Predicting long-term sickness absence from sleep and fatigue

Authors

Errata

This article is corrected by:

  1. Errata: Corrigendum Volume 17, Issue 1, 120, Article first published online: 11 February 2008

T. Akerstedt, Public Health Sciences, Institute for Psychosocial Medicine; Karolinska Institutet, Sweden. Tel.: 46 181 524 82041; fax: 46 181 5537 8900; e-mail: torbjorn.akerstedt@ipm.ki.se

Summary

Disturbed or shortened sleep is prospectively related to disease. One might also expect that sickness absence would be another consequence but very little data seem to exist. The present study used 8300 individuals in a national sample to obtain information on reports of disturbed sleep and fatigue 1 year and merged this with data on long-term sickness absence 2 years later. A logistic regression analysis was applied to the data with adjustments for demographic and work environment variables. The results showed that individuals without registered sickness absence at the start had a higher probability of entering a period of long-term (≥90 days, odds ratio [OR] = 1.24 with 95% confidence interval [CI] = 1.02–1.51) sickness absence 2 years later if they reported disturbed sleep at the start. The figure for fatigue was OR = 1.35 (CI = 1.14–1.60). When fatigue or disturbed sleep was separately excluded the OR increased to OR = 1.44 and OR = 1.47, respectively. Intermediate sickness absence (14–89 days) showed similar but slightly weaker results. The results indicate that disturbed sleep and fatigue are predictors of long-term absence and it is suggested that impaired sleep may be part of a chain of causation, considering its effects on fatigue.

Background

Poor sleep has been prospectively related to mortality (Kripke et al., 2002), cardiovascular disease (Nilsson et al., 2001) and diabetes (Nilsson et al., 2004). One might expect then a relation also to sickness absence, but the evidence is relatively scant even if cross-sectional studies show disturbed sleep among individuals on long-term sickness absence (Doi et al., 2003; Floderus et al., 2005; Ohayon and Smirne, 2002).

In one of the few prospective studies, Ihlebaek et al. (2007) studied a large increase in sickness absence from 1996 to 2003 in Norwegian national samples, but did not find any relation to disturbed sleep or any other changes in health indicators. Vahtera et al. (2006), however, found disturbed sleep after bereavement to predict long-term sickness absence. Also those not bereaved showed a 20% increase in risk of sickness absence if the disturbed sleep was present at the start of the study. Sleep duration was not a predictor.

In Sweden, long-term sickness absence (≥90 days) has doubled from 1993 to 2001 (Lindholm et al., 2005; SCB, 2004; Wikman, 2004) the main reasons being ‘reaction to stress’ or ‘burnout’, whereas musculo–skeletal diagnoses have dominated traditionally. Interestingly, the most obvious change in complaints across the corresponding time period has been disturbed sleep (SCB, 2004). The purpose of the present study was, therefore, to investigate the prospective relation between self-reported disturbed sleep in a national representative sample and the incidence of long-term sickness absence. No previous studies of this type seem to be available.

The term ‘long term’ in relation to sickness absence does not have any agreed-upon definition but recent custom in Sweden has been to use a period of ≥90 consecutive days to denote ‘long term’. In addition, also the sickness absence periods ≥14 days (but <90 days) have been included for comparison. Below that duration the employer takes the economical burden (instead of the health insurance system) and sickness absence figures are not reported to the health insurance system.

Apart from sleeping difficulties, ‘non-restorative’ sleep also constitutes one of the diagnostic criteria of disturbed sleep (AASM, 2005). Essentially, the term ‘non-restorative’ sleep refers to fatigue and similar states that cannot be remedied by sleep. Fatigue is also a very prevalent indicator of ill health in the population (Wessely et al., 1997) and it has been related to subsequent sickness absence in several studies (Bultmann et al., 2005; Huibers et al., 2004; Janssen et al., 2003). The latter studies used rather moderate criteria of long-term absence and did not involve national representative samples. The second purpose of the present study was, therefore, to analyse, in a national representative sample, the predictive power of fatigue in relation to difficulties sleeping in the development of long-term sickness absence. In view of the correlation between sleep and fatigue the present study looked at each predictor separately and as adjusted for each other.

The register data used in this work constitutes a sample of the biannual Swedish Work Environment Survey (AMU) from the year 2001 as this timing is close to the peak of the temporal development of sickness absence. Preliminary cross-sectional data from that study has indicated cross-sectional correlations between long-term sickness absence and fatigue (Lindholm et al., 2005).

