Abstract
Although the number of empirical applications of partial least–squares structural equation modeling (PLS-SEM) in tourism has increased in the last two years, Assaker, Huang, and Hallak have conducted the only assessment on the use of PLS-SEM in four studies and with a limited number of criteria. Thus, this study aims to critically analyze how the PLS-SEM method has been applied in 44 articles published in 11 leading tourism journals from 2000 to 2014 in terms of four key criteria: (1) themes explored and main motivations for using PLS-SEM; (2) characteristics of proposed models; (3) how the models were evaluated; and (4) the use of more advanced analyses within the method. The findings revealed that although applications in tourism have improved in recent years, problematic aspects in the application of PLS-SEM in tourism research still exist. The article provides suggestions on how to improve the use of PLS-SEM in future tourism applications.
Keywords
Introduction
Structural equation models (SEMs) with latent variables emerged from the need to measure multidimensional concepts that are not directly observable (constructs or latent variables), as well as test the relationships among them (Bollen 1989). In recent years, the use of SEM has become widespread in diverse fields of knowledge, including tourism. Indeed, in this field of knowledge the application of the method is already quite extensive. We can find it in empirical studies on the image of tourist destinations, destination loyalty, the quality and competitiveness of destinations, and perceptions about the impacts of tourism, and/or residents’ attitudes toward tourism development (Nunkoo and Ramkissoon 2011; Nunkoo, Ramkissoon, and Gursoy 2013).
However, when applying SEM, two approaches need to be considered. The first is based on the concept of covariance (which leads us to covariance-based SEM or, in its abbreviated form, CB-SEM) (Jöreskog 1973, 1978). The second approach relies on the concept of variance and is referred to as variance-based SEM or, more commonly, partial least squares–structural equation modeling (PLS-SEM) (Wold 1982, 1985). Although both methods share the same roots (Jöreskog and Wold 1982), the CB-SEM involves strict rules and assumptions that, if not met, can compromise the validity of the results. These include (1) the multivariate normality of data; (2) minimum sample size; (3) reflective constructs (i.e., directional arrows progress from the constructs to the indicators); and (4) strong theoretical knowledge about the model being tested (i.e., composition of the latent constructs and the causal relationships among the constructs must be directed by theory). PLS-SEM differs from CB-SEM in that it is not restricted by the same stringent assumptions that characterize CB-SEM. As such, variance-based PLS-SEM emerges as the suggested alternative for dealing with causality problems among latent variables when CB-SEM assumptions cannot be met. PLS-SEM can thus present further flexibility in modeling various phenomena in business disciplines in general, and in the tourism discipline in particular, that could not be modeled otherwise because of violations of the assumptions of traditional CB-SEM.
Typical tourism examples where PLS-SEM would be appropriate include grouping supply-side indicators into a smaller number of destination factors/constituents (e.g., the economy, infrastructure, and environment factors at the destination), on an empirical basis, without any theory behind the authors’ assertions in terms of how these indicators are to be grouped in order to predict tourism demand (see, e.g., Mazanec and Ring 2011).
Another example involves using formative constructs (i.e., relationship progresses from the indicators to the construct, suggesting that indicators used to measure the construct are not correlated and thus contribute differently to form their underlying construct) to model tourism variables/constructs, such as in the case of loyalty (see Song et al. 2011). For example, three indicators are typically used to measure loyalty: positive word of mouth, revisit intention, and recommendation to acquaintances. These indicators do not move together and are assumed to form the loyalty construct, given that a tourist could say positive things about his/her stay in the destination. Nevertheless, this does not necessarily mean that he/she will return to the destination in the future, as this depends on other factors as well, such as available time or money. A final example of PLS-SEM use is in the case of complex models where the number of indicators compared with the sample size is large. This is the case in tourism studies carried out at the macro/country level, where the number of observations is limited because of the number of countries and lack of data available for each country (Assaker and Hallak 2012).
Despite the appropriateness of PLS-SEM to advance research in tourism studies (because of PLS-SEM’s features and modeling flexibility, as previously mentioned), the number of studies using PLS-SEM in tourism has grown in recent years but is still considerably smaller than in other disciplines in the business field, such as marketing (Hair, Ringle, and Sarstedt 2011), strategic management (Hair et al. 2012a), and management information systems (MIS) (Ringle, Sarstedt, and Straub 2012). Researchers in the tourism field still seem to rely predominantly on CB-SEM in their empirical studies and thus might miss opportunities frequently exploited in other disciplines by using the PLS-SEM technique. Potential reasons for the still-limited use of PLS-SEM in tourism, which tends to apply this method cautiously, might stem from researchers’—though well-equipped with a basic understanding of CB-SEM models—lack of understanding and familiarity with PLS-SEM. Indeed, textbooks on multivariate data analysis tend not to discuss PLS-SEM (e.g., Churchill and Iacobucci 2010) or address it only superficially (e.g., Hair et al. 2010). In addition, the topic is seldom found in research methodology course syllabi, where the focus is still on traditional CB-SEM because of the popularity of this method and the widespread availability of CB-SEM software (e.g., Lisrel, AMOS, Mplus; Chin 1998).
