Identifying and treating unobserved heterogeneity with FIMIX-PLS: part I – method

Abstract

Purpose – The purpose of this paper is to provide an overview of unobserved heterogeneity in the context of partial least squares structural equation modeling (PLS-SEM), its prevalence and challenges for social science researchers. Part II – in the next issue (European Business Review, Vol. 28 No. 2) – presents a case study, which illustrates how to identify and treat unobserved heterogeneity in PLS-SEM using the finite mixture PLS (FIMIX-PLS) module in the SmartPLS 3 software. Design/methodology/approach – The paper merges literatures from various disciplines, such as management information systems, marketing and statistics, to present a state-of-the-art review of FIMIX-PLS. Based on this review, the paper offers guidelines on how to apply the technique to specific research problems. Findings – FIMIX-PLS offers a means to identify and treat unobserved heterogeneity in PLS-SEM and is particularly useful for determining the number of segments to extract from the data. In the latter respect, prior applications of FIMIX-PLS restricted their focus to a very limited set of criteria, but future studies should broaden the scope by considering information criteria, theory and logic. Research limitations/implications – Since the introduction of FIMIX-PLS, a range of alternative latent class techniques have emerged to address some of the limitations of the approach relating, for example, to the technique’s inability to handle heterogeneity in the measurement models and its distributional assumptions. The second part of this article (Part II) discusses alternative latent class techniques in greater detail and calls for the joint use of FIMIX-PLS and PLS prediction-oriented segmentation. Originality/value – This paper is the first to offer researchers who have not been exposed to the method an introduction to FIMIX-PLS. Based on a state-of-the-art review of the technique in Part I, Part II follows up by offering a step-by-step tutorial on how to use FIMIX-PLS in SmartPLS 3.

Antecedents of Online Purchase Intention and Behaviour: Uncovering Unobserved Heterogeneity

“…The following procedures were used: 1) a stop criterion of 1.0E-10 and a maximum number of 5,000 iterations; 2) 10 repetitions to avoid convergence in a local optimum; 3) definition of a reasonable segment number; and 4) avoidance in using the mean value replacement (Hair Jr et al, 2016b). To evaluate results, the information criteria consistent AIC (CAIC) and the normed entropy criterion (EN) were considered jointly to determine the number of data segments that should be retained (Sarstedt, Becker, Ringle, & Schwaiger, 2011).…”

Determining the Key Capacities of Effective Leaders in Malaysian Public and Private Focused Universities

“…The results of the evaluations of the model developed based on the aggregate data have been presented in Figure 3. However, as suggested by Hair, Sarstedt, Matthews, and Ringle (2016) and Matthews, Sarstedt, Hair, and Ringle (2016), FIMIX-PLS analysis was considered to detect unobserved heterogeneity within the data. As displayed in Figure 3, the number of the arrows from the exogenous constructs toward the endogenous construct was three in Malaysian focused HEIs model.…”

“…For the purpose of evaluating the solutions and ex post analysis, the guidelines provided by Hair et al (2016) and Matthews et al (2016) were followed. Focusing on the model and based on the contents of the table of fit indices, selecting two-segment, three-segment, and four-segment solutions seemed to be unrealistic due to minimum sample size limitations.…”

The Key Determinants of Leadership Performance Effectiveness in Public Research and Comprehensive Universities: An Advanced PLS-Sem Study in Malaysian Context

“…In addition, the results indicated a 2-segement solution since AIC3, AIC4, BIC, and CAIC values in this solution were minimum and also EN was greater 0.5. These procedures were followed by conducting Ex Post Analysis on the grounds of guiding principles proposed by Matthews et al (2016) and Hair et al (2016). The results, displayed in the following Table 10, show that the data categorized by Leadership Level, as one of the 13 explanatory variables in the dataset, had an overlap of 66 percent with the data partitioned using FIMIX-PLS module of SmartPLS 3.…”

“… Evaluating discriminant validity on the basis of HTMT criterion (Henseler et al, 2015) as a more accurate new criterion to establish discriminant validity in variance-based structural equation modeling.  Performing FIMIX-PLS (Hair, Hult, et al, 2014;Hair et al, 2016;Matthews et al, 2016) and IMPA (Hair, Hult, et al, 2014; to extent the results of PLS algorithm.…”