Introduction
An unusually large value deserves attention, but unusual does not mean wrong. A high claim amount might reflect a recording mistake, a valid large loss, or a class of policies with different coverage. The same number can look extraordinary in one view of the data and ordinary in a relevant subgroup. The task is therefore not simply to find extremes; it is to assess whether each one is credible in context.
The earlier material on boxplots and the 1.5 IQR rule provides ways to flag observations. A flag is a signal to investigate, not a diagnosis. A sound assessment combines the value’s position in the observed distribution with related record details and subject-matter plausibility. The result should distinguish what the data show from what is inferred about the record.
What a Flag Can and Cannot Tell You
A flag describes a relationship between a value and a reference distribution or rule. For example, the 1.5 IQR rule flags values beyond fences constructed from the first and third quartiles. It does not estimate the probability that a value is an error, nor does it establish that a value is impossible. The choice of reference matters: a pooled portfolio can combine different product types, limits or exposure levels, so an observation may be extreme overall but typical among comparable records.
Three explanations should remain open while investigating:
- Recording or processing error. A misplaced decimal, wrong unit, duplicated transaction, or incorrect join may produce a value that does not describe the intended observation.
- Valid rare case. The value is correctly recorded but lies far into the tail of the distribution, such as a genuinely severe loss.
- Population feature. The value belongs to a meaningful group with a different scale or process, such as a policy with a substantially higher coverage limit.
These explanations call for different actions. Correct a confirmed error using a traceable source; retain a verified rare case; and investigate whether a distinct population should be described separately. If the evidence is insufficient, record the uncertainty rather than silently changing the data.
Worked Example: A Flagged Settlement Duration
Method: apply an empirical IQR flag, then check the record and its case details. A hypothetical claims team has settlement durations, in days, of \(8, 9, 10, 10, 11, 12, 13, 14, 15, 16, 18,\) and \(62\). Use the median-of-halves quartile convention. Does 62 days appear to be an error?
Step 1: Calculate the quartiles. The values are already ordered. The lower half is \(8, 9, 10, 10, 11, 12\), so \(Q_1=(10+10)/2=10\). The upper half is \(13, 14, 15, 16, 18, 62\), so \(Q_3=(15+16)/2=15.5\).
Step 2: Calculate the IQR and upper fence. The interquartile range is \(15.5-10=5.5\). The upper fence is \(15.5+1.5(5.5)=15.5+8.25=23.75\) days. Since \(62>23.75\), the observation is flagged. This calculation establishes that it is far from the middle half of these 12 records; it does not establish that it is wrong.
Step 3: Examine related evidence. Suppose the file shows that this claim was assigned a litigation status and that its dated case notes document a court-related delay. The recorded duration is the difference between the recorded opening and settlement dates, and those dates agree with the source record. Those checks support the interpretation that the value is a valid, unusual case.
Conclusion and trap avoided. Retain 62 days as a verified long settlement and document why it was reviewed. The common mistake is to delete every observation beyond an IQR fence. The fence is a screening rule, not a rule for removing valid data. No parametric distribution is assumed in this empirical calculation.
Distributional Context: How Rare Is Rare?
A distributional assessment asks how unusual a value is relative to a clearly specified reference. With an empirical reference, one can count observations at least as extreme as the flagged value. With a parametric reference, one can calculate a tail probability only after stating the model and its parameters. Neither approach, by itself, identifies a data error.
For a continuous random variable \(X\) with cumulative distribution function \(F_X\), the upper-tail probability of a value \(x\) is \(P(X\geq x)=1-F_X(x)\). This describes the probability of a value at least that large under the assumed model. For discrete variables, the probability at the observed value may have positive mass, so the relevant tail is \(P(X\geq x)\), calculated by summing the probabilities at \(x\) and above. In either case, a tail probability is conditional on the model being appropriate for the records being assessed.
