Non-metric statistical analyses: Some metric alternatives (Q1073498)

Two classes of permutation techniques for completely randomized and randomized block designs are considered in this expository paper. The analysis space of these techniques corresponds to the data space in question for only a small subset of these classes (the congruence principle is satisfied when this correspondence holds). In particular, the analysis space associated with most classical techniques among these classes (e.g., one-way analysis of variance and randomized block analysis of variance) is non-metric and does not satisfy the congruence principle. An example demonstrates how a well known statistical method which involves a non-metric analysis space and does not satisfy the congruence principle (the two-sample t test) is unable to detect obvious differences, whereas, a corresponding statistical method which does satisfy the congruence principle provides unambiguous detection of the same differences. Since the obvious differences in question are analogous to the types of differences routinely encountered in many biological, physical and social science disciplines, the use of the well known statistical methods (such as the two-sample t test) has probably prohibited the detection of many major findings in these disciplines. In addition, new multivariate methods are introduced to compare cyclic data.