**3. Hypothesized Models**

The present study aims to explore whether the commitment to learning sustainability plays a mediating role in explaining the intention to learn sustainability. As such, three models are developed: the original model of the TPB (i.e., Model A) and two extended models. In Model A, the three antecedents (i.e., attitudes toward learning sustainability, subjective norms, and perceived behavioral control) are proposed to influence the intention to learn sustainability (i.e., Hypotheses 1–3). In the first extended model (i.e., Model B), in addition to the above three hypotheses, it is further posited that attitudes toward learning sustainability (Hypothesis 4), subjective norms (Hypothesis 5), and perceived behavioral control (Hypothesis 6) are positively related to the commitment to learning sustainability, which in turn is positively related to the intention to learn sustainability (Hypothesis 7). In the second extended model (i.e., Model C), Hypotheses 1–3 are deleted while retaining the four new hypotheses in Model B. By comparing these a priori models with an empirical research design, it is possible to determine whether the inclusion of the variable "commitment" is suitable for explaining behavioral intentions. The hypotheses are listed below:

**Hypothesis 1 (H1).** *Attitudes toward the learning of sustainability are positively related to the intention to learn sustainability.*

**Hypothesis 2 (H2).** *Subjective norms are positively related to the intention to learn sustainability.*

**Hypothesis 3 (H3).** *Perceived behavioral control is positively related to the intention to learn sustainability.*

**Hypothesis 4 (H4).** *Attitudes toward learning sustainability are positively related to the commitment to learning sustainability.*

**Hypothesis 5 (H5).** *Subjective norms are positively related to the commitment to learn sustainability.*

**Hypothesis 6 (H6).** *Perceived behavioral control is positively related to the commitment to learning sustainability.*

**Hypothesis 7 (H7).** *The commitment to learning sustainability is positively related to the intention to learn sustainability.*

#### **4. Research Method**

#### *4.1. Participants and Procedure*

This is a quantitative study. An online survey was conducted and a self-report questionnaire in Chinese was designed to collect quantitative data from a sample of students in a public junior secondary school in Huizhou City, China. The experience of these students in learning sustainability in geography was the focus of this study because geography is one of the main subjects involved in teaching sustainability. As stated in the China's new Geography curriculum standards, the essence of the compulsory geography education is to understand the geographical environment and form geographical skills and SD concepts [31]. Thus, SE has become part of the standard geography curriculum. Furthermore, Huizhou City was selected because it is not a first-tier city and is generally considered to have lower educational performance than first-tier cities, such as Beijing, Shanghai, and Guangzhou. Motivating students' enthusiasm for learning is one of the core missions of teachers across the city. Therefore, this research may help to explore the issues that hinder students' acquisition of sustainability knowledge. Recommendations can then be made to increase students' intentions to learn sustainability.

The questionnaire was divided into two parts. The first part collected personal information (demographic background), such as gender, years of working, teaching subject, etc. The second part collected respondents' perceptions of the target variables (those in the later section of Measures). Ethical clearance was obtained. A total of 181 valid responses were received from a sampling frame of 259 students, representing a response rate of approximately 70%. An analysis of respondents' demographics reveals that approximately 50.3% were male students (*n* = 91) and approximately 49.7% were female students (*n* = 90). Their ages ranged from 12 to 16 (mean = 13.22). Respondents were asked about the number of subjects, other than geography, that were considered having elements of sustainability. The results were quite diverse. Seventy-seven respondents reported 1 subject, followed by 2 subjects (*n* = 68), 3 subjects (*n* = 14), and 4 subjects (*n* = 22). Specifically, most students (*n* = 156) studied sustainability from biology, followed by physics (*n* = 93), social studies (*n* = 58), and chemistry (*n* = 36). To test whether gender, age, the number of subjects with sustainability (i.e., their accumulated experience), and the subjects they taught acted as extraneous variables [41], t-statistic and correlation tests were performed. The results indicate that they were independent of the latent variables, with the exception of the number of subjects with sustainability that was modestly related to subjective norms (r = 0.16, *p* < 0.05) and perceived behavioral control (r = 0.18, *p* < 0.05). Therefore, all demographic variables were not included in further analysis.