Methods

Just over 14 000 members of the population in employment were selected for the survey, which is a supplement to the continuous Labour Force Survey. The respondents provided answers to introductory questions by phone in October and November 2001. Shortly afterwards, they received a postal questionnaire: Most people responded during the same year but a small number of questionnaires were completed in early January 2002. The reference period is generally the recent 3 months, except for a few questions concerning the recent 12 months.

The non-response rate in the postal questionnaire was about 25%; the overall non-response in the telephone phase was about 16% but lower for those employed. Efforts to compensate for non-response effects by adjusting sample weights were made similar to the procedure of poststratified weighting of groups. For the present analysis, data on sick leave were obtained from the Social Insurance Administration for each year 2001–2003 for each individual. These data include all sick leaves after the first 2 weeks of each sickness period, that is, for all sickness paid by the social insurance system, but not for shorter sickness absence periods.

Sickness groups were compared using logistic regression adapted to the complex survey design of stratified sampling. Population attributable risk (PAR) was estimated according to ([OR−1]/OR)% exposed cases, where OR denotes the odds ratio. The results were adjusted for demographic variables as well as for variables reflecting physical and mental work load and for work hours. The reason for the work-related variables was their strong position as predictors of sickness absence (Lindholm et al., 2005; Wikman, 2004).

As regards the outcome ‘sickness absence’, we made calculations on two different outcomes relating to the situation in 2003, sickness absence ≥14 days but <90 days (14–89 days) and ≥90 days, respectively. The analysis of the relationship between different predictors and outcome was restricted to subjects that had no registered sick leave (meaning <14 days) in 2001. The following groups of sickness absence were compared: Those who changed from 0 days in 2001 to 14–89 days in 2003 and those who changed from 0 days in 2001 to ≥90 days in 2003.

The analysis used fatigue and disturbed sleep as main predictors and was adjusted for a number of demographic factors. As the focus was on work environment factors adjustment was made also for factors representing physical work load, mental work load (stress) and work hours.

The following variables were used as main predictors: disturbed sleep because of thoughts of work (1 day per week or more often versus less often) and tired and listless (1 day per week or more often versus less often).

As demographic confounders were used: gender (female versus male), marital status (single versus married/cohabiting), age (16–29 years and 50–64 years versus 30–49 years), socioeconomic group (blue-collar workers: unskilled/skilled; white-collar workers: lower grades, intermediate grades, senior; and self-employed), child/children at home (yes versus no).

‘Awkward work posture’ was a combination of three questions: ‘working bent-over/leaning forward’ one-fourth of the time or more, ‘working in a twisted position’ one-fourth of the time or more and ‘working with your hands above shoulder height’ one-fourth of the time or more. The exposure was considered present if at least one of the questions had a positive response. The reference was ‘all other’. Physically heavy work was a combination of two questions: ‘physically hard work at least 50% of the time’ or ‘lifting 15 kg at a time several times per day at least once in a week’. At least one positive response was needed for a classification as exposed and the reference was ‘all other’. Work hours was categorized into part-time versus fulltime (>35-h week), Overtime work was categorized into yes versus no. Shift work was categorized into yes versus no.

The index job demands was based on four questions: ‘have to skip lunch, do overtime or bring work home’ every week or more often, ‘work is so stressful that you cannot talk or think of other things than work’ half the time or more often, ‘work demands all my attention and concentration’ almost all the time or more often, ‘too much to do’ agree fully or partly. Two or more positive responses were considered high work demands, with ‘all others’ as reference. Job control (low) was constructed from four questions: ‘can decide how fast to work’ half the time or less often, ‘most of the time (or never) not able to decide when tasks should be carried out’, ‘most of the time (or never) not participating in decisions about how my work should be organized’, ‘have too little or no influence over my work’ agree fully or partly. Low control was defined as three out of four positive answers and the reference was ‘all others’. Lack of social support: ‘have the possibility to get support and encouragement from colleagues when work becomes difficult’ never or usually not, ‘Have the possibility to get support and encouragement from my supervisor when work becomes difficult’ never or usually. Two positive answers were required.

Results

Table 1 shows the number of subjects (N) in various categories of each predictor, as well as the number of ‘cases’ with sickness absences with intermediate or long-term duration in these categories.