In fact, as with other statistical methods, PLS-SEM is a complex technique requiring several choices that, if not duly considered, can mislead researchers into improper findings, interpretations, and conclusions (Hair et al. 2012b). Thus, users (researchers) in tourism can only benefit from the unique properties of PLS-SEM if they understand its underlying principles, apply them properly, and report the results correctly. Because of the complexities involved in using PLS-SEM, systematic assessments of how PLS-SEM was applied in prior tourism studies can help disseminate rigorous research and publication practices in the field. They can also provide important guidance and, if necessary, opportunities for course correction in future applications. Although PLS-SEM usage reviews have been carried out across many disciplines in business research, including marketing (e.g., Hair, Ringle, and Sarstedt 2011), strategic management (e.g., Hair et al. 2012a; Hulland 1999), and MIS (e.g., Ringle, Sarstedt, and Straub 2012; Urbrach and Ahlemann 2010), corresponding assessments of PLS-SEM in tourism are still limited. Assaker, Huang, and Hallak(2012) provided the only assessment so far, assessing four studies in tourism. They showed that the PLS-SEM technique had been applied with considerable variability in terms of appropriately handling conceptual and methodological issues. In this context, an update and extension of the Assaker, Huang, and Hallak (2012) case study–based assessment of PLS-SEM specific to the tourism discipline seems timely and warranted.
Therefore, the objective of this article is to critically analyze the use of PLS-SEM in tourism and to provide recommendations for the use of the technique in future tourism research (for explanations of the PLS-SEM method itself, see, e.g., Chin 2010; Hair, Ringle, and Sarstedt, 2011; Henseler, Ringle, and Sinkovics 2009; Henseler, Ringle, and Sarstedt 2012). In particular, this study analyzes 44 empirical studies following the PLS-SEM approach that were published in indexed journals in the Social Sciences Citation Index (SSCI) (Web of Science) from 2000 to 2014 (Thomson Reuters 2014). These studies were divided into two groups: pre-2012 and post-2012. Year 2012 was when the first article assessing PLS-SEM in four tourism articles was published (Assaker, Huang, and Hallak 2012) and when several reviews of PLS-SEM in marketing and strategic management and MIS were published (see Hair, Ringle, and Sarstedt 2011; Hair et al. 2012a; Hair et al. 2012b; Ringle, Sarstedt, and Straub 2012). The studies are analyzed according to four key dimensions: reasons for using PLS-SEM, characteristics of the proposed models, model evaluation and possible use of more advanced applications, and reporting. Where possible, we also indicated best practices as guidelines for future applications and suggested avenues for further research. The current study thus helps identify how the application of PLS-SEM has changed, if at all, over time (i.e., before and after 2012). More importantly, it helps identify which aspects of the method have been more carefully considered and which need to be improved to ensure more rigorous application in the field. Such information can be used to provide suggestions for researchers interested in applying PLS-SEM to future tourism applications by highlighting current areas of concerns.
Partial Least Squares Structural Equation Modeling
Partial least squares (PLS) structural equation modeling (Wold 1982, 1985) is an alternative to traditional SEM that can be applied to model the relationships of causality among the variables. In this partial information method, rather than using the model to explain covariations among indicators (i.e., modeling the measurement errors, as SEM does), PLS maximizes the explained variance of all dependent variables based on how these variables relate to their neighboring constructs. It uses an iterative algorithm in which the parameters are calculated with a series of least squares regressions after explicitly creating construct scores by weighting the sums of items underlying each construct (Chin 1998 ). The term partial thus stems from the idea that the iterative procedure involves separating the parameters rather than estimating them simultaneously (Hulland 1999).
PLS follows a two-step process that starts with an iterative estimation of latent variables scores. The method uses the PLS algorithm to estimate outer and inner weights based on how the structural and measurement models are specified. This requires an iterative procedure in which two kinds of approximations for the latent variables are estimated until the weight estimates converge. The two weight calculations—inside and outside—relate to the inner relations and outer relations. The algorithm starts with an arbitrary initial weight used to calculate an outside approximation of the latent variables. The inner relations among latent variables are then considered to calculate the inside approximations. To perform this approximation, the researcher can choose among three possible scenarios, called weighting schemes: (1) centroid, (2) factor, or (3) path scheme. After the inside approximations are obtained, the algorithm turns again to the outer relations, and new weights are calculated considering how the indicators are related to their constructs: by Mode A (reflective) or Mode B (formative). Mode A implies simple linear regressions between the construct and its reflecting indicators as the construct is assumed to affect each indicator separately. Mode B implies multiple linear regressions between the construct and the set of indicators as the indicators are assumed to affect the construct on a collective basis. The simple or multiple regression coefficients are then used as new weights for the outside approximation. This process continues iteratively until the weights converge—that is, until the change in the outer weights between two iterations drops below a predefined limit. This limit is usually preset within the PLS-SEM software (Henseler, Ringle, and Sinkovics 2009).
