Model Claim Counts with an Exposure Offset

Why Claim Counts Need an Exposure Offset

A policy observed for a full year has more opportunity to generate claims than an otherwise similar policy observed for only part of a year. A model of raw claim counts that ignores this difference can mistake longer observation for higher risk. A claim-count GLM handles the difference by modeling a rate and incorporating each observation’s exposure through a fixed offset.

Here, exposure \(E_i\) is the amount of time at risk for observation \(i\), measured in policy-years or another consistent unit. The response \(N_i\) is the number of claims during that exposure. For example, a policy in force for six months contributes \(E_i=0.5\) policy-years if it was covered throughout that period.

Definition: A Poisson claim-count model with a log-exposure offset specifies \(N_i\mid \mathbf{x}_i,E_i\sim\operatorname{Poisson}(\mu_i)\), where \(\mu_i=E_i\lambda_i\) and \(\log(\lambda_i)=\beta_0+\mathbf{x}_i^\mathsf{T}\boldsymbol{\beta}\). Equivalently, \(\log(\mu_i)=\log(E_i)+\beta_0+\mathbf{x}_i^\mathsf{T}\boldsymbol{\beta}\). The term \(\log(E_i)\) is an offset: its coefficient is fixed at one, not estimated.

The model therefore has two related quantities. The rate \(\lambda_i\) is the expected number of claims per unit of exposure for the given predictors. The mean count \(\mu_i=E_i\lambda_i\) is the expected number of claims over the observation’s actual exposure. The distinction matters whenever policies have different exposure.

Model, Assumptions, and Interpretation

In a standard Poisson GLM, conditional on the predictors and exposure, the count \(N_i\) has a Poisson distribution with mean \(\mu_i\). The model assumes that the conditional mean has the stated log-linear form, exposures are positive and correctly measured, and each unit of exposure contributes proportionally to expected claims. It also assumes conditional independence across observations when the likelihood is formed. The Poisson distribution implies \(\operatorname{Var}(N_i\mid\mathbf{x}_i,E_i)=\mu_i\); this equality may not describe every portfolio.

The proportional-exposure assumption means, for example, that two otherwise identical risks observed for twice as long have twice the expected count. An offset is appropriate when that assumption is defensible and the exposure unit is consistently defined. It does not make zero exposure valid: \(\log(0)\) is undefined, and an observation with no time at risk contributes no claim-count information to this model. Such rows must be handled before fitting.

For a predictor whose coefficient is \(\beta_j\), a one-unit increase in that predictor multiplies the modeled claim rate by \(\exp(\beta_j)\), holding other predictors fixed. For a categorical predictor, the exponentiated coefficient is the rate ratio relative to its reference category. Neither coefficient is a change in expected count independent of exposure; predicted counts also depend on the exposure supplied for the prediction.

Interpretation rule: First calculate or describe the fitted rate \(\widehat{\lambda}_i=\exp(\widehat{\beta}_0+\mathbf{x}_i^\mathsf{T}\widehat{\boldsymbol{\beta}})\). Then obtain the expected count for exposure \(E_i\) as \(\widehat{\mu}_i=E_i\widehat{\lambda}_i\). State the exposure unit when reporting the rate.

Why the Offset Produces a Rate Estimate

The offset has a direct likelihood justification. Consider an intercept-only Poisson model with observations \(N_i\) and positive exposures \(E_i\). Its rate is the same, \(\lambda=\exp(\beta_0)\), for every observation, while the count mean is \(\mu_i=E_i\lambda\). The observations are assumed conditionally independent Poisson counts.

Ignoring terms that do not depend on \(\lambda\), the log-likelihood is \(\ell(\lambda)=\sum_i[N_i\log(E_i\lambda)-E_i\lambda]\). Differentiating with respect to \(\lambda\) and setting the derivative to zero gives the maximum-likelihood estimate:

\frac{\partial \ell}{\partial\lambda} =\frac{\sum_i N_i}{\lambda}-\sum_i E_i=0 \quad\Longrightarrow\quad \widehat{\lambda}=\frac{\sum_i N_i}{\sum_i E_i}.

Thus the fitted common rate is total claims divided by total exposure, not the unweighted average of the counts per row. In a model with predictors, the rate varies with those predictors, but the offset still scales each observation’s expected count by its exposure.

