Confirmatory Factor Analysis and Assessment of Common Method Variance
Following Anderson and Gerbing’s [37
] two-step structural equation modelling procedure, we established discriminant validity among the variables prior to testing the hypotheses. A confirmatory factor analysis (CFA) was therefore conducted with the maximum likelihood estimation procedure of Mplus, version 7.11 [38
] on the three variables examined: economic stress, organizational orientation to employee welfare and workplace bullying. To evaluate the model fit, we considered model chi-square (the higher the values, the worse the model’s correspondence to the data) and used both absolute and incremental fit indexes. Absolute fit indexes evaluate how well a priori model reproduces the sample data. In our study, we focused on three absolute fit indexes: the standardized root-mean-square residual (SRMR), for which values of less than 0.08 are favorable, the root-mean-square error of approximation (RMSEA), which should be less than 0.10 [39
], and the Akaike’s information criterion (AIC), for which lower values indicate a better fitting model [40
]. Incremental fit indexes measure the proportionate amount of improvement in fit when a target model is compared with a more restricted, nested baseline model [41
]. We considered the comparative fit index (CFI), for which values of 0.90 or greater are recommended [41
]. As expected, the hypothesized 3-factor model yielded a good fit to the data: χ2
(17) = 156.90, CFI = 0.92; RMSEA = 0.09; SRMR = 0.05, AIC = 15148.79. Additionally, as shown in Table 1
, this model had a significantly better fit than alternative, more parsimonious models (p
< 0.01), providing evidence of the distinctiveness of the study variables.
However, because data were collected at the same time with self-report scales, common method bias problems may arise and inflate the patterns of relationships among the study variables. Following Podsakoff, Mackenzie, Lee and Podsakoff’s [42
] statistical recommendations, we therefore used the unmeasured latent method factor approach to control for the effects of common method variance. This approach was chosen because it does not require specifying the source of method bias, and it controls for any systematic variance among the items that is independent of the covariance because of the constructs of interest [42
]. Thus, this technique is particularly recommended when the specific source of the method bias is unknown or cannot be measured [40
], as in our study. Accordingly, a common method factor was added to the hypothesized three-factor model, to assess the potential increase in model fit that would be obtained from accounting for the unmeasured method factor. The model provided a better fit to the data than the same model without the method factor: χ2
(9) = 8.88, CFI = 1.00; RMSEA = 0.00; SRMR = 0.01; AIC = 15016.7; Δ χ2
(8) = 148.02, p
< 0.01. Nonetheless, the method factor accounted for 17% of total variance, which is below the average portion of variance (26%) reported in self-report studies [39
]. We can therefore conclude that common method bias is unlikely to be a serious threat in our study.3.2. Hypothesis Testing
displays descriptive statistics, correlations and reliability coefficients for the study variables. As can be seen, the correlation between economic stress and organizational orientation to employee welfare was negative and significant (p
< 0.01). In addition, organizational orientation to employee welfare negatively and significantly correlated with workplace bullying (p
< 0.01). These results provide preliminary support to our hypotheses.
In order to examine the hypothesized theoretical model, we performed Structural Equation Modelling (SEM). SEM offers the advantage of (a) controlling for measurement errors when the relationships among variables are analyzed, and (b) comparing the goodness-of-fit of the hypothesized model with that of alternative models. Accordingly, we tested our proposed structural model and compared it with alternative models [43
]. Additionally, when conducting SEM analyses, we controlled for the effects of gender on both the mediator and the dependents variable. Fit indexes for each tested model are presented in Table 3
The hypothesized model (Model 1), which is a fully mediated model, displayed a good fit to the data: χ2 (24) = 190.54, CFI = 0.91; RMSEA = 0.08; SRMR = 0.05; AIC = 15127.31. Specific inspection of direct relationships further revealed that economic stress was negatively associated with organizational orientation to employee welfare (β = −0.43, p < 0.01), and that organizational orientation to employee welfare in turn was negatively related to workplace bullying (β = −0.42, p < 0.01).
To assess whether the hypothesized model was the best representation of the data, we then compared its fit to that of alternative models. First, we assessed a partially mediated model, which included an additional direct path from economic stress to workplace bullying. This model yielded an adequate fit to the data, but it was not significantly better than Model 1, as revealed by the chi-square difference test (Δ χ2 = 0.0.01, ns). Moreover, the additional direct link between economic stress and workplace bullying was not significant (β = 0.01, ns). These findings hence supported the mediation model (Hypotheses 2 and 3) rather than a direct association between economic stress and workplace bullying (Hypothesis 1).
Next, we compared the hypothesized model against a non-mediated model (Model 3), which only included direct relationships of economic stress and organizational orientation to employee welfare with workplace bullying. Because this model had the same degrees of freedom as the hypothesized model, the statistical significance of the chi-square difference could not be calculated. Accordingly, we used the AIC, instead of the chi-square, to compare the two models. As recommended by Burnham and Anderson [44
], a model is considered as superior if its AIC is lower than another model’s AIC by four or more units. Results showed that the non-mediated model had an AIC more than four units higher than the AIC of the fully-mediated model (Δ AIC = 21.99), suggesting that the former was a significantly poorer fit to the data than the latter.
Finally, Model 1 was compared against a set of alternative models that specified all the possible reverse indirect relationships among the study variables, namely: economic stress → workplace bullying → organizational orientation to employee welfare (Model 4); workplace bullying → organizational orientation to employee welfare → economic stress (Model 5); workplace bullying → economic stress → organizational orientation to employee welfare (Model 6); organizational orientation to employee welfare → economic stress → workplace bullying (Model 7); organizational orientation to employee welfare → workplace bullying → economic stress (Model 8). Again, these models had the same degrees of freedom as the hypothesized model. We therefore compare model fit using the AIC difference test. Results suggested that Models 4, 6, 7 and 8 were not a good fit to the data, as indicated by the significant AIC difference from the hypothesized model, as well as by the poor fit indexes. In contrast, Model 5, which specified a reverse indirect path from workplace bullying to economic stress through organizational welfare, provided an acceptable fit: χ2 (24) = 194.83, CFI = 0.91; RMSEA = 0.09; SRMR = 0.05; AIC = 15131.60. However, this model had an AIC more than four units higher than the fully-mediation model (Δ AIC = 4.30), indicating that the latter was a significantly better fit to the data.
Overall, results from model comparison suggested that Model 1 was the best fitting and most parsimonious model. We therefore retained the hypothesized fully mediated model. In order to assess whether the indirect relationship between economic stress and workplace bullying through organizational orientation to employee welfare was significant (Hypothesis 4), we calculated 95% bootstrapping confidence intervals [45
]. Based on 5000 bootstrap replications, results indicated that economic stress had an indirect positive effect on workplace bullying via organizational orientation to employee welfare (indirect effect = 0.18; 95% CI = 0.12, 0.24). Hypothesis 4 was therefore supported.