Offsets Versus Estimated Exposure Effects

Why the Exposure Coefficient Is a Modeling Choice

In a count model, exposure often represents the amount of time, coverage, or opportunity during which events could occur. A policy observed for two years usually has twice the time at risk of a policy observed for one year. Treating exposure as a known multiplier encodes that proportionality directly. Alternatively, a model can estimate how the expected count changes with exposure. These choices answer different questions and impose different assumptions.

The distinction is not simply whether exposure appears in the model. It appears in both formulations. The choice is whether the coefficient of log exposure is fixed at one or estimated from the data. A fixed coefficient represents a structural rate-model assumption; an estimated coefficient allows the observed exposure relationship to differ from exact proportionality.

Definition: Let $Y_i$ be a count and $E_i>0$ its known exposure for observation $i$. Under a Poisson log-link model with other predictors $x_i$, an exposure-offset specification is $$ Y_i\mid E_i,x_i \sim \operatorname{Poisson}(\mu_i), \qquad \log(\mu_i)=\log(E_i)+\alpha+x_i^\mathsf{T}\beta. $$ The coefficient of $\log(E_i)$ is fixed at one; $\log(E_i)$ is an offset, not an estimated predictor. An estimated exposure-effect model instead uses $$ Y_i\mid E_i,x_i \sim \operatorname{Poisson}(\mu_i), \qquad \log(\mu_i)=\alpha+\gamma\log(E_i)+x_i^\mathsf{T}\beta, $$ where $\gamma$ is estimated. Both formulations assume conditionally independent Poisson counts with conditional mean equal to conditional variance, unless a different response distribution is specified.

In the offset model, holding the other predictors fixed, the expected count is $E_i$ times the modeled event rate. In the estimated-effect model, multiplying exposure by $c$ multiplies the expected count by $c^\gamma$. Thus $\gamma=1$ recovers exact proportionality. A value below one means predicted counts rise less than proportionally with exposure; a value above one means they rise more than proportionally.

What the Offset Assumes

An offset is appropriate when exposure is a known amount of opportunity and the expected count is proportional to it, conditional on the modeled risk factors. For example, if an individual’s event rate is constant over the observation period, doubling that period doubles the expected count. The offset makes this relationship exact rather than asking the data to estimate it.

This can be a useful and principled choice when exposure is measured reliably, the rate interpretation is substantively justified, and there is not enough information to estimate a separate exposure effect robustly. It is not a general correction for differences in observation time. If the rate changes with duration, exposure is measured inconsistently, or longer exposures systematically represent a different population, the proportionality assumption may be unsuitable.

An estimated coefficient relaxes the exact-proportionality restriction, but it does not automatically explain why the relationship differs from one. Differences in policy mix, selection, omitted risk factors, or exposure measurement can all affect the estimate. In particular, an estimated $\gamma$ is not automatically a causal effect of increasing exposure.

Worked Example: Comparing Predictions at Different Exposures

A hypothetical model has no additional predictors. For illustration, both specifications use a baseline expected count of $0.03$ at exposure $E=1$. Compare predictions at exposures $0.25$, $1$, and $4$, first with an offset and then with an estimated exposure coefficient of $\gamma=0.5$. Treat these as two specified mean relationships, not as fits to a particular data set.

Method. With an offset, the mean is $\mu=0.03E$. With $\gamma=0.5$, the mean is $\mu=0.03E^{0.5}$. At $E=0.25$, the offset prediction is $0.03(0.25)=0.0075$, while the estimated-effect prediction is $0.03\sqrt{0.25}=0.03(0.5)=0.015$.

At $E=1$, both predictions are $0.03$, since $1^1=1^{0.5}=1$. At $E=4$, the offset prediction is $0.03(4)=0.12$, while the estimated-effect prediction is $0.03\sqrt{4}=0.03(2)=0.06$.

The predictions are therefore $0.0075, 0.03, 0.12$ under the offset and $0.015, 0.03, 0.06$ under the estimated effect. The estimated coefficient of $0.5$ gives a square-root increase rather than the offset model’s direct proportional increase. The common mistake is to treat $\gamma=0.5$ as if it still meant that doubling exposure doubles the mean; it does not. The model says the multiplier for doubling exposure is $2^{0.5}=\sqrt{2}$, not $2$.

Check. At $E=1$, both formulas return $0.03$, and at $E=4$ the estimated-effect formula returns $0.06$, half the offset prediction, as expected from the square-root relationship.

