Statistics Glossary → Getis-Ord Gi* Hotspot
Getis-Ord Gi* Hotspotanalysis
A local clustering statistic that shows, with a z-score for each location, whether the sum of values at that location and its neighbours is significantly higher (hot spot) or lower (cold spot) than expected across the area.
Concept
A single high value is not a hot spot; what matters is high values occurring together. Gi* places a window around each unit: it sums the values of the unit and its neighbours and compares this local sum with the sum that would be expected if values were scattered at random across the area. The difference is expressed as a z-score: a large positive z marks a hot spot where high values cluster, and a large negative z a cold spot where low values cluster.
The starred Gi* includes the focal unit's own value in the sum; the unstarred Gi looks only at the neighbours. Neighbours can be defined as k nearest neighbours, a fixed distance band, or weights that decline with distance. Significance is computed by an analytic approximation or by permutations in which the values are shuffled among locations. When there is no value column, the points can first be counted into hexagonal cells and Gi* run on those counts, so hot spots can also be sought among event points.
Unlike Moran's I, Gi* separates only clusters of high values and clusters of low values, not high-low mixtures; it does not show individual units that differ from their surroundings. Because each unit is a separate test, a multiple-comparison correction is important. The result is sensitive to the neighbourhood scale, which should rest on a theoretical reason or on the spatial autocorrelation profile by distance (ISA). With counts, differences in population or in the quantity at risk must also be considered.
When to use it
- To map statistically significant areas of high and low values for a variable such as crime, disease, price or density.
- If, besides clusters, outlying units that differ from their surroundings (high-low, low-high) are also sought, use local Moran (LISA); for a single measure of clustering over the whole area, use global Moran's I.
- If a visual density surface is enough, use KDE; if the points are to be assigned to separate clusters, DBSCAN is preferred.
An everyday example
A housing researcher maps the monthly rent of 500 rental flats in a city. There are scattered expensive flats by the sea, but the real question is whether expensive flats are concentrated in particular districts. Gi* sums the rents of the eight nearest flats around each flat and compares them with the city average: an area along the seafront comes out as a strong hot spot with z = 4.8, and a neighbourhood near the industrial zone as a cold spot with z = −3.9. A street by the sea with only one expensive flat surrounded by cheap ones is not counted as a hot spot.
Parameters in MerQur
Classify by / Method
Concept: Classify is a data-exploration step that colours the points by a categorical column before the analysis; Method chooses the spatial analysis to run on the map tab. Gi* tests where high and low values cluster significantly.
Everyday example: Colouring the flats as 'furnished' or 'unfurnished' shows in advance whether furnished flats gather in certain districts. For the question 'where are rents significantly high?', Gi* is chosen.
What it does: Classify by colours the points by category (data exploration); Method selects the map method.
Effect on the result: Does not change the numeric output.
K neighbors (default 8)
Concept: Sets how many nearest neighbours join each point's local sum; this is the scale at which hot spots are sought. A small k finds narrow, sharp clusters; a large k finds broader but less detailed areas.
Everyday example: With k = 8 the area of a few streets around each flat is considered, and small pockets of expensive flats by the sea appear. With k = 20 the scale rises to the district, and the hot spot becomes a broad area covering the whole seafront.
What it does: In the "k nearest neighbours" weights, determines the number of neighbours of each point.
Effect on the result: Changes the Gi* z-scores and the hot/cold counts.
On the validation data: In Boston (analytic) hot 82, cold 112 for k = 8; 113 and 149 for k = 20. At k = 3 the graph is disconnected: MerQur raises k to 5 and writes this in the note (77 and 92).
α (alpha) and multiple-comparison correction (None, FDR (Benjamini-Hochberg), Bonferroni)
Concept: α is the significance threshold a point must pass to count as a hot or cold spot. Because each point is tested separately, without a correction some points turn out significant by chance even without a real cluster; FDR controls the share of false positives among the significant points, Bonferroni the probability of any false positive.
