Socio-demographic and psychosocial predictors of salivary cortisol from older male participants in the Speedwell prospective cohort study

Abstract

Introduction

Associations between measures of socio-economic position and cortisol remain controversial. We examined the association between social class and cortisol reactivity in an aging male population.

Methods

The Speedwell cohort study recruited 2348 men aged 45–59 years from primary care between 1979 and 1982 (phase I) where occupational social class was used to classify socioeconomic position. Men were seen on four more occasions, the last being between 1997 and 1999 (phase 5) when salivary samples were obtained capturing cortisol reactivity to stressors (cognitive test and venepuncture) and circadian variations (awakening and night-time cortisol levels, circadian slope and area under curve) at morning and afternoon clinic sessions. Longitudinal association between social class at phase 3 and log-transformed salivary cortisol measures at phase 5 was assessed using multivariable linear regression adjusted for variables associated with sampling time and age as a potential confounder, stratified by time of clinic session. We also explored possible mediation by psychosocial factors (e.g. work dislike) and health-related factors (e.g. waist-to-hip ratio and high-density lipoprotein cholesterol).

Results

From 1768 living men, 1003 men (57%) attended a clinic at phase five, 854 participants (85% of attendees) returned home cortisol samples (mean age 71.7 years). We found little evidence of association between social class and baseline cortisol (i.e. prior to stress), cortisol response to stressors, and cortisol diurnal variation. However, we found lower social class was associated with higher and delayed post-stress recovery cortisol for participants that visited the clinic in the morning (adjusted β coefficient for manual versus non-manual 0.25 ng/ml; 95% CI: 0.06–0.48; P = 0.008). This association did not appear to be mediated by any of the measured psychosocial or health-related factors.

Conclusion

Our data did not show an overall association between social class and cortisol variability either diurnal or in response to a stressor. Lower social class was associated with a slower time to recover from exposure to stress in the morning, thereby increasing overall cortisol exposure. These findings provide some evidence for a mechanism that may contribute to the association between lower social class and a higher risk of adverse health outcomes.

Highlights

• •Social class is not associated with cortisol diurnal variability.
• •Social class is not associated with cortisol response to stressors.
• •Lower social class is associated with a slower time to recover from exposure to stress in the morning.

1. Introduction

Socio-economic position (SEP) has been defined as “the social and economic factors that influence what position individuals or groups hold within the structure of a society” (Galobardes et al., 2006). Strong associations between SEP and health disparities are widely recognised. Those in lower SEP have poorer physical and mental health and an increased risk for a wide range of diseases (Adler et al., 1994). One of the potential biological mediators of the associations between SEP and health outcome is the hypothalamic-pituitary-adrenal (HPA) axis, the neuroendocrine system that regulates the secretion of the glucocorticoid hormone cortisol (McEwen, 1998). Cortisol is a steroid hormone produced by the adrenal gland in response to stress and has regulatory effects on numerous physiological processes, including metabolic, cardiovascular, immunological and cognitive function (Miller et al., 2007).

In basal, unstressed, conditions cortisol secretion is characterised by a circadian rhythm, with highest circulating levels of hormone in the early morning, upon awakening, and decreasing levels throughout the day with lowest levels at night-time (Adam and Kumari, 2009). Circadian variations of cortisol, normally measured as the difference between awakening and bed-time cortisol, as well as cortisol reactivity to stressors, are associated with health status (Adam et al., 2017, Dowd et al., 2009, Turner et al., 2020). Dysregulation of cortisol secretion, both in resting (unstressed) conditions and in response to stressors, is considered maladaptive and a potential risk factor for many diseases including cardiometabolic disorders (Anagnostis et al., 2009, Rosmond, 2005), cognitive and mental disorders (Lupien et al., 2005), inflammation (Miller et al., 2009) and cancers (Schmidt et al., 2016). Indeed, findings from a study in a sample of middle-aged adults show flattened cortisol circadian slope to be associated with a number of psychological and metabolic characteristics, including increased abdominal adiposity and poor quality of sleep (Lasikiewicz et al., 2008). Furthermore, a systematic review and meta-analysis has shown that a flatter cortisol slope was associated with negative health outcomes for 10 of the 12 outcomes assessed, including depression, fatigue, immune/inflammatory outcomes, BMI/obesity, cancer and mortality (Adam et al., 2017). There is also some evidence that circadian rhythm of cortisol flattens with ageing, with a decrease in morning cortisol levels accompanied by an increase in evening cortisol levels (Adam and Kumari, 2009). Studies that have addressed the association between cortisol circadian rhythm and health outcome in ageing have shown an association between flattened cortisol circadian rhythm and worse physical (Gardner et al., 2013) and cognitive function (Gardner et al., 2019, Tsui et al., 2020), while positive association between elevated cortisol levels and frailty has also been reported (Clegg et al., 2013). Furthermore, elevated cortisol and cortisol to testosterone ratio predict increased cardiovascular mortality (Davey Smith et al., 2005).

A systematic review of studies on the association between circadian rhythm of cortisol and SEP in adults produced conflicting results (Dowd et al., 2009). For example, studies from Cohen (2006) in a sample of young/middle age adults have shown an association between low SEP (measured by income and education) and high cortisol levels in the evening (Cohen et al., 2006b) and through the day (Cohen et al., 2006a); others found that lower social class was associated with flattened circadian slope (Zilioli et al., 2019, Zilioli et al., 2017), or no association between financial strain and evening cortisol levels, nor with circadian slope (Kumari et al., 2010, Rosmond and Bjorntorp, 2000, Steptoe et al., 2005). Furthermore, studies that have assessed the role of potential mediators of the association between measures of SEP and cortisol measures have shown that changes in diurnal cortisol levels could be explained by poor health behaviour such as smoking (Cohen et al., 2006b), short sleep duration and perceived financial insecurity (Kumari et al., 2010).

