Common method bias
Standard method bias refers to the spurious variance arising from the measurement method rather than the measured constructs. This bias can significantly affect the validity of research results, mainly when using self-reported questionnaires. To effectively address this issue, both procedural and methodological improvements were made in this study.
Procedurally, this study adopted the method proposed by Kock (Kock et al., 2021). We collected data from multiple sources, gathering information from different cities and schools to avoid over-reliance on homogeneous respondents and to reduce the potential for common method bias. To ensure the quality of the questionnaire design, we focused on clear and concise wording, avoiding leading questions or ambiguous terms to minimize response bias. Additionally, by providing confidentiality and anonymity assurances, we ensured participants’ privacy, reducing the influence of social desirability or response tendencies and encouraging them to provide truthful and objective answers.
Methodologically, we conducted Harman’s single-factor test on the constructs involved in the theoretical model (Aguirre-Urreta & Hu, 2019). We used principal component analysis with varimax rotation, and the results showed that a single factor explained 24.3% of the total variance. Since the variance explained by a single factor was less than 50%, it can be preliminarily concluded that common method bias is not a significant concern in our data. This indicates that the multidimensionality of the measurements was reasonably ensured, and the differences between the constructs in the data were reflected to some extent.
To further verify this, we also conducted confirmatory factor analysis (CFA), comparing the goodness-of-fit between the single and multi-factor models. The fit indices for the single-factor model (CFI = 0.65, TLI = 0.62, RMSEA = 0.12, SRMR = 0.08) indicated poor fit, whereas the multi-factor model showed significant improvement (CFI = 0.91, TLI = 0.89, RMSEA = 0.05, SRMR = 0.04). This further supports that common method bias had a minimal impact in this study. Additionally, we employed the marker variable technique, selecting a theoretically unrelated variable as the marker variable. The results showed that the correlations between the marker variable and the primary variables were low (all r values below 0.20), indicating that common method bias was insignificant.
In conclusion, through multiple tests and validations, we can confirm that common method bias had a minimal impact on this study’s data, ensuring the research results’ validity and reliability.
Assessing the outer measurement model
In structural equation modeling (SEM) analysis, evaluating the external measurement model is a critical step in ensuring the model’s validity. The external measurement model primarily assesses the relationship between latent variables (constructs) and their corresponding observed variables (indicators) (Hair et al., 2019). This study first evaluated the reliability of the external measurement model through reliability analysis. Specifically, Cronbach’s Alpha coefficient was used to assess the internal consistency of each dimension of the questionnaire. The results showed that the Cronbach’s Alpha values for all latent variables were above 0.7, indicating good internal consistency of the questionnaire. Additionally, the Composite Reliability (CR) for all latent variables in the study was more significant than 0.7, further indicating the high reliability of the measurement model (as shown in Table 1).
Validity analysis is another crucial aspect of evaluating the external measurement model. This study assessed validity through convergent validity and discriminant validity. Convergent validity was evaluated using the Average Variance Extracted (AVE), and the results showed that the AVE values for all latent variables were above 0.5, indicating good convergent validity. Discriminant validity was assessed using the Fornell-Larcker criterion and the cross-loadings method. According to the Fornell-Larcker criterion, the square root of the AVE for a latent variable should be greater than its correlations with other latent variables. The results showed that the square roots of the AVEs for all latent variables were more significant than their correlations with other variables, and the cross-loading test results indicated that each measurement indicator had the highest loading on its corresponding latent variable, confirming good discriminant validity (as shown in Table 2).
Finally, this study evaluated the model fit of the external measurement model. Using the Standardized Factor Loadings method, the standardized factor loadings of each observed variable on its corresponding latent variable were all above 0.6, indicating strong explanatory power of the observed variables for the latent variables. These evaluation results indicate that the reliability and validity of the measurement model are well assured, laying a solid foundation for subsequent structural model analysis. The measurement model in this study reliably and validly captures the critical drivers of university students’ flexible employment, providing robust data support.
