Emotional intelligence, emotional labor, and job satisfaction among physicians in Greece

The sample included 130 physicians, 80 males and 50 females, working at the University Hospital of Patras. Participants were administered a series of self-report questionnaires. The study protocol was approved by the Hospital Research Ethics Committee (119/25.11.10) and participation was voluntary. Response rate was 86.7%.

JS was assessed using the Brayfield & Rothe General Index of Job Satisfaction (e.g. “I am satisfied with my job for the time being”) [45]. The scale comprises 18 items (Cronbach’s alpha equal to .947), answered on a 5-point Likert scale (1=totally disagree to 5= totally agree).

Surface Acting of EL was assessed using the first five items of the Dutch Questionnaire on Emotional Labor D-QEL (e.g. I put on a “mask” in order to express the right emotions for my job) [46] (Cronbach’s alpha equal to .846). All items were answered on a 5-point Likert scale ranging from 1 (totally disagree) to 5 (totally agree).

Since our study was cross-sectional, common method bias could provide an alternative explanation for the correlations observed between measures of different constructs (e.g., among SEA, OEA, UOE, ROE, EL, and JS). Common method bias refers to variance that is attributable to the measurement methods used (e.g., common forms of data collection such as self-report questionnaires, common scale types, similar scale anchors and response format, item content overlap, common rater effects), rather than to the constructs the measures represent [47]. Common method bias is a main source of systematic measurement error. In order to rule out the possibility of common assessment method bias, two statistical techniques were applied: Harman’s single-factor test and the single-method-factor technique [47]. In the former case, all observed variables/ items from all the constructs in the study were included into an exploratory factor analysis (using unrotated principal component analysis as well as principal component analysis with varimax rotation) to determine whether the majority of the variance in the variables could be accounted for by one general factor. If a single factor emerges or one general factor accounts for most of the covariance among the variables, then a significant common method variance effect is present. In the latter case, items were allowed to load on their theoretical constructs, as well as on a latent common methods variance factor, applying confirmatory factor analysis (CFA). The significance of the structural parameters was then examined both with and without the latent common methods variance factor in the model [47]. One of the advantages of this latter approach is that it enables the researcher to account for measurement error in variables.

In order to assess convergent validity of the constructs, the average variance extracted (AVE) was calculated. AVE equals the sum of all squared standardized factor loadings (obtained from CFA) divided by the number of items. Fornell and Larcker [48] suggested adequately convergent valid measures of each latent construct should contain less than 50% error variance (i.e., AVE should be 0.5 or above, signifying that, on average, the variance due to measurement error is less than the variance captured by an underlying factor). Construct reliability (CR) was also used as an indicator of convergent validity. CR was computed from the squared sum of factor loadings for each construct and the sum of the error variance terms for a construct. CR estimates equal to 0.7 or higher suggest good reliability [48]. Moreover, AVE was used to evaluate discriminant validity of the constructs [49]. Discriminant validity of a target factor was established if the squared interconstruct correlations associated with that factor were less than the AVE estimates corresponding to the target factor and all the other factors, suggesting that the target factor had more internal (extracted) variance than variance shared between the factors. Discriminant validity of a target factor was further established when correlations with other constructs were (in absolute value) below 0.7, providing evidence of measure distinctness.

Regarding the relationship between EI and JS, a mediated model was developed and tested, in which EI was posited to positively influence JS both directly and indirectly through EL. Subsequently, a moderated mediation model was tested [50], in which any significant mediated effects were assumed to be moderated by gender. Furthermore, a moderated model was developed and tested, in which a significant interaction between EI and EL in predicting JS was assumed. Prior to analyses, all continuous measures were mean-centered by subtracting the variable’s mean from each case’s value on that variable, whereas gender was coded 0 for men and 1 for women. Age, years at work, position, and days of duty were included as covariates in regression equations examining mediator effects.