In contrast to a vast literature on the association of SEP and circadian measures, there are not many studies that have addressed the association between SEP and cortisol response to and recovery from exposure to stress. A study on the association between life course SES and cortisol response to a laboratory stress challenge found slower cortisol recovery rate in lower social class, when compared to higher social class participants (Le-Scherban et al., 2018). On the other hand, another study showed that lower subjective social status was associated with a faster rates of cortisol reactivity to, and recovery from, the Trier Social Stress Task (Rahal et al., 2020).

Because the majority of studies have explored the concomitant association of social class with cortisol measures, the objective of this study was to estimate the longitudinal association between SEP and cortisol measures both in basal condition and following exposure to stress in an ageing male population with a view to determining whether any such associations could be explained by psychosocial factors and health-related conditions.

2. Methods

2.1. Study design

The Speedwell prospective cohort study was set up in 1979 by the Avon Health Authority to determine risk factors for ischaemic heart disease (Yarnell et al., 1984). Participants were all men aged 45–63 that were registered with 16 general practitioners in 2 health centres in the Speedwell neighbourhood of Bristol, United Kingdom, in 1979. Socioeconomic status and other relevant variables were measured at phase 1–5 as following. At phase 1 (1972–1982) 2348 participants (92% of the sample) were recruited underwent a clinical examination and completed a questionnaire regarding socioeconomic circumstances and other relevant variables including general health, medical history and lifestyle. Participants were re-examined at 4 subsequent follow-ups (phase 2: 1982–1985, phase 3: 1985–1988, phase 4:1988–1991; phase 5: 1997–1999). At phase 5, the surviving men in the cohort (1768) aged 64–80 years were approached to participate by self-completing a questionnaire. Participants were invited to attend a clinic session (day 1), either in the morning or in the afternoon, in which they underwent a battery of psychometric tests in order to assess their cognitive function (cognitive test), and venepuncture in order to collect blood samples for analysis of biological markers. The cognitive tests included the Mini Mental State Examination (MMSE) and the Wesnes CDR Test Package (Wesnes et al., 2016).

During the same visit, salivary samples for assessment of basal levels of cortisol (prior to stressors), as well as cortisol reactivity to stress, were obtained at time 0 (pre-stress baseline, sample 1), after the cognitive test (~45 min after pre-stress baseline, sample 2), and after venepuncture (~35 min after the post-cognitive test sample, sample 3). Participants were also asked to collect one extra salivary sample on returning home (approx. 100 min after the post-venepuncture sample, sample 4) in order to assess cortisol levels following recovery from the above stressors. Participants were asked to collect four more samples on the next day (day 2) within 30 min from awakening (sample 5), at 10:00 (sample 6), 15:00 (sample 7) and 22:00 (sample 8) to assess cortisol secretion and parameters of cortisol circadian rhythm at resting conditions (i.e. not under laboratory conditions). The exact date and time of sample collection was assessed by a self-completed record that was returned with the home-collected samples. A schematic timeline showing the time of each sample collection and exposure to stressors is shown in supplementary Fig. 2.

2.2. Collection of saliva samples and cortisol assay

Participants were instructed not to brush their teeth before collecting the samples and to refrain from eating, smoking and drinking for at least 30 min prior and were asked to record any stressful event prior to each sample collection. Samples were stored in a domestic fridge and posted back to the researchers for processing for cortisol measurement. Salivary total free cortisol was determined by radio immune assay with a rabbit anti-cortisol antibody (BioClin Therapeutics, Inc., CA, US) and ¹²⁵I-cortisol tracer (Amersham, UK). Free from bound cortisol was separated by dextran-coated charcoal and gamma radiation of supernatants counted using a gamma counter to determine the bound portion. Inter and intra assay coefficients were measured across hormone assays using different batches of ¹²⁵I-cortisol and were 19% and 3–12%, respectively. Final data are expressed as ng/ml of saliva. Because high doses of synthetic corticosteroids have suppressive effects on HPA axis activity,¹ any participant under oral treatment with prednisolone at the time of examination was excluded from all the analysis.

2.3. Exposure and other covariates

SEP was measured using occupational social class data collected at phases 1, 2 and 3. Social class was defined from the man's current, or, if not working, previous employment. Social class in the United Kingdom is classified in accordance with the Registrar General's classification (RGSC) and dichotomised as non-manual work, including: I professional (e.g. solicitor), II intermediate (e.g. manager) and III skilled non-manual (e.g. bank clerk) and manual work including: III skilled manual (e.g. electrician), IV semi-skilled (e.g. postman) and V unskilled (e.g. factory worker) as previously reported (Davey Smith et al., 1997, Lowe et al., 2004). To assess the cumulative effect of social class over time, as well as the effect of social class mobility, a summary score representing social class at phase 1–3 was derived by adding the social class value (0 for non-manual, 1 for manual) at different phases, ranged from 0 (for those always in non-manual social class) to 3 (for those always in the manual social class), with a score of 1 and 2 indicating a shift between social class in either direction. Because of small numbers in the middle groups (score 1 and 2), these were collapsed to one group. Only participant with complete social class data at phase 1–3 were included in this analysis. This approach has been previously used looking at both self-rated health (Power et al., 1999) and mortality (Davey Smith et al., 1997) providing evidence of a linear cumulative effect, though clearly there may be critical or sensitive period effects.

The choice of covariates included in the analysis was based on the literature reporting association of salivary cortisol with sociodemographic, health-related and sampling factors. (Adam and Kumari, 2009, Kumari et al., 2010). Covariates related to cortisol secretion and/or SEP were: age at sampling, retirement age, education attainment (measured as leaving school age), marital status at phase 5, time of sample collection, batch of cortisol assay and whether participants experienced any form of stress (i.e. discussion with family member, feeling upset) prior to collecting saliva sample each saliva sample on day 2. Because the time of arrival to the clinic was not recorded, time of clinic visit on day 1 was assessed based on time of sample 1 collection and dichotomised as morning (9:00–11:59) or afternoon (12:00–16:59).

