Operations & Systems

Outliers in Business Data: Investigate the Record Before Excluding It

A support report lists a resolution duration of three hundred sixty hours. Someone suggests removing it before the weekly review.

Open the record first.

An unusual value gives you a question to investigate. It does not supply the answer. The record might contain a unit error. It might describe a genuine complex case. It might belong to a different reporting population.

The review log below is a proposed operating method. All records, evidence, dispositions, and reviewer roles are fictional. The example uses no customer data and establishes no benchmark for acceptable support duration.

Separate a Flag From a Finding

A potential outlier is an observation marked for review. A verified error has evidence showing what is wrong. Keep those states separate in the report and in the underlying log.

NIST's outlier guidance describes observations that differ markedly from others in the sample. It distinguishes labeling for investigation from formal identification. It recommends correction where an error can be established and cautions against simply deleting unexplained unusual observations.

That distinction sets the boundary for this worksheet. The article supplies a record-review process, not a statistical test. Its flags identify questions for a person to resolve.

Start with the reporting question: “How many elapsed hours did the included resolved support cases take from opening to final resolution?” The example covers one fictional queue. Include resolutions from 2026.10.20 00:00 UTC to before 2026.10.25 00:00 UTC. One record represents one case, counted once.

Use the business metric dictionary to state the clock. Here elapsed hours include waiting and hours when the support queue is closed. They are not active labor hours. All timestamps below use UTC.

The comparison group contains eight resolved cases of the same service family. A complex case remains eligible. No policy in this example excludes escalations, long waits, or particular outcomes.

Preserve the Complete Extract Before Reviewing It

Keep the original input and its field labels. Retain the report version, extraction time, population rule, and any conversion applied between source and report. The review must explain how a value entered the summary.

The following fictional extract presents every duration as hours. It contains one unit error and one genuine complex case. Those are facts supplied by the synthetic example, not conclusions inferred from the sizes alone.

Case Imported duration labeled hours Initial review note
S01 2 No initial flag
S02 3 No initial flag
S03 4 No initial flag
S04 5 No initial flag
S05 6 No initial flag
S06 8 No initial flag
S07 360 Check source unit and timestamps
S08 72 Check whether the long duration is genuine

The raw sum is 2 + 3 + 4 + 5 + 6 + 8 + 360 + 72 = 460 hours. There are eight records. The raw mean is 460 / 8 = 57.5 hours.

Sorted raw values are 2, 3, 4, 5, 6, 8, 72, 360. The middle pair is five and six. Their midpoint is 5.5 hours, the raw median.

Call these the as-imported results. Their arithmetic is reproducible, but their measurement is not yet trustworthy. Once the error is confirmed, retain this snapshot as the explanation of the original report. Do not present its mean as the corrected operational result.

Trace the Unit Error to Its Evidence

In the fictional source record, S07 opens at 2026.10.20 08:00 UTC and resolves at 2026.10.20 14:00 UTC. The difference is six elapsed hours.

The source duration field contains 360, with a separate unit field of minutes. The reporting import copied the number into a column labeled hours without conversion. These source fields and timestamps establish the correction.

Calculate 360 / 60 = 6 hours. Correct the reporting value to six hours. Retain the original 360, original unit, corrected value, evidence reference, and reviewer decision in the review log.

A smaller number alone would not justify the change. The evidence is the source unit and matching timestamps. If the timestamp evidence conflicted with the duration field, leave that conflict open until someone resolves it.

Do not delete S07 from the population. The support case is real within this synthetic dataset and satisfies the inclusion rule. Its measurement needs correction; its membership does not.

For a real process, record the approved correction at the appropriate reporting layer. Keep the raw extract intact. An example in a worksheet authorizes no change to a production record or customer system.

Keep the Verified Complex Case

S08 opens at 2026.10.21 09:00 UTC and resolves at 2026.10.24 09:00 UTC. The source history records a specialist escalation and a wait for an additional diagnostic response. The full elapsed difference is seventy-two hours.

Those invented source details verify the duration under this example's elapsed-time definition. They do not establish seventy-two hours of labor. They also do not prove whether the delay was avoidable.

Keep S08 in the primary summary. It belongs to the defined service family and resolution window. Complexity was never an exclusion rule.

The reviewer can add a separate question about the escalation path. That investigation may need a delivery owner and more evidence. Preserve the duration while the causal question remains open.

A report about standard cases could use a separately defined population. It would need an eligibility rule that readers can inspect and apply consistently. Calling every slow case nonstandard after seeing its duration would conceal the experience the original metric claims to describe.

