One point far from the others can change a graph or a statistical result substantially. That makes it a reason to investigate, not an automatic reason to delete. Reading exclusions carefully means asking what was removed, why, when the rule was chosen and how the decision changes the evidence.
Separate a surprising value from a known mistake
NIST distinguishes outlier flagging, formal identification and methods that accommodate unusual observations. It notes that an unusual point may reflect an error, ordinary variation or a scientifically interesting feature; deletion without establishing the reason can lose valid information.NIST/SEMATECH — Detection of Outliers (opens in a new tab)
A transcription of 120 instead of a documented instrument value of 12 is a different problem from an accurately recorded value of 120 that simply differs from the rest. The first can be checked against a source record. The second requires investigation of the measurement and model.
| Finding | Reading question |
|---|---|
| Verified typo | Was the corrected value traced to the original record? |
| Recorded failure | Does it invalidate this measurement or the whole unit? |
| Unusual value | What explains its classification as unusual? |
| Model mismatch | Would another defensible model describe the data better? |
Calling all four situations data cleaning hides their different implications. A paper should explain the actual reason for changing or excluding a value.
Inspect the effect of the removal
Consider an original five-value illustration: 8, 9, 10, 11 and 27 units. Their mean is 13. Removing 27 leaves a mean of 9.5. The numerical change is substantial, but the arithmetic does not tell us whether removal is justified.
If a source record demonstrates that 27 was entered incorrectly, the analysis should use a traceable correction where possible. If 27 is a valid measurement, deleting it because it weakens the preferred conclusion would change the evidence through outcome-dependent selection.
A sensitivity analysis can show the result with and without a disputed observation, clearly identifying why the alternative was examined. It does not convert the excluded version into the correct result simply because that version looks cleaner.
Find the rule and the timing of its application
NC3Rs recommends specifying inclusion and exclusion criteria before an experiment to reduce the opportunity for decisions to be shaped by emerging results. It distinguishes criteria for units, samples and individual data points.NC3Rs EDA — Inclusion and exclusion (opens in a new tab)
Read whether the rule was applied consistently across conditions and whether the decision-maker knew the group or outcome. A rule applied only to inconvenient values in one condition deserves a different interpretation from a documented measurement-quality rule applied throughout.
A prespecified rule is not automatically a scientifically sound rule. It still needs to fit the measurement and research question. Conversely, an unforeseen equipment failure can justify a later decision if its basis and timing are reported transparently.
Formal outlier tests also have assumptions. A value that looks extreme under a normal-distribution model may be less surprising under a skewed distribution; a software flag does not settle which model is appropriate.NIST/SEMATECH — Detection of Outliers (opens in a new tab)
Reconcile the analysed sample with the original sample
Track how many units were assigned, how many measurements were obtained and how many entered each analysis. Losing one repeated reading is different from removing an independent experimental unit, and the reported sample size should reflect the level affected.
Exclusions can change the population or conditions represented by the result. For example, removing all poorly performing preparations may restrict the conclusion to preparations that passed that criterion rather than to the original set.
When reasons or counts are missing, state the reporting gap. Avoid inferring that every unexplained exclusion is misconduct, but do not assume it had no effect on the result either.
A useful summary identifies the excluded observations, the stated rationale, the timing of the rule and any sensitivity of the conclusion. It gives the reader enough information to evaluate the decision without reducing the issue to whether an outlier button was used.
Sources and further detail
- NIST/SEMATECH — Detection of Outliers (opens in a new tab)
Error-versus-variation distinction, flagging/accommodation and distribution assumptions read. The five-value calculation is original; no particular exclusion test is prescribed.
- NC3Rs EDA — Inclusion and exclusion (opens in a new tab)
Advance criteria and unit/sample/data-point distinctions read. No animal procedure or automatic deletion rule reproduced.
Sources checked 19 September 2026. Worked examples are illustrative unless a supplied report is explicitly identified. This article has not undergone independent scientific peer review.