Inspecting the inner structural model
In structural equation modeling (SEM) analysis, evaluating the internal structural model is essential for validating the relationships between latent variables. This study comprehensively assessed the internal structural model to ensure its validity and reliability. The significance and directionality of path coefficients are critical indicators for evaluating the internal structural model. The path coefficients and their significance levels obtained through PLS-SEM path analysis indicate that hypotheses H1a (IM → WLB → EI) and H1c (IM → JS → EI) are significant, suggesting that intrinsic motivation significantly influences employment intentions through work-life balance and job satisfaction, with p-values of 0.01 and 0.000 and t-values of 2.569 and 3.874, respectively, at the 99% confidence level. Hypothesis H1d (IM → SDP) is also significant, with a t-value of 10.754 and a p-value of 0.000. However, hypothesis H1b (IM → EI) failed to pass the significance test, with a p-value of 0.592, indicating that intrinsic motivation does not significantly affect employment intentions. Among the hypotheses related to extrinsic motivation, H2a (EM → WLB → EI) and H2b (EM → JS → EI) are significant, with t-values of 3.833 and 8.742 and p-values of 0.000 for both. Additionally, hypotheses H2c (EM → EI) and H2d (EM → SDP) also passed the significance test, with t-values of 2.24 and 11.08 and p-values of 0.025 and 0.000, respectively, suggesting that extrinsic motivation significantly influences employment intentions through work-life balance, job satisfaction, and self-development perception. The direction of the path coefficients aligns with theoretical expectations, except for H1b, confirming the validity of the research model (as shown in Table 3).
This study systematically assessed mediation effects using PLS-SEM path analysis to explore further the mechanisms underlying the relationships between latent variables. The mediation analysis involved several steps. First, a direct effect model, which excludes mediating variables, was constructed to evaluate the direct path coefficients and their significance between independent and dependent variables. Subsequently, mediating variables were introduced to create a full mediation model. The mediating effects were identified and analyzed by comparing the path coefficients of the direct effect model and the full mediation model. This study applied the bootstrapping technique further to verify the significance of the effects of mediation. This non-parametric resampling method involves multiple (e.g., 5000) random resamplings of the data to generate the distribution of path coefficients. The bootstrapping process calculated the indirect effect values, standard errors, and confidence intervals for mediating paths (e.g., IM → WLB → EI). If the confidence interval for the indirect effect does not include zero, the mediation effect is considered significant. Additionally, the significance and strength of the mediation effects were evaluated through the p-values of the indirect effects, the standardized path coefficients (β values), and the Variance Accounted For (VAF) value. The VAF value, which represents the proportion of the total effect explained by the mediation effect, was used to determine whether the mediation was complete (VAF > 80%), partial (20% ≤ VAF ≤ 80%), or absent (VAF < 20%). The results show that intrinsic motivation significantly mediates employment intentions through work-life balance and job satisfaction. In contrast, extrinsic motivation significantly mediates employment intentions through work-life balance, job satisfaction, and self-development perception.
Additionally, Table 4 reports the model’s explanatory power (R2 values), which reflect the proportion of variance in endogenous variables explained by exogenous variables and are a critical indicator of the model’s explanatory capacity. The results indicate that the R2 value for employment intentions is 0.375, meaning that the model explains 37.5% of the variance in employment intentions. The adjusted R2 value is 0.372, with only a slight difference from the R2 value, indicating a good model fit and high explanatory power. The R2 value for self-development perception is 0.358, with an adjusted R2 value of 0.357, demonstrating a good model fit and strong explanatory power. The R2 values for job satisfaction and work-life balance are 0.208 and 0.218, respectively, with adjusted R2 values of 0.207 and 0.216, indicating moderate explanatory power for these variables.
Furthermore, the model’s predictive relevance was assessed using the Stone-Geisser Q2 value (Table 5), calculated through the blindfolding procedure. A Q2 value greater than 0 indicates good predictive relevance. The results show that the Q2 value for employment intentions is 0.203, and the Q2 value for self-development perception is 0.187, demonstrating high predictive relevance.
