Offset Versus Exposure Covariate

Why the Exposure Specification Matters

A count model for claims must account for how long, or how much, each observation was exposed to the possibility of a claim. A policy observed for four exposure units generally has more opportunity to generate claims than an otherwise similar policy observed for one unit. A log-link model can represent this opportunity in two different ways: use log exposure as a fixed offset, or include log exposure as a covariate whose coefficient is estimated.

These specifications make different assumptions. An offset says that, after controlling for the other rating variables, expected claim count is exactly proportional to exposure. Estimating a coefficient for log exposure allows the data to indicate whether that proportionality is supported. Neither choice is merely a software detail: it changes the model’s mean structure, rate interpretation, and predictions.

Definition: Let $N_i$ be the claim count for observation $i$, $e_i>0$ its exposure, and $x_i$ its other rating variables. Under a log link with an exposure offset, $$ \log E[N_i\mid x_i,e_i]=\log(e_i)+x_i^\mathsf{T}\beta. $$ The coefficient of $\log(e_i)$ is fixed at 1. With log exposure as an estimated covariate, the model is $$ \log E[N_i\mid x_i,e_i]=x_i^\mathsf{T}\beta+\gamma\log(e_i), $$ where $\gamma$ is estimated from the data. The offset specification is the special case $\gamma=1$.

What an Offset Assumes

Write $r_i=\exp(x_i^\mathsf{T}\beta)$. In the offset model, $$ E[N_i\mid x_i,e_i]=e_i r_i. $$ Thus $r_i$ is the expected count per exposure unit, conditional on the modeled rating variables. Doubling exposure while holding those variables fixed doubles expected count. The corresponding expected rate, $E[N_i\mid x_i,e_i]/e_i$, does not change with exposure.

This structure is appropriate when exposure measures the amount of time or risk at which claims can occur and the underlying claim rate is assumed constant over that exposure, after accounting for the model’s other predictors. The offset coefficient is not estimated because the proportionality is imposed as part of the model specification. That choice does not require every policy to have the same rate: the rate may vary with $x_i$. It requires the rate for a given $x_i$ not to vary systematically with $e_i$.

A common count-model assumption is that, conditional on the predictors, observations are independent and $N_i$ has a Poisson distribution with mean $e_i r_i$. The offset mean structure itself is not limited to a Poisson distribution, but the distribution and variance assumptions must match the model being fitted. An offset cannot repair a misspecified distribution or unmodeled dependence.

Interpretation: In the estimated-covariate model, if $a_i=\exp(x_i^\mathsf{T}\beta)$, then $$ E[N_i\mid x_i,e_i]=a_i e_i^\gamma,\qquad \frac{E[N_i\mid x_i,e_i]}{e_i}=a_i e_i^{\gamma-1}. $$ Here $\gamma$ is the elasticity of expected count with respect to exposure: multiplying exposure by $c$ multiplies expected count by $c^\gamma$, holding the other predictors fixed. The expected rate is constant in exposure only when $\gamma=1$.

Why the Offset Coefficient Is One

The fixed value of one follows directly from the meaning of a rate. Suppose that, conditional on the rating variables, the expected number of claims per exposure unit is $r_i$. For $e_i$ exposure units, expected count is $e_i r_i$. Taking logarithms gives the offset form:

\begin{aligned} E[N_i\mid x_i,e_i]&=e_i r_i,\\ \log E[N_i\mid x_i,e_i]&=\log(e_i)+\log(r_i). \end{aligned}

If the log rate is modeled as $x_i^\mathsf{T}\beta$, then $\log(r_i)=x_i^\mathsf{T}\beta$. Consequently, $\log(e_i)$ enters with coefficient exactly one. This is not an empirical estimate of how exposure relates to count; it is the algebraic consequence of assuming a constant rate per exposure unit.

By contrast, if the model estimates $\gamma$, it assumes expected count follows a power relationship in exposure. The estimated model can reveal departures from proportionality, but it also changes the rate: $a_i e_i^{\gamma-1}$. A coefficient below one implies a decreasing expected rate as exposure grows; a coefficient above one implies an increasing expected rate. Such patterns may reflect genuine variation, but they may also indicate omitted variables, exposure measurement issues, or a mismatch between the model and the data.

Worked Example: Comparing Count and Rate Predictions

A hypothetical portfolio model has a fitted base expected rate of $0.12$ claims per exposure unit. A policy has a rating relativity of $1.4$ and exposure $e=2.5$. Compare its predicted count under a fixed offset with a model that uses the same non-exposure rate component but estimates $\gamma=0.8$. For the second model, $a=0.12(1.4)$.

