Offset Versus Observation Weight

Why the distinction matters

A policy observed for one month does not provide the same claim-count information as an otherwise similar policy observed for a full year. A frequency model must account for that unequal exposure. One way is to model claim counts with an exposure offset. A different modeling setup uses an observation weight to change how much a record contributes to fitting. These operations have different meanings: an offset adjusts the predicted count for the amount of exposure, while a weight changes the fitting contribution of an observation.

Confusing them can produce a model that fits a large-exposure policy as though it simply had an unusually large claim count, rather than recognizing that it was observed for longer. The distinction is especially important when records have substantially different exposure amounts.

Definition: In a Poisson claim-count model, let $N_i$ be the claim count for record $i$, $e_i>0$ its known exposure, and $x_i$ its predictor vector. With a log link, an exposure-offset model is $$ N_i\mid x_i,e_i\sim\operatorname{Poisson}(\mu_i),\qquad \log(\mu_i)=\log(e_i)+x_i^\mathsf{T}\beta. $$ The term $\log(e_i)$ is a fixed offset: its coefficient is 1, not an estimated parameter. The modeled claim frequency per exposure unit is $\lambda_i=\exp(x_i^\mathsf{T}\beta)$, so the expected count is $\mu_i=e_i\lambda_i$.

This formulation assumes that, conditional on the predictors and exposure, the count is Poisson with mean proportional to exposure. In particular, the predictors describe the claim intensity during the observed exposure, and the exposure amount is known. The independent-record version also assumes that the counts from different records are conditionally independent. If the exposure unit is policy-year, for example, $\lambda_i$ is expected claims per policy-year and $e_i$ is measured in policy-years.

What offsets and weights do

An offset is part of the linear predictor. It changes the fitted mean count in a known, specified way: doubling exposure doubles the expected count when the risk characteristics stay fixed. It does not tell the model to give a record a freely chosen or arbitrary amount of influence.

An observation weight instead multiplies that record’s contribution to the fitting criterion. For example, a weighted log-likelihood has the form $\sum_i w_i\ell_i(\beta)$, where $\ell_i$ is the record’s log-likelihood contribution. The weight $w_i$ affects the estimating equations. It does not, by itself, place $\log(e_i)$ in the linear predictor or make the record’s expected count proportional to its exposure.

For a Poisson log-link model with a fixed offset, the log-likelihood, apart from terms that do not depend on $\beta$, is

\ell(\beta)=\sum_i\left[N_i\{\log(e_i)+x_i^\mathsf{T}\beta\} -e_i\exp(x_i^\mathsf{T}\beta)\right].

Differentiating with respect to $\beta$ gives the score equation

\frac{\partial \ell}{\partial\beta} =\sum_i x_i\{N_i-e_i\exp(x_i^\mathsf{T}\beta)\}=0.

The known exposure appears in the expected-count term, while each observed count contributes through $N_i$. The model is not treating exposure as an extra multiplier on the observed count. The equation also shows why the intercept-only fitted rate is total claims divided by total exposure: setting all $x_i=1$ gives $\sum_i N_i=\lambda\sum_i e_i$.

Worked applications

Worked Example: Separate Frequency Rates for Two Territories

A hypothetical insurer observes 18 claims over 30 policy-years in Territory A and 20 claims over 50 policy-years in Territory B. Fit a Poisson log-link model with a separate frequency for each territory, using Territory A as the reference level. Find the fitted rates, coefficients, and expected counts. Use the offset method.

Step 1: State the model. Let $x_i=1$ for Territory B and $x_i=0$ for Territory A. The model is $$ \log(\mu_i)=\log(e_i)+\beta_0+\beta_1x_i. $$ Thus Territory A has rate $\exp(\beta_0)$, and Territory B has rate $\exp(\beta_0+\beta_1)$. The observations are assumed to follow independent Poisson distributions with these means.

Step 2: Estimate Territory A’s rate. With a separate rate for each territory, the Poisson maximum-likelihood estimate is claims divided by exposure: $$ \widehat{\lambda}_A=\frac{18}{30}=0.600\text{ claims per policy-year}. $$ Therefore $\widehat{\beta}_0=\log(0.600)=-0.5108$, to four decimal places.

