Several authors reviewing the association between shift work and cardiovascular dysfunction1^–3 concluded that a causal relation might exist between shift work and ischaemic heart disease (IHD). According to Kristensen,1 a “relative risk (RR) of 1.4 must be regarded as the most reasonable estimate” and a later review2 reached the same conclusion. Harrington3 (page 70) noted that: “Causal mechanisms are not well defined but contributing factors include disruption of circadian rhythm, disturbed sociotemporal patterns, and social support, stress, smoking, poor diet, and lack of exercise”.

However the results from individual studies have not been consistent. For example, an OR of 0.9 (90% CI 0.7 to 1.2) for shift workers compared to non-shift workers was found in one UK study with no evidence of increasing risk with increasing duration of shift work.4 In a prospective cohort study5 in 14 different occupational settings in Copenhagen, an adjusted RR of 0.9 (95% CI 0.7 to 1.1) was found. On the other hand, the Helsinki Heart Study,6 found an RR of 1.4 (95% CI 1.0 to 1.9) for shift workers. Adjusted ORs of 1.3 (95% CI 1.1 to 1.6) for men and 1.3 (95% CI 0.9 to 1.8) for women were reported in a study of Swedish citizens.7

Contradictory evidence from different studies may be due to chance or variation in the amount of bias or confounding. For example, it has been suggested5 that some increased RRs might have been positively confounded by social class. Most studies reporting a positive relation between shift work and IHD have not been supported by a “dose-response” relation, which is an important consideration when assessing causality. Furthermore, some findings may reflect the effect of health-related “selection” both into and out of shift work, which could exaggerate, or attenuate, the relation between shift work and IHD.

In this study, we investigated the relation between shift work and IHD mortality, using methods similar to McNamee et al4 but in a different cohort of workers. We were particularly interested in investigating confounding by social class, evidence for a healthy shift worker hire effect (hire effect) or by a healthy shift worker survivor effect (survivor effect), and to control for these as far as possible when necessary.

METHODS

Subjects

A case-referent design nested within a male industrial cohort was used. The study cohort was defined as all men who began work as industrial workers aged 50 years or under between 1 January 1950 and 31 December 1998, at one site of a company producing and reprocessing nuclear fuel. To be eligible for the study, workers had to have worked there for at least 30 days. Job titles at the start of employment were used to exclude office workers and other “white-collar” workers as far as possible so as to restrict attention to “blue-collar” shift and non-shift workers.

Cases were cohort members who died of IHD (ICD 410–414), as determined from the code given by the UK Office of National Statistics on the death certificate, in the period 1950–98 aged 75 or under. For each case, a referent worker who was alive at the time of the case’s death was chosen from the cohort, matched on age and year of starting work with a maximum difference of five years being allowed. In the course of the study, 24 men were found to have been employed in 1948 or 1949 but they were kept in the analysis.

Shift work information

The intention was to ascertain work status (shift work or day work) from company records for each case on every day of his employment and, for referents, from the start of employment until either the end of employment, or to the date of death of the matched case, whichever was the earlier. The majority of work status information was extracted from personnel records where different pay codes had been used for shift and non-shift workers. Other company sources—occupational health records and dosimetry records—were used to fill in any gaps left by personnel information (see McNamee et al4 for details).

Work status information was incomplete for some periods for a small number of workers—in which case it was imputed using a variation of a bias-free method proposed by Weinberg.8 In Weinberg’s method for exposure concentrations, missing periods were filled by the arithmetic mean across all measured referent working periods. In our adaptation of this method, the proportion of total referent employment which was spent in shift work was calculated; gaps were then assigned at random to be shift work or not, with probability of shift work equal to this proportion.

