Comparing Local vs Global Clustering with FIMIX-PLS: Application to Marketing
Résumé
The increasing use of the Partial Least Squares Structural Equation Modeling (PLS-SEM) method to estimate model parameters between latent and manifest variables in marketing research highlights the issue of heterogeneity. Several segmentation methods dedicated to PLS-SEM have been proposed to account for the potential heterogeneity of the data, such as FIMIX-PLS, REBUS, PLS-POS, or more recently PLS-SEM KMEANS. In some situations, heterogeneity may only affect a subpart of the structural model and not impact the full set of path coefficients parameters. As a result, it is possible to identify some values of path coefficients as common to all observations, while other path coefficients show significantly different values as parameters reflecting heterogeneity that rely on the observation’ segments. Our work aims to apply the FIMIX-PLS procedure such that the segmentation affects only part of the structural model. Then, the local partitioning is introduced as a moderating variable in the PLS-SEM applied to all the observations on the basis of the global structural model. We advocate such a strategy to generate more stable and easier to interpret clusters. This method is demonstrated using a marketing case study.