3.1 Reasons for using PLS-SEM

Our review shows that 196 (82.01%) of the 239 studies provide a rationale for using PLS-SEM instead of factor-based SEM or sum scores regression (Table 3). The five most frequently mentioned reasons are: small sample size (114 studies, 47.70%), nonnormal data (76 studies, 31.80%), theory development and exploratory research (73 studies, 30.54%), high model complexity (70 studies, 29.29%), predictive study focus (61 studies, 25.52%), and formative measures (56 studies, 23.43%).

The two main reasons mentioned in our study are similar to those reported by Hair et al. (2012) a decade ago. Given the increasing calls to move beyond data characteristics to motivate the use of PLS-SEM and, instead, emphasize the research objective (e.g., Hair et al., 2019b; 2019c; Sarstedt et al., 2021), this finding raises concerns. While PLS-SEM performs well in terms of statistical power (Sarstedt et al., 2016) and convergence compared to other methods when the sample size is small (Henseler et al., 2014; Reinartz et al., 2009), it is undeniable and axiomatic that small samples adversely affect all statistical techniques, including PLS-SEM (Benitez et al., 2020). In fact, more than two decades ago, Chin (1998, p. 305) warned researchers that “the stability of the estimates can be affected contingent on the sample size.” Cassel et al. (1999) also noted that PLS-SEM analyses should draw on sufficiently large sample sizes to warrant small standard errors. Numerous others, such as Marcoulides and Saunders (2006), Hair et al. (2013), and Rigdon (2016), have clearly advised against relying on the logic of using PLS-SEM to derive results from small samples—unless the nature of the population can justify such a step (e.g., a small population size as commonly encountered in B2B research; Benitez et al., 2020; Hair et al., 2019c). In fact, critics such as Evermann and Rönkkö (2021), Rönkkö and Evermann (2013), and Rönkkö et al. (2016) have repeatedly used PLS-SEM's misapplication as a “strawman” argument to criticize the technique itself, rather than the weak research designs (Petter & Hadavi, 2021; Petter, 2018). Similarly, simulation studies have shown that PLS-SEM does not offer substantial advantages over other SEM methods in terms of parameter accuracy when the data depart somewhat from normality (Goodhue et al., 2012; Reinartz et al., 2009). Reflecting this evidence, guideline articles have emphasized that it is not sufficient to motivate the choice of PLS-SEM over alternative methods primarily on the basis of nonnormality alone (e.g., Hair et al., 2019b, 2019c, 2020b).

Despite their strong emphasis on the data characteristics, our results also attest to marketing researchers' higher maturity in terms of motivating their method choice. For example, compared to the previous period, an increasing number of researchers emphasize model complexity and their study's predictive focus. Both are valid arguments, as PLS-SEM can handle highly complex models with many indicators, constructs, and model relations without identification concerns (Akter et al., 2017). When estimating these models, PLS-SEM follows a causal-predictive paradigm aimed at testing the predictive power of a model developed carefully on the basis of theory and logic (Liengaard et al., 2021). The underlying theoretical rationale, also referred to as explanation and prediction-oriented (EP) theory, “corresponds to commonly held views of theory in both the natural and social sciences” (Gregor, 2006, p. 626). Numerous standard models, such as the various customer satisfaction index (e.g., Eklöf & Westlund, 2002; Fornell et al., 1996) or technology acceptance models (e.g., Davis, 1989; Venkatesh et al., 2003), follow an EP-theoretic approach by aiming to explain the cause-effect mechanisms the model postulates, while also generating predictions that underline its practical usefulness (Sarstedt & Danks, 2021). PLS-SEM's ability to strike a balance between machine learning methods, which are fully predictive but atheoretical by nature (Hair & Sarstedt, 2021), and factor-based SEM, which is purely concerned with theory confirmation (Hair, et al., 2021b), makes PLS-SEM a particularly valuable method for applied research disciplines like marketing.

Researchers' increased focus on predictive purposes is no coincidence. Seminal articles by Evermann and Tate (2016) as well as Shmueli et al. (2016) on the method's predictive performance have opened a new chapter in PLS-SEM-based methodological research. For example, Sharma et al. (2021c) introduced prediction-oriented model comparison to the field by identifying metrics that excel at selecting the model with the highest predictive power and adequate fit from a set of competing models. Liengaard et al.'s (2021) cross-validated predictive ability test and its recent extension (Sharma et al., 2021a) enable researchers to test a model's predictive power relative to that of competing models and on a standalone basis. Other studies have extended prior research by comparing PLS-SEM's predictive power with that of Hwang and Takane's (2004) generalized structured component analysis, identifying situations that favor each method's use from a prediction point of view (Cho et al., 2021). Guideline articles that make these techniques accessible to a broader audience have accompanied these methodological extensions (Chin et al., 2020; Hair, 2021; Shmueli et al., 2019). Our review of the reasons given for the choice of PLS-SEM reflects some of these developments, but we expect a further shift toward emphasizing the method's causal-predictive focus in the future.

