Correcting for Policy-Limit Censoring

Why Policy Limits Censor Severity Data

A payment equal to a policy limit does not generally reveal the ground-up claim amount. If a hypothetical policy has a limit of $25,000 and pays $25,000, the underlying loss could be $25,000 or substantially greater. The payment is therefore a lower bound on severity, not necessarily an exact observation. Treating it as exact makes the observed severity distribution appear lighter in the tail than the ground-up distribution.

This issue matters when estimating an exposure curve from historical claims. A curve built from fully observed ground-up claims can use each claim amount directly. A curve built from limited payments must account for which claims are exact and which are only known to have reached their limits. The correction is not to replace a capped amount with an invented higher amount: it is to use a censored-data method and report estimates only where the observations support them.

Definition: A claim is right-censored at limit $L$ when its ground-up loss $X$ is not observed exactly, but the data establish that $X\ge L$. The recorded payment is $L$. An uncensored claim has an observed ground-up amount $X<L$. In what follows, severity is assumed nonnegative and continuous, so an exact loss equal to the limit has probability zero.

For a claim observed at limit $L$, write $T=\min(X,L)$. The censoring indicator is $\delta=1$ when the ground-up claim is observed exactly and $\delta=0$ when the record is censored at $L$. Thus, an observed amount of $18,000 with $\delta=1$ says the claim was $18,000; an amount of $25,000 with $\delta=0$ says only that the claim was at least $25,000.

The Kaplan–Meier Estimate for Censored Claims

A standard nonparametric method for right-censored observations is the Kaplan–Meier product-limit estimator. It combines the exact claims with the information in censored claims: a censored claim does not count as an event, but it leaves the risk set at its censoring limit. This method requires censoring to be independent of severity, either unconditionally or after conditioning on relevant characteristics. In ratemaking data, that assumption can fail if policy limits are selected in a way related to expected severity.

Definition: At each observed exact claim amount $t_j$, let $Y_j$ be the number of claims still under observation immediately before $t_j$, and let $d_j$ be the number of exact claims at $t_j$. The Kaplan–Meier estimate of the ground-up survival function is $$ \widehat{S}_X(t)=\prod_{t_j\le t}\left(1-\frac{d_j}{Y_j}\right). $$ A censored observation reduces later risk sets but does not appear in the event count $d_j$.

The formula follows from multiplying conditional survival probabilities. Among the $Y_j$ claims still observed just before $t_j$, $d_j$ have an exact claim at that amount. The estimated fraction that survives past that event time is therefore $(Y_j-d_j)/Y_j=1-d_j/Y_j$. To estimate survival past several event times, multiply these successive conditional fractions. A censoring observation contributes no event at its limit; it simply stops contributing to later risk sets.

For a nonnegative loss, the limited expected loss at limit $u$ is the area under the ground-up survival function through $u$:

E[\min(X,u)]=\int_0^u S_X(t)\,dt.

Replacing $S_X(t)$ with $\widehat{S}_X(t)$ gives a nonparametric estimate of limited expected loss, provided the data support the calculation through $u$. Between observed exact claims, the Kaplan–Meier estimate is constant, so the integral is calculated as a sum of rectangles. A nonzero survival estimate after the last exact claim is not evidence that the tail ends there; it means the remaining severity is unresolved by the observations.

Worked Example: A Limited Expected Loss With Varying Limits

A hypothetical dataset contains six claims, all measured in thousands of dollars. The exact observations are $10,000, $20,000, and $40,000. The other three records are censored at policy limits of $15,000, $30,000, and $50,000. Estimate the limited expected loss at $u=$$25,000 using the Kaplan–Meier method.

Method and assumptions. Treat the three exact amounts as events and the three policy-limit records as right-censored. Assume censoring is independent of severity and claim amounts are continuous. Work in thousands of dollars.

Step 1: Calculate the risk sets and survival estimates. At $10,000, all six claims remain under observation, so $Y_1=6$, $d_1=1$, and the survival estimate becomes $1(1-1/6)=5/6$. The claim censored at $15,000 then leaves the risk set. At $20,000, four claims remain, so $Y_2=4$, $d_2=1$, and the estimate becomes $(5/6)(1-1/4)=5/8$. The claim censored at $30,000 is beyond the requested limit. The later event at $40,000 does not affect survival through $25,000.

