Skip to content

Spatial or Network Cluster Detection

Statistical detection tool — instantiates Vulnerability Hotspot Mapping and Hardening

Tests where high-risk units genuinely cluster in space or on a network, screening out the concentrations that are only chance, so hardening targets real hotspots.

Where the index scores units and the heatmap overlays them, Spatial or Network Cluster Detection asks the sharper question: is the apparent concentration real, or would a lump like this appear by chance somewhere on a big map? It scans space or a network for statistically significant clusters, corrects for the fact that you looked everywhere at once, and screens whether the piled-up layers are independent signals or one driver showing up repeatedly. Its defining move is significance under multiplicity: a hotspot is not simply where a score is highest — it is where concentration exceeds what randomness alone would scatter across a large search space.

Example

Pediatric asthma emergency visits look elevated in a few neighborhoods near a rail-and-truck corridor. Eyeballing the map, five or six "clusters" jump out — but across hundreds of blocks, some tight-looking clusters are pure chance. Cluster Detection runs a spatial scan statistic (in the family of Kulldorff's scan statistic and the Getis-Ord Gi* hotspot statistic) that compares each candidate window against what random scatter would produce, keeping only clusters that survive a multiple-comparison correction. Two of the eye-catching clusters wash out as noise; one near the corridor survives strongly. The independence screen then asks whether the co-located elevations — asthma, diesel exposure, aging housing stock — are separate contributors stacking up or all downstream of one driver, flagging that question for deeper decomposition before anyone hardens the wrong layer.

How it works

  • Scan for concentration, not just height. Slide candidate windows across space or across the network's hops and score each on how far its rate exceeds the expected background — a lump, not a single tall unit.
  • Correct for looking everywhere. Calibrate significance against how many windows were examined (Monte Carlo reference distributions, false-discovery control), so a chance cluster somewhere on a large map does not pass as real.
  • Screen independence. Test whether the co-located layers move together because they pile up independently or because a shared driver produces them all, and flag the latter for handoff rather than treating it as several separate risks.

Tuning parameters

  • Scan geometry — window shapes and sizes (circular, elliptical, network-hop). Wider windows catch diffuse clusters but blur their edges; narrow ones sharpen boundaries but miss spread-out concentrations.
  • Significance and correction level — how strict the multiple-comparison threshold is. Stricter cuts false clusters but can drop small true ones.
  • Baseline model — what counts as "expected" (population, historical rate, exposure-adjusted). A wrong baseline invents clusters where risk is only denser, or hides them where the baseline is already high.
  • Independence sensitivity — how aggressively to flag shared-driver patterns for decomposition versus treat the layers as additive.

When it helps, and when it misleads

Its strength is that it separates real concentration from the eye's hunger for patterns, and refuses to let a large search space manufacture hotspots that would not survive a second look.

Its signature misuse is drawing the cluster boundary after seeing the data — the Texas sharpshooter fallacy — which every post-hoc "look how tight this cluster is" invites and which all but guarantees a spurious hit.[1] A misspecified baseline is the other trap, quietly turning population density into false risk. The discipline that guards against both is to pre-specify the scan geometry and baseline, report the multiple-comparison correction openly, and pass surviving clusters to decomposition and validation rather than straight to a fix.

How it implements the components

Spatial or Network Cluster Detection fills the detection-and-verification components — the ones that turn raw scores into a confirmed hotspot:

  • co_location_intersection_map — it locates where elevated layers actually overlap on the grid or network, not merely where each is high alone.
  • clustering_and_multiple_comparison_guard — its core engine: significance testing that corrects for having scanned the whole space.
  • independence_and_shared_driver_check — it screens whether a cluster is independent pile-up or one driver repeated, so the intervention targets the right level.

It does not score the units (harm_or_failure_metric, sensitivity_layer_inventory — Vulnerability Index Construction) or render the visual overlay (exposure_layer_overlay — Layered Risk Heatmap). The dedicated out-of-sample validation is Multiple-Testing Holdout Check, and the deep causal decomposition of a shared driver is Common-Driver Decomposition — this tool only screens for it.

References

[1] The Texas sharpshooter fallacy — firing at a barn wall, then painting the target around the tightest cluster of holes. In spatial analysis it is the temptation to define a cluster's boundary after seeing the data, which manufactures "significance"; a pre-specified scan and an honest multiple-comparison correction are the standard guards. ↩