Finally, and unlike in the previous period, researchers also suggest additional reasons, such as PLS-SEM's popularity and standard use in the field (14 studies, 5.86%), the estimation of moderation effects (12 studies, 5.02%), mediation effects (3 studies, 1.26%) and higher-order constructs (8 studies, 3.35%). Indeed, recent research emphasizes the method's efficacy regarding estimating conditional process models combining mediating and moderating effects in a single analysis (Cheah et al., 2021; Sarstedt et al., 2020a), and clarifies the specification, estimating, and evaluation of higher-order constructs by means of PLS-SEM (Sarstedt et al., 2019). At the same time, fewer studies motivate their choice of PLS-SEM on the grounds of formative measurement—a finding that the studies' model characteristics, which we discuss later, also mirror.

3.2 Data characteristics

Of the 486 models, 476 report the sample size (97.94%). The average sample size (5% trimmed mean = 279.19) and the median (199) of the models in our review are higher than those that Hair et al. (2012) reported for the 1981–2010 period (5% trimmed mean = 211.29; median = 159). Four studies in our review were conspicuous due to their very large sample sizes of between n = 8876 and n = 26,576. At the same time, 85 of the 476 models (17.86%) rely on sample sizes of less than 100 (i.e., the smallest sample size was n = 29). Only 17 of the 476 models (3.57%) fail to meet the ten times rule (Barclay et al., 1995), which is a very rough guideline for determining the minimum sample size required to achieve an adequate level of statistical power (for details and alternatives, see Hair et al., 2022, Chap. 1). Those models that do not meet the ten times rule draw on sample sizes that are, on average, only 10.98% below the recommended level (Table 4).

These results are encouraging compared to those in Hair et al. (2012), who reported a greater share of models estimated with less than 100 observations (24.44%), more models that failed the ten times rule (9.00%), and a higher relative deviation from the recommended sample size (45.18%). Marketing researchers—despite their reporting of PLS-SEM's efficacy regarding estimating models with small sample sizes as motivating their method choice—seem to be more concerned about sample size issues than previously. Critical accounts of the method (Goodhue et al., 2012; Rönkkö & Evermann, 2013) and conceptual discussions (Hair et al., 2019c; Rigdon, 2016) have certainly contributed to this increased awareness. Nevertheless, empirical research almost never discusses the statistical power arising from a concrete analysis, despite the available tutorials on power analyses in a PLS-SEM context (Aguirre-Urreta & Rönkkö, 2015). For sample size determination, researchers can also draw on Kock and Hadaya's (2018) inverse square root method, which is relatively easy to apply and offers a more accurate picture of minimum sample size requirements than the ten times rule does.

Missing values are a primary concern in the data collection and processing phases. Nevertheless, only 41 studies (17.15%) acknowledge missing values in their data, while almost all of them (39 of 41 studies, 95.12%) used casewise deletion to treat them (Table 4). Grimm and Wagner (2020) have recently shown that PLS-SEM estimates are very stable when using casewise deletion on datasets with up to 9% missing values. However, when doing so researchers need to ensure the missing value patterns are not systematic, and that the final model estimation yields sufficient levels of statistical power. None of the relevant studies in our review offer such information. While casewise deletion seems to be the default setting for researchers using PLS-SEM, extant guidelines recommend using mean replacement when less than 5% values are missing per indicator or applying more complex procedures, such as maximum likelihood and multiple imputation (Hair et al., 2022). However, while the efficacy of more complex imputation procedures has been tested in a factor-based SEM context (Graham & Coffman, 2012; Lee & Shi, 2021), their impact on PLS-SEM's parameter accuracy and predictive performance remains a blind spot in methodological research.