Finally, the robustness of the model was evaluated through model fit indices, as reported in Table 6. The Standardized Root Mean Square Residual (SRMR) was used to measure model fit, with a value of 0.091 below the recommended threshold of 0.10, indicating good model fit. Other model fit indices, including the Chi-square value and Normed Fit Index (NFI), confirm that the model’s fit is within acceptable ranges. Overall, based on the comprehensive evaluation of path coefficient significance, mediation effects, R2 and Q2 values, and model fit indices, the internal structural model in this study demonstrates strong explanatory power, predictive relevance, and model fit. Applying the bootstrapping technique and examining mediation effects further validates the complex relationships between latent variables, providing a solid foundation for subsequent theoretical exploration and empirical analysis.
Multi-group comparison
Table 7 presents the results of a Multi-Group Analysis (MGA) conducted to examine the moderating effects of different demographic characteristics, including gender, household registration type, work experience, volunteer experience, and career planning, on the hypothesized paths within the structural model. The MGA method allows for comparing path coefficients across different subgroups, thereby enabling the identification of significant differences in the relationships between variables based on moderating factors. This approach is beneficial in assessing whether demographic characteristics influence the strength or direction of the relationships in the model. This study used Henseler’s MGA approach to evaluate the significance of the differences in path coefficients between groups.
Moderating effect of gender
The analysis revealed a significant difference in the path from external motivation (EM) to employment intention (EI) through work-life balance (WLB) between male and female groups (path coefficient difference = 0.138, p = 0.04). This suggests that external motivation (EM) has a more substantial influence on flexible employment intentions (EI) for women through work-life balance (WLB). This finding aligns with the concepts of autonomy and competence in SDT, which posits that women may be better able to balance work and life under the influence of external motivation, enhancing their intention toward flexible employment. However, no significant differences were observed in the other paths based on gender, indicating that gender-specific effects are limited to this particular pathway.
Moderating effect of household registration type (Hukou)
Household registration type (local vs. non-local students) demonstrated significant moderating effects on the paths from intrinsic motivation (IM) to employment intention (EI) (path coefficient difference = 0.388, p = 0.009) and from IM to EI through job satisfaction (JS) (path coefficient difference = 0.164, p = 0.033). These findings highlight that local students benefit more from the influence of intrinsic motivation (IM) and job satisfaction (JS) on employment intentions than non-local students. This difference may stem from socioeconomic disparities and variations in access to social capital support, as local students may receive more substantial support from their social environment, which enhances their ability to pursue flexible employment intentions. The results also reflect the role of subjective norms from the TPB, as differing social expectations based on household registration backgrounds can significantly influence employment-related behaviors.
Insignificant moderating effects of work experience, volunteer experience, and career planning
The results indicate that work experience, volunteer experience, and career planning have no significant moderating effects on any hypothesized paths. For instance, the paths from intrinsic motivation (IM) to employment intention (EI) and from external motivation (EM) to employment intention (EI) through work-life balance (WLB) did not show significant differences based on these variables. This suggests that these factors influence university students’ flexible employment intentions. However, it is worth noting that these variables may have more pronounced effects over extended periods, highlighting the potential need for longitudinal research to capture their long-term impact.
Overview of multi-group analysis results
Table 7 provides detailed results of the MGA, including path coefficient differences and p-values for each moderating variable across all hypothesized paths. Henseler’s approach was used to test the significance of these differences with a one-tailed p-value test. The findings suggest that while gender and household registration type significantly moderate specific paths, work experience, volunteer experience, and career planning are weaker moderating variables in this model.
In summary, the MGA results provide valuable insights into how demographic characteristics influence the relationships between motivational factors, mediating variables, and employment intentions. By identifying significant differences between subgroups, this analysis highlights the importance of considering demographic diversity in understanding employment-related behaviors among university students.
Artificial neural network analysis
This study employed the Artificial Neural Network (ANN) method to assess the normalized importance of antecedents of endogenous variables (Di, Chen, Shi, Cai, & Zhang, 2024). The ANN analysis utilized the significant exogenous variables identified from the PLS-SEM analysis as inputs for the neural network model. Specifically, constructs such as intrinsic motivation (IM), extrinsic motivation (EM), work-life balance (WLB), job satisfaction (JS), and sustainable development practices (SDP) were used as input variables. These constructs, identified as significant in PLS-SEM, were fed into the ANN model using a sensitivity analysis approach to explore their importance (Cai et al., 2024).
