[Rus13] Beyond the measurement scale: the Non-Metric Partial Least Squares approach

Conférences invitées : CFE-ERCIM 2013 conference, December 2013, pp.1, London, UK,

Mots clés: Partial Least Squares, Optimal Scaling, Categorical variables

Résumé: The Partial Least Squares approach to latent variable path modeling (PLS-PM) is a component-based method that allows investigating the relationships among several sets of variables connected by a path. This method is implemented through a flexible PLS algorithm that, depending on the chosen option, can optimize covariance or correlation based criteria. PLS-PM only applies to quantitative (metric) data, although in many real applications users are interested in analyzing data observed on ordinal or nominal measurement scales. To handle measurement scale heterogeneity in PLS framework, the Non-Metric PLS (NM-PLS) approach has been recently proposed. It extends PLS methods optimizing covariance-based criteria to the treatment of non-metric variables and non-linearity. NM-PLS methods are implemented through PLS-type algorithms that work as optimal scaling procedure: each different value or category is considered as a scaling parameter, which is estimated through a numerical value. NM-PLS algorithms estimate both model and scaling parameters in order to satisfy PLS model criteria. Here I present the main features and properties of the NM-PLS algorithms. Moreover, I extend the standard NM-PLS to PLS criteria related to correlation. Finally, I discuss the effects of the parametric expansion on the variability of the estimates in the NM-PLS approach to multi-block analysis.


@inproceedings {
title="{Beyond the measurement scale: the Non-Metric Partial Least Squares approach}",
author=" G. Russolillo ",
booktitle="{CFE-ERCIM 2013 conference}",
address="London, UK",