Why Classification Comes Before Calculation
A loss record can contain a dollar amount without revealing the exact loss, or it can provide useful information even when no dollar amount is recorded. A reporting cap, a detection threshold, or inspections conducted only at intervals can restrict the possible value of a loss. By contrast, a blank field may provide no restriction at all. These distinctions matter whenever an analyst summarizes or models incomplete loss data.
The classification depends on what the observation process establishes about the actual loss \(X\), not just on how a data field looks. Before using a record, ask: does it identify one value, establish a bound, confine the value to an interval, or leave the value unknown?
“Right” and “left” refer to the direction of the unobserved values on the loss scale. If a record says a loss exceeds a threshold, values above that threshold remain possible: the record is right-censored. If a device only establishes that a loss is below its detection limit, values below that limit remain possible: the record is left-censored. A record bounded on both sides is interval-censored.
Translate the Record into a Set of Possible Losses
A reliable classification method is to write the set of values that remain consistent with the evidence. For an exact observation, that set contains one value. For a censored observation, it contains a range. For a missing loss record, no range narrower than the model’s general support is established by the record.
Determine whether the amount was measured exactly, compared with a threshold, checked only at scheduled times, or not recorded.
For example, “above 2,500” becomes \(X>2{,}500\), while “no more than 1,200” becomes \(X\leq1{,}200\).
A single value is exact; a one-sided bound is left- or right-censored; two bounds give interval censoring; no loss restriction means the loss record is missing.
The direction of the inequality is important. “Less than” or “at most” gives an upper bound and is left-censoring. “Greater than” or “at least” gives a lower bound and is right-censoring. If a reporting rule includes the threshold, use a non-strict inequality; if it excludes the threshold, use a strict inequality. Do not silently replace one rule with the other.
| Record states | Information about \(X\) | Classification |
|---|---|---|
| The measured loss is 740 | \(X=740\) | Exact |
| The loss did not exceed 1,200 | \(X\leq1{,}200\) | Left-censored |
| The loss exceeded 2,500 | \(X>2{,}500\) | Right-censored |
| The loss was above 900 and no more than 1,400 | \(900<X\leq1{,}400\) | Interval-censored |
| The loss field is blank and no other evidence is available | No restriction on \(X\) from this record | Missing |
Worked Classifications
Worked Example: Five Records from a Property Portfolio
A hypothetical property insurer reviews five claim files. File A reports a verified payment of $860. File B states that the covered loss was below $1,200. File C states that the loss exceeded $2,500. File D establishes that the loss was more than $900 but no greater than $1,400. File E has a blank loss field, and the adjuster’s notes contain no amount or threshold information. Classify each file and state the information available about \(X\).
Method. Translate each statement into a condition on the actual loss \(X\), then classify the number of values or range that remains possible.
File A. The verified amount identifies one value: \(X=860\). This is an exact record.
File B. “Below $1,200” gives \(X<1{,}200\), an upper-bounded set of possible losses. This is left-censored.
File C. “Exceeded $2,500” gives \(X>2{,}500\), a lower-bounded set of possible losses. This is right-censored.
File D. The evidence gives both \(X>900\) and \(X\leq1{,}400\), so \(900<X\leq1{,}400\). This is interval-censored.
File E. No loss value or constraint is available, so the loss record is missing. The file does not establish that \(X=0\).
Check. The five classifications are exact, left-censored, right-censored, interval-censored, and missing, respectively. The common mistake to avoid is reversing Files B and C: an upper bound is left-censoring, while a lower bound is right-censoring.
Worked Example: A Detection Threshold and a Reporting Cap
A hypothetical equipment insurer uses a screening system that can detect damage only when the amount exceeds $350. For one unit, the system reports “damage not detected.” For another, the claim system stores only the message “payment reached the $5,000 reporting cap,” and the file confirms that the uncapped covered loss was at least that amount. Classify the loss information for each unit. Explain why neither message should automatically be recorded as an exact loss.
Method. Interpret the system’s stated evidence, rather than substituting the threshold itself for the unknown loss.
Unit 1. Under the stated screening rule, no detected damage establishes \(X\leq350\). The record provides an upper bound, so it is left-censored. It does not establish \(X=350\), nor does it establish \(X=0\).
Unit 2. The file explicitly confirms an uncapped loss of at least $5,000, so \(X\geq5{,}000\). This is a lower-bounded, right-censored record. The cap does not make the actual loss exactly $5,000.