Even though outliers influence the ordinary least squares regressions in PLS-SEM, only seven studies (2.93%) explicitly considered them. In one of these studies, the authors did not try to identify outliers but used an interval smoothing approach to mitigate the effect of corresponding observations. Of the remaining six studies that considered outliers, three did not mention the specific method used to detect outliers, whereas the remaining three studies included brief descriptions of the approaches: univariate and multivariate outlier detection, inconsistent or overly consistent response patterns, and univariate outlier detection using boxplots followed by detecting incongruence in response patterns. Most of these studies (5 out of 6) removed outliers from their dataset; the other study did not identify any outliers. While the presence of outliers might influence results substantially (Sarstedt & Mooi, 2019), our findings indicate that applied PLS-SEM research generally disregards outlier detection and treatment.

Nonnormal data is the second most frequently mentioned reason for using PLS-SEM. Nevertheless, only 24 studies (10.04%) acknowledge the nonnormal distribution of their data with few studies quantifying the degree of nonnormality by means of, for example, kurtosis and skewness (5 studies; Table 4). This is generally unproblematic, since PLS-SEM is a nonparametric method and, as such, robust in terms of processing nonnormal data (Hair et al., 2017b; Reinartz et al., 2009). While Hair et al. (2022, Chap. 2) note that highly nonnormal data may inflate standard errors derived from bootstrapping, Hair et al.'s (2017b) simulation study results suggest that such data do not impact Type I or II errors negatively.

Using discrete variables (i.e., binary and categorical variables) when measuring models is a final area of concern in PLS-SEM applications. Of the 239 studies, only 24 (10.04%) use binary or categorical variables as elements in measurement models (i.e., not as a grouping variable to split the dataset as part of a multigroup analysis), which is considerably less than what Hair et al. (2012) reported (27.94%). In line with Hair et al.'s (2012) recommendations, which raised concerns regarding their general applicability, especially when used as binary single-item indicators of endogenous constructs (e.g., to represent a choice situation), recent studies primarily use binary and categorical variables to perform multigroup analyses. However, research on the handling of discrete variables in PLS-SEM has progressed since 2012. For example, Hair et al. (2019a) document how to process data from discrete choice experiments where binary indicators representing a choice situation (e.g., buy/not buy) measure the constructs. Similarly, Cantaluppi and Boari (2016) extended the original PLS-SEM algorithm to accommodate ordinal variables when researchers cannot assume equidistance between the scale categories (see also Schuberth et al., 2018b). Given these developments, we expect the use of discrete variables in PLS path models, such as when estimating data from choice experiments, to gain traction in the future.

To summarize, while our results point to improvements in researchers' awareness of sample size issues in PLS-SEM use, they devote too little effort toward quantifying the degree of statistical power associated with the model estimation. Missing data are not routinely discussed, and their treatment in PLS-SEM is still an area of potential concern, which methodological research should address. Researchers should also be aware of recent extensions that facilitate the inclusion of discrete variables in PLS path models.

3.3 Model characteristics

The number of constructs and indicator variables that define model complexity and the use of formative measurement models are the key reasons for PLS-SEM's attractiveness (Hair et al., 2022, Chap. 1). Table 5 offers an overview of these and other model characteristics, contrasting our results with those of Hair et al.'s (2012) review of the 1981–2010 period.

The average number of latent variables in the path models is 7.39, which is only slightly lower than in the previous period (7.94). Similarly, the average number of indicators per model is similar in both periods (1981–2010: 29.39; 2011–2020: 29.55), suggesting there has been no change in the model complexity over the years. The average number of structural model relationships, which increased only slightly from 10.56 to 11.90 from the initial period to the most recent, also supports this finding. In addition, we do not observe any substantial changes in the model types considered in the studies; that is, whether the models are focused versus unfocused by comprising a higher versus lower share of exogenous constructs than endogenous constructs. Specifically, the share of focused, unfocused, and balanced models is almost the same and very similar to those reported in Hair et al. (2012). We therefore repeat Hair et al.'s (2012) call to consider focused and balanced models, also because these model types should meet PLS-SEM's prediction goal better.