After the weights converge and latent variables are estimated, the second step of the process calculates the parameters of the structural and the measurement models. The structural coefficients, also known as path coefficients, are calculated using ordinary least squares regression between latent variables (LVs). There are as many regressions as there are endogenous latent variables. The parameters of a measurement model, known as the loading coefficients, are also estimated by least squares regressions by taking into account the mode used (A = reflective or B = formative).
As previously discussed, PLS-SEM can be used to achieve four major purposes. First, PLS is advantageous when the researcher is trying to explore, rather than confirm, theory. It is useful when the phenomenon being investigated is relatively new and the measurement models are at the exploratory stage (Wold 1985). Second, PLS can be used to examine structural models in cases of small samples and when the multivariate normality of the data cannot be supported (Chin and Newsted 1999). Third, PLS modeling allows the unrestricted computation of models composed of “reflective” and “formative” measurement models (Diamantopoulos and Winklhofer 2001). Finally, PLS can examine large, complex models comprising several latent and manifest variables as well as hierarchical models with first-order and second-order latent constructs (Wold 1985). Thus, PLS can overcome identification issues, nonconvergence, limitations, and assumptions associated with CB-SEM (see Vinzi, Trinchera, and Amato 2010).
PLS-SEM Studies in Tourism Research: Journals and themes
Our review focused on the use of PLS-SEM articles in 11 tourism journals: Journal of Travel Research, Annals of Tourism Research, Tourism Management, Journal of Travel and Tourism Marketing, Journal of Sustainable Tourism, Asia Pacific Journal of Tourism Research, International Journal of Tourism Research, Current Issues in Tourism, Tourism Geographies, Tourism Economics, and Tourism Analysis. These journals (except Tourism Analysis) were selected because they were included in the Social Sciences Citation Index (SSCI) (Web of Science) (Thomson Reuters 2014) as leading journals in the field of tourism. Tourism Analysis was included as it is another high-impact journal in the field that publishes significant quantitative research in tourism, making it important to consider this journal when reviewing the use of an advanced statistical technique in tourism (e.g., PLS-SEM) (Nunkoo et al. 2013 ).
In addition to being the leading publication outlets for tourism research, these journals have different foci, aims, and objectives, thereby ensuring that SEM-based articles with different tourism orientations (e.g., marketing, planning, psychology, and economics) were included in this review study. All issues between 2000 (the first year an application was found) and 2014 (March 2014, the time when this article was written) were searched for empirical-based applications of PLS-SEM. We conducted a full-text search in the Thomson Reuters Web of Knowledge and EBSCO Business Source Premier databases using the keywords partial least squares structural equation modeling, partial least squares path modeling, PLS-SEM, and PLSPM. We also looked into the online versions of the journals in order to double-check that we had captured all PLS-SEM articles in the targeted tourism journals. Ultimately, the search resulted in a total of 44 articles from the 11 journals, after removing conceptual articles on PLS-SEM and research articles from the review (see Table 1).
Publications in Tourism Using PLS-SEM.
In particular, Tourism Management, Journal of Travel and Tourism Marketing, and Journal of Travel Research published the largest number of PLS-SEM studies among the reviewed journals (all 3 altogether published 29 articles, accounting for 65.9% of the total publications). In contrast, the Journal of Sustainable Tourism and Tourism Geographies did not include a single PLS-SEM study in the period of time reviewed. Figure 1 shows the (cumulative) number of studies between 2000 (the first year an application was found) and 2014. It is apparent that the use of PLS-SEM has substantially increased over time. Approximately 31.8% (n = 14) of articles were published before 2012, compared to 68.2% (n = 30) between 2012 and 2014 (see Table 1). Regressing the number of studies on the linear effects of time yields a significant model (F = 23.71; p < 0.000) in which the time effect is significant (t = 4.86; p < 0.000), suggesting that the use of PLS-SEM in tourism has accelerated significantly over time.
Number of PLS-SEM Publications in Tourism Journals over Time.
The identified studies explored several themes, which can be grouped into 10 categories. As shown in Table 2, “loyalty and/or satisfaction regarding a destination or tourism organization” is the main theme studied using this method and is clearly more prevalent than the others, being analyzed in 15 (34.1%) of the publications identified. The focus on this theme is stronger in studies before 2012 compared with later studies (57.1% vs. 23.3%). “Perceived quality of organizations, tourist products, or services” occupies the next position in the set of themes and is analyzed in 7 articles (15.9%). “Tourism organizations management” ranks third, being explored in 5 (11.4%) publications. Finally, it is interesting to note that articles published between 2012 and 2014 cover a broader range of themes (all 10 themes) compared to those published prior to 2012 (which only cover 5 of the 10 themes), suggesting more variety in the recent applications of PLS-SEM in tourism and the prevalent use of PLS-SEM to investigate different phenomena within tourism as compared to early applications in tourism (prior to 2012).
