Building Poisson Rate Models With Exposure Offsets

Why Exposure Belongs in a Count Model

A count of events is meaningful only in relation to the amount of time, population, or other opportunity over which the events could occur. For example, observing six incidents in 300 vehicle-years is not equivalent to observing six incidents in 30 vehicle-years. A Poisson rate model accounts for this difference by modeling the expected count as the product of exposure and an event rate.

The model is useful when the response is a nonnegative event count and the observations have different known exposures. Exposure might be policy-years, patient-days, machine-hours, or another positive quantity. The model estimates how risk factors change the rate, while the exposure offset ensures that expected counts scale with the amount of observation. This tutorial is aligned with the SOA Exam SRM (Statistics for Risk Modeling) syllabus.

Definition: For observation $i$, let $Y_i$ be the event count and $e_i>0$ its known exposure, measured in a stated unit. A Poisson rate model specifies $Y_i\mid x_i\sim\operatorname{Poisson}(\mu_i)$, where $\mu_i=e_i\lambda_i$. Here $\lambda_i$ is the event rate per unit of exposure. With a log link and covariates $x_i$, the model is $$ \log(\mu_i)=\log(e_i)+\beta_0+x_i^\mathsf{T}\beta. $$ The known term $\log(e_i)$ is the offset; its coefficient is fixed at one.

This formulation assumes that, conditional on the modeled covariates and exposure, the count has a Poisson distribution, so its conditional mean and variance are both $\mu_i$. It also assumes a suitable rate is constant over each observation’s exposure, conditional on its covariates. For the usual likelihood analysis, counts from distinct observations are conditionally independent. A zero exposure is not entered into this log-offset model because $\log(0)$ is undefined; an observation with zero exposure has no opportunity for events under the model.

Why the Offset Coefficient Is One

The coefficient is fixed because exposure represents the amount of opportunity for events, not an ordinary explanatory variable whose effect is to be estimated. If a rate is $\lambda_i$ events per exposure unit, then the expected count over $e_i$ units is $e_i\lambda_i$. Taking logarithms gives $\log(\mu_i)=\log(e_i)+\log(\lambda_i)$, so the multiplier of $\log(e_i)$ is exactly one.

This is also a proportionality assumption: if all else is unchanged and exposure doubles, the expected count doubles. An estimated exposure coefficient would instead allow a relationship $\mu_i\propto e_i^\gamma$. That may describe a different model, but it is not the standard rate model with an exposure offset. In the offset model, covariates describe the rate, and exposure determines how many opportunities are observed.

\mu_i=e_i\lambda_i,\qquad \lambda_i=\exp(\beta_0+x_i^\mathsf{T}\beta),\qquad \log(\mu_i)=\log(e_i)+\beta_0+x_i^\mathsf{T}\beta

The exposure unit must be consistent throughout. If $e_i$ is measured in years, then $\lambda_i$ is a rate per year. Converting exposure from years to months changes the numerical rate by a factor of 12, but it does not change the underlying expected count when the conversion is done consistently. The offset is the logarithm of exposure in the chosen unit, not the logarithm of a rate or an arbitrary denominator.

Estimating a Common Rate

A useful special case has one common rate $\lambda$ for all observations, with $Y_i\mid e_i\sim\operatorname{Poisson}(e_i\lambda)$. The maximum-likelihood estimate is the total observed count divided by total exposure. This result follows directly from the likelihood.

Write the likelihood.
For independent Poisson counts, the likelihood is the product of their Poisson probabilities, with means $e_i\lambda$.

Take the log and differentiate.
After removing terms that do not involve $\lambda$, the log-likelihood is $\sum_i y_i\log(\lambda)-\lambda\sum_i e_i$. Its derivative is $\sum_i y_i/\lambda-\sum_i e_i$.

Set the derivative to zero.
For a positive total count, solving gives $\widehat{\lambda}=\sum_i y_i/\sum_i e_i$. The second derivative is $-\sum_i y_i/\lambda^2<0$, so this stationary point is a maximum.

