In this study, we analyzed the research model and tested the hypotheses by using the partial least squares (PLS) technique for three reasons. First, the PLS method can specify and test path models with latent constructs (Gefen et al., 2000). Second, the PLS method can be used to address a small sample size (Gefen et al., 2000); this study has (N = 373). Finally, this study employed the PLS method because it is suitable for predictive applications and theory building (Shao et al., 2017; Zhen et al., 2021). In particular, we used SmartPLS version 2.0 for model validation and analyses.
5.1 Measurement model
We tested the quality of the measurement model for reliability, convergent validity, and discriminant validity (Guan and Hsu, 2020). Table 2 shows that Cronbach’s alpha values for all constructs were higher than the general criteria of 0.7, indicating that the items are reliable measures for their perspective constructs. That is, the instruments have good internal consistency reliability. All factor loadings are above 0.8, suggesting positive individual item reliability. These values indicated positive reliability for all constructs.
Furthermore, confirmatory factor analysis was used to evaluate convergent and discriminant validity. Table 2 shows that the composite reliability (CR) for all constructs exceeded 0.8, which was more than the recommended score of 0.7. The average variance extracted (AVE) for all constructs are above 0.5, indicating high reliability and adequate convergent validity. As shown in Table 3, the correlation between the construct and other constructs is lower than the square root of AVE for each construct, suggesting good discriminant validity of the measurement model. Based on these results, our measurement model has sound reliability and validity.
Table 2 Construct reliability and validity
Constructs
|
Items
|
Factor loadings
|
AVE
|
CR
|
Cronbach’s alpha
|
Knowledge
|
K1
|
0.885
|
0.836
|
0.938
|
0.902
|
K2
|
0.929
|
K3
|
0.825
|
Attitude
|
A1
|
0.894
|
0.856
|
0.947
|
0.915
|
A2
|
0.945
|
A3
|
0.935
|
Behavior
|
B1
|
0.903
|
0.789
|
0.918
|
0.866
|
B2
|
0.861
|
B3
|
0.901
|
Experience inertia
|
EI1
|
0.958
|
0.892
|
0.961
|
0.939
|
EI2
|
0.934
|
EI3
|
0.942
|
Learning inertia
|
LI1
|
0.921
|
0.884
|
0.958
|
0.934
|
LI2
|
0.954
|
LI3
|
0.946
|
Information security awareness
|
ISA1
|
0.941
|
0.863
|
0.962
|
0.947
|
ISA2
|
0.938
|
ISA3
|
0.911
|
ISA4
|
0.925
|
Table 3 Correlation analysis of latent variables
|
Mean
|
Std.
|
K
|
A
|
B
|
EI
|
LI
|
ISA
|
K
|
3.551
|
0.771
|
0.914
|
|
|
|
|
|
A
|
3.691
|
0.809
|
0.711
|
0.925
|
|
|
|
|
B
|
3.568
|
0.750
|
0.752
|
0.806
|
0.888
|
|
|
|
EI
|
3.584
|
0.782
|
0.773
|
0.749
|
0.815
|
0.944
|
|
|
LI
|
3.607
|
0.757
|
0.723
|
0.690
|
0.802
|
0.828
|
0.940
|
|
ISA
|
3.597
|
0.772
|
0.764
|
0.721
|
0.828
|
0.819
|
0.845
|
0.929
|
Note: Values in the diagonal area and bold figures pertain to the square root of the AVE
|
5.2 Common method variance (CMV)
This study used a single questionnaire survey to collect data for all latent constructs at one point in time, which may lead to CMV. This study took two steps to detect the CMV problem. First, we used Harman’s one-factor test to conduct an exploratory factor analysis. The results indicated that the first factor only accounted for 45.036% of the total variance. Second, following Podsakoff et al. (2003) and Liang et al. (2007), we assessed CMV in PLS. Table 4 shows the result, which indicates that the proportion of variance in each observed indicator explained by its focal construct exceeded the variance explained by the method factor. Furthermore, the average substantively explained variance of the indicators was 51.2% versus 2.4% for the method constructs, suggesting that CMV was not a major concern in this study.