PLS path models, depending on their measurement model specifications, can be classified as either only reflective, only formative, or a combination of both.1 We find the majority of the 486 models' constructs (374 models, 76.95%) are only reflectively measured, followed by models with both reflective and formative measures (105 models, 21.60%)—a much higher share of reflective measures compared to Hair et al. (2012). Only 7 of the 486 models' constructs (1.44%) are purely formatively specified. Somewhat surprisingly, a large percentage of the models (233 models; 47.94%) lack a description of the constructs' measurement modes, which is much higher than in Hair et al.'s (2012) review (11.90%). Classifying these models ex post by means of Jarvis et al. (2003) guidelines suggests the overwhelming majority (227 models; 97.42%) should be considered as only comprising reflectively specified constructs. While this result suggests that researchers consider reflective measures the default option, the consequences of measurement model misspecification should not be taken lightly. The impact of misspecifications on downstream estimates are generally not pronounced (Aguirre-Urreta et al., 2016), but the primary problem of such misspecifications arises from content validity concerns, since both measurement types rely on fundamentally different sets of indicators (Diamantopoulos & Siguaw, 2006). In line with Hair et al. (2012), researchers should ensure the measurement model specification is explicit to preempt criticism arising from potential misspecification claims. The confirmatory tetrad analysis for PLS-SEM (Gudergan et al., 2008; CTA-PLS; Hair et al., 2018, Chap. 3) provides a statistical test to confirm the choice of measurement model. However, only six studies (2.51%) used this approach to assess the adequacy of their measurement model specification.

The average number of indicators for reflectively measured constructs is 3.85 and 4.28 for formatively measured ones, which is similar to those in the previous period (Table 5). Contrary to their reflective counterparts, formative measurement models need to cover a much broader content bandwidth, as they are not conceived as interchangeable (i.e., highly correlated) measures of the same theoretical concept (Diamantopoulos & Winklhofer, 2001). Consequently, formative measurement models should have more indicators than reflective models do. Our review supports this notion, but the comparably small, albeit significant (t = 2.037, p = 0.044) difference in the average number of indicators is surprising, suggesting that future studies should devote more attention to content validity concerns in formative measurement.

We also find the share of single-item constructs decreased from 46.30% to 36.42% from one period to the next. On the one hand, this finding is encouraging, as it suggests researchers have become aware of single-item measures' limitations in terms of explanatory power. Specifically, Diamantopoulos et al. (2012) show that, due to the absence of multiple indicators of the same concept, the lack of measurement error correction deflates the model estimates, potentially triggering Type II errors in the model estimates (Sarstedt et al., 2016). On the other hand, single items are still too prevalent in PLS path models, which is particularly problematic when measuring endogenous target constructs. Researchers should generally refrain from using single items unless they measure observable characteristics, such as is commonly done in the form of control variables (e.g., income, sales, and number of employees).

To summarize, we observe a clear shift toward reflective measurement in PLS-SEM studies. The various controversies about formative measures' use (Howell et al., 2007; Rhemtulla et al., 2015; Wilcox et al., 2008) probably triggered this development, prompting researchers to “take the safe route” and rely on standard reflective measures. By following this mantra, researchers forego the opportunity to offer marketing practice differentiated recommendations. Contrary to their reflective counterparts, formative measures offer practitioners concrete guidance on how to “improve” certain target constructs—an important consideration in light of the growing concerns about marketing research's relevance for business practice (Homburg et al., 2015; Jaworski, 2011; Kohli & Haenlein, 2021; Kumar, 2017). At the same time, we see some improvement in the model specification practices, with researchers relying less on single-item measures. Nevertheless, the percentage of studies using single items is still high, providing an opportunity for improvement.

3.4 Model evaluation

Since Hair et al.'s (2012) review, research has proposed various improvements in the model evaluation metrics for both measurement and structural models. In the following, we assess whether researchers consider best practices, while acknowledging that some of these improvements have only recently been introduced. We distinguish between reflective and formative measurement models, since their validation—as extensively documented in the literature—relies on totally different sets of criteria (Hair et al., 2021a, 2022; Ramayah et al., 2018; Wong, 2019).

3.5 Advanced modeling and analysis techniques

With the PLS-SEM field's increasing maturation (Hwang et al., 2020; Khan et al., 2019), researchers can draw on a greater repertoire of advanced modeling and analysis techniques to support their conclusions' validity and to identify more complex relationships patterns (Hair et al., 2020a).