Fitting and Using the Model

In R, a typical fit is specified as glm(N ~ territory + vehicle_age + offset(log(exposure)), family = poisson(link = "log"), data = claims). The response is the claim count, and exposure must be positive. The offset is included in the linear predictor with its coefficient fixed at one. It is not an estimated predictor, and entering \(\log(E_i)\) as an ordinary predictor would instead estimate its coefficient from the data.

A disciplined interpretation follows four steps: verify the count and exposure units; identify the fitted rate for the covariate values; multiply that rate by the policy’s exposure to obtain expected claims; and report the prediction in context. If the model’s Poisson assumption is a poor description of conditional variation, coefficient estimates and uncertainty may require further assessment; the offset itself does not resolve that issue.

Define the modeled row.
Confirm that each response is a count over the exposure recorded on that same row.

Fit a log-rate model.
Use the log of positive exposure as an offset, so the model estimates predictor effects on the rate.

Interpret coefficients as rate effects.
Exponentiate a coefficient to obtain a multiplicative rate effect, with other predictors held fixed.

Predict at the required exposure.
Multiply the fitted rate by the policy’s exposure to obtain its expected count.

Worked Example: Fit Rates for Two Territory Groups

A hypothetical insurer summarizes a portfolio into two territory groups. Group A has 12 claims over 20 policy-years; Group B has 24 claims over 20 policy-years. Fit a Poisson model with territory as the only predictor, using A as the reference group. Find the rate ratio for B and the expected claim count for a Group B policy observed for 0.75 policy-years.

Method and assumptions. Treat the two group counts as independent Poisson observations, with means equal to exposure times their territory rates. A two-level factor makes this a saturated rate model: each group has its own fitted rate. The offset is the logarithm of exposure.

For Group A, the fitted rate is \(12/20=0.6\) claims per policy-year. For Group B, it is \(24/20=1.2\) claims per policy-year. With A as the reference, the intercept is \(\log(0.6)\approx-0.5108\), and the B coefficient is \(\log(1.2/0.6)=\log(2)\approx0.6931\). Exponentiating the B coefficient gives a rate ratio of \(2\): the fitted Group B rate is twice the Group A rate.

For the 0.75-policy-year Group B policy, the expected count is \(0.75\times1.2=0.9\) claims. The calculation is in claims, not claims per year, because exposure has been applied.

Check. The fitted group means recover the observed counts: \(20\times0.6=12\) for A and \(20\times1.2=24\) for B. This confirms the fitted rates and shows why the prediction for a shorter exposure is below the annual rate. The common mistake avoided is reporting 1.2 as the expected count for the 0.75-year policy; 1.2 is its annual rate, not its count prediction.

Worked Example: Interpret Coefficients and Predict for a Policy

Suppose a fitted Poisson GLM for a hypothetical portfolio has \(\widehat{\beta}_0=\log(0.2)\), a vehicle-age coefficient \(\widehat{\beta}_{\text{age}}=\log(1.1)\) per year, and an urban-territory coefficient \(\widehat{\beta}_{\text{urban}}=\log(1.5)\), where non-urban is the reference. Find the expected claim count for an urban policy whose vehicle is two years older than the age value represented by zero and whose exposure is 0.75 policy-years. Also find the model-implied probability of at least one claim.

Method and assumptions. Use the fitted Poisson mean \(E\widehat{\lambda}\), with the supplied coefficients treated as fixed fitted estimates. The stated count distribution is Poisson with that mean, conditional on the policy’s predictors and exposure.

The baseline annual rate is \(\exp(\widehat{\beta}_0)=0.2\). A two-year increase in vehicle age multiplies the rate by \(\exp(2\log(1.1))=(1.1)^2=1.21\). Urban territory multiplies it by \(\exp(\log(1.5))=1.5\). Therefore the fitted rate is \(0.2\times1.21\times1.5=0.363\) claims per policy-year.

Applying 0.75 policy-years gives an expected count of \(0.75\times0.363=0.27225\) claims. Under a Poisson distribution with mean \(\mu=0.27225\), the probability of zero claims is \(\exp(-0.27225)\), so the probability of at least one is \(1-\exp(-0.27225)\approx0.2383\), to four decimal places.

