An Outlier Flag Is Not Proof of Bad Data

What an Outlier Flag Tells You

The 1.5-IQR rule identifies observations that lie unusually far from the middle half of a data set. That makes a flag useful: it draws attention to a value worth examining. But the rule only uses the numerical arrangement of the recorded data. It does not know how a value was measured, entered, or produced. A value can be genuinely unusual and accurately recorded; a value can also be an error without crossing either fence.

This distinction matters whenever a summary or comparison is based on recorded measurements. Removing a valid extreme changes the account of what was observed. Keeping an incorrect value as if it were valid can also mislead. The appropriate decision comes from checking the value against its context and source, not from the flag alone.

The rule is a descriptive screening convention, not a test of whether an observation is correct. It does not calculate the probability that a value is an error, and it does not say that an observation beyond a fence must be excluded. Conversely, a value inside the fences is not guaranteed to be correct. An extra zero, a misplaced decimal point, or a unit mistake might produce a value that is not flagged, depending on the rest of the data.

Check the Observation, Not Just Its Position

A useful check begins with the meaning and units of the variable. Ask whether the value could occur in the setting, whether it is consistent with the way the quantity was measured, and whether anything about the record suggests a transcription or coding mistake. “It seems surprising” is a reason to investigate, not enough evidence to change the data.

If possible, compare the data entry with its source: a paper form, instrument output, database field, or contemporaneous log. Check that the correct person or item is attached to the value, that the decimal and units were transferred correctly, and that the value has not been copied into the wrong variable. When appropriate, a repeat measurement or an independent record may help. Each check must be interpreted in context: a repeat measurement can differ from the first because the quantity genuinely changed or because measurement itself varies.

These categories describe the evidence available, not permanent labels attached to a person or measurement. A confirmed error is supported by a source check showing that the recorded entry does not represent the intended observation. A verified unusual value has been checked and is supported as an accurate observation, even if it is rare or extreme. An unresolved value lacks enough evidence to decide between those possibilities. Keep the distinction visible in any analysis or report.

Worked Examples

Why the Flag Cannot Decide

The reason can be seen by separating two questions. The fence calculation asks, “Where does this number sit relative to the middle half of these recorded values?” A source check asks, “Does this recorded number accurately represent the intended observation?” These questions use different information. The first can be answered using the data values; the second requires information about measurement, recording, or context.

For example, the same numerical value could be an accurately measured rare event in one setting and a transcription mistake in another. Its position relative to \(Q_1\) and \(Q_3\) would not reveal which story is true. Even a value that appears physically implausible should prompt a check of units, definitions, and the data source before it is changed. A mistaken unit conversion may be correctable; a genuinely extreme event may instead be what the observation captured.

A defensible correction is tied to evidence. If a source record clearly shows a transcription error, correct the entry and document the correction. If checks support the recorded value, retain it as an unusual observation. If the source is unavailable or conflicting, preserve the value as recorded and state that its status could not be resolved. These choices make clear both what the data show and what the checks were able to establish.

Common Mistake: Deleting Every Flag

Wrong reasoning: “The value is beyond the 1.5-IQR fence, so it must be bad data and should be removed.” This treats a screening rule as a test of data accuracy. It can discard a genuine observation and change the account of the data without evidence.

Correction: A flag identifies a value for review; it does not supply a reason for deletion. Check the context and available source records. If they confirm an error, correct it transparently. If they support the value, retain and describe it. If they do not settle the question, report it as unresolved rather than presenting a guess as a fact.