Censoring Versus Missing Loss Records

Why the Distinction Matters

A loss file can contain an exact amount, a partial observation that bounds the amount, or no usable amount information at all. These cases are not interchangeable. A recorded loss of $1,000 says the loss was $1,000; a record that the loss exceeded $1,000 says only that it was greater; and a blank amount supplies no bound unless other information in the record does so.

The distinction matters because the records support different conclusions about the underlying loss variable $X$. Treating a bound as an exact amount can distort summaries of the loss distribution. Treating a missing amount as zero can add a value that was never observed. Even correctly classified records require assumptions about how censoring or missingness occurred before they can be used to describe a target population.

Definition: An exact observation reports the loss amount $X=x$. A censored observation does not report the exact amount but supplies a known constraint on $X$, such as $X>c$, $X\le c$, or $L\le X\le U$. A wholly missing loss record provides no observed amount and no bound on that amount. A record can still contain other observed information, such as policy type or reporting status.

What a Censored Record Says

Censoring is a feature of what was observed, not a claim that the loss itself has a special value. The direction and endpoint of the bound come from the wording and the recording process. In particular, “exceeded” and “at least” do not express the same inequality.

Right-censored: the loss is known to exceed a threshold. “Exceeded $1,000” implies $X>1000$. “At least $1,000” implies $X\ge 1000$.
Left-censored: the loss is known to be at or below a threshold, for example $X\le c$. If the wording instead says the loss was below $c$, it implies $X<c$.
Interval-censored: the loss is known to lie within a stated interval, for example $L\le X\le U$, subject to the endpoint wording.

The precise distinction between a strict and non-strict inequality should be preserved in the description of the data. For a continuous loss distribution, the probability of equality at a fixed point is zero, so $P(X>c)$ and $P(X\ge c)$ are equal. But that distributional fact does not change what the record literally says, and equality can matter for discrete losses or when interpreting a particular claim.

A wholly missing record is different. It does not imply $X=0$, $X>c$, or any other bound. It may mean that a claim amount was lost during data extraction, was never reported, or could not be matched to a policy. If the observation unit is known to exist, a blank amount is not evidence that the loss did not occur. Conversely, a policy with no claim during the observation period is not automatically a missing severity: it may be a known no-claim outcome in a frequency data set.

Worked Example: Translating the Record Language

A hypothetical adjuster file contains four entries for individual claims: (1) “paid amount: $740”; (2) “amount exceeded $1,000”; (3) “final amount is no more than $450”; and (4) an amount field left blank, with no other note. Classify each entry and state what it establishes about $X$.

Entry 1. The paid amount is reported exactly, so $X=740$. This is an exact observation.

Entry 2. “Exceeded $1,000” is strictly greater than the threshold. It is right-censored and establishes $X>1000$, not $X\ge1000$ and not $X=1000$.

Entry 3. “No more than $450” includes equality. It is left-censored and establishes $X\le450$.

Entry 4. There is neither an amount nor a stated constraint. The amount is wholly missing, so this entry supplies no bound on $X$. It must not be entered as a zero-dollar claim.

Check. The four entries represent one exact value, two bounds, and one amount with no loss information. Only the first entry identifies an exact claim amount.

Why a Bound Is Not an Exact Amount

For a continuous loss variable with density $f_X$, an exact observed amount $x$ provides information through the density at $x$. A right-censored observation at $c$ instead says that the loss fell in the event $X>c$. Its probability is the survival probability $S_X(c)=P(X>c)$. This is the basic reason the two record types are handled differently.

P(X>c)=\int_c^\infty f_X(x)\,dx=S_X(c).

The integral adds the probability over all possible loss amounts above the threshold. Replacing the censored amount with $c$ would retain only one value and discard the fact that any larger loss is consistent with the record. For a left-censored record $X\le c$, the corresponding event probability is $F_X(c)=P(X\le c)$. For an interval $L\le X\le U$, it is $P(L\le X\le U)$; for a continuous distribution this is $F_X(U)-F_X(L)$. These expressions describe what the observation says, not a particular estimator.

A missing record supplies no event about $X$, so it does not supply one of these loss probabilities. It can still be useful indirectly if information about why it is missing and observed characteristics are available, but that requires assumptions about the missingness process. No amount can be recovered from a blank entry alone.

