Content area
Full Text
1. Introduction
Assessing a statistical model’s predictive power is a crucial element of any study. Researchers evaluate theories and the practical relevance of their analyses on the basis of their models’ ability to make falsifiable predictions about new observations. According to Hofman et al. (2017, p. 486), “historically, this process of prediction-driven explanation has proven uncontroversial in the physical sciences, especially in cases where theories make relatively unambiguous predictions and data are plentiful.” In contrast, marketing researchers have generally emphasized prediction less than explanatory modeling, which aims to “test or quantify the underlying causal relationship between effects that can be generalized from the sample to the population of interest” (Shmueli et al., 2016, p. 4553). In other words, marketing researchers’ focus is primarily on assessing whether model coefficients are significant, meaningful and in the hypothesized direction, rather than on testing whether a model can predict new cases.
A similar conclusion can be drawn with respect to applications of partial least squares structural equation modeling (PLS-SEM), a widely used regression-based technique in marketing and other social sciences fields, which estimates relationships in path models with latent and manifest variables (Lohmöller, 1989; Wold, 1985; Hair et al., 2017b). Contrary to covariance-based SEM (Jöreskog, 1978; Rigdon, 1998; Diamantopoulos and Siguaw, 2000), which was only designed for explanatory purposes (Sarstedt et al., 2016b), PLS-SEM is a “causal-predictive” method (Jöreskog and Wold, 1982, p. 270). As such, PLS-SEM overcomes the apparent dichotomy between explanation and prediction. While the method maximizes the amount of explained variance of the endogenous constructs embedded in a path model grounded in well-developed causal explanations (Sarstedt et al., 2017), the PLS-SEM results are well suited to generate out-of-sample predictions. Gregor (2006, p. 626) refers to this interplay as explanation and prediction theory, noting that this approach “implies both [an] understanding of [the] underlying causes and prediction, as well as [a] description of [the] theoretical constructs and the relationships among them”.
While researchers using PLS-SEM routinely stress the predictive nature of their analyses – as evidenced in numerous reviews of PLS-SEM use across a variety of fields (Ali et al., 2018; Hair et al., 2012a; Hair et al., 2012b; Ringle et al., 2019) – model evaluation relies almost exclusively...