Outliers: preserve raw values and treatment consequences
An outlier is a signal to investigate, not an automatic instruction to delete a value. Preserve the raw observation, explain why it is unusual and show how any treatment changes the research result. A spectacular number may be an extraction error, a different economic regime or a genuine event that matters more than the typical observation.
Distinguish three questions
First, is the value recorded correctly? Check units, signs, dates, decimal separators and the original disclosure. Second, is the observation comparable to the intended sample? A business disposal can make a growth rate unusual without making the reported number wrong. Third, how sensitive is the conclusion to this observation? These questions need different evidence and should not collapse into one cleanup label.
The NIST engineering statistics handbook discusses identifying and investigating outliers. Its statistical discussion does not justify deleting every extreme investment observation. A flagging rule is a diagnostic; treatment needs a reason connected to data quality and the analytical question.
A hypothetical growth dataset
Suppose five invented companies report revenue growth of 2%, 3%, 4%, 5% and 86%. The mean is 20%, while the median is 4%. Removing the last company gives a mean of 3.5%. Capping its growth at a prespecified illustrative threshold of 10% gives 4.8%. Those are different statistics under different data treatments, not competing versions of the same raw fact.
| Treatment | Result | What remains unresolved |
|---|---|---|
| Original observations | Mean 20% | Is 86% comparable and correctly measured? |
| Median of originals | 4% | Does a typical issuer answer the question? |
| Exclude last issuer | Mean 3.5% | Is exclusion justified independently of the result? |
| Cap at 10% for illustration | Mean 4.8% | Why this threshold, and what was suppressed? |
The 10% cap is an invented scenario choice, not a recommended cutoff or an empirically estimated percentile. If the 86% value reflects an acquisition, your question may require an acquisition-adjusted measure. If it comes from a unit error, correction is more appropriate than clipping. If it represents a real organic expansion, removing it changes the economic story.
Keep a treatment ledger
For each flagged observation, save the original value, source locator, flagging rule, investigation outcome, derived value and reason for treatment. Distinguish corrected extraction, scope exclusion, winsorization and unresolved review. A blank field should not hide whether the value was unavailable or intentionally removed.
Set rules before inspecting the preferred result where possible. When a surprising finding motivates a new treatment, disclose that sequence and show the untreated result too. Running dozens of thresholds and reporting only the one that supports your thesis can manufacture apparent robustness. The ledger should reveal such analytical choices rather than bury them in code.
Test the economic consequence
Compare the headline statistic, direction of the conclusion and sample weight affected by treatment. Removing one company from five is a substantial sample change even if it looks like one small edit in a spreadsheet. For an exposure-weighted analysis, also show the removed portfolio or revenue weight; observation count alone may understate the effect.
Use robust summaries when they answer a meaningful question, but name them precisely. The median describes the middle observation and does not estimate combined revenue growth. A trimmed mean changes the included distribution. None of these methods repairs a mislabeled period or incompatible accounting definition.
Make the final claim conditional
A useful conclusion could state that the typical issuer grew modestly, while one acquisition-related observation drives the simple average upward and requires separate analysis. That preserves information about concentration. Avoid asserting that growth is strong merely because of an unexplained extreme, or that the extreme is false merely because it is inconvenient.
FAQ
Must an outlier be a data error?
No. It may be a correctly observed rare event or a member of a different population. Investigation comes before treatment.
Is winsorization safer than deletion?
It preserves row count but replaces values with a rule-based limit. That can still materially change results, so preserve raw values and report sensitivity.