Characteristics of Workplace Psychosocial Resources and Risk of Diabetes: A Prospective Cohort Study

Chronic stress is suggested to increase the risk of type 2 diabetes through changes in health-related behaviors, triggering of the immune system response, and a cortisol-induced increase in glucose production in the liver combined with inhibition of insulin production in the pancreas (1–4). Several studies and meta-analyses support this hypothesis, showing that people under work stress have a 10–60% higher risk of developing type 2 diabetes (5–9).

Previous researchers have proposed behavioral, mental, and physiological pathways linking psychosocial working environment to the development of type 2 diabetes (5–9). However, from a prevention perspective, it is important to identify resources at work that can protect the health of employees and counteract the potential health effects of stressful work factors (10). The workplace provides a foundation for social and professional networks. It has been shown that people in the labor market have better physical and mental health than those outside the labor market, although this observation may be partly attributed to health selection mechanisms (11). Specifically, favorable characteristics of psychosocial resources, such as high leadership quality, procedural justice, and social support, at work are related to a lower risk of mental health problems among employees (12).

Empirical evidence of the effects of workplace psychosocial resources on long-term physical health outcomes, including diabetes, is sparse and conflicting (3,13,14). The few prospective studies were based on relatively few cases of diabetes (13,14) or focused on only a single workplace psychosocial resource (3,13,14). This limited focus on specific psychosocial resources may be problematic, as resources tend to cluster in some work teams. In addition, workplace psychosocial resources in different hierarchical domains (e.g., organizational, leadership, group levels) may commonly coexist at work, potentially affecting one another and having synergistic influences on employee health (10). Thus, large prospective studies that examine the coexistence of various workplace psychosocial resources and their association with type 2 diabetes are needed.

The current study aims were to 1) identify clusters of workplace psychosocial resources from four well-established hierarchical domains, including both horizontal and vertical dimensions, and 2) examine, in a prospective design, whether these clusters are associated with the risk of developing type 2 diabetes. Horizontal dimensions include the culture of collaboration and social support from colleagues. Vertical dimensions refer to leadership quality and organizational procedural justice.

We used data from the Finnish Public Sector (FPS) study, an ongoing dynamic cohort study with repeated questionnaire follow-ups every 2–4 years. FPS was established in 1997–1998 and comprises employees in the municipal services of 10 Finnish towns and 21 public hospitals who had a job contract for a minimum of 6 months and were aged between 18 and 65 years at the time of participation (15). We included all participants aged between 40 and 65 years at their first eligible participation from 2000 to 2014 (Fig. 1). To ascertain incident type 2 diabetes during the follow-up, all cases of prevalent type 1 or type 2 diabetes were excluded at baseline (Fig. 1). Ethical approval was obtained from the ethics committee of the Hospital District of Helsinki and Uusimaa (15).

To cover different hierarchical domains of workplace psychosocial resources, we included 1) culture of collaboration, 2) support from colleagues, 3) leadership quality, and 4) organizational procedural justice. All measurements were based on validated items with a five-point Likert scale. Cutoff values were chosen on the basis of previous research practice for categorizing into two or four groups (Supplementary Text 1).

Culture of collaboration (items from the workplace social capital scale) was measured by the mean value of two items (Cronbach α = 0.77): “Do members of the work unit build on each other’s ideas to achieve the best possible outcome?” and “People in the work unit cooperate to help develop and apply new ideas” (16). The population mean value was used as a cutoff for defining a good or poor culture of collaboration. Participants were included if they had responded to at least one of the two items.

Support from colleagues (item from Statistics Finland’s measurement on working climate [17]) was measured by one question: “Open solidarity prevails at our workplace, demonstrated by mutual willingness to help.” Those who answered somewhat agree or completely agree were identified as having support from colleagues.

Leadership quality included four aspects of leadership (three items from The Stress Profile and one item from the relational justice scale [18,19]) from a previous multicohort study by Madsen et al. (20) that measured whether a supervisor 1) cares about the feelings of employees (caring), 2) listens to subordinates’ opinions on important cases (listens), 3) rewards good work effort (appreciative), and 4) informs in good time on decisions taken and their consequences (informative). Collapsing these four items resulted in a scale with high internal consistency (Cronbach α = 0.88). Quartiles of the variable were used to indicate exposure to low, intermediately low, intermediately high, and high levels of leadership quality. The leadership variable was coded as missing if responses to at least two of the four individual items were missing.