Self-reported measures included smoking habits at phase 1, self-reported diabetes at phase 1 and 5, depression, anxiety and use of synthetic steroid (other than oral) at phase 5 and were all classified as yes/no. Psychosocial characteristics measured at phase 5 included work dislike, financial conflict, worries/stress, struggling, social isolation, leisure time and possession of clubs/organisations membership. Reported psychosocial characteristics were assessed by questionnaire and groups were based on the following answers: “not at all”, “only slightly”, “distinctly” and “very much”; all answers had the same weight in the analysis. Participants were asked to self-complete the questionnaire at home prior to the visit to the clinic.

Biological characteristics included clinic and laboratory measures of systolic blood pressure (SBP) and diastolic blood pressure (DBP) at phase 1 and 5, ankle brachial index (ABI, as a measure of peripheral atherosclerosis) at phase 5, height, weight and derived body mass index (BMI) at phase 1 and 5, waist-to-hip ratio (WHR, as measure of abdominal obesity) at phase 5 and blood lipids (e.g. cholesterols, triglycerides) at phase 1. Height was measured on a Holtain stadiometer and body weight was taken on a beam balance. Blood pressure was measured on a Hawksley random zero sphygmomanometer (phase 1) and an automated Omron blood pressure monitor (phase 5). Plasma lipids including total cholesterol, high-density lipoprotein cholesterol (HDL-C), very low-density lipoprotein cholesterol (VLDL-C), and total triglycerides were measured at phase 1 as previously reported (Yarnell et al., 2001). Low-density lipoprotein cholesterol (LDL-C) was derived from measures of cholesterol and triglycerides using the Friedewald equation. Body mass index (BMI) was calculated as weight (kg)/height squared (m²). Ankle Brachial Index (ABI) at phase 5 was calculated by dividing the systolic pressure at the ankle by the systolic pressure at the arm (a ratio less than 0.9 indicating peripheral vascular disease). Waist and hip were measured at phase 5 using a tape measure kept at constant tension by a tension gauge and were used to calculate waist hip ratio (WHR).

2.4. Statistical analysis

All statistical analyses were performed using Stata version 16.1 (StataCorp LLC, College Station, TX, US). We derived two groups of cortisol outcomes: (a) reactivity and (b) diurnal variability. Reactivity was defined as (i) the response to cognitive testing (sample 2 minus baseline sample 1); (ii) the response to venepuncture (sample 3 minus post cognitive testing sample 2); (iii) post-stress recovery (sample 4). These reactivity measures were stratified by time of clinic visit to control for circadian variation in cortisol levels. Diurnal variability was captured by awakening and night-time cortisol levels, as well as cortisol circadian slope and area under the curve (AUC) from the day 2 samples. Circadian slope was calculated by subtracting the values measured at awakening from the values of night time cortisol (measured around 22:00) and dividing this by the number of hours between these two samples, in order to control for awakening time differences as previously described (Adam and Kumari, 2009, Gardner et al., 2019, Gardner et al., 2013). The resulting negative value is a measure of the decline in cortisol level per hour. AUC was calculated from the four cortisol measures, and respective sampling time, of day 2 using the trapezoid rule and was used as a measure of total measured day-time cortisol levels. Samples for which time of collection was missing were excluded from the final analysis.

We examined the characteristics of men from the whole cohort to those participating in phase 5 and who also provided at least one cortisol sample to examine for selection bias. Descriptive statistics followed by multivariable linear regression analysis was used to examine the relationship between measures of cortisol and social class, and the role of potential confounders and mediators. Because the distribution of cortisol measures was positively skewed, natural logarithmic transformations of cortisol data were used, and we report log-transformed cortisol measures. Associations between demographic, biological, clinical and psychosocial characteristics were assessed using chi-square or Fisher exact test for heterogeneity (categorical variables) or linear regression (numerical variables). Derived cortisol measures were calculated from log-transformed cortisol measures on day 2, except for the AUC which was log-transformed after the calculations. Percentage of missing data for each covariate and for cortisol measures was calculated; no missing data imputations were performed.

We ran multivariable linear regression models using social class (manual versus non-manual) as the main exposure and adjusting for cortisol assay batch (to adjust for batch-effect due to high inter-assay CV), age at phase 5 (as cortisol levels decline with ageing (Adam and Kumari, 2009)) and, for cortisol measures at day 1, time of clinic visit to control for circadian variation of cortisol reactivity. Day 1 cortisol measures models were also adjusted for venepuncture outcome (i.e., if venepuncture was successful or not, see supplementary Fig. 1 for specific groups) and time of sample collection. Patients that did not undergo cognitive test were excluded from the analysis of cortisol reactivity to stress. Because of the circadian variation in cortisol secretion, the association of social class with day 1 cortisol reactivity was also tested with an interaction term between social class and time of clinic visit. We examined potential mediators, such as psychosocial and health-related factors, by comparing the coefficients for social class with and without adjustment. We used the difference in coefficient method (MacKinnon et al., 2002). The total effect was estimated by the coefficient for social class on cortisol. This was then repeated adjusting for our potential mediators to estimate the direct effect, and the indirect effect was estimated by subtracting the direct effect from the total effect.

To test the robustness of our results we conducted a number of sensitivity analyses. We derived the rate of recovery from the stress caused by the cognitive testing and the venepuncture by calculating the difference between cortisol samples 4 and 2, and the rate of recovery from the stress caused by the venepuncture by calculating the difference between cortisol samples 4 and 3, both adjusted for the time of collection of post-stress and recover samples. Furthermore, because other methods to calculate circadian slope, alternative to the one we have used in our main analysis, have been reported (Adam and Kumari, 2009, Dowd et al., 2009), we repeated the analysis of social class and circadian slope using multilevel models with a random intercept and random slope. Our model had two levels: cortisol measures (level 1) within participants (level 2). The time axis was centred on the mean time of sample 1 collection (7.50 h). In addition to social class as measure to SEP, we repeated our analyses using education level (measured as school leaving age) as an alternative measure of SEP to see if this did or did not alter any of our findings. We also conducted exploratory analysis to test whether social class was associated with both high and low levels of cortisol at day 2. Cortisol reactivity measures as well as circadian cortisol measures were divided in quartile, with low and high cortisol levels in the first and fourth quartile respectively. We then used multinomial logistic regression with the middle cortisol levels (second and third quartile) as reference, to test whether the odds of having either lower or higher cortisol levels where higher in the lower social class.

