Quantification of intrinsic quality of a principal dimension in correspondence analysis and taxicab correspondence analysis

Vartan Choulakian

DOI: https://doi.org/10.48550/arXiv.2108.10685

2021-08-24

Abstract:Collins(2002, 2011) raised a number of issues with regards to correspondence analysis (CA), such as: qualitative information in a CA map versus quantitative information in the relevant contingency table; the interpretation of a CA map is difficult and its relation with the \% of inertia (variance) explained. We tackle these issues by considering CA and taxicab CA (TCA) as a stepwise Hotelling/Tucker decomposition of the cross-covariance matrix of the row and column categories into four quadrants. The contents of this essay are: First, we review the notion of quality/quantity in multidimensional data analysis as discussed by Benzécri, who based his reflections on Aristotle. Second, we show the importance of unravelling the interrelated concepts of dependence/heterogeneity structure in a contingency table; and to picture them two maps are needed. Third, we distinguish between intrinsic and extrinsic quality of a principal dimension; the intrinsic quality is based on the signs of the residuals in the four quadrants, hence to the interpretability. Furthermore, we provide quantifications of the intrinsic quality and use them to uncover structure in particular in sparse contingency tables. Finally, we emphasize the importance of looking at the residual cross-covariance values at each iteration.

Methodology