In this study a shift worker was defined as someone who did shift work for a period of 30 days or more in total. Over the years of the study there were a variety of shift patterns: three-shift continuous, “forward rotation” with one week on and one week off; three-shift non-continuous (5 days a week, 42-hour shift rotation with weekends off), seven-day double-day shifts (mornings and afternoons) and five-day double-day shifts (mornings and afternoons). We were not able to obtain detailed shift working patterns for individual workers. Workers who never did any of these patterns were defined as day workers. The total number of years spent in shift work for the whole period of employment was computed. Shift workers were further classified by their status at the end of employment into current shift workers—that is, those who were still doing shifts at the end of their employment, and ex-shift workers. For shift working controls, this classification was made on the basis of their work status at the end of employment or the time of death of the matched case, whichever was earlier.

Confounding variables

There was no company policy on the selection of workers as shift workers but we considered the possibility that shift workers might have had a better cardiovascular profile at the start of employment than day workers. All employees had had a pre-employment medical examination, from which height, weight, systolic blood pressure (SBP) and diastolic blood pressure (DBP) were available for nearly all workers. Body mass index (BMI) was computed as weight (kg) divided by the square of height (m). Smoking status at pre-employment was available for more than 62% of workers. Only pre-employment measures of SBP, DBP, BMI, smoking were regarded as potential confounders, as any ill effect of shift work on heart disease might operate through changing some of these, in which case adjustment would be inappropriate.

Job titles or job codes, based on the OPCS Standard Occupational Classification9 10 (SOC), were extracted from personnel files. Of the 1270 job titles at first employment, 186 had not been given codes and were coded by us. Social class was inferred from job codes.9 10 Workers were classified into three social groups, class III (manual and non-manual skilled workers), class IV (semi-skilled), and class V (unskilled). Some workers (34 cases and 26 referents) were found to belong to social class II but to maintain study power, they were included in the analysis which further adjusted for social class.

To investigate any additional evidence of health-related selection into shift work—that is, a healthy shift worker hire effect, the relation between shift work and IHD was examined in several time windows corresponding to different periods of time-since-hire, following the approach of Fox and Collier11 12 for the general healthy worker effect.

To examine the evidence for a healthy shift worker survivor effect, date of death for cases was used to classify them by both time since termination of employment and time since leaving shift work; referents were classified in a similar way but using the date of death of the matching case. Time since leaving shift work was defined as the gap between date of ending the last shift period and the date of the (matching) case’s death. Long-term workers were defined as workers who were employed at least 20 years.

Significance and power

Assuming a prevalence of 0.50 for shift work, a true RR of death from IHD of 1.4 for shift workers and two-sided hypothesis testing with a significance level of 0.10 then, for 80% power, the required sample sizes would be 549 cases and 549 referents. A two-sided hypothesis test at a level of 0.10 is approximately equivalent to a one-sided test with a level of 0.05. Estimated RRs are accompanied by 90% CIs. A 90% CI is narrower than a 95% CI and a two-sided 10% significance level gives a more powerful test than with a conventional two-sided 5% significance level.

Statistical methods

Conditional logistic regression analysis was used to estimate ORs which, with incident density sampling of controls, can be interpreted as RRs.13 14 SBP, DBP, BMI and height were categorised before inclusion in the model. To maintain the power of the study, the small numbers with missing BP or BMI were assigned to missing categories and kept in the analysis. Smoking status was categorised as non-smoker, 1–9, 10–19, ⩾20 a day, smokers but number of cigarettes unknown, and unknown. To control for any residual hire effect, ORs restricted to those who survived 10 years after hire were also calculated.

To examine trend, duration of shift work was classified as “0” (day workers), <1 year, 1–4.9, 5–9.9, and ⩾10 years and included in the model. Duration was also included in the model as a continuous variable and a test for trend based on the regression coefficient with day workers (the zero level) excluded. ORs were also calculated for a subgroup (day and shift workers) with at least 20 years of employment, with shift workers further classified into those who had only worked shifts and a mixed day and shift work group.

Odds ratios were calculated for “active shift workers” (that is, working shifts at time of death) and for “inactive shift workers” further classified by time since leaving shift work (⩾1 day–4.9 years, 5.0–9.9, 10.0–19.9, ⩾20.0) To control for bias due to a survivor effect, we employed two approaches. Firstly we included duration of employment,15 categorised as <1 year, 1–4.9, 5–9.9, 10–19.9 and ⩾20 years as a predictor in the regression models. Secondly, time since termination of employment,16 categorised as active workers (baseline), <1, 1–4.9, 5–9.9, 10–19.9, and ⩾20 years, was included in regression models.