3. Results

3.1. Demographic, biological and clinical characteristics of participants

Of the 1100 men who agreed to complete a questionnaire, 1003 participants also underwent a clinical examination at phase 5 and a total of 854 returned at least one saliva sample (see study flow diagram in supplementary Fig. 1). From these, 16 were excluded from the analysis due to current treatment with oral synthetic corticosteroids. Of the remaining 838 participants with at least one cortisol measure, 732 returned the complete set of 8 samples. We found little evidence of association between missingness of cortisol samples and social class (Odds Ratio [OR]: 0.86; 95% CI: 0.60–1.24; P = 0.43).

We compared participants who returned at least one cortisol sample with the other men in phase 5 who did not, over a wide range of social, clinical and biological variables. All differences were consistent with chance variability except for SBP which was higher at phase 5 for men without a sample (P < 0.001) and higher frequency of psychological “struggling” (P = 0.03). However, there was little evidence of association between social class and providing a sample.

Demographic, biological and clinical characteristics of participants with at least one cortisol measure and time of sample collection at phase 5 of the Speedwell study, stratified according to social class, are shown in Table 1. Out of 838 participants, 343 (40.9%) were in the higher social class (non-manual) and 495 (59.1%) were in the lower social class (manual). Compared to non-manual, manual work participants left school at earlier age (P < 0.001), retired at later age (P = 0.03), had a higher percentage of being non-married (P = 0.04) and being current smokers (P = 0.001), had higher plasma levels of HDL-C (P = 0.05), and lower height and weight at both phases 1 and 5 (height: P < 0.001; P < 0.001; P = 0.001; weight: P = 0.02, at phase 1 and 5 respectively. There were very modest differences for psychosocial characteristics such as work dislike and financial conflict, though non-manual men were more likely to report higher levels of worries/stress (P = 0.01), a greater frequency of leisure time activities (P = 0.01) and belonging to a club or some organization (P < 0.001) (see supplementary Table 3). Percentage of missing data for each variable are reported in supplementary Table 4.

Table 1.

Summary data of demographic, biological and clinical characteristics of men in the Speedwell study phase 1 and 5 restricted to participant with at least one valid cortisol measure at phase 5 analysed according to social class.

		Non-Manuala		Manualb			Total
	Study phase	n	Mean (SD) or %	n	Mean (SD) or %	P-value	n	Mean (SD) or %
Age	5	342	71.6 (4.1)	495	71.7 (4.2)	0.63	837	71.7 (4.2)
School leaving age	5	331	15.2 (2.8)	471	14.2 (0.6)	< 0.001	802	14.6 (1.9)
Retirement age	5	309	62.2 (3.6)	452	62.8 (3.7)	0.03	761	62.5 (3.7)
Marital status (%)
Married	5	278	83.7	372	78.0	0.04	650	80.4
Non-marriedc	5	54	16.3	105	22.0		159	19.7
Smoking habits (%)
Current smokers	1	124	36.2	237	47.9	0.001	361	43.1
Non-/ex-smoker	1	219	63.9	258	52.1		477	56.9
Diabetes (%)
Yes	1	4	1.2	6	1.2	0.95	10	1.2
No	1	339	98.8	489	98.8		828	98.8
Yes	5	24	7.3	35	7.4	0.99	59	7.3
No	5	304	92.7	441	92.7		745	92.7
Depression (%)
Yes	5	29	8.7	52	10.9	0.31	81	10.0
No	5	303	91.3	425	89.1		728	90.0
Anxiety (%)
Yes	5	30	9.1	43	9.1	0.99	73	9.1
No	5	300	90.9	428	90.9		728	90.9
Stress on day 2 (%)
Yes	5	35	12.0	33	8.4	0.13	68	9.9
No	5	258	88.1	359	91.6		617	90.1
Use of steroidsd
Yes	5	25	7.1	51	10.8	0.13	76	9.5
No	5	304	92.4	421	89.2		725	90.5
Plasma lipids (nmol/l)
Total cholesterol	1	335	5.9 (1.1)	474	5.8 (1.2)	0.44	809	5.8 (1.2)
HDL-C	1	333	1.1 (0.3)	472	1.1 (0.4)	0.05	805	1.1 (0.4)
LDL-C	1	328	4.5 (1.2)	459	4.3 (1.2)	0.15	787	4.4 (1.2)
VLDL-C	1	316	0.9 (0.6)	448	0.9 (0.7)	0.59	764	0.9 (0.6)
Triglycerides	1	332	1.6 (1.0)	469	1.6 (0.9)	0.58	801	1.6 (1.0)
SBP (mmHg)	1	342	133.5 (20.3)	495	136.3 (20.4)	0.06	837	135.1 (20.4)
	5	331	160.0 (28.1)	493	161.1 (24.9)	0.57	833	160.7 (26.2)
DBP (mmHg)	1	342	85.1 (13.1)	495	85.8 (12.4)	0.46	837	85.5 (12.7)
	5	340	87.6 (12.7)	493	87.7 (12.9)	0.91	833	87.7 (12.8)
ABI	5	307	1.2 (0.2)	455	1.2 (0.2)	0.13	762	1.2 (0.2)
Height (cm)	1	343	174.1 (6.9)	494	171.8 (6.6)	< 0.001	837	172.7 (6.8)
	5	339	172.3 (7.0)	489	170.1 (6.8)	< 0.001	828	171.1 (6.9)
Weight (kg)	1	342	78.0 (10.0)	492	75.6 (10.7)	0.001	834	76.6 (10.5)
	5	339	78.9 (11.6)	491	76.8 (12.2)	0.02	830	77.7 (12.0)
BMI (kg/m2)	1	342	25.7 (2.8)	492	25.6 (3.2)	0.49	834	25.6 (3.0)
	5	338	26.6 (3.6)	489	26.5 (3.7)	0.90	827	26.5 (3.7)
WHR	5	339	1.0 (0.1)	491	1.0 (0.1)	0.15	830	1.0 (0.1)

Summary data given as number (n) and % for categorical variables or number and mean standard deviation, SD) for numerical variables. P-values are from Chi2 test for categorical variables, and from unadjusted linear regression for numerical variables.