Step 2: Integrate through $25,000. The estimated survival is 1 from 0 to 10, $5/6$ from 10 to 20, and $5/8$ from 20 to 25, with amounts in thousands:

\widehat{E}[\min(X,25)] =10(1)+10\left(\frac{5}{6}\right)+5\left(\frac{5}{8}\right) =10+\frac{25}{3}+\frac{25}{8} =21.458333\ldots

Step 3: Report and check. The estimated limited expected loss is $21,458.33, rounded to the nearest cent. It is between $0 and the $25,000 limit, as it must be. A common mistake is to treat the $15,000, $30,000, and $50,000 capped records as exact claim amounts. That would incorrectly assign ground-up severities to claims known only to be at least those amounts.

Calculating an Exposure-Curve Ratio From Censored Data

Once limited expected losses at two supported limits have been estimated on the same loss basis, their ratio can be used as an exposure-curve ratio. Censoring correction does not remove the need for comparable data: claims should be evaluated on consistent coverage, valuation, and monetary bases. Nor does a ratio make an unsupported tail estimate reliable; each limited expected loss in the ratio must be estimable from the observed data.

Worked Example: Estimating a Ratio at Two Supported Limits

A second hypothetical portfolio has eight claims, with amounts in thousands of dollars: exact claims at 5, 12, 25, and 45; claims censored at limits 8, 18, 35, and 60. Estimate the ratio of limited expected loss at $u=30$ to that at $u=15$.

Method and assumptions. Use the Kaplan–Meier product-limit estimate, assuming independent censoring and continuous severity. The limits 15 and 30 are within the range over which claims remain under observation.

Step 1: Update the risk set at each event. At 5, $Y=8$, $d=1$, so survival becomes $7/8$. The claim censored at 8 leaves six claims. At 12, $Y=6$, $d=1$, so survival becomes $(7/8)(5/6)=35/48$. The claim censored at 18 leaves four claims. At 25, $Y=4$, $d=1$, so survival becomes $(35/48)(3/4)=35/64$. This estimate remains $35/64$ through 30. Events at 45 are not needed for either integral.

Step 2: Estimate limited expected loss at 15. The survival estimates on the intervals from 0 to 5, 5 to 12, and 12 to 15 are 1, $7/8$, and $35/48$, respectively:

\widehat{E}[\min(X,15)] =5+7\left(\frac{7}{8}\right)+3\left(\frac{35}{48}\right) =5+6.125+2.1875 =13.3125.

Step 3: Estimate limited expected loss at 30. From 15 to 25, survival remains $35/48$; from 25 to 30, it is $35/64$:

\widehat{E}[\min(X,30)] =13.3125+10\left(\frac{35}{48}\right)+5\left(\frac{35}{64}\right) =13.3125+7.291666\ldots+2.734375 =23.338541\ldots.

Step 4: Calculate and check the ratio.

\frac{\widehat{E}[\min(X,30)]}{\widehat{E}[\min(X,15)]} =\frac{23.338541\ldots}{13.3125} =1.75313\ldots

The estimated ratio is 1.753, rounded to three decimal places. It exceeds 1, consistent with limited expected loss being nondecreasing in the limit. A common error is to use the event-time survival estimate at 25 for the entire interval from 12 to 30; survival changes at 25, so the integral must be split there.

What Censored Data Cannot Identify

The Kaplan–Meier method corrects for right-censoring only while the risk sets provide information. If every claim is censored at the same limit, the data provide no direct information about the shape of the severity tail above that limit. More generally, an estimated survival curve that remains positive after the last exact claim cannot be integrated indefinitely as though the observed plateau were the true tail. A limited expected loss at a higher limit may require additional information or explicit modeling assumptions; it is not identified merely by recording more capped payments at the same limit.

