Adopting structural equation modeling techniques and using SmartPLS 4.0 software to test hypotheses.Wold's research[39] indicates that partial least squares (PLS) is a technique based on structural equation modeling (SEM), and the reasons for using SmartPLS 4.0 in this study are as follows. First, it is well suited for small sample sizes. In many cases, researchers can obtain limited data. SmartPLS 4.0 performs well with small sample data (less than 200) by employing robust statistical methods to reduce dependency on assumptions about large samples (Ramayah et al.,2018)[40]. Second, the questionnaire data collected often do not follow a normal distribution. SmartPLS 4.0 employs rank-based estimation methods, providing reliable analysis results even for nonnormally distributed data (Hair et al., 2016)[41]. Third, SmartPLS 4.0 offers extensive results and interpretability capabilities. It can generate statistical indicators such as path coefficients, R² values, T values, and P values, assisting researchers in evaluating model fit and hypothesis validation. Finally, SmartPLS provides wide academic support and application in the field of structural equation modeling (Hao et al., 2021[42];Delgosha & Hajiheydari, 2021[43];Sobaih et al.,2022[44]; Purwanto et al.,2021[45]). In recent years, more and more scholars have used SmartPLS as a tool for statistical analysis.
5.1 Confirmatory factor analysis
First, we checked the histograms and kurtosis of all the structures, and the results indicated that our data conformed to normality. Then, we conducted an overall quality check of the data using SmartPLS 4.0 software. We performed a confirmatory factor analysis to assess the reliability and validity of the model's factors.
5.1.1 Reliability
Composite reliability (rho_a) mainly considers the covariance between measurement items, while composite reliability (rho_c) considers factor loading and measurement errors. Composite reliability (rho_c) is often considered more suitable for complex models because it considers the structure of measurement tools more comprehensively. To verify the reliability of our model, we conducted measurements using both methods. First, the Cronbach's alpha values ranged from 0.729 to 0.849. The composite reliability (rho_a) value ranged from 0.737 to 0.945. The composite reliability (rho_c) value ranged from 0.848 to 0.928. These values all exceeded the critical thresholds recommended by SmartPLS 4.0 software (Cronbach's alpha ≥ 0.700, composite reliability (rho_a) ≥ 0.700, and composite reliability (rho_c) ≥ 0.700). The results indicate that the data meet reliability standards and are ready for further analysis.
Second, we evaluated the average variance extraction (AVE) and factor loading to validate the convergence validity of the model. The factor loadings range for all measured variables is 0.704 to 0.889, according to Hair et al.'s suggestion,it should be greater than 0.7[46]. The average variance extracted (AVE) is between 0.597 and 0.865 and is more significant than the cutoff value of 0.50. This implies that more than half of the variability in the surveys can be attributed to their underlying construct, thereby demonstrating robust convergent validity. The results indicated that all indices were sufficiently satisfactory and that our scale has good reliability and validity.The relevant data can be found in Tables 2 and 3.
Table 2 Factor loadings
|
AD
|
AR
|
MR
|
PV
|
PR
|
RP
|
AD1
|
0.803
|
|
|
|
|
|
AD2
|
0.784
|
|
|
|
|
|
AD3
|
0.785
|
|
|
|
|
|
AD4
|
0.759
|
|
|
|
|
|
AD5
|
0.73
|
|
|
|
|
|
AR1
|
|
0.883
|
|
|
|
|
AR2
|
|
0.768
|
|
|
|
|
AR3
|
|
0.77
|
|
|
|
|
MR1
|
|
|
0.884
|
|
|
|
MR2
|
|
|
0.899
|
|
|
|
PV1
|
|
|
|
0.704
|
|
|
PV2
|
|
|
|
0.863
|
|
|
PV3
|
|
|
|
0.848
|
|
|
PR1
|
|
|
|
|
0.957
|
|
PR2
|
|
|
|
|
0.902
|
|
RP1
|
|
|
|
|
|
0.889
|
RP2
|
|
|
|
|
|
0.869
|
RP3
|
|
|
|
|
|
0.731
|
RP4
|
|
|
|
|
|
0.72
|
Note: AD = Adoption, PV = Perceived Value, RP = Risk Perception, PR = Professional reviews, AR = Amateur reviews, and MR = Manufacturer reviews.
Table 3 Construct reliability and validity
|
Cronbach’s Alpha
|
Composite reliability (rho_a)
|
Composite reliability (rho_c)
|
AVE
|
AD
|
0.831
|
0.832
|
0.881
|
0.597
|
AR
|
0.757
|
0.882
|
0.85
|
0.654
|
MR
|
0.742
|
0.745
|
0.886
|
0.795
|
PV
|
0.729
|
0.737
|
0.848
|
0.653
|
PR
|
0.849
|
0.945
|
0.928
|
0.865
|
RP
|
0.823
|
0.869
|
0.88
|
0.65
|
Note: AD = Adoption, PV = Perceived Value, RP = Risk Perception, PR = Professional reviews, AR = Amateur reviews, and MR = Manufacturer reviews.
5.1.2 Discriminant validity
We employed the heterotrait–monotrait ratio (HTMT) method, which measures the discriminant validity between factors by calculating the ratio of the average interconstruct correlations to the average intraconstruct correlations[47]. The maximum HTMT ratio of the sample is 0.686, which is below the maximum value of 0.85 (Henseler et al., 2015)[48]. Furthermore, according to Fornell and Larcker's (1981) research, the square root of the AVE should surpass its maximum correlation with items in distinct structures[49], as evident from the diagonal elements in the corresponding matrix presented in Table 4. Both results indicate that the discriminant validity of our scale was good.
