The joint effect of unemployment and cynical hostility on all-cause mortality: results from a prospective cohort study

The Danish Longitudinal Study on Work, Unemployment, and Health is a prospective population-based study with a baseline postal survey carried out in 2000. The baseline survey was based on a stratified random sample consisting of two population groups: 1) a group of individuals 40 and 50 years old by 1 October, 1999, (response rate 69%, n = 7588), and 2) a group of 37–56-year-old individuals who had been unemployed at least 70% of the time during the period 1 October, 1996 to 1 October, 1999, (response rate 57%, n = 2287). The latter group was included to ensure a sufficient number of unemployed which was one of the main interests of the study. In 1999, the unemployment rate was low with 6% unemployed among both the 40- and 50-year-old. Both samples have initially been drawn from the ‘AKF Longitudinal Register’, which is maintained by the National Centre for Social Science Research (previously AKF) and comprises all Danes aged 15 years or older. Data on non-participation was derived from the registers and showed that non-participants were more often men, non-native-born Danes, persons living on transfer income, and persons with low educational attainment (non-trained or semi-skilled). Out of 9875 participants in the survey, we were able to link 9870 to register data. The Cause of Death Registry, The Danish National Population Register, The National Patient Registry, and the National Drug Prescription Registry were used to identify mortality and covariates. For the present study, an inclusion criterion was those in the workforce. The workforce is defined as those who are working, plus those who are available for work. From the survey, we selected individuals who were working or unemployed at the time of the survey (8733). Of these, 223 (3%) had missing information on hostility, and 100 (1%) had missing information on education. The final study population was 8426. The details of the population selection are available in Additional file 1.

Information on employment status was based on self-reported information on current labour market participation in 2000 and grouped as ‘employed’ or ‘unemployed’. Hostility was measured by the eight-item Cynical Distrust Scale, derived from the Cook-Medley Hostility Scale, originally based on the items from the Minnesota Teacher Attitude Inventory [28]. This scale measures a sceptical approach to other peoples’ sincerity, motives, and beliefs and the showing of a basic mistrust of others. This cynical mistrust dimension of hostility has been most solidly associated with measures of low social position and, consequently, we found it the most relevant measure to include in this specific study. The Cynical Distrust Scale that measured the cognitive component of hostility was factor-analytically derived from the Cook-Medley Scale by Greenglass and Julkunen [29, 30]. In two separate samples of Canadian and Finnish students, they demonstrated that the shorter Cynical Distrust Scale is a sufficiently valid, reliable, and specific measure of cynicism and distrust when compared to the full Cook-Medley Scale [29, 30]. In the present study, we used the following eight items as presented in English by Everson et al. [17]:

1) I think most people would lie to get ahead. 2) Most people inwardly dislike putting themselves out to help other people. 3) Most people make friends because friends are likely to be useful to them. 4) It is safer to trust nobody. 5) No one cares much what happens to you. 6) Most people are honest chiefly through fear of being caught. 7) I commonly wonder what hidden reasons another person may have for doing something nice to me. 8) Most people will use somewhat unfair means to gain profit or an advantage rather than lose it.

We translated the items into Danish and adjusted the wording after a back-translation into English. The Danish version was tested in a large-scale pilot study of the entire questionnaire among a study population (n = 993) drawn from the same sampling frame as the main survey [10]. This work was led by the last author of this manuscript. Response options were: completely agree, somewhat agree, somewhat disagree, completely disagree. In the analyses, the items were summed to obtain a Cynical Distrust Scale score with a range of 0–24. The internal reliability was tested by Cronbach’s alpha, which at 0.89 was satisfactory. In the literature, there is no established threshold for dichotomisation of the Cynical Distrust Scale. To define the cut point for dichotomy, we divided the scale into quartiles as has been applied in the studies by Everson et al. and Stamatakis et al. [17, 31], and we used separate cut points for men (scoring 9+) and women (scoring 7+).

To study the joint exposure to hostility and unemployment, a new composite variable with four categories was defined: 1) unexposed (reference), 2) unemployed, but not hostile, 3) hostile but not unemployed, 4) unemployed and hostile. The joint reference category allows us to compare each combination of hostility and unemployment level according to the same baseline hazard.

Potential confounders for the analyses were identified based on prior knowledge and the method of directed acyclic graphs (DAG) [32]. DAGs provide a graphical tool for identification and selection of relevant confounders in epidemiological studies based on prior substance knowledge. The DAG and overview of included variables are available in Additional file 2. Cohabitation status retrieved from the population registry in 1999 was defined as living with vs. without a partner. Educational attainment was based on self-reported information on the type of vocational training following primary school and categorised into no education (no training—only primary school) or some education (any level beyond primary school). Information on alcohol abuse and depression was retrieved from The National Patient Registry 1980–1999 and the National Drug Prescription Registry 1995–1999, which contains information about outpatient prescription drug use. We used the International Classification of Disease (ICD) to identify diagnoses. We included admissions for alcohol, cirrhosis of the liver (571.09; 571.19 ICD 8 /I85.0-I85.9 ICD 10), and Oesophageal varices (456.09; 456.19 ICD 8). Further, a prescription for alcohol dependence (ATC code N07BB) was used to identify alcohol abuse. Depression was defined as at admissions with bipolar depression, depression, and recurrent depression (ICD8: 296, 298.09, 298.19; ICD-10: F30-F33.9) and/or prescriptions with antidepressants (ATC code N06A). The variable for alcohol abuse was coded as 1 or more indicators of alcohol abuse vs. no indication of alcohol abuse. Similarly, the variable for depression was coded as 1 or more indicators of alcohol abuse vs. no indication of depression vs. no indication.