The ANN analysis used a feedforward-backpropagation algorithm with the sigmoid function as the activation function (Alnoor et al., 2024). A 10-fold cross-validation method was applied to address potential overfitting, as suggested by Aguirre-Urreta and Hu (Aguirre-Urreta & Hu, 2019). Each network was trained on 90% of the data and tested on the remaining 10%, and this process was repeated across 10 different folds. The model architecture included one hidden layer with an optimized number of neurons determined through experimental tuning, ensuring a balance between computational efficiency and model performance. The output layer consisted of a single node, representing the endogenous variable (EI).
Performance metrics, such as the Root Mean Square Error (RMSE), were calculated for the training and testing phases to evaluate the model’s accuracy and consistency. As shown in Table 8, the average RMSE for the training and testing phases was 0.2716 and 0.2733, respectively, indicating stable predictive performance. The low standard deviations of RMSE values across folds (0.0022 for training and 0.0067 for testing) further reflect the model’s robustness and the absence of significant overfitting or underfitting issues. The average testing accuracy of 76.18% demonstrates that the ANN model can make reliable predictions in practical applications.
Comparison of PLS-SEM and ANN models
While PLS-SEM and ANN models were utilized in this study to examine the relationships between exogenous and endogenous variables, their methodological differences and complementary strengths should be noted. PLS-SEM is a statistical method that estimates path coefficients, quantifying the direct effects of exogenous variables (e.g., JS, WLB, IM) on the endogenous variable (EI). In contrast, the ANN model is a machine learning approach that uses non-linear relationships and determines the relative importance of input variables through sensitivity analysis.
Table 10 highlights the consistency between the two methods. For instance, job satisfaction (JS) is identified as the most important variable in both models, with the highest path coefficient in PLS-SEM (0.469) and a normalized importance of 100% in ANN. Similarly, work-life balance (WLB) ranks second in both analyses, further validating its critical role in predicting students’ flexible employment intentions. Although variables such as intrinsic motivation (IM), extrinsic motivation (EM), and sustainable development practices (SDP) have relatively lower impacts, their importance rankings remain consistent across both models. The agreement between PLS-SEM and ANN results demonstrates the robustness of the findings and the reliability of the multi-method approach used in this study.
The sensitivity analysis conducted through ANN provides additional insights into the relative importance of variables, which complements the path coefficient analysis in PLS-SEM. This integration of methods enhances the depth of analysis and strengthens the validity of the conclusions drawn from the study.
Table 8 presents the RMSE values and accuracy across the 10-fold cross-validation for the ANN model. The consistent performance metrics indicate the model’s reliability and precision. Table 9 provides a detailed sensitivity analysis of the ANN model, showing that job satisfaction (JS) and work-life balance (WLB) are the most critical variables in predicting flexible employment intentions. Lastly, Table 10 compares the results of the PLS-SEM and ANN models, highlighting the consistency in the rankings of variable importance across both approaches.
By integrating the results of PLS-SEM and ANN models, this study validates job satisfaction and work-life balance as the primary factors influencing students’ intentions toward flexible employment. These findings underscore the importance of improving job satisfaction and promoting work-life balance to stimulate students’ engagement in flexible employment opportunities.
Fuzzy-set qualitative comparative analysis (fsQCA)
Calibration
Before conducting a fuzzy-set qualitative comparative analysis (fsQCA), it is essential to calibrate the variable data. In this study, the scores from the 7-point Likert scale were calibrated into fuzzy-set scores ranging from 0 to 1. According to Ragin (Ragin, 2008), a fuzzy score of 1 indicates full membership in a fuzzy set, while a score of 0 indicates non-membership. For the 7-point Likert scale, researchers have set different calibration thresholds. Fiss (Fiss, 2011) set “6.75” as full membership, “4” as the crossover point (indicating moderate membership), and “1.25” as non-membership. Similarly, Pappas and Woodside (Pappas, Woodside, (2021)) suggested using “6,” “4,” and “2” as the thresholds for full, moderate, and non-membership, respectively.