SBP: systolic blood pressure; DBP: diastolic blood pressure; HDL-C: high-density lipoprotein cholesterol; LDL-C: low-density lipoprotein cholesterol; VLDL-C: very low-density lipoprotein cholesterol; ABI: ankle-brachial index; BMI: body mass index; WHR: waist-to-hip ratio.

^aIncludes professional (I), intermediate (II) and skilled non-manual (III-NM).

^bIncludes skilled manual (III-M), semi-skilled (IV) and unskilled (V).

^cIncludes single, divorced and widower.

^dExcludes oral steroids (i.e. prednisolone) as these participants are already excluded from the analysis.

3.2. Cortisol measures

Details of cortisol measures at day 1 and day 2 are reported in Table 2. Day 2 wake time data were missing in 57.3% of participants, though missingness was not associated with social class (OR 0.81; 95% CI: 0.62–1.08; P = 0.16) nor with AM cortisol levels (sample 5: β coefficient: 0.01 ng/ml; 95% CI: −0.13 to 0.15; P = 0.89). There was a strong correlation between wake time and time of collection of the first sample on day 2 (sample 5: correlation coefficient=0.93; P < 0.001), therefore in all our models we used the time of collection of sample 5 to adjust for wake time.

Table 2.

Descriptive statistics of day 1 and day 2 cortisol measures and time of sampling of participants from the Speedwell study at phase 5.

		Pre-stress baseline (sample 1, ng/ml)	Post-cognitive test (sample 2, ng/ml)	Post-venepuncture (sample 3, ng/ml)	Post-stress recovery (sample 4, ng/ml)	Response to cognitive test (ng/ml)	Response to venepuncture (ng/ml)
Day 1 AM (9:00–11:59)
	n	495	491	486	476	486	475
Cortisol	Mean (SD)	13.92 (12.25)	13.69 (14.31)	10.96 (13.03)	10.67 (14.73)	-0.16 (13.46)	-2.75 (13.79)
	Median (IQR)	10.68 (5.84, 17.16)	9.77 (5.38, 16.92)	8.00 (4.11, 13.47)	6.35 (3.69, 11.95)	-0.59 (−4.60, 3.66)	-1.80 (−5.71, 0.70)
Time (h)	Mean (SD)	10.38 (0.68)	11.15 (0.68)	11.77 (0.70)	13.53 (1.13)
Day 1 PM (12:00–16:00)
	n	314	311	306	297	303	303
Cortisol	Mean (SD)	10.99 (12.36)	14.08 (22.29)	10.43 (16.62)	8.88 (13.70)	3.35 (18.90)	-3.78 (21.22)
	Median (IQR)	7.04 (4.43, 13.50)	7.79 (4.61, 16.05)	6.15 (3.56, 11.88)	5.04 (2.83, 9.21)	0.28 (−2.35, 4.61)	-1.43 (−4.67, 0.51)
Time (h)	Mean (SD)	14.31 (0.68)	15.11 (0.74)	15.69 (0.76)	17.33 (0.86)
P-value		< 0.001	0.05	0.05	0.004	< 0.001	0.92
		Wake + 30 min (sample 5, ng/ml)	10:00 (sample 6, ng/ml)	15:00 (sample 7, ng/ml)	22:00 (sample 8, ng/ml)	Circadian slope (ng/ml/h)	AUC (ng)
Day 2
	n	794	802	801	795	780	766
Cortisol	Mean (SD)	28.00 (30.28)	15.65 (21.94)	9.96 (14.85)	6.24 (13.79)	-1.50 (2.09)	173.06 (169.63)
	Median (IQR)	19.52 (9.91, 35.33)	10.49 (6.02, 17.70)	6.57 (3.48, 11.41)	3.12 (1.79, 6.05)	-1.02 (−2.04, −0.38)	131.52 (79.38, 213,51)
Time (h)	Mean (SD)	7.51 (0.99)	10.13 (0.45)	15.16 (0.52)	22.10 (0.41)

Descriptive statistics of day 1 cortisol measures stratified by time of clinic attendance (AM and PM). Data are reported as number (n), mean (standard deviation, SD), median (interquartile range, IQR) for cortisol measures and mean (SD) for time of sampling. P-values are from linear regression analysis of the associatione between time of clinic visit and day 1 log-transformed cortisol measures adjusted for cortisol assay batch, time of sample, results of venepuncture and age at phase 5. AUC: area under curve.

On day 1, cortisol levels prior to stressors, following venepuncture and at post-stress recovery were higher in participants that attended clinic in the morning (P < 0.001, P = 0.05 and P = 0.004, respectively), whereas cortisol levels following the cognitive test and in response to cognitive stress from baseline were higher in participants attending clinic in the afternoon (P = 0.05 and P < 0.001, respectively). Day 2 cortisol measures revealed the expected cortisol circadian pattern, with higher levels of hormone in the morning, and a negative circadian slope.

