Outliers may not be the result of actual observations, but rather the result of errors in data collection, data recording, or other parts of the data life cycle. The following can be used to identify outliers for closer examination:
- Outliers may be detected by using Dixon’s test, Grubbs test or the Tietjen-Moore test.
- Box plots are useful for indicating outliers
- Scatter plots help identify outliers when there is an expected pattern, such as a daily cycle
Comparison to related observations:
- Difference plots for co-located data streams can show unreasonable variation between data sources. Example: Difference plots from weather stations in close proximity or from redundant sensors can be constructed.
- Comparisons of two parameters that should covary can indicate data contamination. Example: Declining soil moisture and increasing temperature are likely to result in decreasing evapotranspiration.
No outliers should be removed without careful consideration and verification that they are not representing true phenomena. Description Rationale:
Outliers may represent data contamination, a violation of the assumptions of the study, or failure of the instrumentation. Although outliers may be valid observations it is important to identify and examine their validity.
V. Barnett and T. Lewis, Outliers in Statistical Data (John Wiley & Sons, 2d ed., New York, NY, 1985).
Edwards, D. 2000. Data Quality Assurance. Pages 70-91 in: Ecological Data: design, management, and processing. Michener, W. and Brunt, J., eds. Blackwell Science Ltd. (ISBN: 0-682-05231-7).
Cite this best practice:DataONE Best Practices Working Group, DataONE (July 01, 2010) "Best Practice: Identify outliers". Accessed through the Data Management Skillbuilding Hub at https://dataoneorg.github.io/Education/bestpractices/identify-outliers on Aug 22, 2019