Main Themes in Tourism Research Applying PLS-SEM.
Reasons for Using PLS-SEM in Tourism Research
Given that PLS-SEM is an alternative approach to traditional CB-SEM when the latter reaches its limitations and cannot be used, studies applying PLS-SEM in tourism need to present the rationale behind the selection of this method rather than CB-SEM when this is the case. Of the 44 studies identified, only 2 (4.5%) did not provide such justification. Both of these articles were published in the period before 2012, suggesting that users/researchers in recent years (2012 and onward) are more aware of the features that make the PLS-SEM technique suitable for the model/study at hand; they find it easier now to report their reason for using PLS-SEM, which seems to not always have been the case in earlier years (prior to 2012) when the technique was adopted.
Table 3 displays the main reasons for using the PLS-SEM approach: the predictive focus of the method (31 studies, 70.5%), small sample size (21 studies, 47.7%), nonnormal data (21 studies, 47.7%), and formative nature of the model (15 studies, 34.1%). All of these characteristics have been extensively discussed in the methodological literature on PLS-SEM (see Vinzi, Trinchera, and Amato 2010; Lohmöller 1989). The sample size argument in particular has been the subject of much debate (e.g., Marcoulides and Saunders 2006). A rule of thumb for robust PLS path modeling estimations suggests that the sample size should be equal to the larger of the following: (1) 10 times the number of indicators of the scale with the largest number of formative indicators, or (2) 10 times the largest number of structural paths directed at a particular construct in the inner path model (Chin and Newsted 1999). This represents the largest number of regressions performed during the PLS-SEM iterative process; as such, it would be the logical threshold for sample size to ensure the accuracy and statistical power of the model. Nonetheless, several authors (e.g., Hair et al. 2012b) have argued that very small samples can be a problem when performing PLS-SEM for two reasons. First, they rarely capture more heterogeneous populations (i.e., they fail to capture the heterogeneity in the population from which the sample is drawn), thereby resulting in a larger sampling error. Second, if the sample is too small and the data are asymmetrical, the bootstrap standard errors will be too large, thereby reducing the method’s statistical power. Thus, to be on the safe side in terms of sample size, one might recommend 100 cases with the objective of improving accuracy (see, e.g., Assaker, Huang, and Hallak 2012). Indeed, few studies have systematically evaluated PLS-SEM’s performance when the sample size is small (e.g., Chin and Newsted 1999). More recently, Reinartz, Haenlein, and Henseler (2009) showed that PLS-SEM achieves high levels of statistical power—compared to its covariance-based counterpart—even if the sample size is relatively small (i.e., 100 observations). Similarly, Boomsma and Hoogland’s (2001) study underscored the CB-SEM’s need for relatively large sample sizes to achieve robust parameter estimates. PLS-SEM is therefore generally more favorable with smaller sample sizes and more complex models. However, as noted by Marcoulides and Saunders (2006) as well as Sosik, Kahai, and Piovoso (2009), PLS-SEM is not a silver bullet for use with samples of any size, nor is it a panacea for dealing with empirical research challenges. Researchers need to consider the effects of size, reliability, total number of indicators, data characteristics, and other issues likely to affect the statistical power of the PLS-SEM method (Hair et al. 2012b). In the present article, 42 of the 44 analyzed studies fulfill these conditions of using a sample of 100 cases or more, and only two use samples smaller than 100 (Pike et al. 2011; Zach and Racherla 2011). The average and median sample sizes are 487 and 321 observations, respectively, and studies prior to 2012 use significantly (p ≤ 0.01) smaller samples (mean = 451), in terms of sample mean, than later studies (mean = 562), suggesting that researchers have become more aware of the sample size limitations; as such, they are using larger sample sizes to avoid compromising the sample representativeness and to accurately achieve statistical power.
Motivations for Using PLS-SEM in Tourism Research.
Moreover, a comparison of studies published before 2012 with those published in 2012 and onward shows a fairly consistent pattern, with a focus on prediction, small sample size, nonnormal data, and formative measures being the most prevalent reasons in recent years. This observation is consistent with patterns observed in other business disciplines, although with a different order in relevance, including marketing research (Hair, Ringle, and Sarstedt 2011), strategic management (Hair et al. 2012a), and management information systems (Ringle, Sarstedt, and Straub 2012), suggesting a similar understanding and applicability of the PLS-SEM method by tourism researchers compared with researchers applying PLS-SEM in other disciplines.
Model Characteristics in Tourism Research
Table 4 shows the main features of models using PLS-SEM in tourism research. The first aspects to be highlighted concern the number of latent variables in the inner model and the number of indicators used to measure these variables. On average, the number of latent variables per model is 6.02, and the average number of indicators per model is 24.93. Although both figures are less than the average number of latent variables and indicators in path models reported in other business disciplines, including marketing (7.9 and 29.55, respectively), strategic management (7.5 and 27, respectively), and management information systems (8.12 and 27.42), they still suggest a relatively high level of model complexity compared with studies in the CB-SEM context (e.g., Shah and Goldstein review of CB-SEM reported an average number of latent variables of 4.70, with an average number of indicators of 16.30).