3.3. Social class and cortisol association

We found no evidence of an association between manual social class and day 1 cortisol measures including the pre-stress baseline or reactivity to stress, except for the post-stress recovery cortisol, measured as cortisol levels in sample 4 as well as rate of cortisol recovery from cognitive test and from venepuncture stress, with higher recovery cortisol levels in participant in the manual work group (post-stress recovery unadjusted β coefficient: 0.18 ng/ml; 95% CI: 0.03–0.32; P = 0.02; post-cognitive test cortisol recovery rate unadjusted β coefficient: 0.16 ng/ml; 95% CI: 0.01–0.31; P = 0.03; post-venepuncture recovery rate unadjusted β coefficient: 0.21 ng/ml; 95% CI: 0.07–0.36; P = 0.003; Table 3, model 1). Although we found little evidence of interaction between social class and time of visit on any of the cortisol reactivity measures, our model adjusted for cortisol assay batch, time of clinic visit, time of sample collection, results of venepuncture and age at phase 5, showed discrete evidence of association between social class and post-stress recovery in participant attending clinic in the morning (post-stress recovery adjusted β coefficient: 0.25 ng/ml; 95% CI: 0.07–0.43; P = 0.008; post-cognitive test cortisol recovery rate adjusted β coefficient: 0.16 ng/ml; 95% CI: −0.002 to 0.38; P = 0.05; post-venepuncture recovery rate adjusted β coefficient: 0.20 ng/ml; 95% CI: 0.02–0.39; P = 0.03; Table 3, model 2) but not in participant attending clinic in the afternoon, except for post-venepuncture recovery rate (adjusted β coefficient: 0.23 ng/ml; 95% CI: −0.01 to 0.47; P = 0.06.

Table 3.

Association of social class at phase 1 with Day 1 cortisol measures in participants of the of the Speedwell study at phase 5.

		Pre-stress baseline (n = 809)	Response to cognitive test (n = 789)	Response to venepuncture (n = 778)	Post-stress recovery (n = 773)	Post-cognitive test recovery (n = 752)	Post-venepuncture recovery (n = 752)
Model 1a
Manual-AM	β coeff. (95% CI)	0.02 (−0.15, 0.10)	0.03 (−0.08, 0.13)	-0.05 (−0.16, 0.05)	0.18 (0.03, 0.32)	0.16 (0.01, 0.31)	0.21 (0.07, 0.36)
	P-value	0.70	0.60	0.33	0.02	0.03	0.003
Model 2b
Manual-AM	β coeff. (95% CI)	0.05 (−0.10, 0.21)	-0.02 (−0.15, 0.11)	-0.03 (−0.16, 0.01)	0.25 (0.07, 0.43)	0.19 (−0.002, 0.38)	0.20 (0.02, 0.39)
	P-value	0.50	0.80	0.69	0.008	0.05	0.03
Manual-PM	β coeff. (95% CI)	-0.13 (−0.33, 0.07)	0.10 (−0.07, 0.23)	-0.07 (−0.25, 0.09)	0.08 (−0.15, 0.32)	0.12 (−0.12, 0.37)	0.23 (−0.01, 0.47)
	P-value	0.21	0.26	0.39	0.50	0.33	0.06
	Interaction P-value	0.16	0.29	0.66	0.27	0.67	0.86

β coefficient and P-values are from:

^aModel 1: unadjusted linear regression of log-transformed cortisol measures (ng/ml);

^bModel 2: model 1 adjusted for age at phase 5, time of sample(s), cortisol assay batch, venepuncture results (for response to venepuncture and post-stress recovery cortisol measures) and completion of cognitive test (response to cognitive test, response to venepuncture and post-stress recovery cortisol measures). Model 2 also included an interaction term to test for interaction between social class and time of clinic visit. β coefficient and P-values for cortisol measures in Manual-PM group were derived from the interaction term. Coeff.: coefficient; CI: confidence interval.

Our exploratory analysis using multinomial logistic regression models showed little evidence of association of social class with neither low or high cortisol levels measured at pre-stress baseline, in response to cognitive test and in response to venepuncture and post-stress recovery. However, using the same models, we found some evidence of association of social class with high, but not low, rate of recovery from cognitive test and from venepuncture, both in the unadjusted and adjusted models (supplementary Table 6, models 1 and 2).

In contrast with cortisol measures of stress reactivity, we found little evidence of association between social class and any of the diurnal outcomes measured (Table 4, model 1 and 2) and between diurnal slope and social class in the multilevel linear regression model (supplementary Table 5, model 1 and 2). Furthermore, our exploratory analysis using multinomial logistic regression models showed little evidence of association of social class with either low or high cortisol levels for any of the diurnal cortisol outcomes measured at day 2 (supplementary Table 6, models 2 and 3).

Table 4.

Association of social class at phase 1 with Day 2 cortisol measures in participants of the of the Speedwell study at phase 5.

		Wake + 30 min (n = 793)	Night-time cortisol (22:00; n = 790)	Circadian slope (n = 780)	AUC (n = 765)
Model 1a	β coeff. (95% CI)	0.003 (−0.140, 0.146)	0.003 (−0.144, 0.150)	0.001 (−0.011, 0.012)	0.07 (−0.04, 0.18)
	P-value	0.97	0.969	0.93	0.20
Model 2b	β coeff. (95% CI)	0.004 (−0.140, 0.148)	0.002 (−0.145, 0.150)	0.001(−0.011, 0.012)	0.07 (−0.04, 0.18)
	P-value	0.96	0.98	0.90	0.23

β coefficient and P-values are from

Coeff.: coefficient; CI: confidence interval; AUC: area under curve.

^aModel 1: unadjusted linear regressions of log-transformed cortisol measures (ng/ml).

^bModel 2: model 1adjusted for cortisol assay batch, time of collection of respective samples, time of collection of wake + 30 min sample and age at phase 5.

When we assessed the cumulative effect of social class measured at phases 1–3 on morning post-stress recovery cortisol, we found some weak evidence of association with higher recovery cortisol levels in participant that were in the manual work group in all three phases (unadjusted β coefficient: 0.23 ng/ml; 95% CI: 0.03–0.44; P = 0.03; supplementary Table 6, model 1), but not in these that underwent a change in social class in either direction (unadjusted β coefficient: 0.02 ng/ml; 95% CI: −0.27 to 0.31; P = 0.91; supplementary Table 5, model 1). As shown for social class measured at phase 1 only, such associations were not altered by adjustment for cortisol assay batch, time of sample collection, results of venepuncture and age at phase 5 (always manual: adjusted β coefficient: 0.24 ng/ml; 95% CI: 0.03–0.45; P = 0.03; social class change: adjusted β coefficient: 0.04 ng/ml; 95% CI: −0.25 to 0.34; P = 0.76; supplementary Table 5, model 2).