For example, many researchers use PLS-SEM because of its ability to accommodate higher-order constructs without identification concerns. Analyzing the prevalence and application of higher-order constructs, we find that 71 of the 239 studies (29.71%) included at least one such construct. The majority of these studies considered second-order constructs (65 studies), while the remaining studies included third-order constructs (5 studies) or both (1 study). Analyzing the measurement model specification of the second-order constructs (Wetzels et al., 2009), we find that most of the studies employ Type I (reflective-reflective; 30 studies), Type II (reflective-formative; 26 studies), or both (4 studies). Only five studies employ Type IV (formative-formative), while no study draws on a Type III (formative-reflective) measurement specification. Analyzing the specification, estimation, and evaluation of the higher-order constructs in greater detail, gives rise to concern in two respects. First, 47 of the studies that employed higher-order constructs do not make the specification and estimation transparent.2 Those that do, either draw on the repeated indicators approach (15 studies) or the two-stage approach (9 studies). This practice is problematic as endogenous Type II and IV higher-order constructs cannot be reliably estimated using the standard repeated indicators approach (Becker et al., 2012). Second and more importantly, only eight studies correctly evaluate the higher-order constructs using criteria documented in the extant literature (e.g., Sarstedt et al., 2019). The majority of the 71 studies do not evaluate the higher-order constructs at all (30 studies) or incompletely apply the relevant criteria (21 studies), typically disregarding discriminant validity assessment in reflective (i.e., Type I) and the redundancy analysis in formative higher-order constructs (i.e., Types II and IV). Finally, 12 studies misapply the evaluation criteria by erroneously interpreting the relationships between lower- and higher-order components as structural model relationships. Overall, our findings suggest that researchers need to provide much more care in their handling of higher-order constructs by considering the most recent guidelines (Sarstedt et al., 2019).

The identification and treatment of observed and unobserved heterogeneity is another aspect that has attracted considerable attention in methodological research (Memon et al., 2019; Rigdon et al., 2010; Sarstedt et al., 2022). Our review shows that researchers' use of PLS-SEM also mirrors this development. A total of 115 studies (48.12%) in our review took observed heterogeneity into account by either considering continuous (58 studies 24.27%) or categorical (57 studies 23.85%) moderating variables. The strong increase in the percentage of studies using continuous moderating variables compared to the previous period, when only 7.35% undertook corresponding moderation analyses, demonstrates the growing interest in more complex model constellations. Although a total of 57 studies investigated the categorical moderators' impact by means of multigroup analysis (Sarstedt et al., 2011), only 12 address the issue of measurement invariance by using the MICOM procedure (Henseler et al., 2016b). This practice is problematic as establishing measurement invariance is a requirement for any multigroup analysis. By establishing (partial) measurement invariance, researchers are assured that the group differences in the model estimates are not the result of, for example, group-specific response styles (Hult et al., 2008).

While the consideration of moderators facilitates the identification and treatment of observed heterogeneity, group-specific effects can also result from unobserved heterogeneity. Since failure to consider such unobserved heterogeneity can be a severe threat to PLS-SEM results' validity (Becker et al., 2013), researchers call for the routine use of latent class procedures. Recent guidelines recommend the tandem use of finite mixture PLS (FIMIX-PLS; Hahn et al., 2002) and more powerful latent class techniques, such as PLS prediction-oriented segmentation (PLS-POS; Becker et al., 2013), PLS genetic algorithm segmentation (PLS-GAS; Ringle et al., 2014), and PLS iterative reweighted regressions segmentation (PLS-IRRS; Schlittgen et al., 2016). While FIMIX-PLS allows researchers to reliably determine the segments to extract from data (Sarstedt et al., 2011), techniques such as PLS-POS reproduce the actual segment structure more effectively. Recent research calls for the routine application of latent class techniques to ensure that heterogeneity does not impact the aggregate level results adversely (Sarstedt et al., 2020b). Despite these calls, only 10 of the 239 studies in our review (4.18%) conduct a latent class analysis. All these studies apply FIMIX-PLS, which, while detecting heterogeneity issues well, clearly lags behind in terms of treating the heterogeneity.

Researchers have long noted the need to explore alternative explanations for the phenomena under investigation by considering different model configurations. Such explanations will allow researchers to compare the emerging models by using well-known regression literature metrics (Burnham & Anderson, 2002) to select the best-fitting model in the set. Only 10 of our review studies (4.18%) compare alternative models by using the same data set. By disregarding alternative theoretically plausible model configurations, researchers could fall prey to confirmation biases, since they only look for configurations that support “their” model (Nuzzo, 2015), and miss out on opportunities to fully understand the mechanisms at work (Sharma et al., 2019). Further, very few studies compare multiple models by drawing on a large set of criteria, therefore, applying between one and nine different criteria. However, none of the studies applies the Bayesian Information Criterion (Schwarz, 1978) or the Geweke and Meese (1981) criterion, which prior research identified as superior in PLS-SEM-based model comparisons (Danks et al., 2020; Sharma et al., 2019; Sharma et al., 2021c). With further recent developments in these directions (e.g., based on the cross-validated predictive ability test, CVPAT), we expect the popularity and relevance of the predictive model assessment (Sharma et al., 2021a) and model comparison (Liengaard et al., 2021) to increase.