Table 4 CMV detection
Construct
|
Indicator
|
Substantive factor loading (R1)
|
R12
|
Method factor loading (R2)
|
R22
|
Knowledge
|
K1
|
0.732**
|
0.536
|
0.261*
|
0.068
|
K2
|
0.508**
|
0.258
|
0.106
|
0.011
|
K3
|
0.643**
|
0.413
|
0.139
|
0.019
|
Attitude
|
A1
|
0.773**
|
0.598
|
0.268*
|
0.072
|
A2
|
0.631**
|
0.398
|
0.093
|
0.008
|
A3
|
0.512**
|
0.262
|
0.126
|
0.016
|
Behavior
|
B1
|
0.819**
|
0.671
|
0.225*
|
0.051
|
B2
|
0.754**
|
0.569
|
0.229*
|
0.052
|
B3
|
0.676**
|
0.457
|
0.156
|
0.024
|
Experience Inertia
|
EI1
|
0.867**
|
0.752
|
0.028
|
0.000
|
EI2
|
0.827**
|
0.684
|
0.145
|
0.021
|
EI3
|
0.653**
|
0.426
|
0.126
|
0.016
|
Learning inertia
|
LI1
|
0.635**
|
0.403
|
0.161
|
0.026
|
LI2
|
0.696**
|
0.484
|
0.059
|
0.003
|
LI3
|
0.589**
|
0.347
|
0.097
|
0.009
|
Information security awareness
|
ISA1
|
0.876**
|
0.767
|
0.104
|
0.011
|
ISA2
|
0.741**
|
0.549
|
0.100
|
0.010
|
ISA3
|
0.825**
|
0.681
|
0.155
|
0.024
|
ISA4
|
0.687**
|
0.472
|
0.138
|
0.019
|
Average
|
|
0.708
|
0.512
|
0.143
|
0.024
|
Note: **p < 0.01; *p < 0.05
|
5.3 Correlations and multicollinearity
Table 2 indicates that the correlation values of the five inner constructs are above 0.6. Thus, we needed to compute the variance inflation factor (VIF) to eliminate any potential threat of multicollinearity. The results revealed that the highest VIF score was 4.57 (see Table 5), below 5.0 (House and Raja, 2019). Therefore, multicollinearity was not a major concern in this study.
Table 5 Results of collinearity assessment
Variables
|
Knowledge
|
Attitude
|
Behavior
|
Experience inertia
|
Learning inertia
|
VIF
|
2.86
|
3.03
|
4.57
|
4.30
|
3.61
|
5.4 Hypothesis testing
After confirming that all the measurement items have positive reliability, convergent validity, and discriminant validity, we tested the hypotheses in the research model using structural equation modeling with SmartPLS version 2.0. The results indicated that the model could explain 79.6% of the variance of employees’ ISA. Table 6 summarizes the results of the hypothesis tests.
Table 6 Summary of hypotheses and results
Hypothesis
|
Relations
|
Predicted sign
|
Supported?
|
H1
|
Knowledge → ISA
|
+
|
Yes
|
H2
|
Attitude → ISA
|
+
|
No
|
H3
|
Behavior → ISA
|
+
|
Yes
|
H4
|
Experience inertia → ISA
|
+
|
No
|
H5
|
Learning inertia → ISA
|
+
|
Yes
|
The results indicated the following findings. Knowledge was positively related to ISA, thus supporting H1 (β = 0.184, p < 0.01). Attitude has no significant effect on employees’ ISA, thus rejecting H2 (β = 0.038, p > 0.05). Behavior was positively related to ISA, thus supporting H3 (β = 0.264, p < 0.05).
The results also indicated that experience inertia has no significant effect on employees’ ISA, rejecting H4 (β = 0.070, p > 0.05), and that learning inertia was positively related to employees’ ISA, supporting H5 (β = 0.416, p < 0.001).
Three control variables exist in the research model: gender, age, and education. Considering that the number of control variables exceeded one, we conducted three tests following Liang et al. (2007). Specifically, each test only involved one control variable as an independent variable. When three control variables were included in the research model for testing, the results showed that the coefficients of the three control variables were insignificant (t values were 0.461, 0.996, and 1.235). Therefore, three control variables had no statistically significant effect on employees’ ISA.