For a censored observation at $L$, a model-based likelihood would use the probability of exceeding the limit, $S_X(L)$, rather than the density of an exact claim at $L$. Exact observations contribute density terms. This distinction expresses the same core information as the nonparametric method: a capped payment says “at least this much,” not “exactly this much.”

Worked Example: A Common Limit and an Unidentified Tail

Five hypothetical claims have recorded payments of $4,000, $9,000, $14,000, $20,000, and $20,000. The first three are exact ground-up claims; the last two are censored at a common $20,000 policy limit. Estimate the average limited loss at $12,000 from these observations, and state what can be concluded about the average ground-up severity.

Method and assumptions. For the requested limit of $12,000, each claim’s limited amount is known: the exact claims are observed below $20,000, and either censored claim must have ground-up loss of at least $20,000, which exceeds $12,000.

Step 1: Apply the $12,000 limit to each claim. The five limited amounts, in thousands, are 4, 9, 12, 12, and 12. Their sum is $4+9+12+12+12=49$.

Step 2: Calculate the average. Divide by five claims: $49/5=9.8$ thousand dollars. The empirical limited average is therefore $9,800.00, rounded to the nearest cent. It is between zero and the $12,000 limit.

Step 3: State the limit of the data. The recorded payments sum to $4+9+14+20+20=67$ thousand dollars, so the ground-up sample average is at least $67/5=13.4$ thousand dollars. It cannot be determined exactly: the two censored losses could be greater than $20,000, and the records do not reveal by how much. The common mistake is to report $13,400 as the ground-up average; it is only the average of the recorded payments and a lower bound on the ground-up sample average.

Practical Checks and Common Traps

The estimator is useful only when the claim records encode censoring correctly and the independence assumption is defensible. Before calculating, confirm that a payment at a limit is flagged as censored rather than automatically treated as an exact loss. Also verify that losses and limits use the same dollar basis and that the data refer to a consistent coverage definition. If limit selection is related to severity, a pooled Kaplan–Meier estimate may be biased; conditioning on relevant risk characteristics or using a suitable model may be needed.

Classify each observation.
Record exact ground-up claims as events and capped claims as right-censored at the applicable limit. Do not infer an exact severity from a capped payment.

Check the censoring assumption.
Ask whether the censoring limits are independent of severity, at least after accounting for relevant rating or coverage characteristics. If not, a pooled estimate may not represent the target population.

Build event-time risk sets.
At each exact claim amount, count claims still observed just before that amount. Censored claims leave later risk sets but are not counted as exact events.

Integrate only over supported limits.
Use the estimated survival curve to calculate limited expected loss, splitting the integral wherever the estimate changes. Do not extend a plateau beyond observed support without stating additional assumptions.

These distinctions are central to estimating exposure curves from limited claim data and are aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus. The exam trap is to confuse a capped payment with an exact loss: recognize the censoring, use the information the limit actually provides, and state clearly where the data stop supporting an estimate.

Check Your Understanding

For each problem, treat amounts as thousands of dollars unless stated otherwise. Assume continuous severity and independent censoring where a Kaplan–Meier estimate is requested.

Four claims have an exact event at 5, a claim censored at 8, an exact event at 10, and a claim censored at 14. Find the Kaplan–Meier survival estimate immediately after the event at 10. (Answer: $3/8=0.3750$)
Four claims have exact payments 3 and 7, and two claims censored at a common limit of 12. Find the empirical average limited loss at $u=9$. (Answer: $7$ thousand dollars)
Six observations are an exact event at 4, censored at 7, an exact event at 11, censored at 15, an exact event at 20, and censored at 25. Use the Kaplan–Meier estimate to calculate limited expected loss at $u=18$. (Answer: $14.2083$ thousand dollars)
An insurer assigns lower limits to risks it expects to have more severe claims. Can an unconditional Kaplan–Meier estimate automatically be assumed to correct the resulting censoring? State the key issue. (Answer: No; limit choice is related to severity, so censoring is informative unless relevant differences are accounted for.)
Every claim in a dataset is censored at $40,000. Can the dataset alone identify the ground-up survival curve above $40,000? (Answer: No; it only establishes that censored claims are at least $40,000 and provides no tail shape above that limit.)

What You'll Learn