A small tail probability is evidence that the value is unusual under the model; it is not the probability that the record is erroneous. That latter probability would require a model for how errors arise and evidence about the record. Also consider the portfolio size. If each of \(n\) independent observations has upper-tail probability \(p\), the expected number at least that large is \(np\). A tail probability that is small for one record can therefore still imply that one or more such observations are plausible in a large portfolio.
Worked Example: A Large Claim Under a Lognormal Model
Method: calculate a model-based upper-tail probability, then interpret it in portfolio context. In a hypothetical homogeneous product tier, let claim severity \(X\), measured in thousands of dollars, satisfy \(\ln(X/1)\sim N(0,0.5^2)\). Thus the logarithm of severity in thousands is normal with mean 0 and variance \(0.25\), or standard deviation \(0.5\). A recorded claim is \(X=6\), meaning $6,000. Assume for this calculation that the model applies to each claim in this tier.
Step 1: Standardize on the log scale. Since \(\ln(6)=1.79176\), the standardized value is \(z=(1.79176-0)/0.5=3.58352\), approximately 3.584 standard deviations above the model’s log-scale mean.
Step 2: Calculate the upper-tail probability. Under the stated normal model, \(P(X\geq6)=P(Z\geq3.58352)\), where \(Z\sim N(0,1)\). The standard normal upper-tail probability is approximately \(0.00017\), or \(0.0002\) to four decimal places. This is rare under the assumed severity model.
Step 3: Consider portfolio size. If 10,000 independent claims all followed this same model, the expected number at least this large would be \(10{,}000(0.00017)=1.7\). Using \(p=0.00017\), the probability of at least one such claim is approximately \(1-(1-0.00017)^{10{,}000}=0.817\). Thus, although the value is rare for an individual claim, observing at least one is not surprising in a portfolio of that size under these assumptions.
Step 4: Check the record. Review the payment amount, currency and unit, policy limit, and source documentation. If these confirm a $6,000 payment under the correct tier, the result supports retaining it as a valid tail observation. If the amount was entered in dollars where the field expects thousands, that would be evidence of a unit error.
Conclusion and trap avoided. The model indicates rarity, not error probability. The common mistake is to report \(0.00017\) as the probability that the record is wrong. That interpretation does not follow: \(0.00017\) is the probability of a claim at least this large if the stated model is correct. No error-generation model has been specified.
Related Observations and Subject-Matter Plausibility
A useful review compares the flagged value with records that should be comparable and checks information connected to the same observation. Related fields may reveal a unit mismatch, an impossible date sequence, a duplicated record, or a coverage feature that explains the amount. The comparison should be guided by how the data were generated, not chosen only because it makes an extreme value look more ordinary.
Subject-matter knowledge helps define what is possible and what is merely unusual. A negative count may be impossible under the field definition, while an unusually high positive claim can still be valid. A policy limit can put an upper bound on a payment, but only if the relevant limit and payment definition are known. A value outside a hard operational constraint is stronger evidence of a problem than a value that merely exceeds a statistical fence.
There is also a difference between a broad plausibility check and certainty. A value within an allowed range is not automatically correct; it could still be attached to the wrong policy. Conversely, a value outside an expected range may reveal an undocumented product or process. Preserve enough information to reproduce the review, including the original value, any correction, the evidence consulted, and the reason for the decision.
Worked Example: A High Limit in a Mixed Policy File
Method: use a pooled empirical flag, then compare the observation with its policy class and verify it against the record. A hypothetical file contains policy limits, in thousands of dollars, of \(40, 42, 45, 47, 48, 50, 52, 55, 58, 60, 62, 65, 900, 1{,}000,\) and \(1{,}200\). The last three records are labeled high-limit commercial policies. Use the median-of-halves convention. Assess the 1,200 value.
Step 1: Calculate the pooled fence. There are 15 ordered values, so the median is the eighth value, 55. The lower seven values have median \(Q_1=47\). The upper seven values are \(58, 60, 62, 65, 900, 1{,}000, 1{,}200\), with median \(Q_3=65\). Thus, \(\mathrm{IQR}=65-47=18\), and the upper fence is \(65+1.5(18)=92\). Since \(1{,}200>92\), the value is flagged against the pooled file.