Estimating the Exposure Effect

When exposure is a predictor rather than an offset, its coefficient is estimated alongside the other regression coefficients. For the simple model without other predictors, $\gamma$ describes the elasticity of the mean with respect to exposure: a small percentage change in exposure corresponds approximately to a $\gamma$-times-as-large percentage change in the expected count. This interpretation follows from differentiating the mean $\mu=\exp(\alpha)E^\gamma$: $d\log(\mu)=\gamma\,d\log(E)$.

The estimate can be sensitive to the range of exposure values and to how exposure is distributed across risk groups. If all observations have the same exposure, $\log(E)$ is constant and its effect cannot be distinguished from the intercept. More generally, a credible estimate requires enough exposure variation and a model that accounts adequately for other systematic differences in expected counts.

Worked Example: Estimating an Exposure Power From Grouped Counts

A hypothetical portfolio has 10 independent accounts observed at exposure $E=1$ each and 10 independent accounts observed at exposure $E=4$ each. The total event counts are 20 and 40, respectively. Assume the counts are independent Poisson variables with mean $\mu_i=\exp(\alpha)E_i^\gamma$. Compare the estimated-effect fit with the offset fit.

Estimated-effect fit. Within the first group, the total count has mean $10\exp(\alpha)1^\gamma$. Its maximum-likelihood fitted mean equals the observed total, 20, so the fitted mean per account is $20/10=2$. In the second group, the fitted mean per account is $40/10=4$. The model’s per-account means at exposures 1 and 4 therefore satisfy $2=\exp(\alpha)1^\gamma$ and $4=\exp(\alpha)4^\gamma$.

The first equation gives $\exp(\alpha)=2$, or $\alpha=\log(2)$. Dividing the second fitted mean by the first gives $4/2=4^\gamma$, so $2=4^\gamma$. Taking logarithms gives $\gamma=\log(2)/\log(4)=0.5$. The fitted means per account are therefore 2 at exposure 1 and $2(4^{0.5})=2(2)=4$ at exposure 4. These match the observed group averages.

Offset fit. With $\log(E)$ fixed as an offset and only an intercept, the fitted model is $\mu_i=\exp(\alpha)E_i$. There are $20+40=60$ observed events and total exposure $10(1)+10(4)=50$. The fitted rate is $60/50=1.2$, so $\exp(\hat\alpha)=1.2$. The fitted mean per account is $1.2(1)=1.2$ in the first group and $1.2(4)=4.8$ in the second.

The offset fit predicts a common rate of 1.2 events per exposure unit, whereas the estimated-effect fit predicts group means per account of 2 and 4. In this constructed example the estimated power is $0.5$, not one. This does not establish that exposure itself caused the difference: the group counts could also reflect other differences between accounts. The common mistake is to compare group counts, 20 and 40, without accounting for the unequal numbers of accounts and exposures. The relevant observed averages are 2 and 4 per account, and the model uses exposure to relate those means.

Check. The offset predictions total $10(1.2)+10(4.8)=12+48=60$, equal to the total observed count. The estimated-effect predictions total $10(2)+10(4)=60$ as well.

Why the Intercept Estimate Has This Form

A short likelihood argument explains the intercept estimate in the simple offset model. Assume independent Poisson counts with means $\mu_i=\exp(\alpha)E_i$, with all exposures positive and known. The log-likelihood, apart from terms that do not involve $\alpha$, is

\ell(\alpha)=\sum_i \left[y_i\{\alpha+\log(E_i)\}-\exp(\alpha)E_i\right]+\text{constant}.

Differentiating with respect to $\alpha$ gives $\sum_i y_i-\exp(\alpha)\sum_i E_i$. At an interior maximum, this score equals zero. Therefore $\exp(\hat\alpha)=\sum_i y_i/\sum_i E_i$: the fitted rate is total observed events divided by total exposure. This is a rate per exposure unit, not the unweighted average of individual observed rates.

The result depends on the stated model: a Poisson likelihood, an intercept-only rate structure, positive known exposures, and no additional predictors. With covariates, the intercept generally does not have this simple closed-form estimate because the fitted means also depend on the covariates and their estimated coefficients.