Model Characteristics in Tourism Research.
Table 4 also shows that the average number of latent variables per model is significantly (p ≤ 0.05) larger in articles published before 2012 than in 2012 onward (6.9 vs. 5.6, respectively); however, the average number of indicators remained almost the same in articles published before and after 2012 (25.3 vs. 23.9, respectively). Taking these results jointly, we can argue that model complexity did not decrease over time, given that the average number of indicators used to measure each construct significantly increased (p ≤ 0.05) between articles published before 2012 and articles published in 2012 onward (3.7 vs. 4.3 indicators on average per construct for the period after 2012). The same can be said about the number of inner model relationships being analyzed, which remained more or less the same over time (7.23 vs. 8.21 for the periods before and after 2012, respectively), suggesting again that fewer constructs have been used in PLS-SEM studies in recent years. These studies still incorporated large numbers of inner model relationships and higher numbers of indicators, thereby adding to the complexity of the model being analyzed.
Moreover, an important feature of PLS-SEM is its ability to incorporate both reflective (the relationship progresses from the construct to the indicators, suggesting that indicators are correlated or move together in the same direction) and formative measures (relationship progresses from the indicators to the construct, suggesting that indicators used to measure the construct are not correlated and thus contribute differently to forming their underlying construct) of the latent variables. In this perspective, Table 4 also shows that 27 (61.4%) studies still analyze the causal relationships between constructs with the reflective measurement model only, while 16 (36.4%) articles used a combination of both reflective and formative measurement models, and only one article used solely a formatively measured latent construct. Moreover, of the 16 articles involving formative and reflective measurement models, a significantly (p ≤ 0.05) higher number (i.e., 11) were published after 2012. The only articles involving solely a formative measurement model were published in this time period as well (2012 onward). These observations suggest that, overall, researchers in tourism are still not capitalizing on the full features of PLS-SEM in terms of allowing the assessment of reflective and formative models, although this trend seems to have started changing in recent years, with researchers significantly applying more formative measurement schemes in their models.
Another aspect to note is that 6 studies (13.6%; see Table 4) proposed models with at least one construct measured using a single indicator, with no significant difference (p ≥ 0.05) in the number of articles that have used a single indicator across the two time periods (3 prior to 2012 and 3 in 2012 onward). While single-indicator constructs in CB-SEM generally lead to unidentified models, this problem does not occur when PLS-SEM is used. However, Reinartz, Haenlein, and Henseler (2009) showed that PLS-SEM estimates are only “consistent at large” in smaller samples if multiple indicators per construct are available. Otherwise, the tendency is to underestimate the relationships in the inner model and overestimate the relationships in the outer model (Chin 1998). Thus, authors such as Hair et al. (2012a) have advised caution in using this type of construct in PLS-SEM, and researchers in tourism should be aware of this limitation in future works, as it appears that the use of single-indicator constructs is still a common pitfall within tourism studies.
Evaluating PLS-SEM in Tourism Research
PLS-SEM does not offer a universal goodness-of-fit index. In this regard, Chin (1998) proposed assessing the model by applying various criteria at two moments. The first moment is the evaluation of the outer model—that is, the part of the model describing the relationships between the latent variables and their indicators. The second moment is the evaluation of the inner model—that is, the part of the model describing the relationships across the latent variables that make up the model. As pointed out by Henseler, Ringle, and Sinkovics (2009, p. 298), “It only makes sense to evaluate the inner path model estimates when the calculated latent variable scores show evidence of sufficient reliability and validity.” Underlying this statement is the assumption that it is only worth analyzing the relationships between the latent variables if the researcher feels confident that the indicators properly represent these variables. We proceed by analyzing the forms of outer model (reflective or formative) and inner model validation in the identified empirical studies in tourism.
Reflective Outer Model Evaluation
The evaluation of a reflective outer model should observe the aspects listed in Table 5, that is, examine the reliability (individual and composite) and the validity (convergent and discriminant) of the constructs.
Reflective Outer Model Evaluation in Tourism Research.
Note: AVE = average variance extracted.
The individual reliability analysis, by observing loadings’ magnitude (i.e., the correlation coefficients between the indicators and the corresponding latent variable), is reported in 35 (79.6%) of the 44 studies identified. All the studies conducted in 2012 onward used this analysis, while it was performed in only 5 (35.7%) of the articles prior to this date, suggesting significant (p ≤ 0.05) improvement in the application of individual reliability measures in recent years.
Overall, regarding the composite reliability, the Goldstein rho index (as proposed by Werts, Linn, and Jöreskog 1974) should be observed. This index corresponds to a measure of overall correlation between a construct and its indicators, and it was found in all 30 articles (100%) published in 2012 onward and in almost all articles (13 of 14 or 92.9% of the total) published prior to 2012, suggesting the nonsignificant and consistent use of this composite reliability measure across time.
