Sample and procedure
A questionnaire was sent by post to all (n = 1010) employees (including those on the sick list and on leave of absence) at three different regional organizations of the Swedish National Labour Market Administration (AMV). A majority were working as employment officers in different local employment agencies. The mean age was 48.7 years, ranging from 25 to 65 years. Of the 1010 employees, 602 (59.6%) were women and 408 (40.4%) were men. In all, 792 subjects (78.4% of the total population) responded to the questionnaire.
Two years after the first questionnaire was sent a follow-up questionnaire was distributed to those employees who had responded to the first questionnaire, including those who had left the organization during the period (due to turnover or retirement). The respondents who had retired between the baseline and follow-up (n = 15) were excluded from the subsequent analysis. The follow up questionnaire was answered by 662 subjects, 65.5% (disregarding the retiring subjects who had retired).
The final study population consisted of the subjects who had responded to the first and second-wave questionnaire and who had not retired during the study period (n = 662). The mean age of the final study population was 49.4, ranging from 27 to 64 years: 401 (60.6%) were women and 261 (39.4%) were men.
The Ethics Committee at Linköping University approved the study.
Measures
Demographical variables
Sex and age (at baseline) were used as demographical variables in the subsequent analyses, as earlier studies report that health, in particular physical health, is strongly associated with health, and that women tend to report poorer health than men (see for example, Sullivan and Karlsson, [24]).
Perceived organizational justice
The individual experience of justice was measured by three different self-assessment instruments. Distributive justice was measured by a five-item instrument [25] The response scale was a five-point Likert scale (1 = very fair, 5 very unfair) (example item: "How fair has the organization
1 been in rewarding you when you consider the responsibilities you have?"). Procedural justice was measured by a four-item instrument [26]. The response scale was a five-point Likert scale (1 = strongly disagree, 5 = strongly agree) (example item: "The organization went about deciding to reorganize
2 in a way that was not fair to me"). Interactional justice was measured by six items and five-point Likert response scales (1 = strongly disagree, 5 = strongly agree) [27] (example item: "Your supervisor considered your viewpoint"). The Swedish versions of the three instruments have been used in earlier studies and have showed internal-consistency reliability coefficients (Cronbach's alpha) above .85 [28].
Turnover intentions
Turnover intentions at baseline were measured using the exit subscale from a modified EVLN-typology instrument [19]. In addition to the original validation of the instrument, performed by Hagedoorn, van Yperen, van der Vliert and Buunk [19], the psychometric properties of the Swedish version of the instrument have been tested [28]. The internal consistency was high (Cronbach's alpha: .90) and there was a strong association between exit behavioural response and actual exit behavior, indicating a high degree of predictive validity. The subscale used consists of 6 items (Initial statement: "Would you indicate how likely it is that you would react to problematic events [at work] in the described ways" example item: "Consider possibilities to change job"). The response scale was a seven-point Likert scale (ranging from "definitely not" to "definitely yes").
Job mobility
Information about actual turnover behaviour was provided from the organizations where the respondents were employed. Job mobility was coded as 1: non-mobile (still at original employment), 2: internal mobile (changing workplace but still within the organization) and 3: external mobile (changing workplace and organization).
Health
Overall self-rated health was measured using the SF-36 [29]. The SF-36 is a 36-item instrument measuring eight different health concepts: physical functioning (PF), role limitations due to physical problems (RP), bodily pain (BP), general health perceptions (GH), vitality (VT), social functioning (SF), role limitations due to emotional problems (RE), and general mental health (MH). The first four dimensions are considered as primarily measuring physical aspects of health and the remaining four scales measure mental or psychosocial aspects of health [30]. All scales range from 0 (worst) to 100 (best). A detailed description of items, score derivation, translation and validation for the SF-36 scales is found in Sullivan, Karlsson and Ware [30].
Burnout
The degree of burnout was measured by the Copenhagen Burnout Inventory, CBI, [31]. The inventory consists of three scales measuring different dimensions of burnout: personal burnout (six items), work-related burnout (seven items) and client-related burnout (six items). All items have five response alternatives ranging from 'always/very high degree' (coded as '100') to 'never/very low degree' (coded as '0') with the intervening alternatives coded as '75', '50' and '25', (example item: "How often do you feel tired?"). A summary score for each response dimension was calculated as the average value of the individual item scores. A high score indicates a high degree of burnout.
Statistical analyses
As a first step in the analysis, the distribution of turnover intentions, perceived organizational justice, job mobility and self-rated health and burnout at baseline were analyzed in relation to the demographical variables sex and age, using t-test and ANOVA (corrected for multiple comparisons with Bonferroni correction). As earlier studies have shown that both organizational justice and turnover intentions have a distinct and clear relationship with health, burnout and job mobility, turnover intentions and perceived organizational justice were used as independent or exogenous variables in the tested model.
Secondly, correlations (Spearman's rank correlation coefficients) between sex, age, turnover intentions, job mobility, self-rated health and burnout were computed.
As a third step in the analysis, a structural equation model, SEM, was formulated and tested. The model used age, distributive, procedural and interactional organizational justice (at baseline), turnover intentions (at baseline) self-rated health (at baseline), burnout (at baseline) and job mobility as exogenous variables. The endogenous variables in the analysis consisted of self-rated health (at baseline and follow-up), burnout (at baseline and follow-up) and job mobility.
Self-rated health, at baseline and follow-up, was measured by two different latent variables: physical health indicated by the SF-36 variables PF, RP, BP and GH: and psychosocial health, indicated by the SF-36 variables VT, SF, RE and MH. Sex was deleted in this analysis due to its two-category response format since binary data can be difficult to analyze with SEM [32], requiring either very large sample sizes for asymptotic least squares or integration of the multivariate normal distribution over as many dimensions as there are relatives in the pedigree [33]. Longitudinal relationships between the same variable and correlations between residual variables for health, justice and burnout were inserted in the model.
Incomplete data was handled by using the maximum likelihood estimation approach, i.e. treatment of missing data assumption of multivariate normality, based on the direct maximation of the likelihood of the observed data. This approach has numerous advantages over other methods to treat missing data as listwise or pairwise deletion. Firstly, the ML estimation is theory based and not, as many other methods, ad-hoc solutions. Secondly, where the unobserved values are missing completely at random the deletion approach is consistent but not efficient (in the statistical sense): the ML approach is both consistent and efficient. Where the observed values are only missing at random, deletion estimates could be biased, ML estimates are asymptotically unbiased [32]. Root Mean Square Error of Approximation, RMSEA, was calculated for the tested model. Based on recommendations from Bentler [34] and Marsch, Balla, & Hau [35], the RMSEA were complemented with three relative goodness-of-fit indices: the Non-Normed Fit Index (NNFI), the Incremental Fit Index (IFI), and the Comparative Fit Index (CFI). Values of .90 or higher are considered to indicate a good fit for the relative indices [36]. Byrne [32] proposed an RMSEA of ≤ .05 for good model fit, Hu and Bentler [37] advocate a ≤ .06 limit, and Browne and Cudeck [38] a ≤ .08 limit for acceptable fit.
SPSS version 14.0 and AMOS version 6.0 were used for the statistical analyses.