Step 3: Estimate Territory B’s rate and relativity. $$ \widehat{\lambda}_B=\frac{20}{50}=0.400\text{ claims per policy-year}. $$ The estimated log-rate difference is $$ \widehat{\beta}_1=\log(0.400)-\log(0.600) =\log\left(\frac{2}{3}\right)=-0.4055. $$ The Territory B rate relativity to Territory A is $\exp(\widehat{\beta}_1)=2/3$.

Step 4: Check fitted counts. The expected fitted count in Territory A is $30(0.600)=18$. In Territory B it is $50(0.400)=20$. The fitted counts equal the observed counts, as expected for a model with a separate rate parameter for each territory.

Method and mistake avoided. This is a Poisson model with a log-exposure offset. Dividing each territory’s count by its own exposure gives its rate. Treating the two counts as if they had equal exposure would compare 18 with 20 and miss that Territory B had 20 more policy-years of observation.

Worked Example: Why Weighting the Counts Is Not an Offset

A hypothetical portfolio has three records: record 1 has exposure 1 and 0 claims, record 2 has exposure 1 and 1 claim, and record 3 has exposure 8 and 8 claims. Find the fitted common frequency using the correct count model. Then calculate the estimate produced by the incorrect procedure of fitting a common Poisson mean to the counts while using exposure as an observation weight.

Step 1: Fit the correct model. With one common rate $\lambda$, use $$ N_i\sim\operatorname{Poisson}(e_i\lambda). $$ The likelihood score for $\lambda$ sets total observed claims equal to total expected claims: $$ 0+1+8=\lambda(1+1+8). $$ Thus $\widehat{\lambda}=9/10=0.900$ claims per exposure unit. The fitted expected counts are $1(0.900)=0.900$, $1(0.900)=0.900$, and $8(0.900)=7.200$, which sum to 9.

Step 2: Apply the incorrect weighted-count procedure. If the response remains $N_i$, the model has one common Poisson mean $\mu$, and the counts are weighted by $e_i$, the weighted intercept-only estimate is the weighted average of the counts: $$ \widehat{\mu}_{\text{wrong}} =\frac{\sum_i e_iN_i}{\sum_i e_i} =\frac{1(0)+1(1)+8(8)}{1+1+8} =\frac{65}{10}=6.500. $$ This is a fitted mean count per record under that incorrectly specified model, not a frequency per exposure unit.

Check and interpretation. The correct model predicts 0.900 claims for one exposure unit and 7.200 claims for eight exposure units. The weighted-count procedure instead returns a common mean of 6.500 claims for every record and never makes the mean proportional to exposure. The discrepancy is not a rounding issue; it is a different model.

Method and mistake avoided. Use the exposure offset for counts. Weighting the count response by exposure does not insert exposure into the mean and can let a high-exposure record with a high count dominate the fitted count mean.

When exposure weights can reproduce an offset fit

There is a closely related formulation in which exposure does appear as a weight: model the observed rate $R_i=N_i/e_i$, rather than the count $N_i$, and use exposure as the weight. Under the Poisson count assumptions above, this gives the same coefficient estimates as the offset count model when the weights are used in the corresponding Poisson rate fitting criterion.

The reason follows directly from the count-model score. Since $R_i=N_i/e_i$, the offset-model score can be rewritten as

\sum_i x_i\{N_i-e_i\lambda_i\} =\sum_i e_ix_i\{R_i-\lambda_i\},\qquad \lambda_i=\exp(x_i^\mathsf{T}\beta).

The right-hand side is the exposure-weighted rate score. Thus the two specifications have the same score equations for $\beta$. Under the stated model, the rate $R_i$ has conditional mean $\lambda_i$ and variance $\lambda_i/e_i$: more exposure makes the measured rate more precise. An exposure-weighted rate fit reflects that precision. It is not the same as applying exposure weights to the original count response.

This equivalence requires a rate response $R_i=N_i/e_i$, the corresponding Poisson variance and fitting criterion, and weights proportional to exposure. It does not mean that every kind of GLM weight has the same interpretation or that all software options called “weights” are interchangeable. In particular, frequency weights used to represent replicated observations and other analytic or precision weights may have different definitions. Check the response scale, variance assumption, and weight convention before interpreting a fit.

Worked Example: Matching an Offset Fit With Weighted Rates

A hypothetical insurer records four risks. In Group A, one record has 1 claim over 1 exposure unit and another has 6 claims over 6 units. In Group B, one record has 0 claims over 1 unit and another has 4 claims over 4 units. Fit separate group rates in two ways: using counts with an offset, and using rates with exposure weights. Verify that the estimates agree and calculate the Group B log-rate coefficient relative to Group A.

