The PLS (partial least squares) approach to multidimensional contingency tables (Q760127)

The present paper shows that the basic design for PLS estimation of soft models allows straightforward adaptation to the analysis of the dichotomous items of multidimensional contingency tables. The ensuing model defines one or more latent variables for each margin of the contingency table; the observed items are interpreted as indicators of the corresponding latent variables; each latent variable is estimated explicitly as a weighted aggregate of its indicators; the model has ''outer relations'' between each latent variable and its indicators, and ''inner relations'' between the latent variables; the inner and outer relations are causal-predictive; for the indicators of a latent variable that is explained by an inner relation, substitutive elimination of the latent variable from the outer relations gives causal-predictive relations for the indicators in terms of the explanatory latent variables of the inner relation. Several generalizations of the basic PLS estimation algorithm carry over to multidimensional contingency tables, including feedbacks in the inner relations; hierarchic structure of the latent variables; and latent variables in two or more dimensions.