Common Errors in Severity Weights and Coverage Limits

Why Severity Specification Errors Matter

A severity model is only interpretable when the response matches the quantity the insurer wants to model and the way observations entered the data set. A claim count, a ground-up loss, a payment after a deductible, and an incurred amount may all be recorded in dollars or counts, but they are not interchangeable responses. Confusing them can change the fitted mean, the likelihood, or the meaning of a prediction.

This tutorial focuses on four recurring specification errors: using claim counts as severity exposure offsets, ignoring deductible-based selection, treating capped payments as exact ground-up losses, and mixing paid with incurred targets. The central discipline is to define the observation and response first, then choose weights, offsets, and likelihood contributions that match them.

Definition: A severity response is the monetary amount whose distribution or conditional mean is being modeled. A weight changes how much an observation contributes to a fitting criterion; an offset is a known component of a linear predictor, commonly used when an expected count is proportional to exposure. Neither term changes what the response means.

Claim Counts Are Not Severity Offsets

Suppose individual claim severities are grouped by rating category, and category $j$ has $n_j$ observed claims with sample mean $\bar{x}_j$. If the goal is the overall mean severity across all observed claims, each claim—not each category—has equal weight. The appropriate grouped-data calculation is therefore

\bar{x}_{\mathrm{all}}=\frac{\sum_j n_j\bar{x}_j}{\sum_j n_j}.

For a grouped severity model, claim counts may serve as frequency weights for category means when the fitting method and variance assumptions justify that representation. They are not automatically an exposure offset. In a claim-count model with a log link, an offset such as $\log(e_j)$ can represent known exposure $e_j$ when expected claim count is proportional to exposure. A severity mean, by contrast, does not generally increase just because more claims were observed.

Worked Example: Weighting Grouped Severity Means

A hypothetical portfolio has two categories. Category A has 2 claims with mean severity $1,200; Category B has 8 claims with mean severity $1,800. Estimate the mean across the 10 observed claims. Assume claim severities within each category are independent Gamma variables with shape 2 and scale equal to the category mean divided by 2. Thus the category means are $1,200 and $1,800. No interest or mortality assumption is relevant.

Method. Recover the total dollars represented by each category mean, then divide by the total claim count.

Category A represents $2\times \$1{,}200=\$2{,}400$. Category B represents $8\times \$1{,}800=\$14{,}400$. The total is $\$2{,}400+\$14{,}400=\$16{,}800$, and the total count is $2+8=10$. Therefore,

\bar{x}_{\mathrm{all}}=\frac{\$16{,}800}{10}=\$1{,}680.

The unweighted average of the two category means is $(\$1{,}200+\$1{,}800)/2=\$1{,}500$, which gives each category equal influence rather than each claim equal influence. As a check, the weighted result lies between the category means and is closer to $1,800, the mean for the category containing most claims.

Common mistake avoided: using the count as a log offset in a severity-mean predictor would make the predicted severity change with the number of claims, even if the category and its severity characteristics were unchanged. The count is a weight in this grouped calculation, not a severity exposure offset.

The weighting interpretation has limits. Replacing individual severities with group means and counts is not automatically identical to every individual-level likelihood. It is appropriate only when the model’s likelihood or estimating equations support that aggregation; within-group variation and dispersion information may matter. When individual claim data are available, fitting those observations directly avoids losing that information.

Deductible Selection Changes the Observed Distribution

Let $X$ be the ground-up loss, and suppose a claim is observed only when it exceeds an ordinary deductible $d$. The observed ground-up severities then come from the conditional distribution of $X$ given $X>d$, not from the original distribution of $X$. For a continuous loss distribution, the conditional density is

f_{X\mid X>d}(x)=\frac{f_X(x)}{S_X(d)},\qquad x>d.

The denominator is the probability of selection. Omitting it treats the sample as though losses below the deductible could have appeared. If the recorded response is the payment after the deductible, $Y^L=X-d$ for a claim above $d$, its conditional density is $f_X(y+d)/S_X(d)$ for $y>0$, under the same continuous-loss assumption.

Why the truncation adjustment appears: For any $x>d$, the conditional probability density is the joint density of observing $X=x$ and satisfying the selection rule, divided by the probability of satisfying the rule. Since $X=x>d$ already meets the rule, that density is $f_X(x)$; the selection probability is $P(X>d)=S_X(d)$. Hence the conditional density is $f_X(x)/S_X(d)$.

Worked Example: Correcting for Deductible Truncation

A hypothetical insurer observes only ground-up claims exceeding a $5,000 deductible. The two observed losses are $8,000 and $11,000. Assume ground-up losses follow an exponential distribution with mean parameter $\theta$, so $f_X(x)=\theta^{-1}e^{-x/\theta}$ and $S_X(x)=e^{-x/\theta}$ for $x\geq 0$. No interest or mortality assumption is relevant. Estimate $\theta$ by maximum likelihood, accounting for the selection rule.