Compare the Complete and Reviewed Results

After correcting S07, the reviewed values are 2, 3, 4, 5, 6, 8, 6, 72 hours. In sorted order, they are 2, 3, 4, 5, 6, 6, 8, 72.

The sum becomes 106 hours. The count stays eight. The mean becomes 106 / 8 = 13.25 hours. The central pair remains five and six, so the median remains 5.5 hours.

The sum falls by 360 - 6 = 354 hours. The mean falls by 354 / 8 = 44.25 hours. That change is a measurement correction. No support case resolved faster because the report was corrected.

View Count Sum of elapsed durations Mean Median Disclosed treatment
Complete as-imported extract 8 460 hours 57.5 hours 5.5 hours Contains confirmed unit error
Reviewed primary result 8 106 hours 13.25 hours 5.5 hours S07 corrected; S08 retained; none excluded
Sensitivity view without S08 7 34 hours About 4.86 hours 5 hours Omits genuine complex case for comparison only

The sensitivity view asks how much S08 affects the summary. It is not the official population. Removing seventy-two from 106 leaves thirty-four hours across seven cases. The exact mean is 34 / 7, approximately 4.86 hours.

Do not choose that view because it looks favorable. Show it beside the reviewed primary result and its omitted count. Its purpose is to make the effect of a genuine observation visible.

Adding elapsed case durations does not establish team workload. Cases can overlap and include waiting. Keep that sum here as an arithmetic reconciliation quantity, not a staffing measure.

Copy This Outlier Review Log

Use one row per flagged record. Link to permitted evidence rather than copying sensitive customer content into a general reporting document. The fields are editorial suggestions.

Field What to record
Record identifier Stable case key and retained extract version
Reporting question Measure, unit, comparison group, and time rule
Flag reason Exact value and reason it needs investigation
Source evidence Field names, original unit, timestamps, and controlled evidence reference
Disposition Keep, correct, exclude under a named rule, or unresolved
Original and reviewed values Both values with units; blank if still unknown
Inclusion effect Whether the case remains in the primary population
Reviewer Role or person responsible for the decision, plus review date
Sensitivity note Effect on counts and summaries; any separately labeled comparison
Follow-up Missing evidence, owner, and next review point

A completed fictional log can stay compact:

Record Flag Evidence Disposition Reviewer Sensitivity note
S07 Imported 360 hours Source says 360 minutes; UTC interval is six hours Correct to 6 hours; keep case Reporting reviewer, 2026.10.28 Count stays 8; primary sum decreases 354 hours
S08 Imported 72 hours UTC interval is three days; escalation history retained Keep at 72 hours Service reviewer, 2026.10.28 Separate omission view has 7 cases; primary result keeps 8

Write a reason someone else can reproduce. “Looks wrong” is a flag. “Source unit is minutes; import label is hours” is evidence for a correction.

Leave Unresolved Records Visible

Suppose a third flagged case has no accessible history. Its duration looks unusual, but no source confirms whether it is wrong. Mark the disposition unresolved.

Record what is missing and who can inspect it. If the primary result contains that value provisionally, disclose the unresolved count. If no trustworthy value exists, report the missing measurement count separately. Neither state is equivalent to zero.

NIST notes that formal outlier tests depend on distributional assumptions. Its page's tests focus on approximately normal data. This worksheet applies none of those tests and makes no normality claim about support durations.

An automatic threshold can create a review queue in a proposed workflow. Keep the distinction between a flag and a final disposition. A threshold by itself cannot establish that the source unit is wrong or that an escalation should be excluded.

The CRM data-quality guide covers broader hygiene work. This log handles the narrower question of one observation's evidence and treatment. Do not let a general cleanup label erase the record of a reporting decision.

Carry the Review Into the Next Report

Give the reporting automation process explicit inputs: retained extract, approved conversion rule, reviewed dispositions, and unresolved count. A scheduled refresh should not silently replace those decisions.

Before the next review, compare the population rule and correction history. Check whether the same unit mismatch appears in another incoming record. Keep any proposed upstream fix separate from the evidence supporting this report.

If a genuine exclusion rule changes, show its name, reason, affected count, and effective version. Preserve a bridge to the previous result. A new definition can change a number without changing the underlying service.

The primary result in this example contains eight cases. One value is corrected. One complex case remains. Nothing is excluded to improve the display.

Keep the record until the evidence explains its treatment.

Sources

Topic map: Business reporting and statistics

Articles in this cluster, then the 3 clusters that share the most topics with it.

This cluster: Business reporting and statistics

Neighbor: Spreadsheets, CSV files, and data handoffs

Neighbor: Records, access, and continuity

Neighbor: APIs, webhooks, and automation