Table 1.   Change in sickness absence and number of subjects and cases (and %) in exposed/non-exposed categories
 Change from 0 to14–89 daysChange from 0 to 90+ days
Total N levelN casesPercentTotal N levelsN casesPercent
  1. Ref, reference category for odds ratios; total N, all subjects at that level (including cases).

  2. SEG, socioeconomic group.

Not disturbed sleep68246589.668841051.5
Disturbed sleep154219412.61542422.7
Not fatigue51644598.95164671.3
Fatigue283338413.62833782.8
Male41143077.54066461.1
Female375955814.841131493.6
Age 16–2926111385.31455151.0
Age 30–49 (Ref)41133999.74113631.5
Age 50–64261132812.62611712.7
SEG unskilled workers201326413.12013492.4
SEG skilled workers (Ref)125918214.51259372.9
SEG lower white collar11661048.91166161.4
SEG intermediate white collar19591809.21959291.5
SEG higher white collar482357.348261.2
SEG self-employed12911007.71291120.9
High social support49194859.94919761.5
Low social support326038011.63260732.2
Married/cohabiting545466812.359281943.3
Single184121611.71992653.3
No children430644710.44306821.9
Children387341810.83873671.7
Not heavy physical work54284999.25424821.5
Heavy physical work274436613.32744672.4
Not awkward work posture61966029.76196961.5
Awkward work posture187425313.51874522.8
Day work63806309.96380961.5
Shift work179723513.11797532.9
Fulltime work515955610.855541613.0
Part-time work158828417.91787854.8
No overtime work564566311.756451122.0
Overtime work24821977.92482361.5
Low work demands399141610.43991601.5
High work demands418844910.74188892.1
High work control365345612.53653742.0
Low work control45264099.04526751.7

Table 2 shows the results of the logistic regression using long-term sickness absence as the dependent variable. The univariate analysis shows that both disturbed sleep and fatigue have highly significant ORs for subsequent long-term sickness absence. Adjustment for demographic or work load factors reduces the OR somewhat, but it remains significant. Using only fatigue or only disturbed sleep as predictors yields somewhat higher ORs. The PAR of future ≥90 days of absence was 13.1% for disturbed sleep and 27.4% for fatigue.

Table 2.   Odds ratios (OR) disturbed sleep and fatigue against long-term sickness absence; also corresponding 95% confidence intervals (95% CI) N = 8426
 UnadjustedAdjusted for all predictors, but fatigue excludedAdjusted for all predictors, but disturbed sleep excludedAdjusted for all predictors
  1. Adjusted for: age, gender, marital status, number of small children, socioeconomic group, heavy physical work, twisted work posture, shift work, overtime work, fulltime/part-time work, high work demands, low work influence, overtime work.

0 to 14–89 days
 Disturbed sleep1.43 (1.21–1.70)1.44 (1.19–1.74) 1.26 (1.03–1.54)
 Fatigue1.59 (1.37.1.84) 1.46 (1.24–1.72)1.38 (1.16–1.65)
0 to ≥90 days
 Disturbed sleep1.86 (1.41–2.44)1.86 (1.36–2.55) 1.55 (1.09–2.18)
 Fatigue2.04 (1.58–2.62) 1.90 (1.44–2.61)1.69 (1.23–2.33)

Table 2 also shows the results of the logistic regression using intermediate (14–89 days) absence as the dependent variable. The univariate analysis shows that both disturbed sleep and fatigue show a highly significant OR for subsequent sickness absence. Adjustment for background or work load factors reduced the OR somewhat but it remained significant. Using only fatigue or only disturbed sleep as predictors yielded somewhat higher ORs. The PAR was estimated at 6.8% for disturbed sleep and 16.9% for fatigue.

Removing the work-related variables in the analyses above had only marginal effects, that is, the ORs changed with <0.08 units and all remained significant.

The analyses were repeated for each gender separately. In males, the OR and 95% confidence interval (CI) for long-term sickness absence was 1.19 and 0.84–1.69 for disturbed sleep and 1.41 and 1.04–1.90 for fatigue. For short-term sickness absence, the values were OR = 1.50 (CI = 0.80–2.81) for disturbed sleep and OR = 1.50 (CI = 0.84–2.67) for fatigue.