L(\lambda)=\prod_i\frac{\exp(-e_i\lambda)(e_i\lambda)^{y_i}}{y_i!}, \qquad \ell(\lambda)=\text{constant}+\Big(\sum_i y_i\Big)\log(\lambda)-\lambda\sum_i e_i

For the common-rate model, take the parameter space to be $\lambda\ge 0$. If the total observed count is zero, the maximum occurs at $\widehat{\lambda}=0$. If the parameter space were restricted to $\lambda>0$, the likelihood would approach its supremum as $\lambda\to0^+$, but would not attain a maximum there. This boundary detail matters when stating the estimate precisely.

Worked Example: Estimating a Common Incident Rate

A hypothetical safety study records counts of equipment incidents over three independent groups of machine-hours. The observed counts are 2 incidents in 100 machine-hours, 5 in 200 machine-hours, and 3 in 150 machine-hours. Assume each count is Poisson with mean equal to exposure times a common rate. Estimate the rate and the expected count for a new 250 machine-hours of exposure.

Method. Use the common-rate maximum-likelihood estimate, total count divided by total exposure.

The total count is $2+5+3=10$. The total exposure is $100+200+150=450$ machine-hours. Therefore $$ \widehat{\lambda}=\frac{10}{450}=0.022222\ldots $$ incidents per machine-hour.

For 250 additional machine-hours, the fitted mean count is $250(10/450)=2500/450=5.5555\ldots$, or 5.56 incidents to two decimal places. As a check, the fitted counts for the original groups are $100(10/450)=2.2222\ldots$, $200(10/450)=4.4444\ldots$, and $150(10/450)=3.3333\ldots$. Their sum is 10, matching the observed total.

Common mistake avoided. Averaging the three group-specific rates without weighting would give each exposure the same influence. The likelihood-based estimate instead divides the combined count by the combined exposure.

Adding Risk Factors to the Rate

With covariates, the model assigns each observation its own rate. A coefficient $\beta_j$ changes the log rate by $\beta_j$ for a one-unit increase in the corresponding predictor, holding other predictors fixed. Thus the rate is multiplied by $\exp(\beta_j)$. For a binary predictor, $\exp(\beta_j)$ is the fitted rate ratio for the group coded one relative to the reference group coded zero.

The offset does not change this interpretation. Given fitted coefficients, first calculate the rate from the linear predictor without exposure, then multiply that rate by the new exposure to obtain a fitted expected count. If the model instead is written directly for $\log(\mu_i)$, the offset must still be included.

Worked Example: Comparing Rates and Predicting Counts

A hypothetical insurer models the number of roadside-assistance events per policy-year. Let $x=0$ denote the reference vehicle group and $x=1$ denote a second group. Suppose the fitted coefficients are $\beta_0=\log(0.012)$ and $\beta_1=\log(1.5)$, with $\log(\mu_i)=\log(e_i)+\beta_0+\beta_1x_i$. Calculate the fitted rates and expected counts for 200 policy-years in the reference group and 100 policy-years in the second group.

Method. Use the offset model to calculate each rate, then multiply by the appropriate exposure.

For $x=0$, the rate is $\exp(\log(0.012))=0.012$ events per policy-year. Its fitted count is $200(0.012)=2.4$.

For $x=1$, the rate is $\exp(\log(0.012)+\log(1.5))=0.012(1.5)=0.018$ events per policy-year. Its fitted count is $100(0.018)=1.8$. The fitted rate ratio is $0.018/0.012=1.5$, which agrees with $\exp(\beta_1)=1.5$. The two fitted counts sum to $2.4+1.8=4.2$.

For a new half-policy-year in the second group, the fitted mean is $0.5(0.018)=0.009$. Under the Poisson model, the probability of at least one event is $1-P(Y=0)=1-\exp(-0.009)=0.0089596$, which is 0.0090 to four decimal places.

Common mistake avoided. The coefficient exponentiates to a rate ratio, not an additive increase in the rate. The second rate is $0.012(1.5)$, not $0.012+1.5$.