#### *4.2. Measures*

There are five latent variables in this study. The items that measured these variables were mainly adapted from Ajzen [42] and are listed in Appendix A. Their measures were described as follows:


#### *4.3. Statistical Analysis*

The PLS-SEM was employed to examine both measurement and structural models (see Appendix B for Model A with both latent variables and corresponding measurable items). In the measurement model, the relationship between a latent variable and its respective measurable items was proposed. This involves testing the reliability and validity of the measure, including composite reliability, Cronbach alpha reliability and convergent validity [23].

In the structural model, the relationship between two latent variables (i.e., the relationship between an independent variable and a dependent variable) was proposed. The survey results were interpreted by (1) the adjusted R<sup>2</sup> contribution of all independent variables that explained the variance of their respective dependent variable and (2) the beta coefficient (β) of each independent variable that explained the variance of its respective dependent variable. Compared to the R2 value, the adjusted R2 value is more suitable for comparing various models with the same dependent variable because the adjusted value corrects for the expansion in R<sup>2</sup> coefficients caused by non-significant independent variables in each latent variable block [23,24]. For the test of the structural model, the method suggested by Kock [24] was employed, which will be described in the next section.

#### **5. Results**

Table 1 shows the mean scores, standard deviations, and correlations for the five latent variables. The mean scores of the variables indicate that all variables were positively rated, with mean scores ranging from 4.67 to 5.38 (out of a seven-point scale). The standard deviations of the variables also indicate that the subject scores for each variable tended to be quite close to the mean score. Finally, the table shows that the latent variables were all significantly correlated. Therefore, the hypotheses are worth examining.


**Table 1.** Means, standard deviations, and correlations for the five latent variables.

Notes: AT = attitudes toward the learning of sustainability; SN = subjective norms; PBC = perceived behavioral control; COM = commitment to learning sustainability; INT = intention to learn sustainability; numbers in parentheses are square roots of average variances extracted; \*\*\* *p* < 0.001.

#### *5.1. Test of the Measurement Model*

Measurement biases were assessed through the test of reliability, convergent validity, and discriminant validity of the reflective measures of the latent variables [23]. Table 2 presents the results. First, the internal consistency of the latent variables was good because their composite reliability values ranged from 0.759 to 0.927 and Cronbach's alpha values ranged from 0.758 to 0.928, both of which were above the threshold of 0.7 [24]. Second, the convergent validity of all latent variables appeared to be sufficient because (1) their AVE values were between 0.613 and 0.762, exceeding the threshold of 0.50 and (2) each item has a structure loading above 0.7 for its respective latent variable [23]. Third, the discriminant validity of all latent variables was confirmed by meeting the Fornell–Larcker criterion; that is, the square root of AVE of each latent variable was higher than the correlation coefficients between this latent variable and other latent variables (see Table 1) [23].


**Table 2.** Results for assessing the measurement model.

Notes: AT = attitudes toward the learning of sustainability; SN = subjective norms; PBC = perceived behavioral control; COM = commitment to learning sustainability; INT = intention to learn sustainability; AVE = average variance extracted.

#### *5.2. Test of the Three Structural Models*

Structural models were assessed by means of the full collinearity test, output model fit, coefficient of determination (R2) for each dependent variable, and the standardized beta coefficient (β) for each hypothesized relationship [23,24]. To test for multicollinearity (also known as full collinearity) among the latent variables in the three structural models, this study employed the full collinearity VIF (FCVIF), which could also be used to assess common method biases. The FCVIF identifies both vertical and lateral collinearity involving all latent variables in a structural model, thereby outperforming the "classic" VIF that considers only vertical collinearity [24]. For minor multicollinearity and common method biases, the FCVIF value of a variable should be less than 3.3 for regression-based models and less than 5 for models incorporating measurement errors, such as factor-based PLS-SEM models [43], while this threshold could also be relaxed to 10 for highly correlated variables [24]. Tables 3–5 show that the FCVIF values were all less than 5, except for commitment in models B and C where the value was slightly larger than 5, suggesting that both multicollinearity and common method biases were trivial. For readers' information, the "classic" or vertical collinearity VIF values for the latent variables in the three models ranged from 1.614 to 4.275 (where only one of them was higher than 3.3), all below the threshold of 5 for factor-based PLS-SEM [24].

