Zero and Unequal Exposure in Rate Models

Why Exposure Must Be Handled Explicitly

In a count-rate problem, two records can have different claim counts simply because one was observed for longer or covered more units of risk. A count of three claims over three policy-years is not equivalent evidence about the underlying rate to three claims over one policy-year. The model must distinguish the amount of exposure from the rate of events per unit exposure.

A logarithmic offset incorporates known exposure into a Poisson generalized linear model (GLM) while leaving the rate relationship to be estimated. The key boundary case is zero exposure: its logarithm is undefined, and the data may contain no information about the rate—or may reveal a serious inconsistency.

Definition: Let $Y_i$ be the observed event count for record $i$, and let $e_i$ be its known, nonnegative exposure in a specified unit, such as policy-years. For positive exposure, a Poisson rate model has conditional mean $E[Y_i\mid \mathbf{x}_i]=e_i\lambda_i$, where $\lambda_i$ is the expected event rate per exposure unit. With a log link, the model is $$ Y_i\mid \mathbf{x}_i,e_i\sim \operatorname{Poisson}(e_i\lambda_i),\qquad \log E[Y_i\mid \mathbf{x}_i,e_i]=\log(e_i)+\mathbf{x}_i^\mathsf{T}\boldsymbol{\beta}. $$ The known term $\log(e_i)$ is an offset: its coefficient is fixed at 1, not estimated.

This specification assumes that, conditional on the predictors and exposure, the counts are independent Poisson variables and the expected count is proportional to exposure. In particular, the exposure unit and event definition must be consistent across records. If $e_i=0$, the expression $\log(e_i)$ cannot be evaluated. The model does not justify replacing zero by an arbitrary small number.

Unequal Positive Exposures

For every record with $e_i>0$, the offset makes the expected count equal to exposure multiplied by rate. Thus a record with twice the exposure has twice the expected count when its predictors—and therefore its rate—are the same. The regression coefficients describe the rate, while the offset accounts for the known amount of observation.

For a model with one common rate $\lambda$, the likelihood gives a useful direct result. Suppose the counts are independent with $Y_i\sim\operatorname{Poisson}(e_i\lambda)$, where all $e_i>0$. The log-likelihood, apart from no omitted terms, is $$ \ell(\lambda)=\sum_i\left[y_i\log(e_i\lambda)-e_i\lambda-\log(y_i!)\right]. $$ Differentiating with respect to $\lambda$ gives $$ \frac{d\ell}{d\lambda}=\frac{\sum_i y_i}{\lambda}-\sum_i e_i. $$ Setting this derivative equal to zero yields $\widehat{\lambda}=\sum_i y_i/\sum_i e_i$. The second derivative is $-\sum_i y_i/\lambda^2$, which is negative when there is at least one event, so this stationary point is a maximum. If the total count is zero, the likelihood is maximized at the boundary $\widehat{\lambda}=0$.

Result: Under the independent Poisson model with one common rate and positive exposures, the maximum-likelihood estimate is total observed count divided by total exposure. This is a rate estimate, not an average of the separate record-level ratios.

Worked Example: Estimating a Common Rate With Unequal Exposure

A hypothetical insurer observes three groups of policies. Their claim counts are 1, 3, and 2, and their exposures are 0.5, 1.5, and 2.0 policy-years. Assume independent Poisson counts with one common claim rate. Estimate the rate and calculate the fitted expected count for each group.

Method. Use the Poisson offset model with common rate $\lambda$, so each expected count is $e_i\lambda$. Add counts and exposures: $$ \sum_i y_i=1+3+2=6,\qquad \sum_i e_i=0.5+1.5+2.0=4.0. $$ Therefore, $$ \widehat{\lambda}=\frac{6}{4.0}=1.5\text{ claims per policy-year}. $$

The fitted expected counts are exposure times the estimated rate. For the first group, $0.5(1.5)=0.75$ claims; for the second, $1.5(1.5)=2.25$ claims; and for the third, $2.0(1.5)=3.00$ claims. Their sum is $0.75+2.25+3.00=6.00$, which equals the observed total.