While our review results point to some progress in terms of applying advanced analysis techniques, there is still ample room for improvement. First, researchers need pay greater attention to the specification, estimation and particularly the validation of higher-order constructs. Second, measurement invariance assessment should precede any multigroup analysis (Henseler et al., 2016b), since failure to establish at least partial measurement invariance renders any multigroup analysis meaningless. Third, researchers should routinely apply latent class analyses to check their aggregate level results' robustness. Failure to do so may result in misleading interpretations when unobserved heterogeneity affects the model estimates. Finally, researchers should more often compare theoretically plausible alternatives to the model under consideration. Considering alternative explanations is a crucial step before making any attempt to falsify a theory (Popper, 1959).

3.6 Reporting

Hair et al.'s (2012) review emphasized that the algorithm settings used in the analysis needed to be transparent to facilitate its replicability. Our review shows some improvements compared to the previous period, most notably with regard to documenting the software used for the analysis. Whereas less than half of all the studies in the previous period report the software (49.02%), the percentage is much higher in recent PLS-SEM research (193 studies, 80.75%). Most PLS-SEM studies (166 studies, 69.46%) apply SmartPLS, with PLS-Graph following far behind (18 studies, 7.53%). Other rarely used software programs include XLSTAT (4 studies, 1.67%), WarpPLS (3 studies, 1.26%) and ADANCO (2 studies, 0.84%). In addition, two studies (0.82%) use the plspm package of the R software.

Only five studies (2.09%) mention the parameter or algorithm settings, seven studies (2.93%) the computational options (e.g., weighting schemes), and 12 studies (5.02%) the maximum number of iterations. Although making these algorithm settings transparent is certainly useful, different algorithm settings do not normally induce significant deviations in the estimation results. This, however, differs in bootstrapping settings; Rönkkö et al. (2015), for example, show that the use of sign change options can trigger considerable Type I errors, whereas Streukens and Leroi-Werelds (2016) emphasize the need to draw many bootstrap samples (i.e., at least 10,000) from the original data. In this regard, it is encouraging that 175 studies (73.22%) mention the use of bootstrapping, 141 of which provided additional details of the algorithmic settings, such as the number of bootstrap samples used. None of the studies use jackknifing as an alternative means of deriving standard errors, which is likely due to this method only being implemented in the early PLS-Graph software that only a few researchers still use.

Research has produced various variants of the original PLS-SEM algorithm (Becker & Ismail, 2016). One prominent research stream deals with the original PLS-SEM's extensions, to estimate common factor models. While researchers have introduced various such extensions (Bentler & Huang, 2014; Kock, 2019), consistent PLS-SEM (PLSc-SEM) has become the primary technique in this stream (Dijkstra & Henseler, 2015). The use of PLSc-SEM has triggered discussions with certain researchers, who call for its routine application (Benitez et al., 2020; Evermann & Rönkkö, 2021; Henseler et al., 2016a), while others conclude the method “adds very little to existing knowledge of SEM” (Hair et al., 2019c, p. 570). PLS-SEM users seem to share this skepticism as only five studies (2.09%) apply PLSc-SEM.

Finally, we see no increase in the share of studies reporting indicator covariance or correlation matrices. This practice is problematic, because it hinders the results' replicability and checking whether model evaluation metrics, such as the HTMT, are adequate. On the contrary, studies routinely report construct-level correlation matrices (180 studies, 75.31%), largely due to their relevance for Fornell-Larcker-based discriminant validity assessment.

To summarize, reporting practices regarding the software and algorithm settings used have improved since Hair et al.'s (2012) review. Nevertheless, more studies need to make the software and, particularly, the bootstrap settings transparent. Furthermore, SmartPLS (Ringle et al., 2015) is currently the most widely applied software, likely due to its user friendliness (Sarstedt & Cheah, 2019), features, and functionalities (Memon et al., 2021). However, with new software packages, such as cSEM (Rademaker & Schuberth, 2021) and SEMinR (Ray et al., 2021), as well as the detailed documentation on them (Hair et al., 2021a; Henseler, 2021), we expect the use of R packages to gain traction. Finally, in light of the growing concerns about social science research's replicability (Rigdon et al., 2020), reporting indicator and construct-level correlation matrices should become mandatory.