Real clusters: hot spot analysis (Getis-Ord Gi*), explained simply
A bright patch on a map is not proof of anything - random scatter makes patches too. 'Real clusters' runs the standard hot-spot test (Getis-Ord Gi*) and marks only the areas where there is genuinely more, or less, around than chance would explain. Pubs in London show how.
Colour a map by pub counts and you will see busy patches. Some of those patches are real districts. Some are what random scatter looks like - throw enough pubs at a city and a few will land next to each other by luck alone. The two look identical on an ordinary map.
The Clusters map tells them apart. It runs the standard hot-spot test used across the GIS industry - Getis-Ord Gi*, the thing ArcGIS calls Hot Spot Analysis - and colours only the areas where there is genuinely more, or genuinely less, around than chance would explain. Everywhere else stays grey, because “nothing unusual” is the honest answer nearly everywhere.
Clustersthe field's term: Getis-Ord Gi*
Pubs, tested for real clusters. Orange is a genuinely busy cluster; grey is nothing unusual. The centre, plus a string of town centres - Ealing, Wimbledon, Kingston, Bromley, Romford - come out as real.
What the test actually asks
For every area it looks at the area and its neighbours together, and asks one question: if London's pubs were shuffled at random across the city, how often would a neighbourhood this busy turn up? If the answer is “hardly ever”, the cluster is real. If the answer is “all the time”, it is noise.
Genuinely busy cluster
More pubs around here than chance would explain. Not one busy area, but a neighbourhood that is busy together.
Nothing unusual
The neighbourhood is within what random scatter would produce. This is most of the map, and it should be.
Genuinely quiet
Fewer than chance would explain - a real cold spot. For pubs in London there are none; for other metrics there can be many.
Two details make the result trustworthy rather than merely pretty. The test is two-sided, so quiet pockets can be significant as well as busy ones. And it applies a correction (Benjamini-Hochberg) that removes the roughly 5% of areas that would look unusual by luck alone across 40 000 areas - without it, a city-sized map would be dotted with false alarms.
Switching it on
Clusters is a row in the Map type picker, under the heading Is it real - the group whose maps answer with a verdict rather than a number. Whatever metric you already had comes with you; only the question changes.
Clusters selected in the picker: the plain name, the field's term beside it, and the question it answers. The three-colour thumbnail is the legend you are about to get.
Where to click
Map type
Clusters
Pub
The panel counts the verdicts instead of drawing a spread: 1 632 areas are genuinely busy clusters for pubs, 11 017 are nothing unusual, none are genuinely quiet.
One row is missing here, and that is deliberate
On every other map the measure has a Look at row, choosing between this spot and a 10-minute walk. On Clusters it is simply absent. A cluster test already reads each area's neighbours, so a walk would count the neighbourhood twice - and rather than show you a permanently greyed control on a map where it could never be used, the app leaves the row out. A greyed option tells you something about your data; this would only have told you something about the map you had chosen.
Reading the map
Three colours, no ramp. There is no “slightly hot” here, because the question has a yes-or-no answer: either the neighbourhood beats chance or it does not.
The legend also carries the one honest caveat: areas at the edge of the data are not judged, because their neighbourhood is incomplete.Hovering an orange area: the verdict in words, then the area's own count. This spot has just one pub - it is the neighbourhood around it that is unusual.
Why an area with one pub can be a hot spot
The test is about the neighbourhood, not the single area. An area with one pub whose neighbours all have several is part of a real cluster; an area with six pubs surrounded by none is a lone spike, and Gi* dilutes it. That is exactly the opposite of a top-ten list, and it is why the orange patches look like districts rather than dots.
The maths, honestly
For each area, add up the pubs in the area and its two rings of neighbours. Work out what that total would be on average if the city's values were shuffled. Compare the two, divided by the normal wobble you would expect from shuffling. The result is a z-score: roughly, how many wobbles away from chance this neighbourhood sits.
'Show the maths' is where the field's term lives. On the map itself you only ever see the three verdicts.
The statistic never appears on the map, and that is a rule of the product rather than an omission: a z-score of 2.6 means nothing to most people, but “a genuinely busy cluster, not a fluke” does. The z-score is there under “Show the maths” for anyone who wants to check the working.
Real clusters asks whether a neighbourhood is busier, or quieter, than random scatter would produce. That is the Getis-Ord Gi* test.
Three verdicts, no ramp: genuinely busy, nothing unusual, genuinely quiet. Most of any map is the middle one.
A false-discovery correction removes the areas that would look unusual by luck, and edge areas are never judged.
The 'Look at' row is absent here, because the test already reads the neighbours and a walk would count them twice.
Hover for the verdict in words; open 'Show the maths' for the statistic behind it.
Want the magnitude rather than the verdict? That is the Compared to the average post - the two are designed to be read together.
Quick questions
What is hot spot analysis?
A statistical test (Getis-Ord Gi*) that looks at each area together with its neighbours and asks whether there is more, or less, of something there than a random arrangement of the city's own values would produce. MapBees calls it 'Real clusters'.
Why is most of the map grey?
Because 'nothing unusual' is the honest answer nearly everywhere. Only areas whose neighbourhood is genuinely busier (orange) or quieter (blue) than chance would explain are coloured, after a correction that removes the ~5% of areas that would look unusual by luck alone.
Why is there no 'Within a 10-minute walk' option on the Clusters map?
Because the cluster test already reads each area's neighbours. Adding a walk would count the neighbourhood twice, so the app leaves that row out entirely on this map rather than showing a control that could never be used.
What happens at the edge of the city?
Areas whose neighbourhood runs off the edge of the data are not judged. Their neighbourhood is incomplete, so a verdict there would be about the wrong neighbourhood.
Do I need to know what a z-score or p-value is?
No. The map says 'Genuinely busy cluster', 'Nothing unusual' or 'Genuinely quiet', and the panel counts how many areas fall in each. The statistic is available under 'Show the maths' for anyone who wants it.
See it on a real map
The best way to understand the score is to play with it. The full app is on a free tier, with no sales call.
Builds MapBees's location-intelligence pipeline end to end — census and Overture Maps ingestion, H3 hex scoring, and the map interface — and writes the methodology behind every number on the site.