Step 1: Fit counts with offsets. Group A has $1+6=7$ claims over $1+6=7$ units, so $$ \widehat{\lambda}_A=\frac{7}{7}=1.000. $$ Group B has $0+4=4$ claims over $1+4=5$ units, so $$ \widehat{\lambda}_B=\frac{4}{5}=0.800. $$

Step 2: Fit rates with exposure weights. The two Group A rates are $1/1=1$ and $6/6=1$. Their exposure-weighted mean is $$ \frac{1(1)+6(1)}{1+6}=\frac{7}{7}=1.000. $$ The Group B rates are $0/1=0$ and $4/4=1$. Their exposure-weighted mean is $$ \frac{1(0)+4(1)}{1+4}=\frac{4}{5}=0.800. $$ Both groups’ weighted-rate estimates match their offset-model estimates.

Step 3: Calculate and check the coefficient. With Group A as reference, the Group B coefficient is $$ \widehat{\beta}_B=\log(0.800)-\log(1.000)=\log(0.800)=-0.2231. $$ Exponentiating gives a relative rate of $\exp(-0.2231)\approx0.800$, agreeing with the ratio $0.800/1.000$. The offset model’s expected counts are $1.000$ and $6.000$ for the Group A records, and $0.800$ and $3.200$ for Group B. Within each group, the expected counts sum to its observed total.

Method and mistake avoided. The weighted formulation is valid here because the response is the rate and the exposure is its weight. Applying these same weights to the untransformed counts would be a different fitting problem.

Practical checks and exam traps

When specifying or auditing a frequency model, identify the response before deciding what exposure does. If the response is a count, exposure ordinarily enters as a log offset under the proportional-mean assumption. If the response is a rate, exposure can be used as a weight in the corresponding Poisson rate formulation. Do not silently switch between these response scales.

A useful check for an intercept-only offset model is that the fitted frequency equals total claims divided by total exposure. For a model with rating groups, apply the same check within each group when each group has its own rate parameter. Also verify that a record with twice the exposure and identical predictors has twice the fitted expected count.

Common exam trap: An exposure value does not have a single role independent of the response. In $N_i\sim\operatorname{Poisson}(e_i\lambda_i)$, it scales the mean count. In a rate model with response $N_i/e_i$, it weights the rate information. In neither case should “use exposure” be treated as an instruction to multiply the observed count by exposure and call the result an offset.

Key takeaway: For claim counts, an exposure offset puts $\log(e_i)$ in the linear predictor and makes expected counts proportional to exposure. Exposure weights can yield the same coefficient estimates in the corresponding model for rates, but weighting the count response does not substitute for an offset.

Check Your Understanding

For each problem, identify whether the requested calculation uses a count offset or an exposure-weighted rate, and check the units of the result.

A hypothetical risk has 9 claims over 12 policy-years. Under an intercept-only Poisson model with an exposure offset, what frequency is estimated? (Answer: $9/12=0.750$ claims per policy-year.)
Two hypothetical territories have 12 claims over 24 exposure units and 9 claims over 12 exposure units. With separate rates and the first territory as reference, find both fitted rates and the second territory’s rate relativity. (Answer: Rates $0.500$ and $0.750$; relativity $0.750/0.500=1.500$.)
For a rate model, two records in one rating group have $(N,e)=(2,2)$ and $(3,6)$. Find the exposure-weighted rate estimate and the equivalent count-offset estimate. (Answer: Weighted rate $\{2(1)+6(0.5)\}/8=5/8=0.625$; offset estimate $(2+3)/(2+6)=0.625$.)
An intercept-only count model has three records with exposures $2,3,5$ and counts $1,0,7$. Find the fitted frequency and each record’s fitted expected count. (Answer: Frequency $8/10=0.800$; expected counts $1.600,\ 2.400,\ 4.000$, respectively.)
A colleague uses exposures $1,2,4$ as weights on count responses $0,2,5$, while fitting a common Poisson mean with no offset. What estimate does that procedure produce, and why is it not the count frequency estimate? (Answer: Weighted count mean $\{1(0)+2(2)+4(5)\}/(1+2+4)=24/7\approx3.429$ claims per record; it is not a frequency because exposure was used as a weight on counts rather than as an offset in their expected means.)

What You'll Learn