Moreover, all the analyzed studies assess convergent validity (i.e., the degree of connection between the indicators and their construct, or whether they represent one and the same latent concept), by observing indicators’ statistical significance, and 42 of the articles (95.5% of the total) analyze the average variance extracted (AVE), that is, the variance shared between the indicators and the construct. The use of this approach is quite similar in both periods under analysis. Regarding discriminant validity, the aim is to assess the extent to which different theoretical concepts are sufficiently measured by distinct indicators (therefore, they are indeed measuring different constructs). In tourism research, discriminant validity has been evaluated by applying the Fornell and Larcker (1981) criterion, whereby a latent variable should share more variance with its indicators than with indicators of other latent variables. This aspect can be observed if the AVE of each latent variable is greater than the highest squared correlation between this and the remaining latent variable. This form of discriminant validity assessment was observed in 41 (93.2%) of the identified studies. Again, the use of this approach is equally prevalent in both periods under analysis. Only rarely does discriminant validity analysis consider the observation of the cross-loadings—namely, the loadings of each indicator in the other latent variables. Table 5 shows that only 9 (20.5%) of the studies (all from 2012 onward) provided results of the cross-loadings in their analyses, indicative of a significant (p ≤ 0.01) increase in the use of a cross-loading indicator in recent years. In summary, all of the results presented thus far suggest that tourism researchers demonstrate an in-depth understanding of the reliability and validity measures of the reflective constructs in PLS-SEM.
Formative Outer Model Evaluation
In a formative outer model, the indicators represent independent causes of a theoretical concept (a latent variable) and do not need to be correlated. Thus, in evaluating a formative outer model, there is no need, nor does it make sense, to evaluate the reliability and validity, as in reflective outer models. Rather, in evaluating these models, the first aspect that matters is its theoretical rationality and the expert’s opinion (Diamantopoulos and Winklhofer 2001). Second, it is necessary to observe some statistical criteria, as shown in Table 6.
Formative Outer Model Evaluation in Tourism Research.
Of the 44 studies analyzed, 17 (38.6%) include at least one formative construct. As can be seen, all the studies are concerned with analyzing the weights and/or the loadings of the indicators, while 14 (82.35%) of those 17 studies reported as well loadings/weights statistical significance using the bootstrap procedure. Moreover, the issue of multicollinearity was reported in only 11 of the published articles (64.7%), with significant prevalence (p ≤ 0.01) during 2012 (9 studies reporting multicollinearity). Both statistical significance and multicollinearity measures are important when assessing formative models because, if the outer loading or weights are insignificant, it means there is no empirical support for the indicator’s relevance regarding providing content to the formative index; in this case, it should be considered a strong candidate for removal. The analysis of multicollinearity in a set of indicators linked to a formative construct is also important because it can produce a lack of statistical significance in the weight estimate of a formative indicator (Grewal, Cote, and Baumgartner 2004).
Thus, the above observations suggest that, overall, a formative outer model assessment in tourism discipline leaves much to be desired. Researchers neglect fundamental principles of outer model evaluation such as significance testing and multicollinearity assessment, casting doubts onto measurement quality and thus the studies’ findings. This could be explained by the fact that, unlike PLS-SEM, CB-SEM does not entail the use of formative constructs in causal modeling. With CB-SEM, formative constructs are modeled rarely and require that specific conditions be met. As such, the authors are not fully aware of the methodological analysis of models that encompasses the use of formative constructs (Assaker, Huang, and Hallak 2012). Furthermore, it is important to note that eliminating formative indicators that do not meet threshold levels in terms of their contribution has, from an empirical perspective, almost no effect on the parameter estimates when reestimating the model. Nevertheless, formative indicators should never be discarded simply on the basis of statistical outcomes. In this context, Einhorn’s (1972, p. 87) conclusion that “just as the alchemists were not successful in turning base metal into gold, the modern researcher cannot rely on the ‘computer’ to turn his data into meaningful and valuable scientific information” still holds true today. Therefore, before removing an indicator from the formative outer model, researchers need to carefully check its relevance from a content validity point of view, which was not provided in any of the 17 reviewed articles that used formative models in their analyses.
Inner Model Evaluation
As indicated by Hair et al. (2012b), the PLS-SEM approach focuses on the discrepancy between the observed and the approximated values for the dependent variables and the values predicted by the corresponding models, which implies that the assessment of the quality of the model should be based on the observation of their prediction capacity. Table 7 shows that not all studies addressed the evaluation of the outer model. There are two studies that do not perform this analysis because the models proposed do not contemplate a structural component. The first case concerns the estimation of a second-order hierarchical construct; in the second case, it is a model of confirmatory factor analysis.
Structural Model Evaluation in Tourism Research.
