Key points at a glance
- An unusual value may be an error or a genuine observation.
- Choose inspection and treatment rules independently of preferred results.
- Where useful, show how justified alternatives affect conclusions.
Make unusual observations visible first
An outlier is an observation unusually distant from others. Whether it causes a problem depends on the measurement, subject and analysis. Begin with a view of the distribution and inspect individual records against the original data. A software warning is a prompt to investigate, not a final decision.
NIST describes unusually distant observations and their potential effects on analysis. Combine this statistical view with subject knowledge: a long task time might be a recording mistake, an interruption or a genuinely lengthy task.
Distinguish errors from genuine cases
Check units, decimal separators, imports and questionnaire logic. If a daily-use question appears to contain 900 hours, investigate whether 9.00 was imported incorrectly or a different unit was supplied. Correct a value only when the replacement is supported by evidence.
Original example: A respondent reports 60 hours of study per week. This is unusual but possible. Without further information, you must not silently change it to six hours or remove it merely because it changes the mean.
Investigate the effect on your analysis
Compare suitable summaries and graphics. Means and medians respond differently to extremes. In a model, one observation may be highly influential; use diagnostics appropriate to the model instead of identifying only the largest raw value.
Set out how you will handle confirmed input errors, valid extremes and unresolved cases. Exclusion needs a substantive reason. A robust analysis or additional sensitivity check can show how much your conclusion depends on a particular observation.
Preserve raw data and record corrections or exclusions in a working copy. This keeps the analysis traceable and prevents a later export from silently changing the figures.
Report the decision and its consequences
State the detection rule, investigated cause, decision and affected sample count. If the conclusion changes markedly when a valid extreme is excluded, discuss that sensitivity. Do not hide the difference because one version better supports your hypothesis.
For the invented example, you could report: “The 60-hour value was verified in the original response and retained. We additionally report the median and an analysis excluding that record to examine its influence.” The actual choice depends on your data and discipline.
Final submission context
Formal details can feel like a separate writing task, but they matter just as much for a printed submission. Anything missing, misplaced or inconsistently formatted in the document will also appear in the bound copy.
Use this guide together with your cover page, table of contents, page numbers, source notes and appendices. Prepare the final PDF for printing and binding only after the complete file has been checked.
Practical check before PDF export
- Do headings, chapter structure and page numbers match?
- Are sources, figures and tables included completely?
- Are required elements such as declarations, appendices or the cover page included where required?
- Did you open and check the final PDF after exporting it?
Checklist
- Check unusual values against raw records.
- Inspect units and imports.
- Distinguish valid extremes from errors.
- Study influence and defensible alternatives.
- Document rules, changes and sample counts.
Common mistakes
- Automatically deleting every value beyond a standard cut-off.
- Removing only cases that obstruct a preferred result.
- Presenting an unsupported guess as a correction.
Print and bind your finished thesis
Once you have checked the content and final PDF, configure the printed copies to match your submission requirements.
Configure printing and bindingRelated content
Explore related guidance on thesis structure, sources, formatting and final submission.
Frequently asked questions
May I remove outliers?
Yes, when a substantive and methodological rule supports exclusion, such as a documented recording error. Record the decision and its effect. Being an unusual valid observation is not a universal reason for deletion.
Is a boxplot enough?
It can identify potentially unusual values. It cannot by itself establish the cause, make a disciplinary judgement or assess all analytical consequences.
Must every extreme value appear in the main text?
Report cases that materially affect your rules or interpretation. Detailed records can go in an analysis log or suitable appendix according to the project’s scope.