*3.2. Measures*

The questionnaire included two parts: the background information of the respondents and measurement items of the constructs in the extended TPB framework. All of the scales were drawn from existing research or official documents, and a five-point Likert format was adopted for each item.

For the institutional climate toward COVID-19 prevention on campus, a six-item scale was adapted from the Guidelines on COVID-19 Prevention and Control in Higher Education Institutes recommended by the National Health Commission and Ministry of Education of China [47]. The respondents were asked about the extent to which they agreed with statements regarding the policies, procedures and practices against COVID-19 adopted by their respective universities (1 = strongly disagree, up to 5 = strongly agree).

Three items of the attitudes toward COVID-19-preventive behaviors were revised from Cheng and Ng [48] to assess the tendency of students to see the performance of COVID-19 preventive behaviors as benefits or barriers (1 = strongly disagree, up to 5 = strongly agree).

Three items derived from Sumaedi et al. [49] were utilized to evaluate respondents' subjective norms, namely the perception of social expectations from other important people to engage in COVID-19 preventive behaviors (1 = strongly disagree, up to 5 = strongly agree).

For perceived behavioral control, three items drawn from Prasetyo et al. [22] were used to measure the students' perceptions of their degree of control over the adoption of COVID-19-preventive behaviors.

COVID-19 preventive behaviors were evaluated using seven items obtained by Liu et al. [50], based on the preventive measures officially recommended by the Chinese Center for Disease Control and Prevention. We asked respondents how often they had adopted seven different COVID-19 preventive behaviors during the epidemic period (1 = never, up to 5 = always).

For the moderator, three items of the perceived risk scale were adapted from Ma [51] to measure the respondents' judgments concerning the adverse outcomes of COVID-19. Moreover, to examine the moderating role of perceived risk in the hypothesized path model, we used the median split approach to divide the sample into two subgroups of high and low risk perception students (Md = 3.33). The high risk perception group consisted of 1270 respondents, and the low risk perception group consisted of 1816 respondents. For more precise analysis, we omitted the data from respondents on the median (*n* = 607). We coded this as a dummy variable in the data analyses (0 = low perceived risk, 1 = high perceived risk).

As Table 2 illustrates, the Cronbach's α coefficients of the six scales ranged from 0.710 to 0.942, greater than the threshold level of 0.700 [52]. The mean score of the items ranged from 3.915 to 4.528, the standard deviation varied from 0.656 to 0.972, the absolute values of skewness ranged from 0.049 to 1.857 (less than 3), and the absolute value of kurtosis ranged from 0.224 to 7.209 (less than 10), suggesting that the distribution of all the variables and items was not significantly different from normality and that follow-up data analyses could be performed [53].


#### **Table 2.** Scale items and descriptive statistics.

**Table 2.** *Cont.*


Note: (R) = reversed item; SD = standard deviation.

#### *3.3. Data Analysis*

The hypothesized relationships in the proposed model were examined through structural equation modeling (SEM) based on the maximum likelihood estimation method. The analysis adopted the two-step approach advocated by Anderson and Gerbing [54], namely measurement model evaluation followed by structural model evaluation. The indexes that detected the goodness of fit of the model included the goodness of fit index (GFI ≥ 0.90), comparative fix index (CFI ≥ 0.90), incremental fit index (IFI ≥ 0.90), Tucker–Lewis index (TLI ≥ 0.90), standardized root mean square residual (SRMR < 0.08), root mean square error of approximation (RMSEA < 0.08) and ratio of the chi-square to the degree of freedom (*χ*2/*df* ≤ 5). As *<sup>χ</sup>*2/*df* was vulnerable to the sample size, when all 3693 responses were used, the other fit indexes mentioned above may have reflected the model fit more correctly [55]. We utilized the bootstrapping procedure with 2000 bootstrap samples to obtain bias-corrected estimates of the indirect effects of the institutional climate on preventive behavior (via attitudes, subjective norms and perceived behavioral control) and their associated 95% confidence intervals (CIs). The 95% bias-corrected bootstrap CI excluded zero, suggesting a significant mediation effect. The bootstrapping method has been found to be a more accurate test of mediation effects than other available strategies such as the Sobel test, as it enabled us to prevent type I errors that might have occurred from non-normal distributions of the mediation effects [56]. Furthermore, multigroup SEM analysis was performed to investigate the moderating effect of the perceived risk, which is regarded as a more statistically effective and powerful approach to examine structural invariance [57]. All the aforementioned analyses were conducted using the Amos 23 statistical package.

#### **4. Results**