To assess whether mediation was present in the general theoretical model, the significance of the indirect effects of EI on JS through EL was tested using the bias-corrected bootstrap confidence intervals (CIs). Bootstrapping is a non-parametric resampling method that can be extended to designs involving indirect effects. In the case of simple mediation, indirect effects equal the product of two unstandardized regression coefficients, one representing the effect of EI (independent variable) on EL (mediator), and the other representing the effect of EL (mediator) on JS (dependent variable) controlling for EL [51]. However, using the product of regression coefficients for making inferences about indirect effects, involves implicit assumption that the sampling distribution of the indirect effect is normal. There are reasons to suspect that this assumption does not hold when mediation is present [52]. Thus, bootstrapping has been recommended. To bootstrap the sampling distribution of the indirect effects, the regression coefficients are repeatedly estimated k times with bootstrap samples, each of which contains n cases randomly sampled with replacement from the original sample (that is a given case can be selected multiple times), where n is the size of the original sample. This process yields k estimates of the indirect effects of the independent variable (EI) on the dependent variable (JS). These k values of the indirect effects are then sorted from low to high, thus enabling the specification of the lower and upper bounds of the desired CI [52]. MacKinnon, Lockwood, and Williams [53] conducted simulation studies to examine the accuracy of various tests on mediation effects, and advocated the bias-corrected approach as the best way to test indirect paths in mediation analysis, when normality assumptions appear to be violated. The bias-corrected bootstrap was conducted in SPSS using PROCESS computational tool [54], generating 10,000 bootstrap samples and 95% bias-corrected CIs for indirect effects. Since the percentile bootstrap CIs can be asymmetrical because they are based on an empirical estimation of the sampling distribution of the indirect effect, a correction is applied to the percentile values of the sorted distribution of bootstrap estimates used for determining the bounds of the interval. Hence the term “bias-corrected” is derived from this adjustment made to the percentile values so that the CIs are equidistant from the point estimate.

Hierarchical multiple linear regression analysis was also performed to examine the relationships between a set of independent variables (i.e., SA) and a dependent variable (i.e., JS), controlling for the effects of demographic (e.g., age, gender), work-related variables (e.g., days of duty), and other psychological variables (i.e., EI components) on the dependent variable. Screening of the raw data before they were analyzed included detection of univariate and multivariate outliers (based on the studentized residuals and the Mahalanobis distance). A search was also conducted, focused on residuals, to check for violations of the assumptions of normality, equality of variance (homoscedasticity), and linearity. Independence of error terms and sequential correlation of adjacent errors was tested through the Durbin-Watson statistic. This test statistic can vary between 0 and 4, has an acceptable range of values from 1.50 to 2.50, with a value of 2 meaning that the residuals are uncorrelated. The presence of multicollinearity was detected through inspection of the tolerance (<.10) associated with each independent variable [55].

Regarding examination of common method bias, the results of the exploratory factor analysis revealed not a single factor but seven distinct interpretable factors with eigenvalues greater than 1. The seven factors together accounted for 60.91% of the total variance. The first (largest) factor did not account for most of the variance (23.36% in the unrotated solution and 13.65% in the solution after varimax rotation). Thus, no general factor that accounted for most of the covariance among the variables was apparent. Despite the fact that this procedure is widely used to test common method bias, it has several limitations [47]. Consequently, CFA, as a more sophisticated procedure, was employed to test the hypothesis that a single factor could account for most of the variance in our data. All items were allowed to load on their theoretical constructs, as well as on a latent common methods variance factor (Figure 1). Model fit without the latent common methods variance factor was good: root mean square error of approximation (RMSEA)= .04, non-normed fit index (NNFI)= .97, comparative fit index (CFI)= .97, standardized root mean square residual (SRMR)= .08. These fit indices are compatible with those recommended by Hu and Bentler [56] for a good fit to be present between the hypothesized model and the observed data, that is CFI> .95, NNFI> .95, SRMR< .08, and RMSEA< .06. When a latent common methods factor was added to the model, there was a significant improvement in model fit: Δχ²(33)= 74.84, p< .01 (applying the scaled difference in χ²s test for nested models [57]), CFI= .98, NNFI= .98, SRMR= .07, and RMSEA= .03. Thus, the revised model provided a significantly better fit to the data than the original solution without the common methods factor. However, convergent validity and construct reliability of the common methods factor were not supported (CR= .243, AVE= 4.2%), given that the items did not consistently represent the same latent construct and much more error remained in the items than variance explained by the latent factor structure. Only 2.5% of the standardized factor loadings were above .40 for the common methods factor. Furthermore, only 36% of the factor loadings of the manifest variables on the latent common methods factor were significant at 5% level, not satisfying the convergent validity criteria. While the results of these analyses do not preclude the possibility of common method variance, they do suggest that common method variance is not of great concern and thus is unlikely to confound the interpretations of results.

Regarding convergent validity for the other factors (i.e., apart from the common methods factor), as shown in Table 1, all constructs exhibited CR values above the conventional threshold of .70, ranging from .800 to .947. AVE criterion (≥ .50) was satisfied for all constructs. Regarding discriminant validity of the constructs, we compared the shared variances between paired factors with the average variance extracted of the individual factors. For example, it can be seen that the AVE of .51 for UOE is greater than the shared variance of .14 (i.e., correlation .37²) between UOE and ROE, as well as than the shared variance between UOE and each one of the rest of the constructs. The additional finding that the estimated correlations between the factors were not excessively high (e.g., > .70) further indicated discriminant validity.