Key distinction: Censoring retains partial information about a loss through a constraint. A wholly missing amount retains no direct information about the loss value. Neither record should be substituted by a guessed exact amount without an explicitly justified model.

Assumptions Needed to Use the Records

Correctly classifying a record is only the first step. To use censored observations to learn about a loss distribution, the censoring process must not selectively hide loss amounts in a way that invalidates the analysis. In the standard noninformative-censoring setting, the censoring mechanism is independent of the loss, or is independent conditional on relevant observed covariates. The latter means that after comparing observations with the same covariate values, the chance or timing of censoring does not still depend on the unobserved loss.

This assumption is not guaranteed by calling a record “censored.” For example, if high losses are more likely to exceed a reporting cap, censoring is related to loss size. A method that ignores this relationship may misrepresent the loss distribution. In practice, the recording rule and operational context should be examined; the assumption cannot be established merely from a table of recorded amounts.

For wholly missing amounts, a basic assumption for using the observed complete records as representative is that missingness is unrelated to the loss. A more flexible possibility is that missingness depends on observed variables, such as product type or region, but not on the amount after conditioning on those variables. In that case, the observed values may be informative within comparable groups, provided each group has a nonzero chance of having its amount observed. If missingness continues to depend on the unobserved loss after accounting for observed information, the data alone generally do not identify the loss distribution; additional assumptions, external information, or sensitivity analysis are needed.

These are assumptions about the observation process, not properties of the loss distribution itself. They should be distinguished from the sampling assumptions that make the observed policies or claims representative of the population of interest.

Identify the observation unit.
Decide whether each row represents a claim, a policy, or another unit. Establish whether an amount is expected for every row in the analysis.

Read the amount field and its qualifiers.
Record an exact value only when the amount is known. Translate threshold and interval wording into the correct strict or inclusive constraint.

Separate bounds from blank amounts.
Retain a stated bound as censoring information. Do not invent a bound for an otherwise blank record or replace it with zero.

Assess the observation process.
Ask whether censoring or missingness could depend on loss size, and state what independence or conditional-independence assumption is needed for the intended analysis.

Worked Interpretations

Worked Example: Bounds on a Group’s Total Loss

A hypothetical portfolio has eight claim records, and every claim amount is nonnegative. Three records are exact: $400, $900, and $1,500. One says “exceeded $1,200,” one says “at least $2,000,” one says “no more than $300,” one gives the interval $600 to $800 inclusive, and one amount is blank with no bound. Find the strongest lower bound supported for the total of the eight losses and state whether a finite upper bound is available.

Method. Use each exact amount and each recorded constraint, while allowing every loss to take any value consistent with its record. The blank record is only constrained by the stated nonnegativity assumption.

Step 1: Add the exact amounts. The exact total is $400+900+1500=2800$ dollars.

Step 2: Use the lowest values permitted by the other records. The “exceeded $1,200” record contributes strictly more than $1,200. The “at least $2,000” record contributes at least $2,000. The “no more than $300” record can be as low as $0 under nonnegativity. The inclusive interval can be as low as $600. The blank record can also be as low as $0 under nonnegativity.

Step 3: Combine the constraints. The sum of the exact amounts and the non-strict lower bounds is $2800+1200+2000+0+600+0=6600$. Because one loss is strictly greater than $1,200, the total is strictly greater than $6,600.

There is no finite upper bound: the two right-censored losses have no stated upper limits, and the missing loss has no stated upper limit. The answer is therefore a total loss $> \$6,600$, with no finite upper bound supported by these records. The common mistake is to set the “exceeded $1,200” loss equal to $1,200; doing so would turn a strict lower bound into an unsupported exact amount.

Worked Example: Complete Records and Group-Specific Missingness

A hypothetical insurer has four low-risk policies and four high-risk policies in a study. Among the low-risk policies, three exact losses are observed: $100, $200, and $300; the fourth amount is missing. Among the high-risk policies, one exact loss of $900 is observed and the other three amounts are missing. Calculate the mean of the four observed amounts. Then calculate the group-weighted mean estimate if, as an explicit assumption, observed amounts are representative of missing amounts within each risk group.

Step 1: Calculate the complete-record mean. The four observed amounts sum to $100+200+300+900=1500$. Dividing by four gives $1500/4=\$375$.

