Introduction
A loss record does not always give the final amount. A claim may still be developing when observation ends, or a recording system may report only that the ultimate loss passed a specified threshold. Such a record is incomplete, but it is not uninformative: it rules out some possible values of the loss.
The key is to express exactly what the record establishes. A right-censored observation gives a lower-bound type of statement about the unknown loss. Whether that statement is \(X>c\) or \(X\ge c\) depends on the observation rule and its wording. Confusing the two can incorrectly include or exclude a loss exactly at the threshold.
What the Record Does and Does Not Say
An exact loss record \(X=x\) identifies a single value. A right-censored record at \(c\) identifies a collection of possible values, all above the threshold under the strict convention. For example, if the record establishes \(X>12{,}000\), then \(12{,}500\) and \(30{,}000\) remain possible, but \(12{,}000\) and \(11{,}900\) do not.
The threshold is not the observed ultimate loss. Replacing the unknown loss by \(c\) treats a lower bound as if it were an exact amount. Nor does a right-censored record say that the loss was below the threshold. That would describe the opposite side of the threshold.
The distinction between strict and inclusive inequalities comes from the observation rule, not from the phrase “censored” alone. “The loss exceeded \(c\)” means \(X>c\). “The loss was at least \(c\)” means \(X\ge c\). “The loss was greater than or equal to \(c\)” also means \(X\ge c\). If the record says only “above the threshold,” check how that phrase is defined in the data specification before deciding whether equality is possible.
| What the record establishes | Mathematical information | Is equality at \(c\) allowed? |
|---|---|---|
| The loss exceeded \(c\) | \(X>c\) | No |
| The loss was at least \(c\) | \(X\ge c\) | Yes |
| The exact loss was \(c\) | \(X=c\) | Yes; the value is observed exactly |
| The loss was below \(c\) | \(X<c\) | No |
These statements are about the unknown value for one record. They should not be confused with a population distribution. If \(F_X(x)=P(X\le x)\) is the distribution function and \(S_X(x)=P(X>x)\) is the survival function, then \(S_X(c)=1-F_X(c)\). A record establishing \(X>c\) tells us that this individual’s loss belongs to the set counted by \(S_X(c)\); it does not reveal the individual’s exact value within that set.
Why the Strict Inequality Follows
Suppose a reporting rule follows a loss until a fixed threshold \(c\), and the record states that the ultimate loss passed that threshold. “Passed” means the ultimate loss is larger than \(c\), not merely equal to it. This can be expressed directly in terms of possible values:
The open left endpoint excludes \(c\). Equivalently, the record rules out the event \(X\le c\), so its information is \(X>c\). If the reporting rule instead says only that the loss reached \(c\), equality remains possible, and the set of consistent values is \([c,\infty)\), corresponding to \(X\ge c\).
This reasoning depends on the meaning and precision of the threshold statement. It does not depend on a particular probability distribution for \(X\). No distributional model is needed to translate “exceeded \(c\)” into an inequality. A distributional model would be needed for additional probability calculations, and should not be silently assumed.
Worked Example: Classifying Exact and Threshold-Only Records
A hypothetical administrator records ultimate losses for three files using the following notes: (A) “final loss, $7,260”; (B) “ultimate loss exceeded $9,000, amount unavailable”; and (C) “ultimate loss was at least $9,000, amount unavailable.” For each file, state the mathematical information and whether a loss of exactly $9,000 is consistent with the record.
File A: exact amount. The final loss is observed, so \(X=7{,}260\). This is not a censored record. The threshold question is not relevant to this file.
File B: strict exceedance. “Exceeded $9,000” means \(X>9{,}000\). In set form, the possible values are \((9{,}000,\infty)\). A loss of exactly $9,000 is not consistent with the note.
File C: inclusive lower bound. “At least $9,000” means \(X\ge9{,}000\), with possible values \([9{,}000,\infty)\). A loss of exactly $9,000 is consistent with this note.
Check. The only difference between B and C is the inclusion of the threshold itself. The wording excludes equality in B and permits it in C. Replacing either threshold-only record with an exact loss of $9,000 would add information that was not recorded.
Method and trap avoided: Translate the recorded wording first, then test the boundary value. The common mistake is to treat “exceeded” and “at least” as interchangeable.
Threshold Wording, Precision, and Rounding
A threshold can refer to the actual loss, a recorded loss, or a value constrained to a fixed grid. These are different statements. In particular, rounding an actual amount to the nearest cent does not make the actual loss a cent-grid value. A true amount just above $8,400.00 may still be recorded as $8,400.00.
For instance, under ordinary rounding to the nearest cent, an actual amount of $8,400.001 rounds to $8,400.00. The actual amount is strictly greater than $8,400.00, while the rounded record is not. Conversely, if the model explicitly restricts possible losses to exact cent values, then a loss strictly greater than $8,400.00 must be at least $8,400.01. That conclusion follows from the grid restriction, not from rounding.
