The study was performed within the PART-project. An initial data collection took place in 1998–2000 when a questionnaire was sent to 19742 individuals randomly selected from the population in Stockholm County, 20–64 years of age. The PART Study was approved by the Ethical Committee at Karolinska Institutet as being in accordance with ethical standards (Dnr 96–260 and 01–218). The participants in this project have given written consent after having read a detailed written description of the research. The questionnaire contained questions on potential risk indicators for psychiatric disorder as well as scales to measure well-being, depression and other symptomatology. The response rate was 53%, 58% among women and 48% among men, i.e. 10441 responded. A register-based non-response analysis showed that risk indicators for psychiatric disorders as well as the disorders themselves were more common among non-participants than among participants. However, the relationships between the risk indicators and psychiatric disorders were remarkably similar among participants and non-participants . To the respondents we sent a new questionnaire in 2001–2003 with basically the same questions as in the first questionnaire. Now 84% responded which corresponded to 8613 individuals. Of these, 881 were interviewed regarding psychiatric symptoms and living conditions, 444 were selected because of low well-being defined as ≤ 10 points in the WHO (Ten) Well-being index and 437 were selected because of high well-being defined as >10 points in the same index . Before the interview, which for 80 % of the subjects took place within eight weeks after the questionnaire had been returned, the Well-being index was again filled in.
Participants in the psychiatric interview were asked to deliver saliva samples at awakening (first), half an hour later (second), at lunch (third) and immediately before going to bed (fourth) during a Tuesday, Wednesday, Thursday or Friday in as close connection to the interview as possible. Standard swabs were used. The participants were asked to keep the swabs in their mouth until they were totally swamped by saliva and then to place them in tubes, which were mailed to the laboratory. At the laboratory all samples were centrifuged and frozen. They were analysed later on one occasion. Of the 881 participants, 717 delivered saliva samples. Several types of sampling error occurred, however. Principles were applied for the inclusion of some of these subjects since it was important to keep non-participation as low as possible;:
- Twelve subjects failed to specify sampling time for one or several samples. Ten of these participants had specified sampling time on three of them but failed in one. The missing sample was assumed to correspond to the missing sampling time.
- Twenty-five subjects had failed to deliver at least one of the four samples. Four of them had missed the third sample, and in these cases the average of the saliva cortisol concentration was calculated from the second and fourth samples. This average replaced the missing value.
- Twenty-seven subjects had collected their four samples on two different consecutive days, for instance the fourth sample on day one and the remaining three on day two. All these were accepted.
After these inclusions and exclusions 685 subjects remained for analysis. Extreme outliers were identified. This resulted in the exclusion of seven subjects. Additional checking was made with regard to sampling time. Exclusions were made if sampling had occurred at grossly erroneous hours: For the first sample after 10.55 am, for the second after noon, for the third before 10.15 am or after 4 pm and for the fourth before 6.05 pm or after 3.00 am. Fifty participants were excluded for these reasons. Ninety-three of the remaining subjects were unemployed at the time of the study. They were also excluded. A final condition was that individuals who could not be categorized as exposed or unexposed on one of the demand and control dimensions were removed. Hence, the final study group consisted of 529 subjects (348 women and 181 men).
Cortisol levels in saliva were measured by the Spectria cortisol coated tube radioimmunoassay (RIA) kit, Orion Diagnostica, FI-02101 Spoo, Finland. All samples from each subject or group of subjects were analysed simultaneously in duplicate. The within- and between assay coefficient of variation never exceeded 5.0 and 10.0, respectively. A direct comparison of the cortisol analyses was made with Kirschbaum's laboratory in Düsseldorf. Thirty samples were analysed "blindly" in both laboratories and the results were compared. There was a very high correlation (0.98) but a slight difference in level – with systematically lower levels in the Stockholm laboratory. The difference was 12.5% with 95% confidence limits 1.5–22.3%.
Psychological demands and decision latitude
Demand-control at work was assessed by means of the short Swedish version of the demand-control-support questionnaire [14–17]. There were five questions about psychological demands (for instance "Does your work require that you work fast?") and six questions about decision latitude (for instance "Can you influence what to do at work?" and "Is your work monotonous?"). All responses were graded from 1 to 4. Combinations of psychological demands and decision latitude were divided into four categories based upon medians. The "low strain" group was defined as those below the median for demands and above the median for decision latitude. The "passive" group consisted of those below the median for both demands and decision latitude and the "active" group of those above the median for both dimensions. The "high strain" group finally was defined as participants above the median for demands and below the median for decision latitude.