Check. Unit 1’s possible values lie at or below $350; Unit 2’s lie at or above $5,000. The common mistake is to enter a threshold or cap as if it were a measured loss. Here the wording explicitly includes the boundary in both records, so the inequalities are non-strict.
Worked Example: Loss Discovered Between Inspections
A hypothetical insurer inspects a machine at the end of each month. At the end of month 4, no covered failure has occurred. At the end of month 6, the machine is found to have failed, but the failure time is not known more precisely. Let \(T\) be the time of failure in months after monitoring began. Classify the observation and state the available time information. A separate machine has a completed inspection record through month 5 but no record for later months; no failure status or later observation is available.
Method. For the first machine, combine the last inspection known to be failure-free with the first inspection at which failure is known to have occurred. For the second, use only the evidence actually recorded.
First machine. No failure by month 4 means \(T>4\). Failure has occurred by the month-6 inspection, so \(T\leq6\). Together these give \(4<T\leq6\), an interval-censored failure time. Neither inspection identifies an exact failure time.
Second machine. A completed inspection through month 5, without a recorded outcome after that point, does not establish whether or when failure occurred later. On the stated facts, the later failure-time record is missing; it does not establish \(T>5\). If the record instead explicitly stated that the machine remained failure-free through month 5, that evidence would establish \(T>5\), a right-censored observation.
Check. The first machine’s information is bounded on both sides, while the second machine’s later status is unknown. The common mistake is to treat a missing follow-up record as proof that the unit survived through the last date mentioned; that conclusion requires an explicit failure-free observation at that date.
Why Censored Records Still Contain Information
A censored observation does not identify the exact loss, but it rules out values inconsistent with the observation. This is why it is different from a missing record. The distinction can be expressed using the probability distribution of \(X\). Let \(F_X(x)=P(X\leq x)\) be the distribution function. Under the specified endpoint conventions, the probability of the information supplied by a record is the probability of its corresponding set.
For the interval expression, the justification is that the event \(X\leq b\) consists of the disjoint events \(X\leq a\) and \(a<X\leq b\). Therefore,
This identity illustrates the information content of an interval record: it identifies membership in a range, not a point inside that range. Similarly, a left- or right-censored record identifies a one-sided set of possible values. A missing record supplies no event about \(X\) from which to narrow its possible values. These statements describe the record’s information; they do not by themselves specify an estimation procedure.
Exact observations are different: they report the realized value itself. For a continuous loss model, a single exact value has probability zero under the model, so its probability is not obtained by substituting \(x\) into \(F_X\) as though it were a range. The exact amount is nevertheless observed data. Keep this distinction clear when describing what a record says.
Common Classification Traps
- Reversing the censoring direction. \(X<c\) or \(X\leq c\) is an upper bound and hence left-censoring. \(X>c\) or \(X\geq c\) is a lower bound and hence right-censoring.
- Replacing a bound by its threshold. “At least $5,000” does not mean the loss equals $5,000. It means the loss is in a set beginning at that value.
- Calling every blank field censored. A blank is censored only if other evidence establishes a bound or interval. With no such evidence, the loss record is missing.
- Assuming an unobserved event did not occur. No follow-up record after a date is not the same as a recorded observation that the event-free status continued through that date.
- Ignoring endpoint wording. “Below,” “at most,” “above,” and “at least” encode different inequalities. Preserve the stated boundary rule when it is known.
Check Your Understanding
For each item, classify the record and state the precise information available about the loss or event time.
- A hypothetical repair file gives a verified loss of $1,075, while another says the loss was no more than $640. Classify both records. (Answer: The first is exact, \(X=1{,}075\); the second is left-censored, \(X\leq640\).)
- A sensor records only “above $3,200,” and a second record states “less than $480.” Classify each and identify whether its possible values have a lower or upper bound. (Answer: \(X>3{,}200\) is right-censored with a lower bound; \(X<480\) is left-censored with an upper bound.)
- An inspection at month 7 confirms no failure; the next inspection at month 9 finds a failure. Let \(T\) be failure time. State the information and classification. (Answer: \(7<T\leq9\); interval-censored.)
- A claim system reports that an uncapped loss is at least $8,000. An analyst enters $8,000 as the observed loss. Is that exact? Give the correct classification and condition. (Answer: No. It is right-censored, \(X\geq8{,}000\), not exact.)
- A file has no reported amount. The only additional note says that the loss was below the $900 investigation threshold. Classify the record and state the information; contrast it with a file having no amount and no note. (Answer: The first is left-censored, \(X<900\); the second is missing and supplies no restriction on \(X\).)