Step 2: Calculate the observed mean within each group. For the low-risk group, the observed mean is $(100+200+300)/3=600/3=\$200$. For the high-risk group, its only observed amount gives an observed mean of $900.

Step 3: Weight by the target group sizes. Each group contains four of the eight policies, so each receives weight $4/8=0.5$. The group-weighted estimate is $0.5(200)+0.5(900)=100+450=\$550$.

The complete-record mean is $375, while the group-weighted estimate under the stated within-group representativeness assumption is $550. The difference arises because low-risk policies are much more represented among the complete records. The $550 figure is not established by arithmetic alone: it relies on the assumption that the observed loss amounts represent missing amounts within each group, and the high-risk group estimate is based on just one observed amount. The common mistake is to assume that the four complete records automatically represent all eight policies.

Worked Example: Deciding Whether Censoring Is Informative

A hypothetical administrator records exact losses up to a stated threshold of $2,500. For any larger loss, the file says only “above $2,500.” The administrator also notes that the system applies this rule to every claim, regardless of policy type or claim circumstances. Classify the above-threshold records and assess whether the description supports treating the censoring mechanism as noninformative.

Step 1: Classify the records. Each “above $2,500” entry is right-censored and implies $X>2500$. It is not an exact $2,500 loss.

Step 2: Assess how the threshold is applied. The stated rule is a fixed recording threshold applied to every claim, rather than a rule that chooses which claims to censor based on their unrecorded amount or circumstances. This description supports an assumption that censoring is independent of loss, subject to the accuracy of the stated process and the rest of the study design.

Step 3: State the limitation. The rule does not reveal the actual amounts above $2,500. It supplies a lower bound but no upper bound for those losses. If, contrary to the description, claims were more likely to be censored because of unrecorded severity-related factors, the noninformative-censoring assumption would need to be reconsidered. The common mistake is to treat a uniform threshold as if it had disclosed the amount at the threshold.

Common Mistakes and a Practical Check

Changing the inequality. “Exceeded $1,000” means $X>1000$, while “at least $1,000” means $X\ge1000$. Preserve the wording’s endpoint.
Replacing a censored loss by its threshold. A threshold is a bound, not an observed amount. Losses above it may differ substantially.
Entering a missing amount as zero. Zero is a specific loss value; a blank amount does not imply it.
Assuming every incomplete record is censored. A record is censored only if the data establish a constraint on the amount. A blank with no bound is missing.
Assuming complete records are representative. If the probability of an amount being observed varies with its value, a complete-record summary may be biased for the target population.

Before using an incomplete loss file, write down for each record whether it is exact, censored, or missing; state the corresponding value or constraint; and identify the assumption about the observation process that supports the intended use. This prevents a data-recording limitation from silently becoming a false statement about the loss itself.

Check Your Understanding

For each problem, classify the relevant records or calculate the requested quantity. Preserve strict and inclusive bounds, and state any assumption used.

A hypothetical claims file says “loss exceeded $1,350.” Classify the record and state the exact inequality implied. (Answer: Right-censored; $X>1350$.)
A record states that a loss is “at most $620,” and another states that it is “below $620.” Write the constraint for each record and identify the distinction. (Answer: Left-censored with $X\le620$ for “at most”; left-censored with $X<620$ for “below.”)
Three nonnegative losses are recorded as $250, “at least $700,” and blank with no bound. What is the strongest lower bound on their total, and is a finite upper bound supported? (Answer: Total $\ge \$950$; no finite upper bound is supported.)
A study contains five policies in each of two groups. Two observed amounts in group A are $120 and $180; the only observed amount in group B is $600. All other amounts are missing. Find the complete-record mean and the group-weighted mean under the assumption that observed amounts represent missing amounts within each group. (Answer: Complete-record mean $=(120+180+600)/3=\$300$; group-weighted mean $0.5(150)+0.5(600)=\$375$.)
A system censors every loss above $3,000 and records “at least $3,000.” What inequality does the record imply, and what additional condition is important before using censored records to learn about the loss distribution? (Answer: Right-censored; $X\ge3000$. The censoring mechanism should be noninformative—independent of loss, or conditionally independent given relevant observed covariates.)

What You'll Learn