Severity Weights Are Not Exposure Offsets

Why Severity Weights and Offsets Must Be Distinguished

A claim-severity model predicts the expected payment for a claim, conditional on the modeled rating characteristics and on the definition of the claim being analyzed. In some data sets, the observations are claim amounts; in others, they are averages for groups of policies or exposure units. A modeler may assign weights so that observations representing more volume have more influence on the fit. That does not, by itself, mean the expected claim severity should increase with that volume.

An offset has a different role. In a generalized linear model (GLM), an offset is a known term added to the linear predictor with its coefficient fixed at one. With a log link, an offset therefore multiplies the modeled mean by a known factor. A weight instead affects how strongly an observation contributes to model fitting. Confusing the two can change the target being modeled and produce predictions on the wrong scale.

Definition: For observations $y_j$, a weight $w_j$ is a nonnegative quantity that controls the observation’s contribution to the fitting criterion. For example, a weighted deviance criterion has the form $\sum_j w_j D(y_j,\mu_j)$, where $D$ is the observation’s deviance contribution and $\mu_j$ is its modeled mean. An offset $o_j$ is a known predictor term whose coefficient is fixed at one. Under a link function $g$, the model is $g(\mu_j)=x_j^\mathsf{T}\beta+o_j$. A weight changes the contribution to fitting; an offset changes the specified mean.

What a Weight Means in a Severity Model

The unit represented by each row matters. If each row is one claim, an unweighted claim-level fit gives each observed claim the same fitting weight, subject to any other model assumptions. If each row is a cell-level average, a weight can reflect how much policy or exposure volume that average represents. Then the fit gives greater influence to averages representing more of the population, according to the chosen weighting scheme.

The choice of weights also determines what the fitted result represents. Weighting group averages by policy exposure targets a policy-exposure-weighted average of those group averages. Weighting claim observations equally targets a claim-weighted fit. These need not be the same: groups may have different claim frequencies, and the number of claims in a group need not be proportional to its policy exposure.

In GLM software, “weight” can refer to different conventions, such as prior, frequency, or variance weights. Their precise likelihood or variance interpretation depends on the model and implementation. Integer frequency weights can represent repeated identical observations under an appropriate likelihood; other weights may have a different interpretation. Always identify what a weight represents and how the fitting procedure uses it. Do not infer that a weight is an exposure offset merely because it is measured in exposure units.

A Short Derivation: Weighted Intercept-Only Fit

The difference between weighting and an offset is visible even in an intercept-only model. Suppose the response is modeled by a single common mean $\mu$, and a weighted squared-error criterion is used. The weights affect the objective:

Q(\mu)=\sum_j w_j(y_j-\mu)^2.

Differentiating with respect to $\mu$, setting the derivative to zero, and solving gives the fitted mean:

\frac{dQ}{d\mu}=-2\sum_j w_j(y_j-\mu)=0 \quad\Longrightarrow\quad \widehat{\mu}=\frac{\sum_j w_jy_j}{\sum_j w_j}.

Thus, weights change the relative influence of the observed responses. They do not make the model predict $w_j\widehat{\mu}$ for observation $j$. By contrast, a log-link model with offset $o_j$ specifies $\mu_j=\exp(x_j^\mathsf{T}\beta+o_j)$; the offset is inside the predictor and therefore changes the mean itself.

The weighted-mean result also holds for an intercept-only Gamma likelihood with common dispersion and log link, when the observations contribute with the stated likelihood weights. In that model, the score equation for the intercept is proportional to $\sum_j w_j(y_j/\mu-1)=0$, which gives the same weighted arithmetic mean. This result depends on the stated model and criterion; it is not a universal formula for every GLM or every type of weight.

Worked Example: Weighting Group-Level Severity Averages

A hypothetical insurer has two rating segments. Each row records the segment’s average claim severity and its policy exposure volume. For an intercept-only severity fit, assume the segment averages are treated as independent Gamma observations with common dispersion and likelihood weights equal to the stated policy exposure units. Estimate the common mean using the log-link Gamma model, and interpret the result.

Segment A has average severity $y_A=\$700$ and weight $w_A=2$. Segment B has average severity $y_B=\$1{,}100$ and weight $w_B=6$. The model has one intercept and no offset. Under the assumptions above, the fitted intercept-only mean is the weighted arithmetic mean:

\widehat{\mu} =\frac{2(700)+6(1{,}100)}{2+6} =\frac{1{,}400+6{,}600}{8} =\frac{8{,}000}{8} =\$1{,}000.

