Partial Least Squares: The Better Approach to Structural Equation Modeling?

With the ever-increasing acceptance of the need to empirically validate theories in the social science disciplines (e.g., Sheth, 1971), data and multivariate analysis techniques (e.g., Hair et al., 2010; Hair et al., 2011b; Mooi and Sarstedt, 2011) play a central role in today’s research. The evolution of structural equation modeling (SEM) methods is perhaps the most important and influential statistical development in the social sciences in recent years. SEM is a second generation multivariate analysis technique that combines features of the first generation techniques, such as principal component and linear regression analysis (Fornell, 1982, 1987). SEM is particularly useful for the process of developing and testing theories and has become a quasi-standard in research (e.g., Hair et al., 2012; Ringle et al., 2012; Shook et al., 2004; Steenkamp and Baumgartner, 2000). When estimating structural equation models, researchers must choose between two different statistical methods: covariance-based SEM (CB-SEM; Diamantopoulos and Siguaw, 2000; J€ oreskog, 1978, 1982; Rigdon, 1998) and variance-based partial least squares (PLS) path modeling, also referred to as PLS-SEM (Hair et al., 2013; Lohm€ oller, 1989; Rigdon, 2012; Wold, 1982). These two approaches to SEM differ greatly in their underlying philosophy and estimation objectives (Hair et al., 2011a; Henseler et al., 2009). CB-SEM is a confirmatory approach that focuses on the model’s theoretically established relationships and aims at minimizing the difference between the modelimplied covariance matrix and the sample covariance matrix. In contrast, PLS-SEM is a prediction-oriented variance-based approach that focuses on endogenous target constructs in the model and aims at maximizing their explained variance (i.e., their R value). Although both approaches were developed at about the same time, their subsequent evolution has been far from parallel. CB-SEM experienced many methodological advances and became a broadly used approach in the social sciences (e.g., Baumgartner and Homburg, 1996; Medsker et al., 1994) due to the early development of the LISREL program in the 1970s (J€ oreskog and S€ orbom, 1996). In contrast, PLS-SEM software was not available for many years until Lohm€ oller (1984) introduced the LVPLS program in the 1980s. It was the late 1990’s when Chin’s (1998) scholarly work and the availability of graphical user interfaces for the LVPLS program (e.g., PLS Graph; Chin, 2003) stimulated key applications in marketing (e.g., customer satisfaction model studies; Fornell et al., 1996) and management information systems research (e.g., technology acceptance model studies; Gefen and Straub, 1997). Today user-friendly software tools such as SmartPLS (Ringle et al., 2005), as well as the need for more flexibility in applying statistical techniques, have revived the PLS-SEM method for applied researchers (Hair et al., forthcoming; Hair et al., 2012). New textbooks on how to use PLS-SEM (e.g., Hair et al., 2013) will further disseminate the method in university courses on the Masters and Ph.D. level, as well as in industry. Some 35 years after Herman Wold (1974, 1975) introduced PLS-SEM as a soft modeling approach that overcomes the strict assumptions of CB-SEM, it is enjoying increasing popularity across various disciplines. In fact, PLS-SEM is experiencing widespread application as a method in both academic research and practice (e.g., Hair et al., forthcoming; Hair et al., 2012; Lee et al., 2011; Ringle et al., 2012). Likewise, methodological research has presented a wide range