Offsets For Known Denominators

Why a Known Denominator Becomes an Offset

Ratemaking often targets a quantity per unit of something already known: claim counts per vehicle-year, claims per thousand dollars of payroll, or aggregate losses per insured unit of time. The numerator is observed for each risk or group; the denominator specifies how much exposure produced that numerator. A model for the total numerator must account for the fact that more exposure generally produces a larger expected total, even when the underlying rate is unchanged.

For a log-link model, the denominator enters in a precise way. It contributes a known term to the linear predictor, rather than a coefficient that the model estimates. The key assumption is that expected numerator is proportional to the denominator, with the proportionality rate described by the modeled risk characteristics.

Definition: Let $Y_i$ be a nonnegative numerator for observation $i$, and let $e_i>0$ be its known denominator in specified exposure units. If the expected numerator satisfies $E[Y_i\mid x_i,e_i]=e_i r_i$, where $r_i$ is the expected numerator per denominator unit, then a log-link model with $\log r_i=x_i^\mathsf{T}\beta$ has $\log E[Y_i\mid x_i,e_i]=\log e_i+x_i^\mathsf{T}\beta$. The term $\log e_i$ is an offset: it is included with coefficient fixed at 1.

Here $x_i^\mathsf{T}\beta$ represents the modeled log rate, and $e_i$ must be positive for the logarithm to be defined. The units of $e_i$ determine the units of $r_i$. For example, using payroll measured in thousands of dollars models claims per thousand dollars of payroll; using payroll in dollars instead changes the numerical rate and the offset, but not the underlying expected total if the units are handled consistently.

The offset does not mean that every denominator belongs in every model. It encodes a specific structural assumption: doubling $e_i$ doubles the expected numerator while leaving the rate $r_i$ unchanged. If exposure has an additional effect on the rate, or if the expected numerator is not proportional to the denominator, this simple offset formulation is insufficient. Also, the log-offset equation follows from the log link; for a different link $g$, it is generally not true that $g(e_i r_i)=\log e_i+g(r_i)$.

Deriving the Offset and Interpreting the Model

The derivation starts with the target, not with a software convention. Suppose the expected numerator is exposure multiplied by a rate:

E[Y_i\mid x_i,e_i]=e_i r_i

A log-link model for the rate sets $r_i=\exp(x_i^\mathsf{T}\beta)$. Substitution and taking logarithms gives:

\begin{aligned} E[Y_i\mid x_i,e_i] &=e_i\exp(x_i^\mathsf{T}\beta),\\ \log E[Y_i\mid x_i,e_i] &=\log e_i+x_i^\mathsf{T}\beta. \end{aligned}

This is why the coefficient of $\log e_i$ is fixed at 1: it comes from the assumed multiplication by $e_i$, not from an estimated association. The expected rate is recovered by dividing the expected numerator by exposure:

\frac{E[Y_i\mid x_i,e_i]}{e_i} =\exp(x_i^\mathsf{T}\beta).

Thus, the same model provides two different predictions. The rate prediction depends on the rating variables through $x_i^\mathsf{T}\beta$; the expected total additionally scales with the known denominator. This distinction matters when comparing risks with unequal exposures or translating a rate indication into an expected claim count or loss total.

For a count response, a common distributional specification is $Y_i\mid x_i,e_i\sim\operatorname{Poisson}(e_i r_i)$. The offset determines the conditional mean; the Poisson assumption additionally specifies the variance as equal to that mean. The offset relationship can also be used with other response distributions when their mean structure supports it. It does not, by itself, establish a particular variance or distributional assumption.

Worked Models

Worked Example: Claim Counts per Vehicle-Year

Method and mistake avoided. Use a Poisson log-link model for counts with vehicle-years as the known denominator. The common mistake is to treat the predicted rate as the expected count and forget to multiply by exposure.