In the remaining studies (42), the structural model is assessed by observing the signal, magnitude, and bootstrap associated with each path coefficient. The coefficient of determination, R2, is shown and discussed in 41 of those 42 studies (93.2% of all articles). The significance, magnitude, and bootstrap of each path coefficient are shown and discussed in all 42 studies (95.5% of all articles). With respect to the other values that must be observed in a structural model, Table 7 shows that their use in tourism research has been sparse and, in some cases, nonexistent, although there has been significant improvement (p ≤ 0.01) in recent years, as the Stone–Geisser index (Q2, used to measure predictive relevance in terms of the indicators, not just the constructs) was reported in 11 (36.7%) of the articles published in 2012 onward compared to 3 (21.40%) of the articles published prior to 2012.
Although the goodness-of-fit (GoF) index is not included in the table (because it is not a universally usable index in the evaluation of a structural model), it is worth mentioning significant differences (p ≤ 0.05) in the reporting of this index over time, with two studies prior to 2012 and four studies published later reporting this index for models with reflective constructs. Overall, these observations suggest that a sounder understanding of the PLS-SEM inner model estimation is needed among tourism researchers, especially in terms of advanced criteria that could further help support the validity of the model.
Reporting and Advanced Analyses Using PLS-SEM
According to Chin (2010), studies that use PLS-SEM should report information about the target population and sample, data distribution, conceptual model (inner and outer), research hypotheses, and results that allow the validation of the model and testing of the hypotheses. As can be seen, the majority of studies examined in the tourism field report this information except for the analysis of the data distribution.
Hair, Ringle, and Sarstedt (2011) add that studies should also mention the more technical aspects, such as the resampling technique (e.g., bootstrapping procedure or blindfolding), the number of iterations, and the weighting schemes (centroid, factor, or path-weighting scheme for inner model; mode A or B for outer models) used in the estimation procedure. However, none of the studies analyzed provided all this information, although some of them indicated the resampling technique (7 or 15.9% of the total studies), the number of bootstrap samples (6 or 13.6% of the total articles), and the weighting scheme (2 or 4.5% of the total articles). All of these previous articles were published in 2012 and onward, suggesting a significant improvement (p ≤ 0.01) in reporting PLS-SEM results in recent years.
The same authors also stress the importance of reporting the software used to estimate the model. On this point, most of these publications (25 or 56.8% of the total studies) used the software SmartPLS (Ringle, Wende, and Will 2005), 6 (13.6%) used the PLSgraph (Chin 2003), 2 (4.5%) each used the software XLSTAT-PLSPM (Addinsoft 2013) and VisualPLS (Fu 2006), and 1 (2.3%) used the software WarpPLS (Kock 2011). In the remaining articles (8 studies or 18.2% of total), no indication is given on the software used.
Another notable aspect relates to the type of analysis presented: standard or advanced. In general, the studies conduct a standard analysis in the sense that a structural model with latent variables is proposed, with a formative or reflective outer model, and a set of research hypotheses is tested. However, more advanced analyses were identified in 14 (31.8%) studies, including the analysis of observed heterogeneity by testing moderating effects through the product of indicators method (4 studies, 9.1%) or via multigroup analysis (3 studies, 6.8%), the estimation of hierarchical models (5 studies, 11.4%), and the detection of unobserved heterogeneity by using PLS-FIMIX (2 studies, 4.5%). Table 8 shows the distribution of these articles in the two periods with the majority (10 of 14 published in 2012 onward), suggesting that although the use of such advanced applications has significantly grown in recent years (p ≤ 0.01), researchers in tourism still do not exploit the fully advanced potential of PLS-SEM and thus miss opportunities to further substantiate the appropriateness of findings and to improve their analyses.
Advanced Analyses Using PLS-SEM in Tourism Research.
Conclusion
The use of PLS-SEM has increased in business and tourism research. PLS-SEM, as an alternative technique to traditional CB-SEM, offers greater flexibility with regard to data assumptions, sample size, analyzing complex models that have both “formative” and “reflective” constructs, and testing predictive relationships between constructs. These characteristics make PLS-SEM particularly useful for tourism researchers and could be better harnessed by tourism researchers as a research tool. However, as with any statistical technique, PLS-SEM requires researchers to make several choices, which, if made incorrectly, can have substantial consequences on the validity of the results. Although reviews of PLS-SEM usage have been carried out across many disciplines in business research (e.g., Hair, Ringle, and Sarstedt 2011, 2012; Hulland 1999; Urbrach and Ahlemann 2010; Ringle, Sarstedt, and Straub 2012), corresponding assessments of PLS-SEM in tourism are still limited. To fill this gap, this study reviewed 44 articles that implemented the PLS-SEM approach to SEM in the field of tourism and were published in journals indexed in the SSCI and in Tourism Analysis for the period of 2000 and 2014 in terms of four key areas of the technique (as noted earlier in the text) to make recommendations for future use of PLS-SEM in tourism.