A physician examined the medications and coded them into three groups: Medications likely to have an effect on cortisol level (group I), medications which might have an effect on cortisol levels (group II) and medications with no suspected effects on cortisol levels (group III).
Substitution therapy with anabolic/androgenic steroids, estrogenes, gestagenes, gonadotropic releasing agents, hypophyseal hormones and pancreatic hormones as well as those on antihormonal therapy (anti estrogen and androgen) were regarded as belonging to group I. In the first step of the analyses of the effects of medications, these subjects (group I) were excluded (n = 54, 53 women and 1 man). Exclusion of subjects in group I had very small effects on mean levels and no effects on the significance levels in the comparisons between the mean saliva cortisol concentration in the different job-strain groups. In the following, those subjects in group I are included in the presentation of the results and in subsequent tables.
Group II medicines included a large number of medications that could potentially be of importance such as serotonin inhibitors as well as histamin 2 receptor inhibitors and similar medications for high production of acid in the stomach, glucocorticoids for local bowel treatment of gastrointestinal diseases, medication acting on the central nervous system for the treatment of obesity, beta receptor blockers, anticonception pills, ergot alkaloids and selective HT1 receptor inhibitors for the treatment of migraine, dopaminergic medication for the treatment of parkinsonism, anti-depressive medication (mono-amine reuptake inhibitors and selective serotonin uptake inhibitors as well as mono-amine inhibitors) and adrenergic medication for bronchial asthma and anti-histamine medications.
Confounding from medications in group II was tested by means of the introduction of this kind of medication as a confounder in statistical analysis. This was shown to be of very small importance. Accordingly we did not adjust for intake of these medications in the final analyses.
Other potential confounders
Full-time or part-time employment, severe life events , age, smoking, obesity, alcohol consumption measured by AUDIT (Alcohol Use Disorders Identification Test) , depression according to the Major Depression Inventory , and well-being, from the questionnaire in the second data collection were analysed as potential confounders of the relationship between job strain and saliva cortisol activity. However, all of them were shown to have very marginal and never any statistically significant confounding effects on relationships between job strain category and saliva cortisol. Therefore these analyses will not be presented.
The participants provided repeated and mutually dependent cortisol samples and since our interest was on the comparison of the different demand-control categories with each other at each point in time we used a mixed model approach specifying a marginal means model. Specifically, we were interested in possible significant differences in saliva cortisol between demand control categories at half an hour after awakening and moreover in possible significant differences between demand control categories with regard to change in cortisol (CAR = Cortisol awakening response) between awakening and half an hour after awakening. The reason for these limited research questions was that we wanted to contribute to the understanding of how cortisol changes throughout the day may differ in different demand control categories as has been investigated by many others [2, 4–8, 21]. Thus, even when overall main effects of demand control category or the interaction effect, time*job strain was not significant, we still wanted to test specific relationships at these time points. Before using the mixed model procedure a check for normality of data was done which revealed that cortisol values were markedly skewed to the right. After a log transformation was performed the data were normally distributed. To be able to interpret the results an exponentiation was applied on the log-transformed estimates of cortisol for every demand/control group and point in time for men and women respectively. Hence, the remaining estimates are geometric means which could be interpreted as median values from a log-normal distribution of cortisol . The time intervals between the different measurements varied somewhat between subjects in our study also after exclusion of subjects who had provided samples at times that deviated substantially from the specified scheme. These remaining differences had no effects on the results. A number of different variance-covariance patterns were fitted to the data. An unstructured pattern gave the best fit (lowest Akaike information criterion, AIC, than other patterns)  since it took into account different variances at specific time points and different covariances between each time point. This option was used in all analyses presented in this paper. REML (Restricted maximum likelihood) was used as the estimation method when trying different types of variance-covariance matrices. Maximum likelihood was used as the estimation method when building models. Separate models were built involving candidate confounding variables and demand-control category.
All the analyses were done in the SAS computer package version 9.1 using the MIXED procedure. The LSMEANS statement gives the possibility to estimate all median values for every group at every moment in time. The ESTIMATE statement gives the possibility to estimate the difference between groups at specific moments in time and estimates of differences in the mean changes between groups from one moment in time to another. In the REPEATED statement heterogeneity was specified as an option enabling different unstructured variance-covariance matrices for separate groups .