RESULTS

There were 635 men who died with cause of death coded as ICD 410–414. Age at death ranged from 27–75 years with a median of 63. Thirteen per cent of cases were employed for less than a year and 25% for 20 years or more. Twenty one per cent died while still employed by the company and 30% died more than 20 years after leaving. The median age at start of employment was 35 years for cases, and also for referents (table 1) as expected by matching. Median duration of employment was 10.5 for cases and 8.3 for referents whose employment history was “censored” at the date of death of matching case as necessary.

There was a mean follow-up—from first employment to death, or end of 1998—of 25.5 years per worker or 32 443 years in total for the 1270 workers. Work status was inferred from pay codes for 87% of workers, from pay code and other sources for 4.5% of workers and from other sources alone for 5.2%; it was imputed for the remainder (3.5%).

In total, 55% (n = 694) of subjects were defined as shift workers. Twenty three per cent (n = 295) of subjects worked only shifts and 32% (n = 399) worked both periods of days and shifts. Twenty per cent (n = 139) of shift workers worked less than two years of shift work, whereas over 42% worked for 10 years or more.

Among referents, the median duration of employment was 13.5 years for shift workers compared with 3.5 years for day workers, with corresponding values for cases of 14.4 and 5.1 years respectively. There was a strong correlation (r = 0.64, p<0.001) between duration of employment and duration of shift work.

Association between IHD mortality and shift work

The prevalence of shift work (at least 1 month) was 56% (n = 354)) among cases and 54% (n = 340) among referents. The crude OR of death from IHD for shift workers compared with day workers was 1.09 (90% CI 0.91 to 1.32). There was little evidence that shift workers were healthier than day workers at the start of employment (table 2); shift worker referents had slightly higher blood pressure than day worker referents but a lower proportion were overweight. More shift workers were known to smoke cigarettes than day workers. When pre-employment SBP, DBP, BMI, smoking and height were included in a regression model with work status, the OR for all shift workers increased to 1.10 (90% CI 0.91 to 1.34). The addition of duration of employment hardly changed the OR (OR = 1.09, 90% CI 0.89 to 1.32).

Association between IHD mortality and social class

The crude ORs of death from IHD for social class IV and V versus social classes III were 1.17 (90% CI 0.91 to 1.51) and 1.46 (1.07 to 1.98) respectively (table 3). Inclusion of work status (shift vs day), five risk factors and duration of employment for IHD hardly changed these results, suggesting that social class was an independent predictor of IHD in this cohort.

Social class is strongly related to lifestyle factors and shift workers were more likely to be in social classes IV and V at recruitment (table 2). When social class was included in a regression model with the five risk factors and employment duration, the OR decreased from 1.10 to 1.03 (90% CI 0.83 to 1.28).

Duration of shift work

The median (mean) years of shift work was 8.5 (11.1) and 6.6 (10.0) for cases and referents respectively. Overall, there was no evidence of increasing risk with increasing duration of shift work; however, if those who did shift work for less than one year were ignored, a pattern of increasing OR of death from IHD with increasing duration of shift is seen: the ORs, adjusted for six risk factors, were 0.79, 1.20 and 1.26 for 1–5 years, 5–10 years and ⩾10 years of shift work compared with day workers) (table 5, column 5). However none of the ORs was significantly different from one and an overall test for linear trend of increasing risk with increasing duration of shift was not significant (p = 0.22)

The healthy shift worker hire effect

A hire effect or bias would result if there was a selection mechanism (active or passive) whereby less healthy workers were less likely to take up shift work. Although we were able to control for several indicators of cardiovascular health at recruitment, including social class, this was possibly insufficient to control fully for any bias. The effect of any such selection process would be expected to decline with time since hire.11 There was a slightly reduced risk of death from IHD among shift workers compared to day workers in the first 10 years after employment (table 4).