Review of the above-mentioned articles demonstrated that the use of this approach has accelerated in recent years, with 30 of the 44 articles (68.2% of the total articles) published in 2012 or 2014 suggesting that, while there may have been some initial unawareness or even apprehension of the use of PLS-SEM compared with the traditional CB-SEM approach, currently it is well accepted in the most important journals in the field (e.g., Tourism Management, Journal of Travel Research, and the Journal of Travel and Tourism Marketing); moreover, the technique is increasingly being applied to examine a large variety of themes in tourism in comparison to earlier years.
Second, the review also demonstrated consistent understanding of the reasons behind using PLS-SEM, with focus on prediction, small sample size, non-normal data, and formative measures being the most prevalent reasons over time; suggesting that tourism researchers show full awareness of the basic features of the PLS-SEM technique, and, in recent years, they have even become aware of more advanced features/types of analyses that are available within PLS-SEM software such as multigroup analysis, interaction effects, and response-based segmentation using FIXMIX that helps cope with data heterogeneity.
Third, a review of the model characteristics suggests that researchers in tourism should pay closer attention to model-specification issues, particularly when using formatively measured constructs in PLS-SEM. In formative constructs, indicators are assumed to capture the entire construct domain (or at least major parts of it). This issue needs to be further stressed in future applications, as it was not identified in any of the 17 articles that used either formative only or a combination of formative and measurement models in our review. Similarly, researchers in tourism should be careful when using single-item measures, which likely lag behind multi-item scales in terms of predictive validity. Their use should be avoided in PLS-SEM in most situations, especially in light of the technique’s predictive focus. Again this issue was not brought up in any of the six articles that used single-item latent variables. More careful thought should also be given to data characteristics. Although PLS-SEM performs well with small samples and nonnormal data, researchers should not be careless in implementing these advantages. Small sample sizes and skewed data easily increase sampling error, yielding inflated bootstrap standard errors. When this occurs, the technique’s statistical power is reduced, offsetting one of PLS-SEM’s major advantages. However, as shown in our review, recent articles have relied on larger sample sizes with a mean sample size of 562 for articles published in 2012 onward compared with a mean of 451 for articles published prior to 2012, thereby showing tourism researchers’ greater awareness with respect to this issue.
Fourth, researchers should make greater use of model-evaluation criteria, especially when assessing the quality of formatively measured constructs and the quality of inner models. Our review showed that current practice still indicate gaps in this regard, casting doubt on the validity of some of the measures. Similarly, researchers should make use of the full range of criteria available to assess the model’s predictive capabilities, such as the cross-loading measures, effect size redundancy index (f2) or the Stone–Geisser index (Q2). Tourism researchers need to understand that these measures are by definition not indicative of model fit in a covariance-based sense, and any effort to interpret them as such should clearly be rejected.
Finally, besides basic PLS-SEM analyses, our review demonstrated that researchers still have to take further advantage of a much larger set of methodological extensions of the PLS-SEM method, ranging from evaluation techniques such as blindfolding (e.g., Chin 1998) and confirmatory tetrad analysis (e.g., Gudergan et al. 2008) to interaction effects and PLS-SEM multigroup analyses (e.g., Sarstedt, Henseler, and Ringle 2011) and response-based segmentation approaches, such as finite mixture-PLS (e.g., Rigdon, Ringle, and Sarstedt 2010). Results from this study showed that the majority of research articles in the discipline still do not fully exploit such potentials and thus miss opportunities to further substantiate the appropriateness of findings and improve their analyses. Although results show that the use of such advanced applications (10 of the 14 advanced analyses articles were published in 2012–2014) has significantly grown in recent years, it is important for tourism researchers to stay updated in terms of the latest developments in PLS-SEM by consulting statistical journals (e.g., Structural Equation Modeling: A Multidisciplinary Journal; Multivariate Behavioral Research) that frequently publish innovations on this technique. Tourism scholars should also refer to recently available textbooks on PLS-SEM (see, e.g., A Primer on Partial Least Squares Structural Equation Modeling [PLS-SEM] by Hair et al. 2014), which contain further comprehensive and detailed explanation of the different types of analyses available within PLS-SEM. The present review and identified recommendations on the use of PLS-SEM in tourism should be particularly useful for tourism researchers adopting PLS-SEM in their studies in the future. Hopefully, these recommendations will help enhance the quality of empirical research in tourism and advance theory and practices in the future.
Limitations and Future Research
As with all research, the present study is not without limitations; for example, it assessed applying PLS-SEM based on information reported in published articles. In some cases, the authors of the reviewed articles may have made appropriate decisions and discussed decisions with referees during the review process but did not include such material in their articles, especially with regard to the suitability of PLS-SEM as it related to their work and model evaluation. This information, however, cannot be known just by reviewing the article. Another limitation is the small number of articles (44) examined as compared with the number of articles examined in other disciplines (e.g., 204 in marketing; 109 in management information systems). As the use of PLS-SEM becomes more common in tourism studies, a more comprehensive review can be prepared, and the rationale for using PLS-SEM and assessing its use in the context of tourism research will be more effective.
Footnotes
Declaration of Conflicting Interests
The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This article was partially financed by FCT–Foundation for Science and Technology.