Data on sickness absence are suggested to be a good measure of health in working populations, when health is understood in terms of physical and social functioning.1^–3 Associations between sickness absence, chronic conditions and early retirement on medical grounds have been reported in various settings4^–12 and, recently, three large studies showed that long-term absences are related to future mortality.3 13 14 Thus, sickness absence appears to be a predictor of future health in working populations.

There have been repeated calls for increased understanding of prognostic factors that can be used to estimate the chance of recovery from a disease or the chance of an unfavourable disease outcome.15 We hypothesise that sickness absence may be such a prognostic marker for employees with chronic conditions. The future course of a disease is the net effect of a wide range of factors including premorbid biological disposition, disease severity, co-morbidity and health-related behaviours, as well as demographic and socioeconomic circumstances. Sickness absence may well capture the full array of diseases employees experience during their job contract and it is also known to be associated with health behaviours, socioeconomic circumstances and other exposures.1 16^–20

For chronic conditions of major public health relevance, such as cancer, coronary heart disease (CHD) and diabetes, mortality is a widely used prognostic outcome. A limitation in previous studies of sickness absence is that they provide little information on whether the level of sickness absence is predictive of death among those with a common chronic condition. If this were the case, sickness absence may prove a useful additional indicator of prognosis for such conditions. We sought to examine this question in the current study. We hypothesised that sickness absence is a good prognostic indicator because absence records provide day-to-day data on functioning in occupational settings and thus capture the presence of co-morbidity, illness severity and the effects of risk factors.

METHODS

Study population

We used data from the GAZEL cohort which was established in 1989 and comprised 20 625 employees aged 35–50 from France’s national gas and electricity company Electricité de France-Gaz de France (EDF-GDF).21 Since baseline, less than 1% of GAZEL cohort members have been lost to follow-up. In the present study, we included those alive and working for EDF-GDF in 1990–1992 and excluded those who had died, retired or left the company before 1 January 1993 (n = 1390), were non-responders in any of the years 1990–1992 (n = 1270) or had missing data on age, sex or socioeconomic position (n = 17). Thus, a total of 17 948 participants (13 077 men, 4871 women) aged 37–51 at baseline were included in the analysis.

Sample characteristics

Baseline variables obtained from EDF-GDF records included sex, age and a three-level employment grade (high, intermediate and low) derived from the French National Statistics Institute (INSEE; see http://www.insee.fr/fr/nom_def_met/nomenclatures/prof_cat_soc/html/L03_N1.HTM).

Assessment of sickness absence

All employees at EDF-GDF are covered by the same sickness insurance scheme entitling them to compensation for work incapacity due to disease or injury. Every sickness-absence episode requires a medical certificate. If a sick-leave spell lasts for more than 1 year a certificate from an occupational health physician is required every 6 months. The benefit is equal to normal salary and, like the old age pension, is provided by the company. After 5 years of sickness absence the person is granted a disability pension with benefits that are at the same level as for an old age pension. There is no limit to the duration of a sick-leave spell and persons are not transferred to a disability pension before the end of the 5 years even if the inability to work is permanent.

We obtained computerised sickness-absence records from 1 January 1990 to 31 December 1992. These records included the first and last dates of all absences. For each employee, we computed the total number of sickness-absence days and classified this based on the distribution into categories of 0, 1–28 (4 weeks) and >28 days annually. We also computed for each employee the number of sickness-absence episodes lasting 1–7 days, 8–14 days and >14 days (long absence episodes). These indices of sickness absence were further categorised by frequency, as in previous studies.3

Assessment of chronic conditions

Prevalent common chronic conditions, known to predict mortality, were derived from the annual GAZEL questionnaires for 1990, 1991 and 1992 using a checklist of common chronic conditions including cancer, diabetes, depression, chronic bronchitis or asthma, CHD or cerebrovascular disease, and chronic hypertension. Participants were deemed to have all the conditions checked in any of the three surveys.