Regarding correlations among the factors, as seen in Table 1, higher doctor EI (UOE dimension) was significantly correlated with more JS (r= .42, p< .001), while doctors who demonstrated higher levels of SA derived less JS (r= −.39, p< .001). Hypotheses 1a and 2a were supported. In addition, EI (SEA dimension) was negatively correlated with SA (r= −.20, p< .01), supporting Hypothesis 3. Correlations among the four dimensions of EI were statistically significant (Table 1).

We then proceeded to evaluate a moderated mediation model in which the (previously found significant) indirect effect of SEA on JS through SA was presumed to be moderated by gender. The statistically significant interaction between SEA and gender in the model for SA implied that the indirect effect of SEA on JS was moderated by gender (B= .521, SE= .258, p= .046) (Figure 2). The positive sign of the interaction was consistent with the interpretation that the indirect effect was larger for male physicians (effect= .3386, SE= .151) than female physicians (effect= −.217, SE= .262). Given this significant interaction, it made sense to probe the indirect effect by obtaining bootstrap confidence intervals for these conditional indirect effects. The conditional indirect effect for male physicians was significantly different from 0, at 5% level, given that this interval did not contain 0 (95% CI= .093, .681). This effect was not statistically significant for female physicians (95% CI= −.900, .099). Hypothesis 1b was supported via the indirect SEA-JS path through SA; on the contrary, Hypothesis 2b was not supported, as gender did not moderate the direct effect of SA on JS. No other significant moderated mediation was found for other EI components.

Regarding the moderating role of SA in the relationship between EI and JS, neither the interaction between ROE and SA, nor the interaction of UOE with SA were significant, but the interaction between SEA (and OEA) and SA was. The effects of SEA and OEA on JS were not significant, while their interaction with SA was significant (B= −.120, SE= .045, p= .008 and B= −.157, SE= .068, p= .023, respectively). There was also a significant negative relationship between SA and JS. Using the Johnson-Neyman technique, we found that the moderating effects of SA were significant at 5% level for any value of it greater than 4.401 for SEA, and for values below −1.567 for OEA. Moderation analysis for UOE and ROE revealed no statistically significant results.

The interaction between SEA and SA, depicted visually in Figure 3, was interpreted to mean that, among those physicians with high scores in SA (> 4.401), there was a significant effect of SEA on JS such that physicians with low SEA scores had relatively higher JS compared with those with high SEA scores. Regarding the interaction between OEA and SA, this was interpreted as indicating that among those physicians with low SA scores, there was a significant effect of OEA on JS such that physicians with low OEA scores had relatively low JS compared with those with high OEA scores. Hypothesis 4b was supported.

Hierarchical- sequential multiple regression analysis was conducted to find out the significant predictors of JS, after controlling for the effects of demographic and work-related variables. Four steps were considered: the demographic variables (e.g., age, gender) were entered first as independent variables, the work-related variables (e.g., days of duty) were added in the second step, the EI variables were entered afterwards, while SA was entered in the fourth step.

Table 2 shows that in the model including SEA as an independent variable, the only variable that appeared to have unstandardized regression coefficients significantly different from zero in predicting JS was SA. The negative coefficient for the SA variable (B= −1.138) means that survey physicians who have high SA scores have lower JS, controlling for the other independent variables. The model R ² when only the demographic variables were in the model was equal to .027. There was a statistically significant improvement (R ² change= .182) in the relationship between the set of independent variables and the dependent variable, when the SA variable was included. The proportion of variance in the dependent variable explained by the independent variables was 23.2%. The Durbin-Watson statistic was equal to 2.00, a value within the acceptable range of values from 1.50 to 2.50. Multicollinearity was not detected, as none of the independent variables had a small (<.10) tolerance value.

Table 3 shows that in the model including UOE as an independent variable, the only variables that appeared to have unstandardized regression coefficients significantly different from zero in predicting JS were SA and UOE. The positive coefficient for the UOE variable (B= .894) means that survey physicians who have high UOE scores have high JS as well, controlling for the other independent variables. The model R ² when only the demographic variables were in the model was equal to .027. There was a statistically significant improvement (R ² change=.165) in the relationship between the set of independent variables and the dependent variable, when the SA variable was included. The proportion of variance in the dependent variable explained by the independent variables was 33.9%. The Durbin-Watson statistic was equal to 1.84, a value within the acceptable range of values from 1.50 to 2.50. Multicollinearity was not detected, as none of the independent variables had a small (< .10) tolerance value. In addition, it should be noted that SA was the only significant predictor of JS, when the other dimensions of EI, namely OEA and ROE which both had positive regression coefficients in the equation, were introduced as independent variables. Thus, Hypothesis 2c was supported.