3.4. Education and cortisol association

We found little evidence of association between education and any of the cortisol measure assessed in our sensitivity analysis (supplementary Table 8), except for weak evidence of association between school leaving age and cortisol circadian slope (unadjusted β coefficient: 0.003 ng/ml; 95% CI: 0.0002–0.005; P = 0.03; supplementary Table 8, model 1; adjusted β coefficient: 0.003 ng/ml; 95% CI: 0.0003–0.005; P = 0.03; supplementary Table 8, model 2), where the unit of exposure was years of age attending school, so more educated subjects had smaller circadian drop. This was also seen in the crude multilevel linear regression model (unadjusted β coefficient: 0.002 ng/ml; 95% CI 0.0002–0.004; P = 0.03; supplementary Table 9, model 1) but was attenuated after adjustment for age and cortisol assay batch (adjusted β coefficient: 0.001 ng/ml; 95% CI −0.0002 to 0.0003; P = 0.64; supplementary Table 9, model 2).

3.5. Mediation analysis

We conducted mediation analysis to evaluate whether the association between social class and cortisol post-stress recovery in the morning could be explained by psychosocial factors and health-related conditions. We examined the association between potential mediators and morning post-stress recovery cortisol further (supplemental Table 9). We found some evidence of a positive association with HDL-C measured at phase 1 (β coefficient: 0.33; 95% CI: 0.09–0.57; P = 0.007) and a negative association with WHR measured at phase 5 (β coefficient: −2.76; 95% CI: −4.321 to −1.200; P < 0.001), and weak evidence of a positive association with work dislike measured at phase 5 (β coefficient: −0.25; 95% CI: −0.53 to 0.04; P = 0.09).

We undertook a number of multivariable models adjusting for education as a potential confounder and other covariates that were found to be associated with morning post-stress recovery cortisol further (supplemental Table 8), including WHR, HDL-C and dislike for work as potential mediators. Compared to the base model (Table 5, model 1), the model adjusting for education showed a modest attenuation of effect (from 28.2% to 24.3%; Table 5, model 2). Similarly, a small attenuation of effect was observed when adjusting for HDL-cholesterol (26.5%; Table 5, model 3) or work dislike (22.6%; Table 5, model 4), and with the fully adjusted model (25.1%; Table 5, model 7) suggesting that our potential mediators had little effect on the total effect of social class on morning post-recovery cortisol.

Table 5.

Role of confounders and mediators in the association of social class with post-stress recovery cortisol.

	Model 1a	Model 2b	Model 3c	Model 4d	Model 5e	Model 6f	Model 7g
n	476	476	457	457	474	413	394
β coefficient	0.248	0.250	0.218	0.235	0.247	0.204	0.225
(95% CI)	(0.065, 0.432)	(0.065, 0.434)	(0.030, 0.405)	(0.046, 0.423)	(0.067, 0.428)	(0.011, 0.397)	(0.030, 0.419)
% difference	28.2	28.4	24.3	26.5	28.1	22.6	25.1
(95% CI)	(6.7, 54.1)	(6.7, 54.3)	(3.1, 49.9)	(4.7, 52.6)	(6.9, 53.4)	(1.1, 48.7)	(3.1, 52.0)
P-value	0.008	0.008	0.02	0.02	0.007	0.04	0.02

β coefficient and P-values are from linear regression of the associatione between social class and log-transformed cortisol measures (ng/ml); % difference relative to non-manual class is calculated from the β coefficient geometric mean.

HDL: High-density lipoprotein; WHR: waist-to-hip ratio.

^aModel 1: Unadjusted model including only participans with valid data on time of sample, venepuncture results, psychometric test, cortisol assay batch and age at phase 5.

^bModel 2: Model 1 adjusted for age, time of sample, results of venepouncturte and cortisol assay batch.

^cModel 3: Model 2 adjusted for education; unadjusted model of reference includes only participants with valid data on education (β coeficient=0.218).

^dModel 4: Model 2 adjusted for HDL cholesterol; unadjusted model of reference includes only participants with valid data on HDL cholesterol (β coefficient=0.258).

^eModel 5: Model 2 adjusted for WHR; unadjusted model of reference includes only participants with valid data on WHR (β coefficient=0.227).

^fModel 6: Model 2 adjusted for work dislike; unadjusted model of reference includes only participants with valid data on work dislike (β coefficient=0.198).

^gModel 7: Model 2 adjusted for HDL cholesterol, WHR and work dislike; unadjusted model of reference includes only participants with valid data on HDL cholesterol, WHR ratio and work dislike (β coefficient=0.224).

Results of pair-wise correlations between the main cortisol outcomes and the covariates included in our models are reported in supplementary Table 10.

4. Discussion

The results from our main analysis revealed that there was no strong association between social class or educational level, as measures of SEP, and acute cortisol reactivity or diurnal variability. However, lower social class was associated with a higher morning post-stress cortisol and hence a delayed return to basal cortisol levels. This was hardly attenuated in our mediation analysis despite conditioning on HDL-C, WHR and work struggles. Cholesterol is the precursor of all steroids, and HDL-C represent 75% of cholesterol source for cortisol synthesis in the adrenal gland (Bochem et al., 2013), thus a positive association may reflect greater bioavailability. The negative association between WHR was somewhat surprising as we would have predicted, if anything that this would have shown a positive association. Previous studies have shown an association between abdominal obesity and both dysregulation of circadian HPA axis activity and hyperresponsiveness to psychological and physiological stress (Rosmond et al., 1998), and higher abdominal adiposity has been reported to be associated with lower morning cortisol levels (Ljung et al., 2000).

Similar to some but not all previous studies, we found no strong association between social class with pre-stress baseline or other measures of reactivity or diurnal variability. This suggests that elevated post-stress recovery cortisol in low social class participants that were subjected to stressors (cognitive testing and venepuncture) in the morning may be the result of dysfunctional recovery from stress and not due to any underlying disrupted circadian cortisol pattern in these participants.