For females, the long-term sickness absence showed OR = 1.62 (CI = 1.07–2.45) for disturbed sleep and OR = 1.86 (CI = 1.27–2.71) for fatigue. Intermediate-term sickness absence showed OR = 1.32 (CI = 1.04–1.69) for disturbed sleep and OR = 1.37 (CI = 1.12–1.65) for fatigue.

The results were similar when either of the two predictors was removed.

Discussion

The results demonstrate that reporting disturbed sleep or fatigue in 1 year is associated with an increased risk of long-term sick leave 2 years later. The effect was less pronounced for intermediate-term sickness absence. The reason for this is not clear but, it is likely that intermediate-term sickness absence may have a more mixed set of causes than long-term absence.

Fatigue has previously been related to subsequent short or intermediate-term sickness absence (Bultmann et al., 2005; Huibers et al., 2004; Janssen et al., 2003) and burnout, as an extreme form of fatigue, is linked to prospective sickness absence (Borritz et al., 2005). In addition, we here also demonstrate that the longer-term sickness absence is predicted by fatigue. Logically, it seems reasonable that an inability to muster energy would be associated with a risk of being absent from work and it has been suggested that sickness absence may be a form of energy conservation (Hobfoll, 1998). This may also be an explanation of the considerably higher PAR for fatigue than for disturbed sleep. As discussed below, effects of the latter would probably have to be mediated by some intervening variable related to health perception.

Sleep has not been studied as a predictor of sickness absence before, but fatigued patients on long-term sick leave showed pronounced physiological sleep disturbances in terms of more microarousals, reduced sleep stages 3 and 4, and lower sleep efficiency (Ekstedt et al., 2006). Burnout patients on sick leave also report extreme sleep reduction towards the end of the process of becoming burned out (Ekstedt and Fagerberg, 2004). It is not clear to what extent such observations can be generalized to long-term sickness absence per se. However, sleep and fatigue often correlate highly in cross-sectional studies (Åkerstedt et al., 2004; Hossain et al., 2005) and reduced sleep duration involves gradual accumulation of sleepiness/fatigue (Van Dongen et al., 2003). Furthermore, experimental sleep reduction or studies of insomniacs show clear elevations of fatigue-inducing pro-inflammatory cytokines (Vgontzas et al., 2002, 2004a,b). Proinflammatory cytokines, in turn, appear to be important components of the sickness experience (Dantzer, 2001). Thus, even if much data are still missing it seems possible to hypothesize that disturbed sleep may be involved in the development of fatigue, which in turn, may lead to long-term sickness absence. At least, the possibility of such a link is worth further investigation.

Sleep and fatigue showed significant ORs even when work environment variables were adjusted for and removal of the latter only affected the OR marginally. Thus, other factors must have contributed. This could include, for example, family problems, or the presence of disease at the start of the study, even if the latter should have been limited through exclusion of subjects with registered sickness absence. Still, short-term sickness absence may have been present but was not available for analysis.

When the analyses were carried out separately for each gender the results were somewhat weaker and the prediction for disturbed sleep was not significant in males. However, this my have been caused by the strong reduction in N and the lower incidence of long-term sickness absence in men as well as the lower prevalence of disturbed sleep and fatigue. Possibly, a larger sample will yield significant results.

The present study has its strength in the prospective design and in the sample being nationally representative. Also, the dependent variable was obtained from official health insurance registers, which regulate the economical compensation for loss of salary. Thus, the absence data are probably reliable. However, limitations like the lack of a more sophisticated measure of disturbed sleep prevent analysis of what aspects of sleep that may have been most important. Also, the selection of potential confounding factors may not have been optimal. It would, for example, have been interesting to have had included variables such as smoking, body mass index (BMI) and alcohol consumption as covariates. However, these were not available in the survey database. This, however, does not need to affect the conclusion that fatigue and disturbed sleep are strong predictors of long-term sickness absence.

The implications of the present results may be that intervention/treatment to ameliorate sleep difficulties (Morin et al., 2005) or fatigue (Evengard and Klimas, 2002) might be able to reduce the amount sickness absence beyond any effects of improved working conditions. However, this remains to be demonstrated in future studies.

In summary, the results suggest that both disturbed sleep and fatigue predict new cases of long-term and intermediate-term sickness absence.

Acknowledgements

This work was supported by the Swedish Research Council for Working Life and Social Science and The Alecta AB Insurance Group.

Ancillary