Statistical method paper using historical whisky descriptions to illustrate comparison of distance matrices; does not validate a sensory prediction engine.
Citation And Scope
Legendre, P., and Lapointe, F.-J. (2004). Australian & New Zealand Journal of Statistics 46(4):615–629. All fifteen PDF sheets reviewed; Table1 and Figure4 on PDF13 (printed627) visually checked.
Method
PDF2–8: CADM tests congruence among distance matrices using Kendall's W and a permutation test. Objects must correspond across matrices. Permute object labels jointly in rows and columns, not independent distance entries. The Friedman chi-squared transformation is a statistic here; its significance comes from the permutation procedure, not an ordinary chi-squared table. A significant overall test means at least some matrices are congruent, not that all modalities agree. Follow-up tests and pairwise rank Mantel comparisons address different questions; multiplicity and directional alternatives matter. PDF7–11: simulation results assess Type I error and power under the authors' generated-data scenarios. They do not guarantee performance for arbitrary future Lab data.
Whisky Example
PDF11–14: 109 whiskies from 108 distilleries described by 68 binary attributes, drawn from Michael Jackson's 1989 guide. Groups contain 14 color, 12 nose, 8 body, 15 palate and 19 finish variables. This is coded published description, not a new blinded panel, chemical assay or blending experiment. Jaccard distance excludes joint absences; an unmentioned descriptor is not evidence of a measured nondetection. Table1: W=.22658. Printed P=.000 reflects rounding with 9,999 permutations, not a literal zero probability. Significant pairwise correlations are small (roughly .034–.084). Statistical detection does not establish useful predictive accuracy. The finish follow-up P=.476 is a failure to reject incongruence, not proof that finish is independent of every other modality.
Critical Appraisal And Application
The discussion sometimes interprets nonsignificance too strongly. Its suggestion that swallowing is necessary to appreciate finish was not tested by this analysis and must not become a Lab requirement. Retain nose, body, palate and finish as distinct observations. If future research compares chemical and sensory distances, specify objects, scaling, metric, missing-data rules, permutation design, effect sizes and multiplicity first. Consider CADM only as an advanced analytical option with statistical review, not an MVP recommendation algorithm. Do not use a global congruence result to justify collapsing all observations into one quality or similarity score.
Research Connections
Non-detection is not sensory equivalence
Sensory notes are observer-conditioned observations, not product constants
No access