We note that the above association was only seen in participants that attended clinic in the morning, though cortisol reactivity was higher in the afternoon, irrespective of social class. Diurnal responsiveness of glucocorticoids to stress has been well characterized in animal models (Spiga and Lightman, 2020), whereas only few studies have addressed this in human. For example, one study has shown that cortisol secretion in response to low to moderate intensity physical exercise is higher in the morning, whereas cortisol response to high-intensity exercise is higher in the evening (Scheen et al., 1998). Furthermore, one other study has shown no differences in cortisol responses to the Trier Social Stress Test (TSST) between morning and evening (Kudielka et al., 2004). Human response to stress is highly subjective and context dependent (Cohen et al., 2007). In our study, participants were exposed to a combination of psychological and physical stress, each of these activating the cortisol response via different central pathways (Herman et al., 2003). It is noteworthy that the stressors in our study, cognitive tests and venepuncture may be weaker activators of the system, compared to more potent stressors that are known to robustly activate the HPA axis, such as the TSST (Dickerson and Kemeny, 2004), which may show a more pronounced differential response to Social class differences. Furthermore, it has been shown that stress can induce both hyperactivity and hypoactivity of the HPA axis (Ben-Shlomo et al., 2014), which may explain the lack of association between social class and circadian measures of cortisol and reactivity to stress in our study however, our exploratory analysis revealed little evidence of association of social class with having both high and low cortisol levels in all our cortisol measures outcomes. In addition, underlying its circadian rhythm, cortisol secretions are characterized by a highly dynamic pulsatile pattern of release, with high pulses of hormone in the morning (Veldhuis et al., 1989). Therefore, it is also possible that, due to pulsatile secretion of cortisol in the morning, the stress response in participants attending clinic in the morning, may have been masked by the already-elevated levels of cortisol. From our data we cannot establish to what degree each stressor contributed to the lack of recovery in manual social class participants as the stressors differ in nature (psychological stress versus physical stress) affecting activation of the HPA axis via different neuronal pathways (Herman et al., 2003). Our findings are consistent with a previous study in which adult SEP (measured as combination of education, annual household income and wealth), was associated with a slower cortisol recovery following a cognitive test challenge (Le-Scherban et al., 2018). Similarly, the same study found no association between SEP and cortisol response to cognitive stress.

As part of our sensitivity analysis, we used education, measured as school leaving age, as alternative indicator of SEP. We found little evidence of association between education and pre-stress baseline or other measures of reactivity to stress response. However, we did find weak evidence that more years of education was associated with a smaller decline in cortisol over the day, suggesting a less dynamic HPA axis though this was attenuated in our multi-level model adjusting for age and cortisol assay batch.

The present study has some important limitations. Our analysis was performed using a cohort of male-only participants from a high-income country with an age range of aged 64–80 years. While sex differences in HPA axis has been shown both in human studies and animal research (Heck and Handa, 2019), studies including both male and female participants revealed no sex difference in the association between SEP and measures of cortisol, including evening cortisol (Bann et al., 2015), cortisol awakening response (Wright and Steptoe, 2005) and average daily cortisol (Cohen et al., 2006a). Cortisol response to stress has also been shown to be sexually dimorphic, with higher cortisol response to psychological and psychosocial stress in man (Reschke-Hernandez et al., 2017), and evidence suggests that such differences dependent on the type of stressors (Stroud et al., 2002), with no differences in cortisol response between males and females observed in response to physical activity (Kirschbaum et al., 1992). A further limitation of our study, which is commonly observed in older population studies, is the potential for survival bias, that is participants in lower SEP may have had worst health outcome and did not survive, or were not fit enough, to participate in phase 5 of the study. If higher cortisol levels were also associated with survival, then survival bias would have resulted in an under-estimate of the true association and hence contribute to a null finding. Furthermore, cortisol levels have been shown to decline with ageing, therefore replication of this study in a cohort including both male and female participants with a wider range of age and ethnicity will be needed in order to generalize our findings to other populations. Our diurnal measures are based on only a single day and measures over multiple days would reduce measurement error though this is unlikely to have been differential by SEP. Furthermore, it is noteworthy that the inter-assay coefficient of variation in the cortisol radioimmune assay is 19%, thus affecting the accuracy of cortisol measurement across assays. Although we did adjust for cortisol batch in our models, a more accurate cortisol assay may have resulted in more precise cortisol measures and increased the quality of the evidence of association between cortisol measures and SEP. Finally, a major limitation is that we have undertaken many hypothesis tests and hence some “significant” results may reflect a type I error due to multiple testing and reflect chance rather than causality. We did not undertake any p-value adjustment (e.g. Bonferroni) as this approach also has its problems and may be overly conservative (Perneger, 1998). In our results section we have focussed on statistical significance, and we are aware of the problems of using an arbitrary 0.05 threshold (Sterne and Davey Smith, 2001, Wasserstein and Lazar, 2016) instead p-values should be interpreted as a continuum where decreasing values give stronger evidence against the null hypothesis.

The lack of marked social differences in cortisol profiles imply that this pathway cannot be a major source of social inequalities in health and disease. However, people in lower social class are more at risk to be exposed to natural daily stressors (Almeida et al., 2005). If our findings are causal, then there may be an interaction between social class and external stressors so that poorer individuals who are exposed to morning stressors may have a greater delay in recovery leading over time to chronic activation of the HPA axis. Over stimulation of the HPA axis can have different effects depending on the life course of this system (Ben-Shlomo et al., 2014), leading to altered feedback-loop mechanisms regulating the return of cortisol to normal levels after exposure to stress (Herman et al., 2016). These observations would be consistent with the hypothesis that dysregulation of cortisol reactivity to stress may contribute to the association between chronic stress exposure and health outcomes (Turner et al., 2020).

Declaration of Competing Interest

All authors declare that they have no conflicts of interest.

Appendix A. Supplementary material

Supplementary material.

.