5.1 Structural equation model
The results of the structural equation modeling with the antecedents of S-O-R are presented in Table 8. In the case of users, 20 of the 27 hypotheses were supported. First, both perceived satisfaction and perceived trust—two mediators of our model—had a positive effect on usage intention (H1, β = 0.689, t-value = 9.362, p < .001; H2, β = 0.160, t-value = 1.968, p < .05). Additionally, perceived performance risk, which had a negative impact on perceived satisfaction, perceived trust, mobility, and social image, had a positive relationship with perceived satisfaction (H3: β = 0.433, t-value = 6.451, p < .001; H4: β = 0.171, t-value = 3.147, p < .01; H8: β=-0.115, t-value = 2.793, p < .01; H15: β = 0.227, t-value = 3.820, p < .001). Mobility, environmental value, and social image appeared as preceding variables with positive effects on perceived trust, excluding perceived physical risk, which had a negative effect on perceived trust (H5, β = 0.109, t-value = 1.993, p < .05; H11, β=-0.189, t-value = 3.935, p < .001; H14, β = 0.148, t-value = 2.466, p < .05; H16, β = 0.525, t-value = 12.761, p < .001). Finally, perceived performance risk appeared to be a leading variable that positively affected perceived physical risk (H12, β = 0.303, t-value = 5.230, p < .001). The antecedents of S-O-R, government regulation, and innovativeness were positively related to mobility and environmental value (H17, β = 0.286, t-value = 3.496, p < .001; H21, β = 0.130, t-value = 2.211, p < .05; H22, β = 0.329, t-value = 4.949, p < .001; H26, β = 0.463, t-value = 8.275, p < .001). Furthermore, innovativeness had a positive impact on PEOU (H23, β = 0.386, t-value = 7.344, p < .001). Government regulation also showed positive effects on perceived performance and physical risks (H19, β = 0.274, t-value = 3.887, p < .001; H20, β = 0.395, t-value = 4.972, p < .001), respectively. Finally, innovativeness had a negative effect on perceived performance risk but a positive effect on social image (H24, β=-0.227, t-value = 3.112, p < .01; H27, β = 0.419, t-value = 7.168, p < .001).
In the case of non-users, 19 of the 27 hypotheses were supported. Unlike the user case, only perceived satisfaction had a positive effect on usage intention (H1, β = 0.729, t-value = 10.948, p < .001). Additionally, perceived trust, PEOU, and social image each had a positive relationship with perceived satisfaction (H3, β = 0.452, t-value = 5.904, p < .001; H6, β = 0.107, t-value = 2.316, p < .05; H15, β = 0.365, t-value = 5.172, p < .001). Mobility, environmental value, and social image appeared as preceding variables with positive effects on perceived trust, excluding perceived performance risk, which had a negative effect on perceived trust (H5, β = 0.158, t-value = 2.791, p < .01; H9, β=-0.139, t-value = 2.990, p < .01; H14, β = 0.121, t-value = 2.054, p < .05; H16, β = 0.591, t-value = 10.954, p < .001). Finally, perceived performance risk appeared to be a leading variable that positively affected perceived physical risk (H12, β = 0.271, t-value = 4.477, p < .001). Among the antecedents of S-O-R, government regulation had a positive relationship with mobility and environmental value (H17, β = 0.167, t-value = 2.452, p < .05; H21, β = 0.104, t-value = 1.969, p < .05). Another antecedent, innovativeness, also verified the positive associations with mobility, PEOU, environmental value, and social image (H22, β = 0.323, t-value = 3.980, p < .001; H23, β = 0.353, t-value = 4.816, p < .001; H26, β = 0.477, t-value = 7.548, p < .001; H27, β = 0.488, t-value = 7.576, p < .001). Government regulation had positive effects on perceived performance risk and perceived physical risk (H19, β = 0.221, t-value = 4.001, p < .001; H20, β = 0.405, t-value = 7.422, p < .001). However, innovativeness had a negative effect on perceived performance risk and perceived physical risk but a positive effect on social image (H24, β=-0.223, t-value = 3.347, p < .001; H25, β=-0.205, t-value = 2.980, p < .01).
(Insert Table 8 here)
5.2 Artificial neural network analysis
PLS-SEM assumes and examines the relationship between variables as a linear relationship; it has limitations in examining non-linear relationships. In this study, through the analysis of variance linear test, eight non-linear relationships out of 27 relationships were confirmed (Ooi and Tan, 2016). The artificial neural network (ANN) is a method to derive information through a process similar to that which the human brain uses to derive information from the human neural network (Chong, 2013). The ANN method was used to address the limitation in examining non-linear data to support the findings of PLS-SEM (Kalinic et al., 2019; Leong et al., 2013); however, because of their “black-box” operation, ANN is inappropriate for hypothesis testing (Wong et al., 2020; Lee et al., 2020). Thus, many scholars have utilized dual-stage analysis (Hew et al., 2017; Teo et al., 2015) to take advantage of both methods. First, PLS-SEM was performed to test the hypotheses as a linear analysis, and after path analysis, significant factors were utilized as the inputs for the ANN model, assessing the extent to which these external variables affect the outcomes (Leong et al., 2020).
Additionally, we utilized a deep ANN architecture that enables deep learning because only a single hidden layer was indicated by Huang and Stokes (2016) as a shallow type of ANN. Furthermore, Wang et al. (2017) argued that a deep ANN architecture, which can learn complex non-linear relationships through the use of two or more hidden layers, should be used instead of a shallow ANN to improve the accuracy of non-linear models. In our study, for the ANN analysis, we used IBM SPSS Statistics 26 and a 10-fold cross-validation approach to minimize overfitting, with 90% of the data used in the training phase and 10% used for testing (Hew et al., 2018). The input and hidden layers utilized multilayer perceptron and sigmoid activation functions, respectively (Sharma et al., 2019). Based on the significant path identified in the PLS-SEM analysis, three models were created for users and two models for non-users (users: A, B, C; non-users: D, E) to identify the relative importance of predictors of explained variables in the S-O-R framework.
(Insert Fig. 2 here)
The accuracy and relevance of the ANN analysis have been conventionally confirmed by low root mean square error (RMSE) values and the number of non-zero synaptic weights related to hidden neurons (Teo et al., 2015). Table 9 shows the RMSE values, which are close to zero, indicating that our models exhibit an excellent level of accuracy from the perspective of prediction. Additionally, Fig. 2 shows the non-zero synaptic weights in each ANN model, indicating that our ANN models have achieved predictive relevance (Hew et al., 2018).
(Insert Table 9 here)
Sensitivity analysis was applied to determine the input parameters that were most influential and had the greatest impact on the explained variables based on their normalized importance (Chen et al., 2020; Hew et al., 2017). The results of the sensitivity analysis are summarized in Table 10.
First, in the case of users, perceived trust showed the greatest importance for perceived satisfaction, followed by social image, mobility, and perceived performance risk. Social image was the most important predictor of perceived trust, followed by perceived physical risk, environmental value, and mobility. Finally, perceived satisfaction had a greater influence on usage intention than perceived trust.
However, in the case of non-users, perceived trust was the greatest influence on perceived satisfaction, followed by social image and PEOU. Social image was suggested to be the most powerful leading variable for perceived trust, followed by environmental value, perceived physical risk, and mobility.
(Insert Table 10 here)
(Insert Table 11 here)