Method. First calculate the non-exposure rate component: $$ a=0.12(1.4)=0.168. $$

Fixed offset. The expected count is exposure times the rate: $$ E[N]=2.5(0.168)=0.420. $$ The implied expected rate is $0.420/2.5=0.168$ claims per exposure unit.

Estimated exposure coefficient. Here $E[N]=a e^\gamma$. Calculate the exposure power: $$ 2.5^{0.8}=\exp(0.8\log 2.5) =\exp(0.7330326)\approx 2.0814. $$ Therefore, $$ E[N]=0.168(2.5^{0.8})\approx 0.168(2.081383)=0.349672. $$ The implied rate is $0.349672/2.5\approx 0.139869$ claims per exposure unit.

The offset predicts $0.420$ claims and a constant rate of $0.168$. The estimated-coefficient model predicts approximately $0.349672$ claims and a lower rate at this exposure because $\gamma-1=-0.2$. The comparison holds the other model component fixed so that the effect of the exposure specification is isolated. Counts and rates are rounded to six decimals at the final step.

Estimating and Interpreting the Exposure Effect

When log exposure is a covariate, it must be included as $\log(e_i)$, not as the raw exposure $e_i$, if the intended model is a power relationship $e_i^\gamma$. The coefficient $\gamma$ is estimated alongside the other regression coefficients. Its uncertainty should be considered, particularly if conclusions depend on whether it differs materially from one.

A formal comparison can test $H_0:\gamma=1$. For example, a Wald statistic is $$ Z=\frac{\widehat{\gamma}-1}{\operatorname{SE}(\widehat{\gamma})}. $$ Under the usual large-sample approximation and the null hypothesis, this statistic is approximately standard normal. A test result is evidence about the data under the model; it does not prove the exposure assumption true or false in every setting. A confidence interval for $\gamma$ also communicates the range of exposure effects compatible with the fitted model.

Worked Example: Testing Proportionality and Reading a Prediction

Suppose a hypothetical log-link count model estimates $\widehat{\gamma}=0.50$ with standard error $0.20$. For a policy with the other predictor contribution fixed at $\exp(x^\mathsf{T}\widehat{\beta})=0.06$, compare the offset and estimated-covariate predictions at exposure $e=4$. Then test whether the fitted exposure coefficient differs from one using a two-sided Wald test.

Offset prediction. The mean is $e(0.06)$, so at $e=4$, $$ E[N\mid\text{offset}]=4(0.06)=0.24. $$ The expected rate is $0.24/4=0.06$ per exposure unit.

Estimated-covariate prediction. The mean is $0.06e^{0.50}$, where the exposure is raised to the coefficient, not multiplied by the coefficient: $$ E[N\mid\text{covariate}]=0.06(4^{0.50}) =0.06(2)=0.12. $$ Its expected rate is $0.12/4=0.03$ per exposure unit.

Wald test. Under $H_0:\gamma=1$, $$ Z=\frac{0.50-1}{0.20}=-2.50. $$ The two-sided standard-normal p-value is approximately $0.0124$. At the 5% significance level, this is evidence against $\gamma=1$ under the fitted model. The covariate-model rate at exposure four is half its rate at exposure one because it varies as $0.06e^{-0.50}$. The calculations use the stated estimates and standard error; the p-value is rounded to four decimals.

Portfolio Consequences of the Choice

A coefficient other than one can affect rating comparisons across records with different exposure, even when their other predictors are identical. It can therefore change portfolio expected counts and expected rates, not just the prediction for a single policy. The following example calculates both totals using unrounded predictions before rounding the results.

Worked Example: Comparing Portfolio Totals Under Two Specifications

A hypothetical portfolio contains three groups. Group A has exposure 1 and base rate component $0.08$; a second Group A segment has exposure 4 and the same component. Group B has exposure 2 and a rating relativity of $1.5$, so its component is $0.08(1.5)=0.12$. Compare total expected count and overall expected rate under an offset and under an estimated exposure coefficient $\gamma=0.75$. The total exposure is $1+4+2=7$.

Offset predictions. Multiply each segment’s exposure by its rate component: $$ \begin{aligned} \text{Group A, }e=1 &: 1(0.08)=0.080000,\\ \text{Group A, }e=4 &: 4(0.08)=0.320000,\\ \text{Group B, }e=2 &: 2(0.12)=0.240000. \end{aligned} $$ The total expected count is $0.080000+0.320000+0.240000=0.640000$. The overall expected rate is $0.640000/7=0.091429$ claims per exposure unit.