Check. The mean is positive and less than one, as expected for a moderate rate over three-quarters of a year. The at-least-one probability is below the mean here and is approximately 0.2383. The common mistake avoided is applying the two-year age effect just once: the rate multiplier is \(1.1^2\), not \(1.1\).

Worked Example: Diagnose a Model That Omits the Offset

Two hypothetical rating cells are observed: the first has 5 claims over 10 policy-years, and the second has 9 claims over 30 policy-years. Assume a common Poisson claim rate across the cells. Find the fitted common rate and the expected count for a new observation with 4 policy-years. Explain what an intercept-only Poisson model without an offset would estimate if each cell were treated as one equally weighted row.

Method and assumptions. The cell counts are independent Poisson variables, each with mean equal to its exposure times the common rate. Use the offset-model maximum-likelihood estimate derived above: total claims divided by total exposure.

Total claims are \(5+9=14\), and total exposure is \(10+30=40\) policy-years. Thus \(\widehat{\lambda}=14/40=0.35\) claims per policy-year. For 4 policy-years, the fitted expected count is \(4\times0.35=1.4\) claims.

Without the offset, the intercept-only model treats each row as one comparable count observation. Its Poisson maximum-likelihood estimate of the mean count per row is \((5+9)/2=7\) claims per row. That is not the annual claim rate: it does not account for the cells having different exposures. In particular, it cannot correctly use the 4-policy-year exposure to obtain the intended prediction of 1.4.

Check. The offset estimate satisfies \(40\times0.35=14\), matching the total observed claims. The no-offset estimate of 7 is a mean per row under a different model, not a rate per policy-year. The common mistake avoided is labeling the unadjusted row mean as an annual frequency.

Common Offset Traps

Confusing rate and count. A rate of 0.4 claims per policy-year implies 0.2 expected claims over 0.5 policy-years, not 0.4 expected claims.
Estimating the offset coefficient. The coefficient on \(\log(E_i)\) is fixed at one in an offset model. Treating it as an ordinary predictor changes the model and removes the required proportional scaling assumption.
Using the wrong exposure unit. If the rate is per policy-year, a six-month observation has exposure 0.5, not 6. Keep the count period and exposure unit aligned.
Taking the log of zero exposure. Zero exposure cannot be passed into \(\log(E_i)\). Check whether such a row should be excluded from this count model or corrected in the source data.
Assuming the offset fixes every model problem. It adjusts expected counts for exposure; it does not guarantee that predictors, the Poisson variance assumption, or other aspects of the model are adequate.

Key takeaway: A claim-count GLM with a log-exposure offset models the claim rate through its predictors and converts that rate into an expected count using the observation’s exposure. Exponentiated coefficients describe rate multipliers; policy-level expected counts require the additional exposure multiplication.

This modeling approach is aligned with the SOA Exam PA (Predictive Analytics) syllabus. In an applied analysis, communicate both the rate’s unit and the exposure used for a policy-level prediction, so a reader does not confuse an annualized rate with a count over a shorter period.

Check Your Understanding

For each problem, state the rate or count quantity requested and keep the exposure unit explicit.

An intercept-only Poisson model has 18 claims over 60 policy-years. Find the fitted annual claim rate and the expected count for 2 policy-years. (Answer: 0.3 claims per policy-year; 0.6 expected claims)
A fitted model has baseline rate 0.08 claims per policy-year and a category coefficient of \(\log(1.25)\). Find the category’s rate and expected count for 0.4 policy-years. (Answer: 0.1 claims per policy-year; 0.04 expected claims)
A policy’s fitted Poisson mean is 0.6 claims over its observed exposure. Find the probability of at least one claim, to four decimal places. (Answer: \(1-\exp(-0.6)=0.4512\))
A saturated territory model has 6 claims over 12 policy-years in the reference territory and 15 claims over 20 policy-years in another territory. Find both fitted rates, the second territory’s rate ratio, and its expected count over 1.2 policy-years. (Answer: 0.5 and 0.75 claims per policy-year; rate ratio 1.5; expected count 0.9)
Two cells have 8 claims over 16 policy-years and 12 claims over 24 policy-years. Under a common-rate Poisson model, find the fitted rate and expected count over 3 policy-years. What would the intercept-only model without an offset estimate per equally weighted cell row? (Answer: 0.5 claims per policy-year; 1.5 expected claims; 10 claims per row without the offset)

What You'll Learn