Organizational procedural justice was identified using a modified version of Moorman’s scales (Cronbach α = 0.90) (21), defined as perceived fairness of managerial procedures (22), and asked whether procedures in the workplace are designed to 1) collect accurate information necessary for making decisions, 2) provide opportunities to appeal or challenge the decision, 3) hear the concerns of all those affected by the decision, and 4) generate standards so that decisions can be made with consistency. We divided the distribution of responses into quartiles to indicate exposure to low, intermediately low, intermediately high, and high levels of procedural justice. The variable was coded as missing if responses to at least two of the four items were missing.

All participants were linked to nationwide health and population registers by their unique personal identification numbers in Finland. We used all available information at various historical time points to capture incident diabetes. Type 2 diabetes was identified as a primary or secondary diagnosis in electronic health records by ICD-8 and ICD-9 code 250 and ICD-10 code E11 and in death registers. Participants were also linked to records of filled prescriptions (medication register) containing Anatomical Therapeutic Chemical codes. We used A10A (insulins and analogs), A10B (blood glucose–lowering drugs, excluding insulins), and A10X (other drugs used in diabetes) to identify participants with type 2 diabetes. Insulin treatment can indicate both type 1 and type 2 diabetes, but because all participants were free from diabetes at baseline at the age of ≥40 years, we assumed that the majority of individuals starting insulin treatment had type 2 diabetes. All prevalent type 1 diabetes events (e.g., ICD-8 and ICD-9 code 249, ICD-10 code E10) were excluded at baseline when insulin treatment (Anatomical Therapeutic Chemical code A10A) was applied as a criterion.

Confounders were identified using directed acyclic graphs (a formalized diagram outlining assumptions about how variables are interconnected) informed by previous literature (Supplementary Fig. 1) (23). Main confounders included age, sex (men/women), marital status (married or cohabiting, single, separated or divorced, widowed), educational level (≤9 years, 10–12 years, ≥13 years), type of employment contract (permanent/nonpermanent), preexisting comorbidities, and preexisting diagnosed mental disorders. Information on these variables (except self-reported marital status) was extracted from the national register in Finland. Preexisting comorbidities (calculated from Charlson comorbidity index) and preexisting diagnosed mental disorders were detected using ICD codes from the national patient register when the diseases were recorded before the baseline (Supplementary Table 1).

We also considered health-related behaviors, including smoking (current smoker/nonsmoker), risky alcohol consumption (yes/no), physical inactivity (yes/no), other clinical factors (i.e., BMI and symptoms of mental health problems), and other work-related factors (i.e., job demands [high/low], and occupational grade [high, medium, low]) (Supplementary Text 2). However, because of the unclear temporality between these factors and our baseline measurement of psychosocial resources, these factors were considered more as mediators than confounders and were adjusted for in a supplementary analysis but not in the main analysis.

We first conducted a latent class analysis to identify clustering of psychosocial resources among the baseline participants (i.e., the first eligible participation) (24). The classes observed at the first eligible participation were robust over time and could be extrapolated regardless of the participation year (Supplementary Fig. 2). We decided on a four-class model by combining the criteria of a smaller Bayesian information criterion value (i.e., the model fit), distribution of class membership probabilities, class sizes, and interpretability of the identified patterns (Supplementary Fig. 3) (25). A previous study suggested a categorization of vertical (relations with employers and supervisors) and horizontal (relations with coworkers) social capital, providing theoretical support for our interpretation of the latent classes (26).

Hazard ratios (HRs) were estimated using the Cox proportional hazards model, with follow-up length as the underlying timescale to examine the association between clusters of resources and type 2 diabetes. Participants were censored if they developed type 2 diabetes or died or at the end of the follow-up (31 December 2016), whichever came first. The proportional hazard assumption was checked both graphically using a log-log plot and statistically by including interaction terms between time and covariates. No obvious violation was detected. The model was first minimally adjusted for sex and age and then additionally adjusted for country of birth, marital status, educational level, type of employment contract, comorbidity score, and diagnosed mental disorders. In a supplementary analysis, the association between each individual type of resources and type 2 diabetes was assessed. To calculate the absolute effects of public health relevance, we estimated the corresponding incidence rate difference (IRD) by using the Aalen additive hazards model (27). We calculated the prevented fraction for the population (PFP) using the following equation: P_d (1 − HR) / (1 − [1 − HR][1 − P_d]), where P_d is the prevalence of the resource clusters among cases of diabetes (28). PFP estimates the proportion of a disease outcome that has been prevented as a result of the presence of a protective factor, assuming that it is a causal relation.