Everyday example: If each of the 500 flats is tested at the 5% level, about 25 could come out significant even with no real pattern. With FDR one can be more confident that most of the hot spots left on the map are real.
What it does: Determines the significance threshold and the correction method; since each point is tested separately, limits false positives.
Effect on the result: Changes the hot/cold spot classes; z does not change.
On the validation data: Hot spots 63, 82, 95 for alpha 0.01, 0.05, 0.10. At alpha 0.05 hot/cold 67 and 87 with FDR, 23 and 31 with Bonferroni.
Permutation (0 analytic, 99, 999, 9999) and Seed
Concept: The permutation test keeps the focal point's value fixed while shuffling the other values randomly among locations, building the distribution of 'what Gi* would be without a pattern'; the analytic method relies on a normal approximation. The smallest attainable p value is 1/(M + 1); the seed lets the same shuffles be reproduced.
Everyday example: If the rent sum around a seafront flat is exceeded in none of 999 random shuffles, p = 0.001. Rerunning the report with the same seed finds the same hot spots; with a different seed the class of a few borderline flats may change.
What it does: When 0, analytic p; when greater than 0, the folded p by conditional permutation (smallest 1/(M + 1)). The seed makes the same result reproducible. Permutation is not used with inverse distance.
Effect on the result: Affects the p values and hence the hot/cold counts; z does not change.
On the validation data: At 999 permutations (seed 20261012) hot 82, cold 127; 64 and 110 at 99 permutations (64-87 and 103-135 over 20 seeds).
Aggregate ([Parameters] check box) and Grid (hexagons along x)
Concept: With no value column, event points are counted into hexagonal cells and Gi* is computed on these cell counts; cells without points enter as 0. This answers the question 'where are events significantly dense?'; the grid number sets the cell size, and hence the scale of the analysis.
Everyday example: If the researcher studies the locations of residential burglaries instead of rents, the events have no value. They are counted into hexagons, and areas where high-burglary cells lie side by side come out as hot spots.
What it does: Counts events without a value column: points are counted into hexagons, empty cells inside the convex hull are added as 0, and Gi* is computed on cell counts with hexagonal contiguity (like ArcGIS Optimized Hot Spot).
Effect on the result: Turns the unit of analysis from a point into a cell; the value column is not used.
On the validation data: For grid 15, 199 cells, 21 hot and 0 cold; for grid 30, 757 cells, 71 and 0.
Neighbourhood (k nearest neighbours, Fixed distance band, Inverse distance (within band)) and Distance band (km)
Concept: The neighbourhood defines which units join the local sum: k nearest neighbours gives every unit the same number of neighbours; a fixed distance band takes everyone within a set distance with equal weight; inverse distance gives more weight to closer units within the band. The band should reflect the distance over which the phenomenon operates.
Everyday example: If rents are thought to be affected mainly by homes within 1 km, the band is set to 1 km. In the dense centre this gives a flat dozens of neighbours, and in sparse outer districts only a few; k nearest neighbours instead takes the same number of neighbours everywhere.
What it does: The distance band gives weight 1 to neighbours with d <= threshold; inverse distance uses 1/d within the band. If no threshold is entered, the smallest band giving every point at least one neighbour is used. Distances are measured on the local equirectangular plane (km).
Effect on the result: Changes the neighbourhood structure, hence all Gi* values.
On the validation data: Boston automatic band 3.968 km; fixed distance band neighbourhoods are built at the thresholds 30.01, 60.01 and 90.01.
Gi* / Gi (check box)
Concept: Gi* includes the focal unit's own value in the local sum and answers 'are this unit and its surroundings high together?'. Gi leaves the focal unit out and answers 'are this unit's surroundings high?'.
Everyday example: A single very expensive villa can, under Gi*, produce a high sum with its surroundings because of its own value. Gi looks only at the villa's neighbours; if they are ordinary rentals, that point does not come out as a hot spot.