Check. The exposure-weighted fitted rate is $6.00/4.0=1.5$, matching the estimate. The common mistake is to average the three raw ratios, $1/0.5=2$, $3/1.5=2$, and $2/2=1$, obtaining $1.67$. That gives the short-exposure record the same influence as the longest record and is not the maximum-likelihood estimate.

Rates With Predictors

When rates differ across predictor values, the offset model estimates those differences while preserving the exposure multiplier. For example, with a binary category indicator $x_i$, the model can be written as $$ \log E[Y_i\mid x_i,e_i]=\log(e_i)+\beta_0+\beta_1x_i. $$ For $x_i=0$, the expected rate is $\exp(\beta_0)$. For $x_i=1$, it is $\exp(\beta_0+\beta_1)$. Their rate ratio is $\exp(\beta_1)$; expected counts for either category are obtained by multiplying the corresponding rate by exposure.

Worked Example: Comparing Category-Specific Rates

A hypothetical equipment portfolio has two risk categories. Category A has 8 reported breakdowns over 4 exposure-years; category B has 9 breakdowns over 3 exposure-years. Fit a log-linear model with a separate rate for each category, then find the rate ratio for B relative to A and the expected count for a category B record with 1.25 exposure-years.

Method. In a two-category model with a separate rate in each category, the maximum-likelihood rate within each category is its total count divided by its total exposure. For A, $$ \widehat{\lambda}_A=\frac{8}{4}=2\text{ breakdowns per exposure-year}. $$ For B, $$ \widehat{\lambda}_B=\frac{9}{3}=3\text{ breakdowns per exposure-year}. $$ The estimated rate ratio is $3/2=1.5$. In the model with A as the reference category, $\widehat{\beta}_0=\log(2)$ and $\widehat{\beta}_1=\log(3)-\log(2)=\log(1.5)$, approximately $0.4055$. Exponentiating the category coefficient returns the rate ratio: $\exp(0.4055)\approx1.5$.

For the specified category B record, its fitted expected count is its exposure times the B rate: $$ 1.25(3)=3.75\text{ breakdowns}. $$ This is an expected count, not a prediction that a fractional number of breakdowns will be observed.

Check. The category totals implied by the fitted rates are $4(2)=8$ and $3(3)=9$, as observed. The common mistake is to interpret $\widehat{\beta}_1=0.4055$ as an increase of 0.4055 breakdowns. It is a log rate ratio; the multiplicative rate change is $\exp(\widehat{\beta}_1)=1.5$.

What Zero Exposure Means

A record with zero exposure has no observed time or units at risk under the stated exposure definition. If its count is also zero, it provides no information about the rate: a Poisson variable with mean zero is zero with probability one. Its likelihood contribution is therefore constant with respect to the rate parameters. It is mathematically harmless to recognize this fact, but software using $\log(e_i)$ cannot calculate its offset.

A zero-exposure record with a positive count is different. Under the stated model, its Poisson mean is $0$, so the probability of a positive count is zero. This is not a difficult numerical case to repair; it is an incompatibility between the recorded count, exposure, and event definition that should be investigated.

Key rule: For a log-exposure offset, use only records with positive, correctly measured exposure in the fitted rate likelihood. A zero-exposure, zero-count record contributes no rate information and can be excluded from that likelihood. A zero-exposure, positive-count record must be investigated rather than silently discarded or assigned an arbitrary positive exposure.

Check the exposure and count.
Confirm that both refer to the same unit, time window, and event definition. A positive count paired with zero exposure is a data or modeling inconsistency under the assumed rate model.

Determine why exposure is zero.
If the record truly had no time at risk and its count is zero, it has no information about the rate. If the exposure is zero because of rounding or a recording error, retrieve or correct the defensible exposure value.

Fit only a coherent offset specification.
Construct $\log(e_i)$ only for positive exposures. Do not add an arbitrary small constant merely to make the logarithm computable.