In sensitivity analyses, we 1) applied a 1-year washout period to address the possibility of reverse causation; 2) restricted the follow-up length to the first 4 years of follow-up to identify possible immediate effects; 3) additionally adjusted for BMI, alcohol consumption, smoking, physical activity, and symptoms of mental health problems; and 4) additionally adjusted for occupational grade and job demands. We explored the potential effect modification of age-group, sex, educational level, and occupational grade on the association between psychosocial resources and type 2 diabetes (29). Multiplicative interactions and additive interactions were tested using the Cox proportional regression and additive hazards models, respectively.

We used R package poLCA version 1.4.1 for latent class analysis, SAS 9.4, Proc phreg for the Cox proportional hazards model, and R package timereg version 1.9.3 for the Aalen additive model. Risk estimates were expressed as HR and IRD per 10,000 person-years and their 95% CIs.

We identified four latent classes of workplace resources (Fig. 2 and Supplementary Table 2). Among the 49,835 baseline participants, 24% were grouped into the unfavorable resource class, experiencing unsupportive psychosocial resources for all four resources. The favorable vertical class (29%) was characterized by intermediate to high levels of procedural justice and leadership quality and low levels of coworker support and culture of collaboration. The favorable horizontal class (18%) was characterized by high levels of coworker support and culture of collaboration and low to intermediate levels of procedural justice and leadership quality. The individuals in the favorable vertical and horizontal class (29%) reported relatively high workplace psychosocial resources across all dimensions.

Characteristics of the participants in each latent workplace resource class are presented in Table 1 and Supplementary Table 3. Resource clusters seemed to differ in terms of age, sex, educational level, marital status, holding a permanent job contract, occupational grade, and perceiving high job demands. Level of self-reported symptoms of mental health problems and BMI and proportion of diagnosed mental disorders, being physically active, drinking alcohol excessively, and smoking also varied across the resource classes.

During a mean follow-up time of 10.9 years, we identified 2,148 incident cases of type 2 diabetes (incidence 39 per 10,000 person-years) recorded among 49,835 participants who were initially diabetes free in FPS.

Figure 3 shows the relationship between baseline workplace psychosocial resources and type 2 diabetes. After adjusting for main confounders, favorable vertical (HR 0.87 [95% CI 0.78; 0.97], IRD −6 per 10,000 person-years, PFP = 4.3%), favorable horizontal (HR 0.77 [95% CI 0.67; 0.88], IRD −10 per 10,000 person-years, PFP = 4.3%), and favorable vertical and horizontal (HR 0.77 [95% CI 0.68; 0.86], IRD −11 per 10,000 person-years, PFP = 7.0%) resource classes were associated with a lower risk of developing type 2 diabetes compared with the unfavorable resource class. When analyzing each of the four types of resources individually, we found some associations between individual resources and type 2 diabetes, but they almost disappeared when adding all individual resources into the same model (Fig. 3 and Supplementary Fig. 4).

To address reverse causality arising from an underlying disease process affecting the perception of the work situation already at baseline, we excluded cases of diabetes within the 1st year of follow-up. The results remained almost unchanged (Supplementary Fig. 5). Restricting the follow-up to the first 4 years also showed similar effect estimates (Supplementary Fig. 5). Adjustments for lifestyle or work-related factors did not materially affect the magnitude of the effect, but the adjustment for self-reported symptoms of mental health problems slightly attenuated the association (Supplementary Fig. 6). There was an additive interaction between age-groups and resource (P_interaction = 0.03), showing stronger associations among employees aged ≥55 years (Supplementary Fig. 7B). No interactions were found for sex, educational level, and occupational grade on either a multiplicative or an additive scale (Supplementary Fig. 7C).

In a large longitudinal study, we identified a clear pattern of clustering of workplace psychosocial resources. While 29% of the participants experienced favorable psychosocial resources across all domains, one in four employees worked in a low-resource psychosocial working environment characterized by low levels of both horizontal and vertical psychosocial resources. The first group had a lower risk of developing type 2 diabetes, especially among older employees. The latter group was at higher risk of diabetes and more likely to experience, for example, poorer self-reported mental health at baseline. The patterns of workplace psychosocial resources were stable across study waves, and the findings appeared not to be biased by reverse causation or confounding by other factors.