Mortality follow-up

Follow-up for all-cause mortality from 1 January 1993 to 25 February 2007 used data from the EDF-GDF human resources department and the retirement fund services, which provides comprehensive mortality tracing for all participants resident in France and those who have emigrated but receive a pension from EDF-GDF.

Statistical methods

Each participant was followed either until death or to the end of the study period. We calculated Cox proportional hazard models to examine the associations of sickness absence and chronic conditions with mortality. Hazard ratios and their 95% confidence intervals (CI) were adjusted for age, sex and employment grade. Associations between sickness absence and mortality were additionally adjusted for chronic conditions. Those with no sickness absence or free from the chronic condition of interest were used as the reference. To confirm that observed associations were not confounded by an excess rate of sickness absence shortly before death, in a subsidiary analysis we excluded those employees who died during the first 5 years and started mortality follow-up from 1 January 1998. We calculated receiver operating characteristic (ROC) curves and their respective areas under the curve (AUCs) to further assess the ability of sickness absence to predict mortality.22 AUCs and 95% CI were used to test the null hypothesis that the theoretical AUC is 0.5. The status of sickness absence as a long-term prognostic factor for mortality was studied in analyses confined to participants with chronic conditions. For each chronic condition the AUC is used to examine the ability of sickness absence measures to distinguish between persons with poor versus good prognosis. ROC statistics were calculated with Stata 9.1 (Stata, College Station, TX, USA) and all other analyses with the SAS v 9.1.3 statistical software (SAS Institute, Cary, NC, USA).

RESULTS

Population characteristics

The mean age of the 13 077 men and 4871 women was 45.0 years (SD 3.4, range 37–51 years) in 1990. A total of 25% were in a high employment grade, 57% an intermediate grade and 18% a low grade. Between 1990 and 1992, we identified 620 405 sickness-absence days. These came from 32 583 sickness-absence episodes in 53 807 person-years. According to the annual surveys in 1990–1992, the 3-year prevalence of cancer was 1.2%, diabetes 2.4%, depression 14.4%, chronic bronchitis or asthma 9.0%, CHD or cerebrovascular disease 3.5%, and hypertension 13.1%. In total, 22.3% of participants had at least one chronic condition and 7.1% more than one condition. During the mean follow-up period of 13.9 years (SD 1.5, range 0.008–14.2 years) from 1993 to 2007, 803 (4.6%) participants died.

Chronic conditions and mortality

All chronic conditions are associated with increased mortality. The hazard ratios (95% CI) from Cox models adjusted for age, sex and employment grade are: 5.23 (3.76 to 7.29) for cancer, 1.43 (1.00 to 2.05) for diabetes, 1.49 (1.24 to 1.80) for depression, 1.77 (1.45 to 2.15) for chronic bronchitis or asthma, 2.07 (1.59 to 2.70) for CHD or cerebrovascular disease, and 1.46 (1.22 to 1.74) for chronic hypertension. The AUCs vary between 0.51 (95% CI 0.50 to 0.51) for diabetes and 0.53 (95% CI 0.52 to 0.54) for chronic bronchitis/asthma and hypertension.

Sickness absence and mortality

Associations between the sickness-absence measures and mortality are presented in table 1. After adjustment for age, sex and socioeconomic position, a large number of sickness-absence days (defined as more than 28 days per person-year) versus no sickness-absence days is associated with a 2.65-fold excess mortality, the corresponding hazard ratio for a large number of sickness-absence episodes lasting >14 days (defined as five or more episodes per 10 person-years) being 2.10. Both these associations remain after adjustment for all chronic conditions. Excluding participants who died in the first 5 years of follow-up (n = 206) reduces these associations, but they remain statistically significant. Rates for absence episodes lasting 1–7 or 8–14 days are weaker predictors of mortality. Thus, subsequent analyses are confined to the total number of sickness-absence days and the number of long absence episodes (>14 days). These measures are not exclusive of each other: 396 (2.2%) participants were in the high category for absence days, 594 (3.3%) in the high category for long absence episodes, and 1255 (7.0%) in the high category for both.