Step 2: Check comparable records and field meaning. The other two high-limit records are 900 and 1,000, also far above the pooled fence. Suppose the policy schedule confirms that the three records have high limits and that 1,200 means a $1.2 million coverage limit, not a $1,200 payment. The three values are consistent with their recorded policy class and amount definition.
Step 3: Decide what the flag means. The pooled fence identifies an upper tail, but the records provide a coherent population explanation: the file combines ordinary limits with a high-limit class. The 1,200 value is not thereby proven error-free, but the available evidence supports retaining it and describing the classes appropriately.
Conclusion and trap avoided. Retain the verified policy limit and document the class distinction. The common mistake is to treat a pooled distribution as if every record came from one homogeneous population, then delete all values outside its fence. No parametric distribution is assumed here; this is an empirical flag followed by a contextual review.
A Defensible Review Process
Use a consistent sequence so that a statistical flag leads to a reasoned decision rather than an automatic edit. The steps below apply whether the initial signal came from a boxplot, a histogram, a distributional model, or a basic data check.
Confirm the unit, definition, time period and relevant population. A value cannot be assessed properly if, for example, one record is in dollars and the others are in thousands.
State the rule or reference that marked the value. Distinguish a visual impression, an empirical fence and a model-based tail probability; they do not make identical claims.
Review source values, related fields, data-entry history and genuinely comparable observations. Seek evidence that could support or contradict a proposed explanation.
Correct only when evidence supports a specific correction. Retain a verified rare case, investigate a potential group difference, or mark the case unresolved if the evidence is inconclusive.
Keep the original value and record the flag, checks, evidence and disposition. This makes the analysis auditable and prevents a later reviewer from mistaking an undocumented change for an observed fact.
The key judgment is whether the value is inconsistent with what the field is meant to measure, not whether it is inconvenient for a summary. Removing a valid extreme can distort estimates of totals, tail risk or model performance. Keeping a confirmed error can distort them as well. When an analysis is sensitive to the decision, report that sensitivity or explain how the unresolved record was handled.
Check Your Understanding
For each problem, separate what the numerical rule shows from what the contextual evidence supports.
- A hypothetical file of claim durations has \(12, 14, 16, 18, 20, 22, 24, 26, 28, 30, 32, 47\) days. Using the median-of-halves convention, calculate \(Q_1\), \(Q_3\), the IQR and the upper fence. Is 47 flagged? (Answer: \(Q_1=17\), \(Q_3=29\), IQR \(=12\), upper fence \(=47\); 47 is not above the fence, so it is not flagged by the strict upper-fence rule.)
- In 400 comparable records, three observations are at least as large as a particular flagged value. What is the empirical proportion at least that large? Does that proportion establish an error? (Answer: \(3/400=0.0075\); no, it describes the observed tail proportion and does not establish that any record is erroneous.)
- Suppose a hypothetical measurement follows \(X\sim N(50,10^2)\), and one observation is 80. Calculate its standardized value and approximate upper-tail probability. State what that probability means. (Answer: \(z=(80-50)/10=3\); \(P(X\geq80)\approx0.0013\), meaning the upper-tail probability under the stated normal model, not the probability of a data error.)
- A recorded policy limit is 1,500,000, while most limits in the file are below 100,000. The source schedule confirms the policy belongs to a high-limit product and records the same amount. What is the appropriate conclusion? (Answer: the value is extreme in the pooled file but supported as a valid high-limit record; retain and document it, while considering the product class when describing the data.)
- A payment field is documented as dollars, but one flagged record is 250,000,000. The policy schedule shows a maximum payable amount of 250,000, and the source transaction is 250,000 dollars. What should be checked and, if confirmed, how should the record be handled? (Answer: investigate a likely unit or entry error against the source; if the source confirms $250,000 and the field should be dollars, correct the erroneous field to 250,000, retain the original value in the audit trail, and document the correction.)