Description
Flags a location or unit whose value differs sharply from its neighbours' values, even when it is unremarkable against the whole data set.
A spatial outlier is a point that disagrees with its neighbourhood rather than with the global distribution. For each location, compute the difference between its value and the mean of its neighbours' values, standardise these differences across locations (the Zs statistic), and flag those beyond about 2. Graphical checks support the test: the variogram cloud plots the difference between each pair of points against their distance, and the Moran scatterplot plots each standardised value against the average of its neighbours, so local disagreements stand out. Shekhar and Vatsavai show that a global test (mean plus two standard deviations) flags a different point from the neighbourhood test. The method depends entirely on how 'neighbour' is defined: geographic distance, adjacency, or any similarity structure such as related pages, adjacent steps or nearby stores.
Attributes
- Discipline
- Statistics & Modelling
- Loss type
- Unmigrated
- Stage
- Post-launch
- Data
- Quantitative
- Effort
- Hours
Steps
- Define neighbours for each unit: geographic adjacency, k nearest locations, or another structure such as linked pages or comparable stores.
- Compute each unit's value minus the mean of its neighbours' values.
- Standardise these differences across all units to get the Zs statistic.
- Flag units with |Zs| above about 2 for investigation.
- Confirm with a Moran scatterplot or variogram cloud and compare with a global outlier test.
- Investigate flagged units for instrument faults, local conditions or genuine local behaviour.
Sources
- Handbook of Data Mining (Ye, 2003)
Also known as
- Neighbourhood-based spatial outlier detection
- Spatial outlier
- Zs statistic
- Moran scatterplot
- Variogram cloud
Categories
- Analytics & Behaviour
- Statistics & Modelling