Pj. Hardin et Jm. Shumway, STATISTICAL SIGNIFICANCE AND NORMALIZED CONFUSION MATRICES, Photogrammetric engineering and remote sensing, 63(6), 1997, pp. 735-740
When assessing map accuracy, confusion matrices are frequently statist
ically compared using kappa. While kappa allows individual matrix cate
gories to be analyzed with respect to either omission or commission er
ror rates, kappa is not used to compare individual matrix categories w
ith respect to both rates concurrently. When this concurrent compariso
n is desired, the matrices are typically normalized and then scrutiniz
ed on a cell-by-cell basis by inspection. While no parametric test of
significance exists for such a cell-by-cell examination, sampling dist
ributions for these main diagonal entries can be estimated by repeated
subsampling of the original sample data (i.e., bootstrapping), allowi
ng inferences to be made about the population. In this research, the p
rocedure for estimating the sampling distribution of normalized cell v
alues is described. Three methods for determining the standard error o
f normalized cell value sampling distributions are also outlined. Usin
g these sampling distributions and their attendant standard error, the
statistical comparison of cell values from two normalized confusion m
atrices is illustrated. One illustrated method requires a mild paramet
ric assumption, whereas the other is completely nonparametric. Neverth
eless, the two distinct bootstrap methods produce nearly identical res
ults.