All-cause mortality was assessed in the nationwide Cause of Death Registry between baseline and 2014 and the follow-up time ended at the date of death, or 31 December 2014, whichever came first.

Cox proportional hazards models with age as the underlying time scale were used to analyse the data. All variables met the proportional hazards assumption. Initially, we estimated hazard ratios (HR) and 95% confidence intervals (CI) for all-cause mortality according to hostility, unemployment, and the joint effect of the two. Level of education and cohabitation status, depression, and alcohol abuse were included in the models and considered as possible confounding factors. By using age as the underlying time scale, we ensured thorough adjustments for age.

All analyses were conducted using SAS version 9.3. The ‘proc phreg’ procedure was used for survival analysis. The synergy index (SI) was used to assess deviation from additivity, which indicates a causal interaction where some cases of the outcome would not have occurred if it had not been for the presence of both hostility and unemployment. The SI calculates the ratio between the combined effect and individual effects of the variables. SI can go from 0 to infinity. SI = 1 means no interaction or exactly additivity, whereas SI > 1 implies positive interaction and SI < 1 implies negative interaction. In this study, SI > 1 is interpreted as evidence in favour of the psychosocial vulnerability model [33]. We followed the guidelines provided by Anderson et al. to calculate regression coefficients and covariance matrix, which was used as an input to calculate the synergy index and 95% CI in excel [33].

Using self-reported psychosocial exposures and subjective health outcomes in observational studies has been called into question [37, 38], which points to the issue of reverse causation via the mechanism that perceived poor health also generates hostility [39]. We consider the register-based outcome measure a major strength in this study, as mortality rules out concerns of reverse causation. Furthermore, to reduce confounding by poor health, we excluded all individuals not working due to illness the year before baseline. Most confounders were also measured using register data, which means that information is complete. However, there are also limitations. Alcohol abuse and depression are detectable only in registers if the individual had been either hospitalised with a main or contributory diagnosis indicating these or had purchased prescription medicines to treat depression or alcohol dependence. These drugs are not sold over the counter. Hence, all treated individuals are in the registers, but some cases of alcohol abuse may never be treated and represent a source of unmeasured confounding (see Additional file 1). Hostility was measured once, just before the start of follow-up. However, hostility is considered to be stable from early adulthood [40]. Hostility in early life is likely to influence several confounding factors throughout the life course such as depression and cohabitation (see DAG in Additional file 2). Unemployment was also measured only once. It is likely that, for example, unemployment and alcohol abuse affect each other sequentially over the life course. We adjusted for alcohol abuse and depression which occurred before the time of the survey. In this way, these factors may be considered confounding factors. However, we realise that this does not capture the complexity of associations between covariates and exposures. For simplicity, this was displayed only in the diagram using T1 to indicate time before survey and T2 to indicate time at survey or after survey. Other health behaviours may be on the causal pathway between hostility and unemployment and mortality, for instance, smoking status, which in contrast to alcohol abuse, is unlikely to affect employment. The relationship between hostility and neuroticism has been shown as possibly two simultaneous dimensions of a disease-prone personality [39, 41] and is thus confounding due to other personality characteristics that may have influenced the relationship between hostility and mortality. Unfortunately, we had no data on other relevant personality factors available in our dataset.

The study population was restricted to individuals eligible for the workforce. Consequently, we cannot generalise our findings regarding psychosocial vulnerability to individuals who are not part of the workforce. As shown in Additional file 1, a relatively large number had missing information on employment. In a sensitivity analysis, we found that those with missing information on employment more often had a low educational level or missing information on both and that their risk of mortality during follow-up was higher than that of the employed but lower than the mortality of the unemployed and sick (data not shown). We did not study causes of death, but we think that the study does contribute to the knowledge of how unemployment interacts with hostility in a broad population exposed to unemployment.

Our theoretical point of departure was the psychosocial vulnerability model and an assumption that hostility is a stable trait in the adult population, but other models of the relations between hostility, social conditions, and health have been suggested, among these the social context model. According to this model, adverse social conditions including unemployment may antecede not only health problems but also hostility [11]. The micro-sociological approach introduced by Thomas Scheff may be of some help in understanding how this causation works. According to Scheff, the individual experiences the social world in emotionally loaded categories which have bodily correlates such as pride, shame, anger, and distrust. The way these emotions are experienced is not uniform throughout the social structure [42]. Further longitudinal studies are needed to investigate relations between hostility, unemployment, and other social factors.