Why are controls needed?

They represent the spatial distribution against which case concentration is judged.

Does a small p-value prove a common cause?

No. It indicates incompatibility with the specified null model, subject to the sampling and analysis assumptions.

Yes, but different values examine different neighborhood scales. Selection and multiple testing must be handled transparently.

AI Glossary / Statistics and probability / Advanced

Cuzick - Edwards test

Definition

The Cuzick - Edwards test is a statistical test for spatial clustering of cases relative to a comparison population represented by controls. A commonly used statistic counts cases among the k nearest neighbors of each case in the combined case-control locations. It asks whether case labels cluster more than expected under an appropriate null model, while allowing the underlying population to be unevenly distributed.

What it means

Many maps look clustered simply because people, businesses, or equipment are concentrated in particular places. A test that assumes a uniform background can mistake ordinary population density for unusually concentrated cases.

The Cuzick - Edwards approach uses controls drawn to represent the relevant population at risk. The comparison concerns the labels attached to locations, rather than whether all locations are uniformly scattered. The choice of controls is therefore fundamental. Controls from an unrelated population cannot reliably correct the background distribution.

The original work develops related statistics. This entry focuses on the k-nearest-neighbor case-count statistic and a conditional random-label interpretation.

How it works

Combine case and control locations, choose a distance measure, and identify each point's k nearest other points. For every case, count how many of those neighbors are also cases. Sum those counts to obtain T_k.

The neighbor relationships are directed: one point can be another's nearest neighbor without the reverse being true. Mutual case-neighbor relationships contribute in both directions. This makes the statistic different from simply counting distinct close pairs.

Under an unstratified random-label null, hold locations and the number of cases fixed and compare the observed statistic with values obtained by permuting case labels. Matching or stratification can require restricted permutations. Select k and handle tied distances explicitly. Trying many k values and reporting only the most favorable result requires an appropriate multiplicity strategy.

Three practical examples

A spatial study compares event locations with controls sampled from the same underlying population to distinguish excess clustering from population concentration.
A reliability analyst investigates whether failed devices are unusually near other failed devices relative to comparable installed devices, after defining a defensible control-selection scheme.
An AI-assisted mapping workflow uses the test as a formal check on an apparent geographic pattern while retaining human review of locations, eligibility, and confounding factors.

Think of it this way

Think of red and blue pins on a map. Many pins may lie in a city because most eligible units are there. The question is whether red pins tend to have red neighbors more often than expected if the same number of red labels were reassigned among the existing pin locations. The pins stay put; their labels supply the comparison.

Worked example

Consider an original one-dimensional teaching example with six locations at 0,0.1,1,3,3.1,7. Label the first three locations as cases and the last three as controls. Use ordinary distance and k = 1; there are no nearest-neighbor ties.

The nearest-neighbor links are 0 to 0.1, 0.1 to 0, 1 to 0.1, 3 to 3.1, 3.1 to 3, and 7 to 3.1. Each of the three cases has a case neighbor, so T_1 = 3.

There are 20 equally likely ways to assign three case labels among six fixed locations under the stated null. Exact enumeration gives T_1 values of zero, one, two, and three in 6,6,6,2 assignments respectively. The one-sided probability of a statistic at least as large as observed is therefore 2/20 = 0.10.

The null mean is 1.2, also obtained as 3 times 2/5. Although the observed statistic exceeds that mean, this small example does not reject at a preselected 0.05 level. The enumeration is exact for this toy random-label model, not a real spatial investigation.

Illustrative case study

In this fictional case, Meridian Devices sees a dense group of equipment failures on an AI-generated map. The group lies in a district containing many of the company's installations, so analysts avoid judging the pattern from case locations alone.

They select controls from eligible installations during the same observation period, verify coordinates, and define the distance metric and neighbor counts before testing. They also examine whether device age and operating conditions differ geographically and whether the comparison should be stratified.

Any statistical signal becomes a reason for further investigation, not proof of a local environmental cause. This invented case illustrates the design questions around the test without claiming an actual failure cluster.

Common misconceptions and mistakes

The test is not a clustering algorithm that assigns every observation to a discovered group. A global significance result does not by itself locate cluster boundaries or establish a mechanism.

Do not replace controls with uniformly generated map points unless that corresponds to the intended population model. Uneven detection, reporting, or sampling can produce apparent case clustering even when underlying risk is unchanged.

Approximate locations and duplicated coordinates can alter nearest-neighbor relationships. A documented tie or location-uncertainty method is necessary. Also, k indexes a number of neighbors, not a fixed geographic radius; the corresponding physical scale varies with local density.

Business use

For business use, the method can support investigation of geographically concentrated incidents when appropriate case and comparison data exist. Examples include equipment failures or service events, but the statistical design must reflect exposure, eligibility, and observation period.

Define the action that a signal would trigger and assess location quality before analysis. Protect sensitive location information in any published map. Report the tested scale, control source, null model, and limitations so decision-makers understand what the result establishes. A statistically unusual pattern can still have an ordinary operational explanation that the test itself cannot identify.

Check your understanding

Check your understanding: Why does the mutual pair at 0 and 0.1 contribute twice?

Answer: Each case has its own directed neighbor relationship.

Why is the example's p-value 0.10 rather than zero?

Answer: Two possible label assignments attain the observed maximum.

Would a dense city automatically imply case clustering?

Answer: No; the comparison must account for the underlying population distribution.

Related terms

Spatial statistics
Nearest neighbor
Case-control study
Permutation test
random labeling
population at risk
spatial confounding
multiple testing

Learn next

Recalculate the example using two neighbors per point, then examine how stratified permutations change a test when case probability differs across known groups. Study location uncertainty before applying nearest-neighbor methods to aggregated coordinates.

Nearest neighbor

Sources

Sources: Cuzick and Edwards, Spatial Clustering for Inhomogeneous Populations, publisher summary, https://academic.oup.com/jrsssb/article/52/1/73/7027908. CDC, historical methods appendix, https://www.cdc.gov/mmwr/preview/mmwrhtml/00001798.htm. Jacquez, location-uncertainty extension abstract, https://pubmed.ncbi.nlm.nih.gov/8017404/. Full original derivations were not available in the retrieved publisher text. The exact enumeration and fictional case are original.

Previous: Cutting-plane method Next: Cyber kill chain