Assessment of the structural model

The R²value (i.e., coefficient of determination) is a well-known measurement to evaluate, estimate and specify the structural model in PLS-SEM. It is imperative to explain the variance and predictive accuracy of the endogenous constructs by the independent variables. The ranges of the R²value are from 0 to 1. If the value is close to 1, high predictive accuracy is accounted (Bauer & Matzler, 2014; Gotz, Liehr-Gobbers, & Krafft, 2010; Hair et al., 2017). As a rule of thumb, in marketing research, R² values of 0.25, 0.50, 0.75 are considered as weak, moderate and substantial, respectively (Hair et al., 2011; Henseler et al., 2009). A greater value is expected, but an R² value of 0.20 is reported as influential in behavioral and performance measure studies (Hair et al., 2017).

Table 13. Coefficients of determination R²

R²value R2adj

BMS 0.178 0.165

CBA 0.267 0.216

CBP 0.092 0.084

FP 0.307 0.289

MP 0.092 0.077

RPT 0.244 0.232

SCA 0.081 0.074

In Table 13, all the latent endogenous constructs are shown by the rows while the columns indicate the R² and R²adj. The R² value explains the variance of the endogenous variable, while R²adjcompare the SEM output and explanatory power of the model with the multiple numbers of independent variables and the distinct data sets (Hair et al., 2017; Henseler et al., 2016). The general nature of R²adj is low compared to the regular R². However, this study considers only R² to evaluate the predictive power of the structural model instead of R²adjbecause the prior studies regarded the R²value only (Bauer & Matzler, 2014; Wong, 2013).

Table 13 shows that FP has the highest R² value of 0.307 compared to other endogenous variables. The 2nd most common variances have been explained by the CBA and RPT, which are 0.267 and 0.244, respectively. The variance of BMS is 0.178 while the R²values of CBP, MP, and SCA are moderately low (Henseler et al., 2016; Ringle et al., 2015). Though there is a different level of R²values, this study confirms that the predictive power of the exogenous variables is adequate.

However, the R²is not a bias-free estimation, and it is not the only measurement of model power accuracy (Hair et al., 2017).

Therefore, to evaluate the structural model by the R² value, Hair et al. (2017) proposed that it is necessary to check effect size f² to ascertain whether an omitted independent variable has a considerable impact on the endogenous construct. Reviewers and journal editors have suggested that effect size f²should also be used to assess the model.

Table 14. Effect sizes

f²effect size

BMS CBA CBP FP MP RPT SCA

AM 0.023

BMS 0.053 0.033

CBA 0.004 0.024 0.088

CBE 0.078 0.001 0.207

CBP 0.055

CI 0.004

ED 0.042

FP

MO 0.096

MP 0.302

RPT 0.026 0.101

SCA 0.051 0.049

TE 0.047

Effect size f²frequently measures the strength or magnitude of the relationship between the latent variables (Wong, 2013). Large, medium and small effects are represented by the values of 0.35, 0.15 and 0.02, respectively. Moreover, the value of below 0.02 illustrates that there are no effects among the constructs (Chin, 1988; Hair et al., 2017). The table shows that the construct MP has a remarkable effect on FP with the value of 0.302. Also, there are closely medium-level magnitudes of (CBE->RPT; 0.207), (RPT->CBP; 0.101), (CBE->BMS;

0.078) and (MO->BMS; 0.096). The rest of the effects are low while the bold values indicate that there are no effects between the constructs (Ringle et al., 2015). The empirical assessment shows that the relationships of all the constructs, with the exception of three, have a certain amount of magnitudes.

The predictive relevance of endogenous variables is also important for statistical reporting. Q² (i.e., Stone-Geisser) is used as an indicator of the predictive relevance. If the Q²value is more than 0, it means that a particular endogenous latent variable has predictive relevance in the path model (Hair et al., 2017). The blindfolding process initiates the Q² value for an explicitly recommended omission distance from 5 to 10 (Hair et al., 2012; Henseler et al., 2009;

Tenenhaus, Vinzi, Chatelin, & Lauro, 2005).

Table 15. Predictive relevance and the blindfolding Q²

Table 15 presents seven endogenous variables in the rows while the columns indicate the Q² values. In this study, all the endogenous variables have certain predictive relevance since the Q² values are above 0 (Ringle et al., 2015).

Furthermore, the predictive relevance Q²and effect size f²have been explained in the hypothesis testing.

Global goodness of fit

There are several model fitness criteria though not all of them might be useful in SmartPLS. Most of the fitness measures are still in the early stage to estimate the threshold level. Another reason is that PLS was designed for exploratory research instead of theory testing. However, the researchers endeavored to test the theory in view of various model fitness measurements in SmartPLS (Hair et al., 2017).

Hence, this study considers the global goodness of fit to recognize model misspecification and to judge the fitness of empirical data. Previous studies proposed some model fitness measurements such as goodness of fit (GoF), standardized root mean square residual (SRMR) and the root mean square residual covariance (RMStheta) (Hair et al., 2017).

Tenenhaus et al. (2005) and Wetzels, Odekerken-Schroder, and Oppen-Van (2009) anticipated that GoF is the relevant global substantial fit measure in PLS-SEM. Following the earlier suggestions, Bauer and Matzler (2014) also used the GoF measurement in the acquisition research. On the other hand, Henseler and Sarstedr (2013) conceptually and empirically tested the GoF measurement, and they found that GoF is problematic in SmartPLS. Similarly, some scholars have suggested that GoF should not be used to measure the global goodness of fit (Hair et al., 2017; Henseler & Sarstedr, 2013). Henseler et al. (2016) and Wong (2013) also stated that model estimation by means of GoF is meaningless, questionable and inconclusive because it is still in an early stage of fit measurement. Therefore, this study turns to the standardized root mean square residual (SRMR) to measure model fitness because SRMR is also a reliable measure in covariance-based SEM (CB-SEM).

Q²value

BMS 0.107

CBA 0.165

CBP 0.045

FP 0.190

MP 0.055

RPT 0.123

SCA 0.039

After evaluating the efficiency of SRMR, Henseler et al. (2014) also confirmed that SRMR provides the absolute global goodness of fit to avoid model misspecification. The threshold level of SRMR is 0.08, which is highly restricted by the CB-SEM though it is more flexible in PLS-SEM because CB-SEM aims to minimize the discrepancy while PLS-SEM is applied for model estimation. This study considers the 0.08 threshold level in the saturated model to test the approximation of the model fitness because the saturated model refers to free connection among the constructs (Henseler et al., 2016). In the saturated model, the SRMR value of this study model is 0.078; it shows that there is a high level of approximation of the model fitness. RMStheta is not reported in this study because it is not well developed yet to measure the model fitness. There are some other model fitness measures such as the exact model fit test, Chi², degree of freedom, NFI, and NNFI that are also not considered by this study because those measures are not perfect fit measures in SmartPLS (Dijkstra & Henseler, 2015;

Hair et al., 2017; Henseler et al., 2014).