Worked Example: Handling Zero-Exposure Records

A hypothetical service program records four sites: site A has 2 incidents and 1.0 service-year of exposure; site B has 0 incidents and 0 exposure; site C has 4 incidents and 2.0 service-years; and site D has 1 incident and 0.5 service-year. Assume one common Poisson incident rate. A separate record from the same data extract reports 2 incidents and 0 exposure. Explain what can be fitted and what must be checked.

Method. First separate the zero-exposure records from those with positive exposure. Site B has both zero exposure and zero incidents. It carries no likelihood information about the rate and cannot be given a finite log offset. The three positive-exposure sites have total count $2+4+1=7$ and total exposure $1.0+2.0+0.5=3.5$ service-years. Thus the common-rate estimate from the coherent records is $$ \widehat{\lambda}=\frac{7}{3.5}=2\text{ incidents per service-year}. $$ Their fitted expected counts are $1.0(2)=2$, $2.0(2)=4$, and $0.5(2)=1$, respectively.

The additional record with 2 incidents and zero exposure cannot be treated as another ordinary Poisson observation in this model: its assumed mean would be $0$, making the observed positive count impossible. Check whether exposure was recorded in the wrong unit, rounded down from a positive amount, or defined over a different period from the incidents. Do not simply include it with an artificial offset, and do not remove it without recording the reason for the treatment.

Check. The three positive-exposure records have fitted counts summing to $2+4+1=7$, the observed total used in the estimate. The common mistake is to replace zero exposure by a small value such as 0.001 and treat the positive count as valid. That creates an enormous implied rate and changes the analysis according to an arbitrary choice rather than resolving the inconsistency.

Implementation and Practical Checks

In R, the model form for positive exposures is typically written as glm(count ~ predictors + offset(log(exposure)), family = poisson, data = ...). Before fitting, check whether exposure is positive and whether every positive count has a coherent exposure. A zero-exposure, zero-count row should be handled explicitly rather than passed to log(). A zero-exposure, positive-count row calls for investigation before fitting.

The fitted mean from an offset model is an expected count. To report a rate, divide that fitted mean by the corresponding positive exposure, or interpret the linear predictor after removing the offset. Keep the unit visible: a rate per policy-year and a rate per policy-month differ by a factor of 12. A model may be numerically fit while still representing the wrong rate if the exposure unit or period is inconsistent.

The Poisson specification also assumes that the conditional count distribution is Poisson. If that assumption is inadequate, uncertainty estimates and model fit may be affected; changing how zero exposure is coded does not solve a distributional problem. The central discipline is to define exposure first, use it as a known multiplier for positive exposure, and treat zero-exposure records according to what they mean in the data. These distinctions are aligned with the SOA Exam PA (Predictive Analytics) syllabus.

Check Your Understanding

For each problem, identify the exposure treatment as well as the requested numerical result.

A hypothetical portfolio has 4 claims over 1.5 policy-years and 7 claims over 4.0 policy-years. Under one common Poisson rate, calculate the maximum-likelihood rate and the fitted expected count for a further record with 0.5 policy-year of exposure. (Answer: $11/5.5=2$ claims per policy-year; expected count $=1$.)
A category has 5 events over 2.5 exposure-years, and a reference category has 6 events over 2.0 exposure-years. Calculate the two category rates and the event-rate ratio for the first category relative to the reference. What is the log-link coefficient for the first category relative to the reference? (Answer: rates $2$ and $3$; ratio $2/3$; coefficient $\log(2/3)\approx-0.4055$.)
A record has zero exposure and zero events. Explain whether it can be passed directly to an offset calculation using $\log(e_i)$, and whether it informs the common rate. (Answer: no, because $\log(0)$ is undefined; it contributes no information about the rate and may be excluded from that likelihood.)
A record reports 3 events and zero exposure. Under the Poisson count-rate model, what is the probability of this observation, and what should be done before fitting? (Answer: its probability is zero under a Poisson mean of zero; investigate the count, exposure, units, and event period rather than assigning an arbitrary positive exposure.)
A group has 8 claims over 730 policy-days. Convert the exposure to policy-years using 365 days per year, then calculate the observed rate per policy-year. (Answer: exposure $=730/365=2$ policy-years; rate $=8/2=4$ claims per policy-year.)

What You'll Learn