Estimated-covariate predictions. Use $a e^{0.75}$ for each segment. The needed powers are $$ 1^{0.75}=1,\qquad 4^{0.75}=2.828427,\qquad 2^{0.75}=1.681793. $$ Thus the respective expected counts are $$ \begin{aligned} 0.08(1)&=0.080000,\\ 0.08(2.828427)&\approx 0.226274,\\ 0.12(1.681793)&\approx 0.201815. \end{aligned} $$ Using the unrounded powers, the total expected count is approximately $0.508089$, and the overall expected rate is $0.508089/7\approx0.072584$ claims per exposure unit. The estimated coefficient below one reduces expected rates at higher exposures, so the total is below the offset total. Counts and rates are rounded to six decimals.

Model Choice, Validation, and Common Traps

The offset is usually the natural specification when exposure is a known measure of time at risk and the intended claim frequency is a rate per exposure unit. Estimating $\gamma$ is useful when there is a substantive reason to examine deviations from proportionality or when the data support an exposure pattern not captured by the offset assumption. It should not be used simply because a flexible model can estimate another coefficient.

In either case, examine model fit and prediction behavior across exposure levels. In the covariate model, check whether the fitted exposure relationship is plausible over the range used for pricing. Extrapolating a power relationship far beyond observed exposures can produce questionable rates. In the offset model, an exposure-related pattern in residuals or observed-to-expected results may indicate that the fixed proportionality assumption needs review, though it can also signal other misspecification.

Common exam trap: Do not add both a fixed offset $\log(e_i)$ and an estimated covariate $\gamma\log(e_i)$ while interpreting $\gamma$ as the total exposure exponent. Together they give $$ \log E[N_i]=x_i^\mathsf{T}\beta+(1+\gamma)\log(e_i), $$ so the total exponent is $1+\gamma$. If the desired total exponent is an estimated value $\theta$, use $\theta\log(e_i)$ as a covariate without also adding the offset, or parameterize the offset-plus-deviation coefficient as $(\theta-1)\log(e_i)$. Another frequent error is to include raw $e_i$ under a log link and assume this is equivalent to including $\log(e_i)$; it is not.

State the target.
Decide whether the intended mean is a constant rate times exposure, or whether the rate is allowed to vary systematically with exposure.

Write the linear predictor.
Use $\log(e_i)$ with coefficient fixed at one for an offset, or estimate its coefficient when exposure is a covariate.

Translate the model back to the rate.
For an estimated coefficient, divide the fitted mean by exposure; the resulting rate varies as $e_i^{\gamma-1}$.

Check implications and fit.
Compare predictions over relevant exposure levels and ensure that the coefficient and its interpretation match the chosen specification.

The central distinction is simple but consequential: an offset imposes unit elasticity of expected count with respect to exposure, while a log-exposure covariate estimates that elasticity. For a model intended to predict claim frequency per exposure unit, the offset directly encodes a constant rate; the covariate model instead allows the rate to depend on exposure and should be interpreted accordingly.

Check Your Understanding

For each problem, distinguish the exposure specification from the other rating effects and check the implied expected rate.

A hypothetical offset model has a base expected rate of $0.14$ claims per exposure unit. A risk relativity is $1.25$, and a record has exposure $3.5$. Find its expected count and expected rate. (Answer: expected count $0.6125$; expected rate $0.1750$ claims per exposure unit.)
A model with log exposure as a covariate has $\exp(x^\mathsf{T}\beta)=0.09$, $\gamma=0.8$, and exposure $4$. Calculate the expected count and rate. (Answer: expected count approximately $0.272829$; expected rate approximately $0.068207$ claims per exposure unit.)
A fitted exposure coefficient is $1.12$ with standard error $0.05$. Calculate the Wald statistic for $H_0:\gamma=1$ and the approximate two-sided p-value. State the conclusion at the 5% level. (Answer: $Z=2.40$, p-value approximately $0.0164$; reject $H_0$ at 5%.)
For a record with other-predictor component $0.04$, compare expected count and expected rate at exposure $9$ under an offset and under a covariate model with $\gamma=0.5$. (Answer: offset count $0.36$ and rate $0.04$; covariate count $0.12$ and rate approximately $0.013333$ claims per exposure unit.)
In a covariate model, the fitted mean is $0.10$ at exposure 1 and $0.30$ at exposure 4, with all other predictors held fixed. Find the implied exposure coefficient $\gamma$ and the expected rate at exposure 4. (Answer: $\gamma=\log(3)/\log(4)\approx0.7925$; expected rate $0.30/4=0.075$ claims per exposure unit.)

What You'll Learn