Method. Use the density conditional on exceeding the deductible. For an observed $x_i>5{,}000$, its conditional density is

\frac{f_X(x_i)}{S_X(5{,}000)} =\frac{\theta^{-1}e^{-x_i/\theta}}{e^{-5{,}000/\theta}} =\theta^{-1}e^{-(x_i-5{,}000)/\theta}.

The two observed excesses above the deductible are $\$8{,}000-\$5{,}000=\$3{,}000$ and $\$11{,}000-\$5{,}000=\$6{,}000$. The likelihood is

L(\theta)=\theta^{-2}\exp\left(-\frac{\$3{,}000+\$6{,}000}{\theta}\right) =\theta^{-2}e^{-\$9{,}000/\theta}.

The log-likelihood is $\ell(\theta)=-2\log\theta-\$9{,}000/\theta$. Differentiating gives $\ell'(\theta)=-2/\theta+\$9{,}000/\theta^2$. Setting this equal to zero yields $2\theta=\$9{,}000$, so the estimate is $\hat{\theta}=\$4{,}500$. The second derivative at this value is negative, confirming a maximum. The estimate is the average observed excess, $(\$3{,}000+\$6{,}000)/2=\$4{,}500$, as expected from the exponential distribution’s memoryless property.

If truncation were ignored, the ordinary sample mean would be $(\$8{,}000+\$11{,}000)/2=\$9{,}500$. That calculation incorrectly treats the sample as if losses at or below $5,000 could have been selected. The conditional-likelihood answer is checked by noting that it estimates the mean excess among claims selected above the deductible, not the average of the observed ground-up amounts.

A Payment at a Limit Is Censored, Not Necessarily Exact

With no deductible and a per-loss payment limit $u$, the payment is $Y^L=\min(X,u)$. If a recorded payment is below $u$, it identifies the ground-up loss as that amount. A recorded payment equal to $u$, however, tells us only that the ground-up loss is at least $u$. Under a continuous ground-up model, the likelihood contribution at the cap is the tail probability $P(X\geq u)=S_X(u)$. More generally,

P(Y^L=u)=P(X\geq u)=S_X(u)+P(X=u).

The final equality uses the convention $S_X(u)=P(X>u)$. Thus $P(X\geq u)$ equals $S_X(u)$ only when there is no atom at $u$. For a continuous model this condition holds, but it should not be assumed for a distribution that may have point mass at the limit. Treating each capped payment as an exact ground-up loss understates the information in the observation: the loss could be larger than the limit.

Worked Example: Fitting a Limit-Censored Exponential Severity

A hypothetical policy has a $10,000 per-loss limit and no deductible. Two payments are observed: $4,000 and $10,000. Assume ground-up losses are independent exponential variables with mean $\theta$, a continuous distribution with density $f_X(x)=\theta^{-1}e^{-x/\theta}$. No interest or mortality assumption is relevant. Estimate $\theta$ using the limit information.

Method. The $4,000 payment is below the cap, so its likelihood contribution is the exact density $f_X(\$4{,}000)$. The $10,000 payment is capped, so its contribution is $P(X\geq \$10{,}000)$. Continuity means $P(X=\$10{,}000)=0$, making this probability $S_X(\$10{,}000)$. Therefore,

L(\theta)=\left(\frac{1}{\theta}e^{-\$4{,}000/\theta}\right)e^{-\$10{,}000/\theta} =\theta^{-1}e^{-\$14{,}000/\theta}.

The log-likelihood is $-\log\theta-\$14{,}000/\theta$. Its derivative is $-1/\theta+\$14{,}000/\theta^2$, which is zero when $\hat{\theta}=\$14{,}000$. The second derivative at the estimate is negative. The censored estimate can also be checked from the likelihood: one exact loss contributes $4,000 and the one capped observation contributes a threshold of $10,000, for a total of $14,000 in the exponential likelihood.

A common incorrect calculation treats both payments as exact ground-up losses and reports their mean, $(\$4{,}000+\$10{,}000)/2=\$7{,}000$. That ignores the possibility that the capped loss exceeded $10,000. Under a model with an atom at the cap, the capped observation’s contribution must instead use $P(X\geq \$10{,}000)$, including that atom; it is not automatically $S_X(\$10{,}000)$.

Paid and Incurred Amounts Are Different Targets

A paid amount is the amount disbursed by a valuation date. An incurred amount commonly combines cumulative paid amounts with a case reserve at that date. An ultimate loss is the eventual total loss after development is complete. Those targets answer different questions, so observations labeled paid, incurred, and ultimate should not be pooled as if they were measurements of one response.

