Advanced Statistics
Extreme Observation or Data Error? Influence Diagnostics and Robust Analysis
Do not delete an observation because it changes significance; investigate errors, influence, leverage and robust sensitivity analyses.
Admin
June 20, 2026
8 min read
An outlier is not automatically an error
An extreme observation may represent a real subgroup, error or model misspecification. Decisions need prespecified rules.
Practical framework
Verify source and units.
Separate extremeness from influence.
Use leverage, Cook distance and residual plots.
Compare primary and robust sensitivity results.
Reporting
Report detection rule, affected observations, rationale and impact.
Common mistakes
Deleting every boxplot point.
Removing to obtain significance.
Assuming every outlier is erroneous.
References
Accessed 20 June 2026. General methodological information.