Exposure Units and Fitted Probabilities

Predicted means are expected counts, not event probabilities. For a Poisson count with fitted mean $\widehat{\mu}$, the probability of at least one event is $1-\exp(-\widehat{\mu})$. This conversion is appropriate only when the question asks for an event probability and the Poisson count model applies. It does not make the rate itself a probability: a rate can exceed one when events can occur repeatedly within an exposure unit.

Worked Example: Converting Months of Exposure

A hypothetical monitoring model estimates a rate of $0.020$ events per year for a specified risk profile. A device is observed for 18 months. Calculate the fitted expected count and the probability of at least one event, assuming a Poisson count with mean equal to rate times exposure.

Method. Convert the exposure into years, the unit used for the rate, before calculating the mean.

The exposure is $18/12=1.5$ years. The fitted mean is $\widehat{\mu}=1.5(0.020)=0.030$ events. Therefore $$ P(Y\ge 1)=1-P(Y=0)=1-\exp(-0.030)=0.029554\ldots $$ which is 0.0296 to four decimal places.

As a check on the time conversion, using months instead gives a rate of $0.020/12=0.0016667$ events per month and mean $18(0.0016667)\approx0.030$. The resulting probability is unchanged.

Common mistake avoided. Multiplying 18 directly by a per-year rate treats 18 months as 18 years and would produce an incorrect mean of 0.36.

Common Traps in Offset Models

Several errors can produce plausible-looking but substantively wrong results. Check the definition and units of exposure before interpreting coefficients or reporting a fitted count.

Using counts as if exposure were equal. Unequal exposures require the offset or an equivalent rate-model formulation; raw counts alone are not comparable rates.
Estimating the standard offset coefficient. In $\log(\mu_i)=\log(e_i)+x_i^\mathsf{T}\beta$, the coefficient of $\log(e_i)$ is fixed at one. Treating it as an ordinary predictor changes the model.
Putting the wrong quantity in the offset. The offset is $\log(e_i)$, where $e_i$ is exposure. The event rate is modeled through the intercept and covariates.
Mixing units. A rate per year must be multiplied by years, not months. A consistent unit conversion leaves the fitted mean unchanged.
Calling the fitted mean a probability. The fitted mean is an expected count. A probability such as the chance of at least one event requires a probability calculation under the assumed count distribution.

A final audit is simple: identify the count, state the exposure unit, write $\widehat{\mu}_i=e_i\widehat{\lambda}_i$, and check that the resulting expected count scales proportionally if exposure changes while the risk profile stays fixed.

Check Your Understanding

For each problem, use the stated exposure units and Poisson rate model. Check the final answer against the units and, where applicable, against the proportionality of count and exposure.

A common-rate model observes 0 events in 30 service-hours, 2 in 50 service-hours, and 4 in 120 service-hours. Find the maximum-likelihood rate. (Answer: $6/200=0.0300$ events per service-hour.)
A fitted model has $\log(\mu_i)=\log(e_i)+\log(0.01)+x_i\log(2)$. For $x_i=1$ and exposure $e_i=40$ units, find the fitted rate and mean count. (Answer: rate $0.020$ per unit; fitted mean $0.800$.)
A risk has a fitted rate of $0.060$ events per year and is observed for nine months. Find the expected count and the probability of at least one event. (Answer: mean $0.045$; probability $1-\exp(-0.045)=0.0440$ to four decimal places.)
A model has $\beta_0=\log(0.015)$, $\beta_1=-0.5$, and one binary covariate $x$, with the usual log-exposure offset. For $x=1$ and exposure 80 units, calculate the rate and fitted mean. (Answer: rate $0.015\exp(-0.5)=0.009098$ per unit; fitted mean $80(0.009098)=0.7278$.)
Two otherwise identical observations have exposures of 25 and 75 device-years. Their fitted rate is $0.016$ events per device-year. State the offset difference, the two fitted means, and their ratio. (Answer: offset difference $\log(75)-\log(25)=\log(3)$; means $0.400$ and $1.200$; the second fitted mean is three times the first.)

What You'll Learn