Table 4 Discriminant validity and correlation matrix.
|
AD
|
AR
|
MR
|
PV
|
PR
|
RP
|
AD
|
0.813
|
|
|
|
|
|
AR
|
0.089
|
0.825
|
|
|
|
|
MR
|
-0.217
|
0.314
|
0.893
|
|
|
|
PV
|
0.439
|
-0.004
|
-0.382
|
0.700
|
|
|
PR
|
0.303
|
0.196
|
-0.057
|
0.146
|
0.929
|
|
RP
|
-0.270
|
0.280
|
0.483
|
-0.400
|
0.161
|
0.837
|
Note: Bold-faced diagonal elements are the square roots of the AVEs. In constructs,the offdiagonal elements are the correlations. AD = Adoption, PV = Perceived Value, RP = Risk Perception, PR = Professional reviews, AR = Amateur reviews, and MR = Manufacturer reviews.
5.2. Results
5.2.1 Model fit
We tested the standardized root mean square residual (SRMR) and calculated the normalized fit index (NFI) to evaluate the overall fit of the structural equation model, as shown in Table 5. According to Hu and Bentler (1998)[50], the SRMR is considered an absolute fit index, with a standard threshold of SRMR < 0.08, and our model perfectly meets this criterion. Additionally, the calculation results for the normalized fit index (NFI) adhere to the standard set by Schumacker (2004)[51], with NFI > 0.9. Therefore, the model fitting in this study can be considered good.
Table 5 Model fit
|
Saturated model
|
Estimated model
|
SRMR
|
0.069
|
0.077
|
NFI
|
0.918
|
0.916
|
5.2.2 Direct effects
To verify the proposed model, we have decided to employ partial least squares structural equation modeling (PLS-SEM). We conducted a bootstrapping analysis with 5000 subsets to test our model at a significance level of 0.05. To evaluate the overall predictive ability of the model, we utilized path coefficients and the significance of the main structures within the model. The results are presented in Table 6. For the direct impact on adoption, risk perception (β=-0.528, p<0.001) has the most significant and negative influence, and perceived value (β=0.264, p<0.001) has a notably positive impact. Additionally, concerning the direct impact on risk perception, manufacturer reviews (β = -0.457, p < 0.001) have the most significant influence, professional reviews (β =- 0.078, p < 0.05) have a slightly weaker yet still significant impact, while amateur reviews have a minimal and insignificant impact. Finally, in terms of the direct influence on perceived value, the results show that manufacturer reviews (β = 0.234, p < 0.001) negatively influence perceived value, followed by amateur reviews (β = 0.14, p < 0.01), with a modest but still significant negative impact. However, professional reviews do not have a major influence on perceived value. In summary, most of the path coefficients of the hypotheses are significant, indicating that H1, H2, H3a, H4b, H5a, and H5b are supported. Only two hypotheses, H3b and H4a, are rejected.
Table 6 Results of Hypothesis Testing
Dependent
Variable
|
Hypothesis
|
Path
|
ß
|
P Value
|
Hypothesis Supported
|
AD
|
H1
|
PV→AD
|
0.264***
|
0.000
|
Yes
|
|
H2
|
RP→AD
|
-0.528***
|
0.000
|
Yes
|
PV
|
H3b
|
PR→PV
|
0.083n/s
|
0.07
|
NS
|
|
H4b
|
AR→PV
|
0.14**
|
0.009
|
Yes
|
|
H5b
|
MR→PV
|
0.234***
|
0.000
|
Yes
|
RP
|
H3a
|
PR→RP
|
-0.078*
|
0.04
|
Yes
|
|
H4a
|
AR→RP
|
0.065n/s
|
0.125
|
NS
|
|
H5a
|
MR→RP
|
-0.457***
|
0.000
|
Yes
|
Note: *p < 0.05; **p < 0.01; ***p < 0.001; n/s = not significant. AD = Adoption, PV = Perceived Value, RP = Risk Perception, PR = Professional reviews, AR = Amateur reviews, and MR = Manufacturer reviews.
5.2.3 Mediation effect
We hypothesized that perceived value and risk perception would mediate the impact of manufacturer reviews, professional reviews, and amateur reviews on adoption. To determine the presence of mediation effects, we calculated the variance accounted for(VAF) to determine the strength of the mediating effect relative to the total effect. Moreover, when the VAF is greater than 20% or less than 80%, partial mediation effect exists (Hair et al., 2016)[52]. The outcome indicate that risk perception significantly mediates the relationship between manufacturer reviews and adoption, with a VAF value of 24.41%, exceeding the defined threshold. Similarly, risk perception is an important mediator of the indirect effects of professional reviews on adoption (VAF = 86.36%). In addition, among the indirect effects of manufacturer reviews on adoption, perceived value also has a significant mediating effect (VAF=75.59%). In other words, perceived value played an increasingly important mediating key in the indirect effect of manufacturer reviews on general acceptance compared with risk perception. However, we find that the relationship between amateur reviews or professional reviews and adoption is not moderated by perceived value and that the relationship between amateur reviews and adoption is not moderated by risk perception.