However, in this study, we adopted the method proposed by Ong and Johnson (Pappas, Woodside, (2021)), which uses each variable’s mean and combined standard deviation for data calibration. Specifically, scores more than one standard deviation above the mean of each variable were considered full membership, mean scores were considered moderate membership, and scores more than one standard deviation below the mean were considered non-membership. Greckhamer (Greckhamer et al., 2018) explained that the practical determination of fsQCA calibration thresholds should be “half-conceptual, half-empirical” because the thresholds should reflect “both kind and degree differences among cases”. Therefore, we used Ong and Johnson’s method to reduce the potential bias caused by the non-normal distribution of data, which can be an issue when using fixed calibration thresholds such as “6,” “4,” and “2.”
Analyzing necessity and sufficiency
Necessity analysis tests whether a particular variable is a necessary condition for an endogenous variable (Ragin, 2008). In this study, EI (Employment Intention) was set as the outcome variable for necessity analysis. At the same time, IM (Intrinsic Motivation), EM (Extrinsic Motivation), WLB (Work-Life Balance), JS (Job Satisfaction), and SDP (Sustainable Development Practices) were considered possible antecedent conditions determining university students’ flexible employment intention. Ragin (Ragin, 2008) suggests that a consistency value exceeding the threshold of 0.8 is acceptable, with a value greater than 0.8 indicating that the variable is necessary for the outcome variable. Meanwhile, coverage represents the percentage of the dataset that a particular variable explains for EI. Table 11 presents the results of the necessity and sufficiency analysis. According to the data in the table, the consistency values for IM, EM, WLB, JS, and SDP all exceed 0.6, indicating that they are effective antecedent conditions for determining EI.
Analyzing fsQCA results
This study constructed a truth table that lists all possible configurations leading to university students’ flexible employment intention (EI). Following Fiss(Fiss, 2007), we adopted a consistency threshold 0.8 for the configurations and a PRI consistency threshold 0.5. Given that the study collected 1,270 valid responses, the minimum case frequency was set at 3 to exclude combinations that only reflect specific cases (Ragin, 2008). The truth table enabled us to summarize the final configuration paths that explain the outcome variable. Three types of solutions were returned using the Quine-McCluskey algorithm: parsimonious, intermediate, and complex. However, in this study, we followed the recommendations of Chuah (Chuah et al., 2021) and only interpreted the intermediate solution to simplify the research hypotheses and achieve better interpretability. Table 12 presents the final fsQCA results, identifying each configuration’s core and peripheral conditions. In the final fsQCA results table, a black smiley face (☻) indicates a causal condition; a crossed circle (⊗) indicates the absence or negation of a condition, while a blank space indicates that the population studied has no preference regarding the presence of the condition.
Additionally, a large black smiley face represents a core condition, and a small black smiley face represents a peripheral condition (Pappas, Woodside, (2021)). According to the table, the fsQCA generated five solutions with a total consistency of 0.7798 and a total coverage of 0.7387. Thus, these five solutions explain a significant portion of EI. Pappas, Woodside, (2021) suggest that testing the predictive validity of the solutions can enhance the value of fsQCA, as it assesses the model’s ability to predict the outcome variable in different samples. We randomly divided the data sample into subsample and holdout sample to test predictive validity. The first step was to use the subsample to generate results (solutions) with the fsQCA algorithm. The second step involved modeling the combinations derived from the subsample as a variable and comparing it with the outcome variable using the holdout data. Configuration 1 involved only Job Satisfaction (JS); Configurations 2 and 3 both included Work-Life Balance (WLB) but differed in other factors; Configurations 4 and 5 emphasized Extrinsic Motivation (EM) and Work-Life Balance (WLB). By exploring different combinations of conditions to explain complex causal relationships, this study highlighted the importance of combinations of factors rather than the independent effect of a single factor. This approach provides valuable insights into understanding the complex dynamics of university students’ flexible employment intentions.
