Statistical Outlier Detection
Methods using statistical distribution measures to identify observations unusually far from a reference pattern.
Statistical outlier detection measures whether an observation lies unusually far from, or in a sparsely populated part of, a reference distribution. An outlying value is not automatically a transcription error, a faulty sensor reading or evidence of abuse. The result depends on the chosen model and scale.
Tukey's IQR fences
A familiar exploratory convention is:
lower fence = Q1 − 1.5 × IQR
upper fence = Q3 + 1.5 × IQRWith Q1=10, Q3=18 and IQR=8, the fences are −2 and 30. A value of 35 would be flagged for investigation. The calculation does not establish that the measurement is invalid or should be removed.
When a normal-reference score misleads
A standardized score z=(x−μ)/σ describes distance from the mean in standard-deviation units. Skewed, multimodal or heavy-tailed distributions can make a fixed |z| cutoff too sensitive or too permissive. Median-based and median-absolute-deviation approaches can be more resistant in some settings. Zero variance and zero robust scale require deliberate handling.
Statistical score versus operational evidence
Measurement resolution, physical limits, source integrity and temporal dependence belong to a separate assessment. Independent univariate thresholds can miss anomalies that exist only through relationships among variables. A box plot displays candidate extremes, while IQR supplies their scale. Learning-based detection methods are addressed separately in Anomaly Detection.