For example, a model trained on paid-to-date amounts describes payment as of the specified valuation date, including the effect of claim development and settlement timing. A model trained on incurred amounts describes the recorded paid-plus-reserve measure, which can change as case estimates are revised. If the objective is ultimate severity, partially developed paid or incurred amounts need an appropriate development or estimation treatment; simply relabeling them as ultimate does not make them so.

Worked Example: Keeping Paid and Incurred Summaries Separate

For two hypothetical open claims at a valuation date, paid amounts are $4,000 and $6,000; corresponding incurred amounts are $9,000 and $10,000. The incurred figures include case reserves. For context, suppose ultimate ground-up severities are modeled as Gamma with shape 2 and scale $5,000, giving mean $\$10{,}000$. The two records are not assumed to be fully developed ultimate observations. No interest or mortality assumption is relevant.

Method. Compute each response’s mean separately. Total paid is $\$4{,}000+\$6{,}000=\$10{,}000$, so mean paid-to-date is $\$10{,}000/2=\$5{,}000$. Total incurred is $\$9{,}000+\$10{,}000=\$19{,}000$, so mean incurred is $\$19{,}000/2=\$9{,}500$. The implied case reserves are $\$9{,}000-\$4{,}000=\$5{,}000$ and $\$10{,}000-\$6{,}000=\$4{,}000$, for total reserves of $9,000. This checks because paid plus reserves is $\$10{,}000+\$9{,}000=\$19{,}000$, the incurred total.

The Gamma mean of $10,000 is a stated model assumption for ultimate severity; it is not obtained by averaging either of these partially developed responses. Pooling the four dollar amounts would produce a mean of $(\$4{,}000+\$6{,}000+\$9{,}000+\$10{,}000)/4=\$7{,}250$, but that number has no clear severity interpretation: it treats two paid observations and two incurred observations as four comparable claims.

Common mistake avoided: combining paid and incurred responses in a single severity fit without defining a common target. First choose the target and valuation basis; then include only compatible observations or explicitly model the differences.

A Practical Specification Check

Before fitting or interpreting a severity model, write down how each record was generated. A short audit often reveals that the observed values are selected, transformed, censored, or measured at different development stages.

Name the response.
State whether the model concerns ground-up loss, payment per loss, payment per payment, paid-to-date, incurred, or another explicitly defined amount.

Describe selection and coverage.
Record deductibles, limits, and claim eligibility rules. Use a conditional distribution for truncated observations and a censoring contribution for capped observations when the data only establish a threshold.

Assign weights and offsets by purpose.
Use claim counts as weights for grouped severity information only when the fitting method supports it. Use an exposure offset in a count model only when its mean structure justifies proportionality to exposure.

Check target consistency.
Verify that paid, incurred, and ultimate amounts refer to the same valuation basis before pooling or comparing them.

These checks matter because a numerical fit can converge even when its response or likelihood is wrong. A fitted coefficient is not automatically an estimate of the intended severity quantity: its meaning follows from the data definition and the model specification together.

Key takeaway: Claim counts can weight grouped severities without acting as severity offsets; deductible selection requires a conditional model; a limit payment represents $X\geq u$, not necessarily $X=u$; and paid, incurred, and ultimate amounts must remain distinct unless a model explicitly reconciles them.

Check Your Understanding

For each problem, identify the response and use a calculation or likelihood consistent with how the observations were recorded.

A hypothetical portfolio has 3 claims in class A with mean severity $2,000 and 9 claims in class B with mean severity $1,400. What is the mean across all 12 claims? (Answer: $1,550.)
Ground-up losses follow an exponential distribution with mean $\theta$. Claims are observed only above a $2,000 deductible, and observed ground-up losses are $5,000, $8,000, and $10,000. What is the conditional maximum-likelihood estimate of $\theta$? (Answer: $5,667 to the nearest dollar.)
With no deductible, a hypothetical policy pays up to $12,000. For a continuous exponential ground-up model, what likelihood contribution does a payment of exactly $12,000 provide? (Answer: $S_X(\$12{,}000)=e^{-\$12{,}000/\theta}$.)
A limit-censored exponential sample has one exact payment of $3,000 below a $9,000 cap and one payment at the cap. What is the maximum-likelihood estimate of the exponential mean? (Answer: $12,000.)
Three open claims have paid amounts of $2,000, $5,000, and $7,000, and incurred amounts of $6,000, $8,000, and $10,000. Find both sample means and state whether pooling the six responses gives a paid or incurred mean. (Answer: mean paid is $4,667; mean incurred is $8,000; the pooled mean is $6,333 and is neither a paid nor an incurred mean.)

What You'll Learn