Determining the Number of Segments in Fimix-Pls

Abstract

Response-based segmentation using the finite mixture partial least squares approach (FIMIX-PLS) has recently received increasing research attention in the marketing and management disciplines. Different approaches for latent class detection have been proposed which generalize, for example, tree-like structure (Sánchez and Aluja 2006), finite mixture (Hahn et al. 2002), fuzzy regression (Palumbo et al. 2008), distance-based (Esposito Vinzi et al. 2007) and genetic algorithm approaches (Ringle et al. 2009a) to PLS path modeling. Sarstedt’s (2008a) review of PLS segmentation techniques characterizes FIMIXPLS as the primary choice for latent class detection and segmentation tasks within a PLS context. FIMIX-PLS was introduced by Hahn et al. (2002) and allows for a simultaneous estimation of model parameters and segment affiliations of observations. The approach has performed favourably in different simulation studies (Ringle et al. 2009b) as well as empirical applications (e.g. Hahn et al. 2002; Ringle et al. 2009c) and has been made available to researchers by the software application SmartPLS (Ringle et al. 2005). An unresolved problem in the application of FIMIX-PLS concerns the issue of model selection, i.e. the determination of the number of segments underlying the data. Unlike other latent class procedures, the FIMIX-PLS algorithm allows for the computation of several statistical model selection criteria (Sarstedt 2008b). These are well-known from finite mixture model literature and their performance has been discussed in several simulation studies in different context. However, despite the growing popularity of FIMIX-PLS and the importance of an accurate model selection, up to date, no such simulation study exists for this procedure. Consequently, the scope of this paper is to examine the performance of several model selection criteria for determining the number of segments in FIMIX-PLS by conducting a Monte Carlo simulation study.