As an indication of the generalisability of these absence measures, taking >28 sickness-absence days annually is associated with excess mortality in men (age- and employment grade-adjusted hazard ratio 2.91, 95% CI 2.33 to 3.63) and women (1.64, 0.99 to 2.65), in those below (3.07, 2.18 to 4.33) and above (2.45, 1.90 to 3.04) the age of 45 years, and in the low (4.01, 2.60 to 6.19), intermediate (2.35, 1.80 to 2.93) and high (1.71, 0.79 to 3.69) employment grades. Corresponding hazard ratios for having more than five long absence episodes per 10 person-years varied between 1.56 and 2.65. All p values for interactions between the two sickness-absence measures and age, sex or grade exceeded 0.05.

Sickness absence and disease prognosis

Taking >28 sickness-absence days annually is associated with an increased risk of death in specific groups, such as those suffering from cancer, depression, chronic bronchitis or asthma; in participants with hypertension, the age-, sex- and grade-adjusted hazard ratios varied between 2.0 and 4.9 (table 2). Further adjustment for co-morbidity strengthens this association among the cancer patients, reduces the association among those with chronic bronchitis/asthma or hypertension, and has little effect among those with depression; in all cases the association remains statistically significant. Findings for taking more than five long absence episodes per 10 person-years and subsequent mortality are largely similar to those for the number of sickness-absence days (table 3). Both these measures of sickness absence are less consistently associated with mortality in patients with diabetes, CHD or cerebrovascular disease.

According to the AUC statistics (tables 2 and 3), the total number of sickness-absence days discriminates mortality risk for all chronic conditions except diabetes and CHD/cerebrovascular disease. The AUC statistic is particularly good at distinguishing poor versus good prognosis for cancer (AUC 0.73, 95% CI 0.63 to 0.82). Again the findings for the number of long absence episodes largely replicate these results, except that long absence episodes are able additionally to predict mortality for those with CHD or cerebrovascular disease.

Generalisability and long-term prediction of death

Men and women with cancer may have very different risks of death as women mainly get breast cancer, which has a much better prognosis than lung cancer, a common cause of early death in men of working age in France.23 In spite of this, there is no sex difference in associations between sickness-absence days, long sickness-absence episodes and mortality among those with cancer (p for interaction >0.35) or other chronic conditions (p>0.54 for diabetes, p>0.26 for depression, p>0.74 for chronic bronchitis or asthma, p>0.58 for CHD/cerebrovascular disease and p>0.70 for hypertension). In the same vein, we found no strong evidence of differences in these associations by age group (p for interaction >0.07) or socioeconomic position (p>0.37). Thus, the associations between sickness-absence measures and mortality in employees with chronic conditions seem to be generalisable across demographic and socioeconomic groups.

To examine whether sickness absence is a long-term predictor of mortality among employees with chronic conditions, we excluded those who died during the first 5 years of follow-up (table 4). High numbers of sickness-absence days and long sickness-absence episodes are strongly associated with mortality among employees with cancer and chronic bronchitis or asthma. Among hypertensive employees, the association remains similar to that observed before the exclusions but because of lower numbers does not reach statistical significance. The magnitude of the association between absence days and mortality is slightly reduced among depressive employees while that of long absence episodes is considerably reduced.

Co-morbidity

Of the participants, 4844 (27.0%) have one chronic condition, 1136 (6.3%) two chronic conditions, 194 (1.1%) three conditions and 26 (0.1%) four conditions. The associations of sickness absence with mortality are not different for participants with one chronic condition compared to those with co-morbidity (ie, two or more chronic conditions) (table 5).

DISCUSSION

The principal finding of this study is that data on sickness absence may provide important prognostic information for several common conditions of major public health relevance. A high level of sickness absence was associated with a two- to fivefold excess mortality among employees with cancer, depression, chronic bronchitis, asthma and hypertension. This association was robust for two simple measures of sickness absence, the total number of annual sickness-absence days and the number of sickness-absence episodes longer than 14 days. There was no evidence of significant variation in the magnitude or direction of these associations between men and women, young and old employees, those at different levels of the socioeconomic hierarchy, or those with and without co-morbid conditions.

