2$, multiplies the predicted mean by

2^{0.7}\approx1.6245.

For example, at exposure $1$, the mean is $0.10$, and at exposure $2$, it is approximately $0.1625$, which is $1.6245$ times as large. Check. The per-policy-year frequency is higher at $e=0.5$ than at $e=2$, consistent with $\mu/e=0.10e^{-0.3}$. The common mistake avoided is multiplying the mean by $2(0.7)=1.4$ when exposure doubles; the correct multiplier is $2^{0.7}$.

Common Exam Trap and Model Checks

A free exposure coefficient does not simply make the usual rate estimate more flexible. It changes what “rate” means. With a fixed offset, expected claims per exposure unit are constant within a rating cell. With a free coefficient, that quantity varies with exposure unless $\alpha=1$. Consequently, a single rate indication per rating cell may not summarize predictions from the free-coefficient model without specifying the exposure level.

A frequent error is to see an estimated coefficient below one and conclude that doubling exposure doubles the frequency rate. The correct calculation uses the exponent: doubling exposure multiplies expected count by $2^\alpha$. Another error is to interpret $\alpha<1$ as proof that longer exposure causally reduces risk. It may instead reflect changing risk composition, selection, or a different observation process. The fitted coefficient describes the specified model and data; it does not by itself identify the cause of the pattern.

When considering a free coefficient, check that exposure is positive, its unit is defined consistently, there is meaningful variation in exposure, and exposure is not merely standing in for omitted rating differences. Compare the fitted scaling with the proportional assumption and assess whether the free model’s predictions remain sensible for the exposure range where they will be used. If exposure represents accumulated time under a stable risk process, the fixed offset is generally the natural specification. This topic is aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus.

Check Your Understanding

Work through each problem, paying attention to whether the requested quantity is a count, a rate, or an exposure-scaling factor.

A hypothetical Poisson model has a base frequency of $0.06$ claims per machine-year. A machine’s rating relativity is $1.25$, and its exposure is $0.8$ machine-years. Under proportional scaling, what is its expected claim count? (Answer: $0.06\times1.25\times0.8=0.060$ claims.)
In a Poisson model, $\mu=\lambda e^\alpha$. Exposure increases from 1 to 9 units, and the fitted coefficient is $\alpha=0.5$. By what factor does the predicted mean count change, holding all other predictors fixed? (Answer: $9^{0.5}/1^{0.5}=3$.)
A homogeneous cell has 15 claims across 6 exposure units. Under an independent Poisson model with fixed proportional exposure, calculate the maximum-likelihood estimate of the frequency per exposure unit and the fitted mean for 2 exposure units. (Answer: $\widehat{\lambda}=15/6=2.5$ claims per exposure unit; fitted mean $=2.5\times2=5$ claims.)
A free-coefficient model is $\mu=4e^{0.8}$. Calculate the expected count at $e=4$, and then calculate the multiplier in expected count if exposure doubles from 4 to 8. (Answer: $4(4^{0.8})\approx12.13$ claims; multiplier $2^{0.8}\approx1.741$.)
Two independent Poisson cells have exposures $e_1=1$ and $e_2=9$, with observed counts 5 and 15. In the model $\mu_j=\lambda e_j^\alpha$, choose $\lambda$ and $\alpha$ so that each fitted mean equals its observed count. What is the fitted exposure coefficient, and what mean does the free model predict at exposure 4? (Answer: $\lambda=5$, $\alpha=\log(3)/\log(9)=0.5$, and $\mu(4)=5\sqrt{4}=10$ claims.)

What You'll Learn