In analysing geographical data, one is often interested in finding clusters
with suspiciously large or small values of the variable under consideratio
n. Moran's I tests for general deviations from the null hypothesis of indep
endent observations. A new test is proposed based on triplets, essentially
three districts of the region with a common corner. Subregions composed of
triplets are tested for significant deviation from the general mean. Using
triplet clusters serves to reduce the number of potential suspicious subreg
ions and to enforce a certain compactness. The rest needs approximations fo
r the total number of triplet clusters and for a combinatorial expression.
Despite these approximations, the actual error probability is close enough
to the nominal one to make the test useful in situations like data mining.
The power of the test is compared with that of Moran's I: the latter one is
better for homogeneous counter-hypotheses (as was to be expected and is de
sired) whereas the triplet test is better for small local deviations from t
he independence assumption. (C) 2001 Elsevier Science Ltd. All rights reser
ved.