Derivation: In the estimated-effect model, the log-likelihood contribution involving $\alpha$ and $\gamma$ is $\sum_i[y_i(\alpha+\gamma\log E_i)-\exp(\alpha)E_i^\gamma]$. Its score for $\gamma$ is $\sum_i (y_i-\mu_i)\log(E_i)$. Unlike the offset case, the data must determine how residual counts vary with log exposure. Setting this score to zero generally requires numerical maximum-likelihood estimation; there is no general formula that sets $\hat\gamma$ equal to one.

Testing the Proportionality Restriction

The offset model is the estimated-effect model subject to the restriction $\gamma=1$. If both models use the same response observations, likelihood, and other predictors, they are nested: the offset model is obtained by fixing one parameter in the estimated model. A likelihood-ratio statistic can assess whether the data provide evidence against that restriction. Under the null hypothesis $\gamma=1$, regularity conditions, and a sufficiently large sample, the statistic is approximately chi-square with one degree of freedom.

Worked Example: Interpreting a Likelihood-Ratio Comparison

For a hypothetical count data set, an offset model has maximized log-likelihood $-128.4$, and the otherwise comparable model with estimated $\gamma$ has maximized log-likelihood $-127.1$. Test the restriction $\gamma=1$ using a likelihood-ratio test. Assume the models are nested, use the same observations and Poisson likelihood, and the large-sample chi-square approximation is appropriate.

Method. The likelihood-ratio statistic is twice the difference in maximized log-likelihoods: $2[-127.1-(-128.4)]=2(1.3)=2.6$. The estimated-effect model has one additional parameter, so under the null the reference distribution is chi-square with one degree of freedom. The upper-tail probability is $P(\chi^2_1\ge 2.6)\approx 0.1069$.

At a 5% significance level, $0.1069>0.05$, so this test does not reject $\gamma=1$. The result is not proof that the exposure relationship is exactly proportional; it says this comparison does not provide sufficiently strong evidence against that restriction at the chosen level. The common mistake is to report $0.1063$ as the tail probability for 2.6 on one degree of freedom; the approximate upper-tail value is $0.1069$.

Check. The larger maximized log-likelihood is $-127.1$, so the likelihood-ratio statistic is nonnegative, as required. The models differ by one estimated parameter, giving one degree of freedom.

Choosing and Communicating the Specification

Use an offset when proportionality is a defensible part of the data-generating structure: equal rates per exposure unit are expected after accounting for modeled risk factors. Use an estimated exposure effect when there is a substantive reason to investigate departures from that relationship and the data contain enough exposure variation to estimate it. A likelihood-ratio test can help compare the restriction $\gamma=1$ with the less constrained model, but the result should be considered alongside subject-matter reasoning, model fit, and predictive performance.

The data cannot validate every assumption. For example, if longer observation periods occur mostly among a different class of policies, an estimated $\gamma$ may partly reflect that composition. If exposure is measured with error, either specification can be affected. Also, a Poisson model imposes equality of conditional mean and variance. If that assumption is poor, the usual model-based standard errors and likelihood-ratio approximation may be unreliable; the choice between an offset and an estimated exposure coefficient does not by itself solve such a distributional problem.

Key takeaway: An offset asserts that expected count is proportional to exposure, fixing the log-exposure coefficient at one. Estimating that coefficient relaxes the assertion and lets the data estimate an exposure power. The estimated value describes the fitted relationship; it does not by itself explain its cause.

Check Your Understanding

For each problem, identify whether the exposure relationship is fixed or estimated, and check the arithmetic before interpreting the result.

A hypothetical offset model has a fitted rate of $0.08$ events per unit of exposure. What is the expected count at exposure $3$? (Answer: $0.24$.)
In a model with $\gamma=0.5$, exposure increases from 1 to 9 while other predictors stay fixed. By what factor does the expected count change? (Answer: $9^{0.5}=3$.)
A Poisson offset model has 36 events over 30 total exposure units and an intercept only. What is the maximum-likelihood fitted rate per exposure unit? (Answer: $36/30=1.2$.)
An estimated-effect model has $\hat\gamma=1.2$. If exposure doubles, what is the fitted mean multiplier, rounded to four decimals? (Answer: $2^{1.2}\approx 2.2974$.)
Comparable nested models have maximized log-likelihoods $-50.0$ for the offset model and $-48.0$ for the model estimating $\gamma$. Find the likelihood-ratio statistic and its degrees of freedom. (Answer: $2[-48.0-(-50.0)]=4.0$, with 1 degree of freedom.)

What You'll Learn