To our knowledge, the current study is the first to investigate the association between clustering of workplace resources and type 2 diabetes. Our findings are in line with a previous study in 10,308 English civil servants that showed that low workplace social support was associated with a higher risk of type 2 diabetes (3). However, among 6,784 Danish health care workers with 253 incident cases of diabetes, the intermediate level of leadership quality was more protective of incident type 2 diabetes than low and high levels of leadership quality (13). Another study of 3,752 Canadian female employees with 259 incident cases of diabetes found that a low level of workplace social support was more protective regarding the development of type 2 diabetes compared with high workplace social support (14). The previous Danish and Canadian studies included far fewer incident cases of diabetes, did not consider censoring for death or migration, and might have overadjusted for mediators (e.g., self-reported depression) compared with the current study.

In light of previous studies on diabetes etiology, an effect of workplace psychosocial resources on type 2 diabetes seems plausible. Among English civil servants, relational occupational justice was associated with a 25% lower risk of developing the metabolic syndrome in men (30). Using the same data set, circulating inflammatory marker interleukin-6 was found to mediate 10% of the longitudinal association between workplace social support and onset of diabetes (3). Additionally, a Japanese study showed an association between lower levels of supervisor support and a higher risk of developing insulin resistance (4). Social support also has been found to lead to a smoother blood pressure reaction, higher oxytocin levels, and a lower level of cortisol and inflammation in response to an acute psychological stressor (31). These physiological mechanisms support the hypothesis that workplace psychosocial resources may be important in preventing type 2 diabetes.

Our descriptive results showed that favorable psychosocial resources at work correlate with better self-reported mental health. The additional adjustment of self-reported mental health slightly attenuated the association, implying a potential pathway through poor mental health. However, the association between mental health and type 2 diabetes was not confirmed in a previous multicohort study (32), leaving the pathway through mental health somewhat equivocal.

Another plausible pathway is via changes in health-related behaviors. Our study showed that health-related lifestyle tends to differ according to resource class. Stronger leadership at work could make work-related lifestyle interventions more successful (33). A longitudinal association has been found between a high level of social capital and an increased probability of smoking cessation (34). Cross-sectional evidence has shown that favorable psychosocial working conditions correlate with lower alcohol consumption and higher physical activity (35,36).

Our study adds to the existing literature by looking into the combination of workplace psychosocial resources across vertical and horizontal dimensions. It is interesting that no obvious associations were found between resources and type 2 diabetes when addressed individually, while a clear association was found when patterns of psychosocial resources and diabetes were considered together. Future research may be required to disentangle potential interactions among these resource elements.

Given the annual diabetes incidence of 39 per 10,000 people among Finnish public sector employees, the observed associations, if causal, suggest a yearly reduction of up to 10 per 10,000 people for favorable horizontal resources and a yearly reduction of 11 per 10,000 persons for favorable vertical and horizontal resources compared with unfavorable resources. The importance of favorable vertical and horizontal resources was further supported by a PFP of 7%, indicating that if our finding is causal and the favorable vertical and horizontal resources were to be replaced by the unfavorable resources, the incidence of type 2 diabetes would be 7% higher. Nevertheless, the calculation of PFP was based on the prevalence of resources and should be interpreted with caution.

Moreover, we found an interaction between age and resources on the additive (P = 0.03) but not multiplicative (P = 0.25) scale, suggesting that the absolute effect of resources on diabetes was greater among employees aged >55 years than among those who were younger. This difference between interaction measures is not surprising. When both age and resources are associated with diabetes development, absence of a multiplicative interaction implies the presence of an additive interaction (37). Additive interaction may be of better public health relevance because it estimates the number of additional diabetes cases prevented in one group compared with another. For older employees, workplace psychosocial resources may contribute to selection, optimization, and compensation strategies aimed at reducing age-related losses while keeping age-related gains (38). Previous literature has suggested health benefits as a consequence of intervention in any resource (10). The additive interaction implied that such benefits may be most pronounced for organizations with a large number of older workers.

