Statistics Glossary → Outlier
Outlierconcept
An observation that lies clearly apart from the rest of the data and may reflect a measurement error, a different process or a genuine but rare case.
For a single variable the most common rule is Tukey's boxplot rule: values more than 1.5 × IQR below the first quartile or above the third quartile are possible outliers, and those beyond 3 × IQR are extreme outliers. The z score (|z| > 3) can mask outliers because the mean and standard deviation are themselves pulled by them; the modified z score based on the median and MAD is more robust. In multivariate data, observations that look normal one variable at a time but unusual in combination are found with the Mahalanobis distance, and in data with varying density with methods such as LOF or Isolation Forest.
Flagging a value as an outlier is not by itself a reason to delete it. First check whether it is a data entry or measurement error; if it is a real observation, run the analysis with and without the outliers and report how sensitive the results are. In regression, a point extreme in the outcome (a large residual) differs from a point extreme in the predictors (leverage); what really changes the results are influential points that combine both. Robust methods (the median, robust regression) are less sensitive to outliers.
An everyday example
In an apartment block, monthly water use ranges from 8 to 15 m³, but one flat shows 140 m³. It might be a meter-reading error (140 instead of 14), a burst pipe, or a laundry business run from the flat. Delete it without checking and a real leak could be missed; keep it without checking and the average of the 12-flat building jumps from 12 to about 23 m³.
In MerQur
In Descriptive Statistics, the 'Extreme values / outliers' option lists the five highest and lowest values and reports observations beyond 1.5 hinge spreads using Tukey's hinges. The 'Method' option of Anomaly Detection offers IQR, z score, MAD, Isolation Forest, LOF, Mahalanobis, MCD, Grubbs, Rosner ESD and STL remainder criteria. In multiple linear regression 'Extra output: Influential observation diagnostics' gives the Bonferroni outlier test, and robust regression gives MCD-based outlier diagnostics; DBSCAN / HDBSCAN counts points that belong to no cluster as noise, and the K-medoids (PAM) algorithm in k-means clustering and the Cressie-Hawkins estimator in the variogram are more robust to outliers.
Analyses where this concept appears
Descriptive StatisticsMultiple Linear RegressionK-Means ClusteringDBSCAN / HDBSCANRobust Regression (M-estimator)STL Seasonal DecompositionAnomaly DetectionLOESS (Local Regression)Variogram + Kriging (Spatial Interpolation)
Related concepts
Interquartile range (IQR) and Tukey fencesMedian absolute deviation (MAD) and modified z-scoreMahalanobis distance