For a log link, the fitted intercept is $\widehat{\beta}_0=\log(1{,}000)$, approximately $6.9078$. The predicted severity is $\exp(6.9078)$, approximately $\$1{,}000$, with the displayed logarithm rounded to four decimal places. The result is the exposure-weighted fit to the two segment averages under the stated model; it is not a prediction that either segment’s severity is multiplied by its weight.

An unweighted fit would give $(700+1{,}100)/2=\$900$, so the weighting choice has a material effect on the common fitted mean. The common mistake is to treat the weight 6 as a multiplier on Segment B’s mean, predicting $\$6{,}600$. That is not what this weighted fit specifies: the weight affects Segment B’s contribution to estimation, while the model still has one common mean.

Why an Exposure Offset Is Different

For a log-link GLM with a known offset $o_j$, the mean is

\mu_j=\exp(x_j^\mathsf{T}\beta+o_j) =\exp(x_j^\mathsf{T}\beta)\exp(o_j).

An offset of $\log(e_j)$, for a positive known quantity $e_j$, therefore gives $\mu_j=e_j\exp(x_j^\mathsf{T}\beta)$. This is a modeling assumption that the expected response scales in direct proportion to $e_j$, conditional on the other predictors. It is not a way to say that observations with larger $e_j$ should count more in the fitting process.

In a severity model, multiplying expected claim severity by policy exposure would usually change the modeled quantity from severity per claim toward an amount that scales with exposure. Such a specification requires a response and model for which that scaling is appropriate. It should not be introduced merely because policy exposure is available or because exposure was used as a weight.

Worked Example: Comparing a Weight With an Offset

A hypothetical analyst has two segment-level severity averages: $y_1=\$600$ with exposure weight $w_1=1$, and $y_2=\$1{,}000$ with exposure weight $w_2=4$. Compare (a) a weighted intercept-only Gamma fit with common dispersion and no offset, and (b) an unweighted Gamma fit that incorrectly uses $\log(e_j)$ as an offset in the severity mean, where $e_1=1$ and $e_2=4$. For part (b), use a common intercept $\beta$, so $\mu_j=e_j\exp(\beta)$.

(a) Weighted severity fit. The common fitted severity is

\widehat{\mu} =\frac{1(600)+4(1{,}000)}{1+4} =\frac{600+4{,}000}{5} =\$920.

Both segment observations are fitted to the same mean of $\$920$, because this model has only an intercept. The exposure weights influence the fit; they do not make the Segment 2 mean four times as large.

(b) Offset model. This is a different model: its means are $\mu_1=\exp(\beta)$ and $\mu_2=4\exp(\beta)$. For an unweighted Gamma likelihood with common dispersion, the intercept score equation is proportional to $\sum_j(y_j/\mu_j-1)=0$. Substituting $\mu_j=e_j\exp(\beta)$ gives

\frac{600}{\exp(\beta)}+\frac{1{,}000}{4\exp(\beta)}-2=0, \qquad \exp(\beta)=\frac{600/1+1{,}000/4}{2} =\frac{600+250}{2} =\$425.

The fitted means are therefore $\widehat{\mu}_1=1(425)=\$425$ and $\widehat{\mu}_2=4(425)=\$1{,}700$. These differ sharply from the weighted intercept-only estimate of $\$920$, because the offset imposes proportional scaling of the predicted severity. The common mistake is to put $\log(4)$ in the severity predictor just because Segment 2 has four times the exposure weight. The correction is to use weights when the intended effect is influence on fitting, and an offset only when the mean itself is assumed to scale by the known factor.

As a check, the offset model’s two fitted means have the required ratio $1:4$, while the weighted intercept-only model has equal fitted means across segments. The models answer different questions; neither calculation can be substituted for the other without changing the specification.

Choosing the Unit Before Applying Exposure Weights

A subtle error occurs when exposure volume is attached repeatedly to claim-level records even though the intended target is a policy-exposure-weighted average of group severities. In that case, each claim in a high-exposure segment may receive the segment’s full exposure weight. The segment’s influence then depends on both its exposure and its number of observed claims, which may not be the intended weighting scheme.