The items on colleague support, culture of collaboration, leadership, and procedural justice were measured by standard scales or questionnaire items at the individual level. We applied a flexible data-driven approach for categorizing workplace resources, allowing us to account for measurement error while not relying on predefined characterizations (25). The latent class approach also provided class membership probabilities, which indicate the estimates’ certainty. The categorization of latent classes in our study were based on the highest class membership probability for each person. This procedure could have resulted in some degree of misclassification of the workplace resources, which may have attenuated the results slightly. Furthermore, resources were only measured at baseline. Without considering timing, duration, and changes of the exposure, the association may be underestimated. While this study included a broad range of commonly studied group-, leader-, and organizational-level resources, several important workplace resources were not covered, including control, rewards, ethical culture, predictability, autonomy, and job security. Although a high correlation might exist among resources, future studies mapping a wider range of workplace resources in relation to diabetes risks are warranted.

Cases of type 2 diabetes were ascertained using register-based data with the best available information from Finland’s certified health reimbursement. These data were independent of the exposure measurement and allowed for a virtually full follow-up. People <40 years of age were excluded to reduce the possibility that taking metformin was due to polycystic ovarian syndrome and the possibility of being treated by insulin because of type 1 diabetes. Therefore, in our study, we believe to have captured the majority of the individuals diagnosed with type 2 diabetes. However, because type 2 diabetes is more likely to be treated in primary care and to appear in the health register later than the first time of the diagnosis, we may have slightly underestimated the association between resources and type 2 diabetes. Nevertheless, because we included any health reimbursement owing to type 2 diabetes, including medication prescription and hospital admissions, this underestimation is probably minimal.

The detailed assessment of the baseline workplace psychosocial resources is a strength of this study. Because these resources tend to coexist in the real-life setting, our study provides evidence on various combinations of resources that may facilitate the development of complex multilevel workplace interventions. Our study took advantage of the large sample size with a long enough follow-up to allow for type 2 diabetes to develop. We also had sufficient power to investigate the existence of reverse causation by excluding people diagnosed with diabetes within 1 year after baseline. Furthermore, the study population was representative of Finnish public sector employees. However, caution should be taken when generalizing our results to private sector employees and cultural settings with different perceptions of workplace psychosocial resources.

In conclusion, we identified four distinct patterns of psychosocial resources at work, which tended to cluster in specific groups of individuals. Favorable horizontal and vertical psychosocial resources were associated with the lowest risk of developing type 2 diabetes. These findings suggest that measures aiming at improving workplace psychosocial resources might help to preserve the health of employees. Future studies are needed to understand the effect of the timing of exposure and the mechanistic pathways.

Acknowledgments. The authors thank Martin Claeson from the Stress Research Institute for managing the data for analysis.

Funding. This project is supported by the Arbejdstilsynet (the Danish Working Environment Research Fund) grant 13-2015-09. T.X. was supported by Forskningsrådet om Hälsa, Arbetsliv och Välfärd (the Swedish Research Council for Health, Working Life and Welfare) grant 2020-00040. R.R. was supported by Arbejdstilsynet grant 10-2016-03. R.R., J.V., and M.K. were supported by NordForsk (the Nordic Research Programme on Health and Welfare) grant 70521. J.V. was supported by Academy of Finland grants 321409 and 329240. L.L.M.H. was supported by Forskningsrådet om Hälsa, Arbetsliv och Välfärd (Swedish Research Council for Health, Working Life and Welfare) grant 2019-01318. M.K. was supported by Academy of Finland grant 329202, and Työsuojelurahasto (Finnish Work Environment Fund) grant 190424 and, outside this work, Medical Research Council grant S011676 and National Institute on Aging grant R01-AG-056477.

Duality of Interest. A.J.C. is an employee at Novo Nordisk A/S. No other potential conflicts of interest relevant to this article were reported.

Novo Nordisk A/S had no role in the study design, analyses, and results interpretation.

Author Contributions. T.X. performed the analysis, interpreted data, and drafted the manuscript. T.X., A.J.C., R.R., L.L.M.H., H.W., and N.H.R. contributed to the conception and design of the study. J.P. prepared the data set for analyses, ran the statistical codes, and provided suggestions on the revision of codes. T.L. contributed to the application of statistical methods. J.V. and M.K. contributed to the access of relevant data. All authors contributed to the critical revision of and approval to submit the manuscript. T.X. is the guarantor analyses of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.