**Table 3.** Results for assessing the structural model A.


Notes: Numbers in parentheses are FCVIF values. AT = attitudes toward the learning of sustainability; SN = subjective norms; PBC = perceived behavioral control; INT = intention to learn sustainability; β = beta coefficient; R2 = coefficient of determination; FCVIF = full collinearity variance inflation factor. \*\*\* *p* < 0.001; \*\* *p* < 0.01; \* *p* < 0.05.


**Table 4.** Results for assessing the structural model B.

Notes: Numbers in parentheses are FCVIF values. AT = attitudes toward the learning of sustainability; SN = subjective norms; PBC = perceived behavioral control; COM = commitment to learning sustainability; INT = intention to learn sustainability; β = beta coefficient; R<sup>2</sup> = coefficient of determination; FCVIF = full collinearity variance inflation factor. \*\*\* *p* < 0.001; \*\* *p* < 0.01.

**Table 5.** Results for assessing the structural model C.


Notes: Numbers in parentheses are FCVIF values. AT = attitudes toward the learning of sustainability; SN = subjective norms; PBC = perceived behavioral control; COM = commitment to learning sustainability; INT = intention to learn sustainability; β = beta coefficient; R<sup>2</sup> = coefficient of determination; FCVIF = full collinearity variance inflation factor. \*\*\* *p* < 0.001; \*\* *p* < 0.01.

Tables 3–5 show the test results for the three structural models (A, B, and C), respectively. In these models, each dependent variable was significantly explained by the corresponding independent variable(s), as indicated by their respective adjusted R<sup>2</sup> values. In a regression model, the β value indicates whether an independent variable is significantly related to a dependent variable. Since the β value is standardized, the higher the β value, the stronger the relationship between the two variables. Figures 1–3 illustrate the results of the three hypothesized models. Moreover, if the hypothesized relationship is found to be significant, a hypothesis is supported, and vice versa. In Table 3, attitudes toward learning sustainability, subjective norms, and perceived behavioral control explained 73.6% of the variance in learning intentions of sustainability, while in Tables 4 and 5, attitudes, subjective norms, and perceived behavioral control explained 79.1% and 79.7%, respectively, for the variance in commitment to learning sustainability. Furthermore, in Table 4, attitudes, subjective norms, perceived behavioral control, and the commitment to learning sustainability explained 66.4% of the variance in the intention to learn sustainability, while in Table 5, the commitment to learning sustainability explained 70.1% of the variance in the intention to learn sustainability.

#### *5.3. Comparison of the Three Structural Models*

To compare the three structural models to find the best-fit model, three indicators were used: average path coefficient (APC), average R2 (ARS), and average adjusted R2 (AARS) [24]. They measured the explanatory power of a model, and the best-fit model should have the largest values of these indicators [24]. While each of the three structural models showed a very good data fit, Model C (APC: 0.461, *p* < 0.001; ARS: 0.752, *p* < 0.001; AARS: 0.749, *p* < 0.001) outperformed Model A (APC: 0.319, *p* < 0.001; ARS: 0.740, *p* < 0.001; AARS: 0.736, *p* < 0.001) and Model B (APC: 0.281, *p* < 0.001; ARS: 0.733, *p* < 0.001; AARS: 0.728, *p* < 0.001). This shows that Model C was the best fit for the data.

**Figure 1.** The results for the original model A. \*\*\* *p* < 0.001; \*\* *p* < 0.01; \* *p* < 0.05.

**Figure 2.** The results for Model B. A solid line means a significant relationship, while a dotted line means a non-significant relationship. \*\*\* *p* < 0.001; \*\* *p* < 0.01.

**Figure 3.** The results for Model C. \*\*\* *p* < 0.001; \*\* *p* < 0.01.

Regarding the influence of each independent variable in Model C, attitudes toward the behavior (β = 0.531, *p* < 0.001), subjective norms (β = 0.303, *p* < 0.001), and perceived behavioral control (β = 0.172, *p* < 0.01) were significantly related to the commitment to learning sustainability, supporting H4, H5, and H6, respectively. The study has also found that the commitment to learning sustainability was significantly related to the intention of learning sustainability (β = 0.838, *p* < 0.001), supporting H7. Finally, as this study supports Model C, H1, H2, and H3 were removed and would not be explained.