What it does: If checked, Gi* (focal point included, ArcGIS Hot Spot); if unchecked, the focal point's own value is removed from the sum (Gi).
Effect on the result: Changes the z-scores.
On the validation data: 75 hot and 107 cold for Gi (Gi*: 82 and 112).
Gi_Bin (check box)
Concept: Gi_Bin divides the z-scores into classes by confidence level: ±3 a hot or cold spot at 99%, ±2 at 95%, ±1 at 90% confidence, 0 not significant. The map thus shows not just 'significant / not significant' but also the strength of the evidence.
Everyday example: On the rent map the core of the seafront may come out in class +3 and its edges in class +1. For 'the most reliable hot spots' the council can look at class +3 only.
What it does: Gives the ArcGIS Gi_Bin classes: +-3 = 99%, +-2 = 95%, +-1 = 90% confidence hot/cold (by the corrected p), 0 = not significant.
Effect on the result: Shows colour classes by confidence level on the map; z and p do not change.
On the validation data: In Boston, in the analytic run (hot 82, cold 112) -3: 76, -2: 36, -1: 13, 0: 286, +1: 13, +2: 19, +3: 63; in the run with 999 permutations (hot 82, cold 127) -3: 100, -2: 27, -1: 15, 0: 253, +1: 29, +2: 27, +3: 55. The permutation p is folded one-tailed (esda), while the analytic p is two-sided; the counts differ for this reason and because of the skewed distribution.
ISA (check box)
Concept: ISA (incremental spatial autocorrelation) computes global Moran's I over a series of increasing distance bands. The distance at which z peaks shows the scale at which clustering is most pronounced and is a starting suggestion for the distance band.
Everyday example: If the ISA plot shows z peaking at about 1.5 km, clustering in rents is most pronounced at that scale. The researcher sets the distance band to 1.5 km and runs the hot-spot analysis at that scale.
What it does: Computes the global Moran's I z (randomization) in 10 distance bands; the distance at which z peaks is recommended for the distance band.
Effect on the result: Does not change the Gi* result; produces an extra table for choosing the band.
On the validation data: In Boston the peak is 6.227 km; the ISA table is also produced for the lattice data.
Value (none, numeric column)
Concept: Sets the numeric variable on which Gi* is computed. It is required when aggregation is off; when aggregation is on, point counts are used and no value column is needed.
Everyday example: Choosing the monthly rent column makes the analysis look at clustering of rents. Had rent per square metre been chosen, the effect of large flats simply looking expensive would have been removed.
What it does: Selects the variable analysed for Gi*; it is required when aggregation is off (if not selected, a "a value column must be selected" warning).
Effect on the result: Determines all Gi* z-scores.
On the validation data: CMEDV in the default run.
Example on real data
Data: Harrison and Rubinfeld 1978, spData boston
Variables: 506 census tract centroids (latitude/longitude), value CMEDV (house value); seed 20261012
- Question: Do areas of high and low house value (CMEDV) in Boston cluster in space?
- Method: Getis-Ord Gi*; k = 8 nearest neighbours + the point itself, binary weights; p from 999 conditional permutations.
- Result: 82 hot spots (high-value cluster) and 127 cold spots (low-value cluster) are significant at p < .05; 297 points are not significant.
- Effect size: The z values range from -4.99 to 5.27; the strongest clusters (Gi_Bin ±3, p < .01) are 55 hot and 100 cold points.
- Limitation: 506 separate tests are run; if no correction (FDR) is selected some points may be significant by chance. The permutation p values depend on the seed.
- Report line: Gi* (k = 8, 999 permutations) identified 82 hot and 127 cold spots (p < .05).
Related concepts
Permutation p-valueSignificance level (α)Type I and Type II errorsLocal indicators of spatial association (LISA) and Gi*Spatial weights matrix (W)Spatial autocorrelation (Moran's I, Geary's C)z-score (standardization)Inverse distance weighting (IDW)
The numbers in the examples are MerQur's output on independent, publicly available validation data; see the MerQur Technical Note for details.