To the best of our knowledge, this is the first systematic examination of sickness absence as a potential tool for the estimation of disease prognosis. AUCs for sickness-absence measures were 0.7 among employees with cancer, 0.6 among those suffering from chronic bronchitis or asthma, and slightly lower for depressive and hypertensive employees (but 95% CIs did not include 0.50). These figures provide moderately strong support for the ability of sickness absence to distinguish individuals at increased mortality risk within specific chronic conditions (AUC for a perfect predictor is 1.0). The magnitude of the predictive power is comparable to those of two established, widely-used clinical predictors, the Framingham risk score and the metabolic syndrome in relation to incident CHD and type 2 diabetes (AUCs between 0.6 and 0.7).24 25

Exclusion of deaths from the first 5 years of follow-up barely affected the association between sickness absence and mortality in people with cancer, chronic bronchitis/asthma and hypertension but reduced this association in depressive employees. Depression contributes to increased risk of sickness absence, but the status of depression may fluctuate over time and thus the association between sickness absence and mortality may be particularly open to effect dilution among depressive employees. The reason why sickness absence is predictive of mortality in the depressed is not clear. Co-morbid chronic diseases seem an unlikely explanation as adjustment for all of the chronic conditions we studied hardly attenuated the association between sickness absence and mortality in this group. However, unhealthy behaviours, such as smoking, substance abuse and self-harm, are a potential candidate because they contribute to both sickness absence and mortality, and there is a considerable variation in such behaviours among depressed people.26^–29

We found no attenuation of the sickness absence–mortality association among cancer patients after the effect of other diseases was taken into account. In contrast, this adjustment reduced the association among hypertensive employees and those suffering from chronic bronchitis or asthma. Indeed, the prognosis is obviously worse when hypertension is secondary to CHD, stroke or diabetes, or when hypertension is combined with depression. Similarly, these and other conditions are likely to contribute to sickness absence and mortality among people with chronic bronchitis or asthma.

Our study on all participants, irrespective of whether they had chronic conditions or not, adds to evidence suggesting that long-term sickness absence can be used as a global measure of health differentials between employees.3 4 9 13 With over 800 deaths, this is the largest study of sickness absence, disease and mortality to date and the observed sickness absence–mortality associations in the total cohort are in keeping with results reported in recent smaller scale studies.3 13 Moreover, associations remained after excluding the first 5 years of follow-up, indicating that excess sickness absence shortly before death is unlikely to explain our findings.

Methodology

Assessment of common chronic diseases based on self-reports is a potential source of misclassification bias. However, according to previous studies, the agreement between self-reports and medical records is high for diabetes (total agreement 97%), hypertension (88%), myocardial infarction (98%) and stroke (98%),30 and 87% of self-reported asthma patients have been verified as having chronic asthma.31 Self-reported cancer and depression are likely to be affected by under-reporting because of the associated stigma, lack of knowledge or denial.32^–34 However, as we focussed on those people who did report these conditions, there is probably little risk of misclassification in the analysis of prognosis.

We obtained sickness absence information from the employers’ records, but such data may be difficult to obtain for cohorts with employees from a large number of different workplaces. Previous studies on working populations suggest that the agreement between self-reported and recorded sickness-absence days is generally good and in the Whitehall II study self-reported and recorded data predicted mortality equally well.2 14 This suggests that it should be possible to use self-reports of sickness-absence days when absence records are not available. Further research is now needed to confirm the validity of self-reported sickness-absence for specific chronic conditions, as the possibility of using self-reports would simplify the collection of sickness-absence data in epidemiological studies of general populations.

Conclusion

Data from a large contemporary cohort of French employees suggest that information on sickness absence may be helpful when the prognosis for common chronic conditions at a population level is being estimated. Our findings provide a basis for considering increased use of data on sickness absence when researching prognosis in working populations.