A hypothetical insurer models claims for a vehicle class using $\log r_i=\beta_0+\beta_1x_i$, where $x_i=1$ for the class in question and $0$ otherwise. The fitted baseline rate is $0.04$ claims per vehicle-year, and the class relativity is $1.5$. For a risk with $125$ vehicle-years, find the predicted rate and expected claim count. Assume the count is Poisson. Interest and mortality assumptions are not relevant to this claim-count calculation.

The fitted coefficients are $\beta_0=\log(0.04)=-3.218876$ and $\beta_1=\log(1.5)=0.405465$. For $x_i=1$, the fitted log rate is $-3.218876+0.405465=-2.813411$, so the rate is $\exp(-2.813411)=0.04(1.5)=0.06$ claims per vehicle-year.

The offset is $\log(125)=4.828314$. Therefore the log of the expected count is $4.828314-2.813411=2.014903$, and the expected count is $\exp(2.014903)=7.5$. Equivalently, $125(0.06)=7.5$. Under the stated Poisson assumption, the count distribution is Poisson with parameter $7.5$, so its mean and variance are both $7.5$.

Check: Dividing the expected count by the denominator gives $7.5/125=0.06$, the modeled rate. No rounding beyond the displayed values is needed.

Worked Example: Estimating a Pooled Payroll Rate

Method and mistake avoided. Fit an intercept-only Poisson rate model with payroll measured in thousands of dollars. The common mistake is to use payroll in dollars in the calculation while reporting a rate per thousand dollars.

Two hypothetical business groups have payroll exposures of $240$ and $160$ thousand dollars, with $11$ and $7$ claims, respectively. Assume independent Poisson counts with means $e_i r$, where $r$ is the common claim rate per thousand dollars of payroll. Find the maximum likelihood estimate of $r$, then calculate each group’s fitted expected count. Interest and mortality assumptions are not applicable.

The log likelihood, omitting terms that do not depend on $r$, is $\ell(r)=11\log(240r)-240r+7\log(160r)-160r$. Its derivative is $d\ell/dr=11/r-240+7/r-160=18/r-400$. Setting the derivative equal to zero gives $\widehat r=18/400=0.045$ claims per thousand dollars of payroll. The second derivative is $-18/r^2<0$, so this stationary point is a maximum.

The fitted expected count for the first group is $240(0.045)=10.8$. For the second group it is $160(0.045)=7.2$. These are the parameters of the respective fitted Poisson distributions. Their sum is $10.8+7.2=18$, the observed total count.

Check: The total exposure is $240+160=400$ thousand dollars, so the pooled fitted rate is $18/400=0.045$ claims per thousand dollars. If exposure had instead been recorded in dollars, it would total $400{,}000$, and the numerical rate would be $18/400{,}000=0.000045$ claims per dollar; the fitted expected counts would be unchanged.

Worked Example: Aggregate Loss per Insured Unit

Method and mistake avoided. Model the expected aggregate loss as exposure times a log-linked loss rate. The common mistake is to report the modeled loss rate as if it were the total expected loss.

A hypothetical portfolio segment has $80$ insured units of exposure. Its fitted expected aggregate loss rate is modeled as a base of $\$325$ per unit, multiplied by a territory factor of $1.2$ and a coverage factor of $0.9$. Calculate the expected aggregate loss. For a distributional illustration, suppose the aggregate loss $L$ for this segment is Gamma with shape $\alpha=4$ and scale $\theta=E[L]/4$, so its mean is $4\theta$ and variance is $4\theta^2$. Assume a positive aggregate response appropriate for this Gamma specification. Interest and mortality assumptions are not relevant.

The fitted rate per insured unit is $\$325(1.2)(0.9)=\$351$. In a log-link formulation, the rate predictor is $\log(325)+\log(1.2)+\log(0.9)$, while the exposure offset is $\log(80)$. Thus the expected aggregate loss is $80(\$351)=\$28{,}080$.

