Merging nearby points into one cluster marker erases whatever pattern existed inside that cluster

What it is

Interactive maps commonly merge nearby points into a single cluster marker, replacing per-point drawing with "how many points are in here". Clustering solves overplotting, but it crushes the shape of what it contains: one cluster can come from a uniform scatter, from two tightly packed subclusters, or from a ring with a hollow center—and on screen they are the same cluster. The aggregation step itself systematically hides within-cluster structure.

Why it happens

Aggregation is a spatial grouping plus a summary: all points inside the cluster collapse into one position and one count. Only the first moment (the total) survives; variance, multimodality, and holes are discarded. Two distributions that fall inside the same boundary with the same total produce identical markers. The boundary placement also decides which points group together—shift it slightly and a subcluster splits across two clusters or merges into a neighbor. This is the modifiable areal unit problem (MAUP) surfacing on interactive maps: part of the visible pattern belongs to the data, part belongs to the aggregation parameters. Zooming into a cluster recovers the detail, but recovery never happens on its own, and the default view always shows the aggregated state.

Where it stops holding

When the task only asks "where is there more or less", aggregation is a reasonable simplification, and it is an acceptable trade whenever within-cluster differences are small relative to between-cluster differences. It becomes fatal only when conclusions depend on distribution shape: identifying subcenters, spotting holes, or judging dispersion. Interactive maps provide a drill-down exit, but readers may not realize they need it—the default view says "there is a cluster here", not "this is what the cluster looks like".

Applying it

• Give cluster markers an expansion affordance (click to expand, zoom to split) so within-cluster structure can be recovered on demand.
• For clusters that matter, add an in-place summary (a mini histogram or small scatter) that restores the shape information the aggregation discarded.
• State the current aggregation level in the legend or caption so readers know they are looking at an aggregated view.
• Verify by sampling several clusters and comparing readings before and after expansion; any cluster whose conclusion changes is a region where aggregation has already distorted the judgment.

Related

• Same group: U12.03.1 Point maps systematically undercount in high-density regions because of overplotting · U12.03.3 Cluster markers must encode how many points they contain · U12.03.4 Hexagonal binning or density surfaces can replace direct point overlay · U12.03.5 Zoom-dependent aggregation radii make the same data support different conclusions
• Nearby: U12.03.5 Zoom-dependent aggregation radii make the same data support different conclusions · U8.06.1 Aggregation masks differences between subgroups
• Search terms: cluster · modifiable areal unit problem · aggregation bias