Worked Example: Avoiding Double Weighting of Claim-Level Data

A hypothetical portfolio has two segments. Segment A has two observed claim severities, $\$500$ and $\$700$, and 10 policy exposure units. Segment B has one observed claim severity, $\$1{,}500$, and 2 policy exposure units. First calculate a claim-weighted mean from the individual claims. Then calculate the policy-exposure-weighted mean of the two segment averages. Finally, show what happens if every claim in a segment is incorrectly given the segment’s entire exposure as its weight.

Claim-weighted mean. The three individual claim observations have equal weight, so

\frac{500+700+1{,}500}{3} =\frac{2{,}700}{3} =\$900.

Policy-exposure-weighted mean of segment averages. Segment A’s average is $(500+700)/2=\$600$, and Segment B’s average is $\$1{,}500$. Weighting those two averages by their policy exposures gives

\frac{10(600)+2(1{,}500)}{10+2} =\frac{6{,}000+3{,}000}{12} =\$750.

Incorrect repeated segment weights on claim rows. Giving each of A’s two claims weight 10 and B’s claim weight 2 produces

\frac{10(500)+10(700)+2(1{,}500)}{10+10+2} =\frac{5{,}000+7{,}000+3{,}000}{22} =\frac{15{,}000}{22} \approx \$681.82.

This last calculation weights Segment A’s exposure twice, once for each observed claim, and therefore does not equal the policy-exposure-weighted mean of the segment averages. The three results answer different questions: $\$900$ is claim-weighted, $\$750$ is policy-exposure-weighted across the segment averages, and $\$681.82$ uses repeated segment weights on claim rows. The common mistake is to treat the segment’s full exposure as a weight on every claim and assume the result still represents an exposure-weighted segment average. The correction is to define the observation unit and target first, then assign weights at that level.

Practical Checks and Exam Traps

Before fitting a severity GLM, write down the response and the unit of observation: for example, individual claim payment, average payment per segment, or another explicitly defined severity measure. Then state what each weight represents and whether it should affect the fitting criterion. Separately, state the offset’s predictor-scale value and why the conditional mean should scale by that known factor.

Do not read a weight as a multiplier on the fitted mean. A weight can change coefficient estimates through the fitting criterion, but the mean is still determined by the link and linear predictor.
Do not assume that policy exposure and claim count are interchangeable. Exposure-weighted severity and claim-weighted severity can differ when claim frequency varies across groups.
Do not give a group’s full volume weight to every claim in that group without checking the target. This can make the group’s influence depend on its number of claims as well as its exposure.
Do not add $\log(e_j)$ as a severity offset just because $e_j$ was used as a weight. An offset asserts proportional scaling of the mean; it is not a second way of expressing influence on fitting.
Check the software’s weight convention. A frequency weight, a precision weight, and a prior weight can have different interpretations even if they appear in a similar input field.

Key takeaway: A severity weight answers, “How much should this observation contribute to fitting?” An offset answers, “By what known factor should the modeled mean change?” Specify the observation unit, target population, and model assumptions before using either.

Check Your Understanding

For each problem, identify whether the stated quantity acts as a weight or an offset and calculate the requested result.

Two segment-average severities are $\$800$ and $\$1{,}200$, with likelihood weights 2 and 3. Under an intercept-only Gamma model with common dispersion and no offset, what is the fitted mean? (Answer: $\$1{,}040$)
Under an intercept-only log-link Gamma model with common dispersion and no offset, observations are $\$400$ and $\$1{,}000$, with likelihood weights 2 and 3. Find the fitted mean and its log-link intercept, rounded to four decimals. (Answer: fitted mean $\$760$; intercept $\log(760)\approx 6.6333$)
A log-link severity model has base predicted mean $\$300$. If a justified offset is $\log(0.5)$ for one observation and $\log(2)$ for another, calculate the two predicted means. (Answer: $\$150$ and $\$600$)
Segment A has claims of $\$400$ and $\$800$ with 6 exposure units; Segment B has one claim of $\$1{,}000$ with 2 exposure units. What is the exposure-weighted mean of the segment averages? (Answer: $[6(600)+2(1{,}000)]/8=\$700$)
Two segment-average severities are $\$900$ and $\$1{,}200$, with weights 2 and 5. First fit an intercept-only Gamma model with those likelihood weights and no offset. Then, separately, fit an unweighted Gamma model with log-link offset $\log(e_j)$, where the corresponding $e_j$ values are 2 and 5, and a common intercept. What are the fitted common mean in the weighted model and the two means in the offset model? (Answer: weighted mean $\$1{,}114.29$; offset-model means $\$690$ and $\$1{,}725$)

What You'll Learn