The Gamma scale parameter is $\theta=\$28{,}080/4=\$7{,}020$. Its variance is $4(\$7{,}020)^2=4(\$49{,}280{,}400)=\$197{,}121{,}600$. The model’s mean is therefore $\$28{,}080$, not $\$351$; $\$351$ is the expected loss per insured unit.

Check: Dividing $\$28{,}080$ by $80$ recovers $\$351$ per unit. The Gamma parameters imply mean $4(7{,}020)=\$28{,}080$, as required.

Conditions, Units, and Common Traps

The offset is most useful when the denominator is known for each observation and the modeled mean should scale in direct proportion to it. A denominator may be a duration, a count of insured units, or an amount of payroll, but its definition and units must be consistent across observations. A rate “per vehicle-year” and a rate “per vehicle-month” are different numerical quantities; converting one to the other requires converting exposure units as well.

A denominator of zero cannot be entered as $\log e_i$. An observation with zero exposure contributes no exposure-based numerator information under this model and must be handled consistently with the data and response definition; it cannot simply be passed to a log-offset calculation. Likewise, an offset should not be estimated as an ordinary coefficient if the intended assumption is proportional scaling. Estimating that coefficient would instead ask the data to determine how the mean changes with exposure.

The offset is also distinct from a severity weight or an observation weight. It changes the modeled mean numerator through a known multiplicative exposure factor. A weight affects how an observation contributes to estimation or variance calculations; it does not automatically impose the mean relationship $E[Y_i]=e_i r_i$.

Common exam trap: If $e_i=50$ and the modeled rate is $0.08$, the expected numerator is $50(0.08)=4$. Using $0.08$ as the expected numerator omits the exposure offset. Conversely, multiplying $4$ by $50$ again applies exposure twice. The rate is $4/50=0.08$, and the expected total is $4$.

Before using an offset, verify the numerator, the denominator unit, and the intended rate. Then check whether the mean assumption genuinely implies proportional scaling with exposure. This sequence prevents a correct-looking log equation from modeling the wrong quantity.

Name the numerator and denominator.
Specify what is counted or measured, and state the exposure unit, such as vehicle-years or thousands of dollars of payroll.

State the mean relationship.
Confirm that the expected numerator is the known denominator multiplied by the modeled rate: $E[Y_i]=e_i r_i$.

Apply the log link and check predictions.
Use $\log e_i$ as a fixed offset, predict the rate, multiply by $e_i$ to obtain the expected total, and divide back by $e_i$ as a consistency check.

Check Your Understanding

For each problem, identify the rate and expected numerator carefully; all scenarios and data are hypothetical.

A Poisson model predicts $0.025$ claims per equipment-year before a class relativity of $1.4$. A risk has $90$ equipment-years. What are its fitted rate and expected claim count? (Answer: $0.035$ claims per equipment-year and $3.15$ expected claims.)
Two hypothetical groups have payroll exposures of $300$ and $200$ thousand dollars, with $9$ and $11$ claims. Under a common Poisson rate per thousand dollars, what is the pooled fitted rate and the fitted expected count for each group? (Answer: $0.04$ claims per thousand dollars; $12$ and $8$ expected claims, respectively.)
A log-link aggregate-loss model predicts a base rate of $\$410$ per policy-year, with multiplicative factors $1.25$ and $0.8$. A segment has $60$ policy-years. What is its expected aggregate loss? (Answer: $\$24{,}600$.)
A model’s annual claim rate is $0.12$ per vehicle-year. For a policy with $0.5$ vehicle-years, what is the offset and expected claim count? (Answer: offset $\log(0.5)\approx-0.6931$; expected count $0.06$.)
Suppose a fitted log-mean model is $\log E[Y_i]=\log e_i+\beta_0+\beta_1x_i$, with $\exp(\beta_0)=0.03$ and $\exp(\beta_1)=1.6$. For $x_i=1$ and $e_i=250$, find the rate and expected numerator. (Answer: rate $0.048$ per denominator unit and expected numerator $12$.)

What You'll Learn