Worked Example: Strict Exceedance with a Rounded Record
A hypothetical loss system stores dollar amounts to the nearest cent. Its underlying rule says the actual loss \(X\) strictly exceeded $8,400.00. A recorded amount is $8,400.00. Determine whether the strict exceedance and the recorded amount can both be correct. Then compare this with a system in which actual losses can take only exact cent values.
Step 1: Keep the actual and recorded amounts distinct. The strict rule applies to the actual amount: \(X>8{,}400.00\). The stored figure is a rounded value, not necessarily \(X\) itself.
Step 2: Check a possible actual value. Take \(X=8{,}400.001\). This satisfies the strict rule because \(8{,}400.001>8{,}400.00\). Rounded to the nearest cent, it is recorded as $8,400.00. Thus the two statements can both be correct.
Step 3: Apply the different cent-grid assumption. If actual losses are constrained to exact cent values, the possible amounts around the boundary are $8,400.00 and $8,400.01. The first fails \(X>8{,}400.00\); the next is $8,400.01. Under that assumption, the smallest possible actual loss satisfying the rule is $8,400.01.
Check. The rounded-data conclusion and the cent-grid conclusion differ because their assumptions differ. Rounding permits actual values between cent marks; a cent-grid restriction does not.
Method and trap avoided: Identify whether the threshold applies to the underlying loss or the rounded record, and whether actual values are restricted to cents. The common mistake is to infer a minimum actual loss of $8,400.01 solely because the system rounds to cents.
Using the Inequality in a Data Description
A careful description separates three items: the unknown quantity, the threshold, and the rule that connects them. For example, “the ultimate loss \(X\) was not observed exactly, but the record establishes that it exceeded \(c=\$15{,}000\)” translates to \(X>15{,}000\). This is more precise than writing “loss = $15,000,” which incorrectly states that the exact loss was observed.
When a data set contains several records, apply the rule to each record rather than assuming all incomplete records have the same interpretation. One record may give an exact value, another may establish a strict lower bound, and a third may establish an inclusive lower bound. In the absence of further information, the mathematically correct record preserves only what is known.
A useful check is to test boundary and non-boundary values. For a record \(X>c\), a proposed value \(c\) must be ruled out, and values below \(c\) must also be ruled out; values above \(c\) are consistent. For \(X\ge c\), the boundary value \(c\) is also consistent. This simple test catches many inequality errors before any later statistical procedure is applied.
Use \(X\) for the ultimate loss, rather than treating the threshold as the loss itself.
“Exceeded” or “greater than” gives \(X>c\); “at least” or “greater than or equal to” gives \(X\ge c\).
Determine whether the statement applies to the underlying loss, a rounded amount, or a value restricted to a specified grid.
Ask whether \(X=c\) is consistent with the record. If it is, the inequality is inclusive; if not, it is strict.
Common Mistakes and Why They Matter
- Writing \(X=c\) for a threshold-only record. A threshold gives a bound, not an exact ultimate loss. Preserve the set of possible values.
- Reversing the inequality. “The loss exceeded \(c\)” rules out values at or below \(c\), so it means \(X>c\), not \(X<c\).
- Ignoring equality. The difference between \(>\) and \(\ge\) is exactly whether a loss equal to the threshold is permitted.
- Assuming rounding preserves strictness. A rounded record at \(c\) does not prove that the underlying loss equals \(c\), nor does a strict inequality for the underlying amount force its rounded value to exceed \(c\).
- Assuming a distribution to resolve missing detail. Without a stated model or additional data, a right-censored record does not determine how far above \(c\) the loss lies.
For empirical loss analysis, the inequality is the usable information. Replacing it with a made-up exact value can distort summaries and subsequent estimates. Preserving the distinction between the unknown loss and its censoring threshold is therefore essential. This treatment of incomplete loss information is aligned with the SOA Exam ASTAM (Advanced Short-Term Actuarial Mathematics) syllabus.
Check Your Understanding
For each hypothetical record, translate the wording into a mathematical statement and check whether the boundary value is allowed.
- A file says the ultimate loss “exceeded $18,500,” but gives no final amount. Write the information about \(X\), and state whether \(X=18{,}500\) is possible. (Answer: \(X>18{,}500\); equality is not possible.)
- A report says a loss “was at least $2,750.” Write the inequality and state whether $2,750 is allowed. (Answer: \(X\ge2{,}750\); equality is allowed.)
- A system rounds actual amounts to the nearest cent. The actual loss is known to exceed $6,100.00. Can the recorded value be $6,100.00? Give a numerical actual loss that demonstrates your conclusion. (Answer: Yes; for example, \(X=\$6{,}100.001\) rounds to $6,100.00 while exceeding $6,100.00.)
- Actual losses are restricted to exact cent values, and a record establishes that \(X>\$3{,}249.99\). What is the smallest possible actual loss? (Answer: $3,250.00.)
- A file states “loss reached $11,000, and the rule does not say whether it went higher.” What is the strongest lower-bound statement justified by this wording, and why is \(X>11{,}000\) not justified? (Answer: \(X\ge11{,}000\); equality is not ruled out.)