This crude pattern could be interpreted as consistent with a small hire effect similar to that found in another cohort.4 However, when social class was added to the model, the pattern was in the opposite direction, more consistent with an unhealthy shift worker selection bias. Regardless of the direction of “selection-in” bias, Fox and Collier11 suggested that exposure effects be estimated from follow-up after discarding the early years after hire. Here, the OR for shift work restricted to the period 10 years after hire and adjusted for social class, the five risk factors and duration of employment was not very different (OR 1.04, 90% CI 0.83 to 1.30) from that without this restriction.

A further analysis was restricted to 312 long-term workers employed at least 20 years in the company: 95 did only day work, 150 mixed day and shift work, and 67 only ever shift work. The OR of death from IHD among all shift workers in comparison with day workers in this group was 0.88 (90% CI 0.48 to 1.60), and was 0.66 (90% CI 0.26 to 1.64) and 0.97 (90% CI 0.51 to 1.86) respectively for only ever shift work and mixed shift and day work respectively.

Healthy worker and healthy shift worker survivor effects

Healthier workers are more likely to remain in employment compared to those who develop disabling conditions; this is the healthy worker (employment) survivor effect, which will attenuate dose-response relations when dose is based on cumulative exposure. As discussed by others,11 16 17 the process underlying the healthy worker survival effect tends to lead to variation in mortality rates with time since termination of employment: observed risks are relatively low during active employment, highest shortly after termination of employment and thereafter decreasing with time since termination.

Given cardiovascular health problems, there might also be a tendency for shift workers to move from shift to day work rather than leave work entirely. Then one might expect to see relatively few deaths from IHD in those currently working as shift workers and a corresponding excess in the early years after leaving shift work, which would disappear as time since leaving shift work increased.4 With either phenomenon—movement to day work or leaving employment because of cardiovascular health problems—there would be a healthy shift worker survivor effect with accompanying attenuation of the true relation between duration of shift work and risk. The (non-significant) reduction in risk (OR 0.66) reported above for long-term shift workers compared to long-term day workers is consistent with a shift worker survivor effect. To look for further evidence, we examined patterns of mortality by time since leaving shift work in the whole study sample.

Time since leaving shift work and long-term shift workers

Of the 694 shift workers 582 (84%) were inactive and 112 (16%) active at the time of their death (or matching case’s death for referents). The adjusted OR of death from IHD for active shift workers compared with day workers was 1.10 (90% CI 0.70 to 1.74) when restricted to those who survived at least 10 years after first employment. It rose to 1.39 (90% CI 0.89 to 2.17) among inactive shift workers in the first five years after leaving shift work. There was no such excess of death beyond five years in inactive shift workers except for a non-significant excess in workers who left shift work after 10–20 years (results not shown).

Duration of shift work and risk, corrected for survivor effect

Richardson et al16 suggested the inclusion of time since termination of employment in regression models as a method for removing bias due to survivor effects on duration-response relations, while others15 suggested inclusion of duration of employment. Here, inclusion of duration of employment made little difference (table 5); for example the adjusted OR for 10 or more years of shift work became 1.28 (90% CI 0.94 to 1.74) instead of 1.26. Inclusion of time since termination of employment has only slightly more effect, with the OR for 10 or more years becoming OR 1.32, 90% CI 0.96 to 1.80. When the data were restricted to those who survived at least 10 years after hire, the OR changed slightly to 1.29 (table 5).

DISCUSSION

In this study a 4% excess risk of death from IHD was found in shift workers compared with day workers (OR 1.04, 90% CI 0.83 to 1.30), after adjusting for pre-employment SBP, DBP, BMI, smoking, height, social class, duration of employment and when restricted to workers who survived at least 10 years after hire. The corresponding excess risk would be 11% (OR 1.11, 90% CI 0.90 to 1.37) if no adjustment was made for social class. No relation between duration of shift work and risk of death from IHD was found. The best estimates of the OR of death from IHD for the different levels of duration of shift work, <1, 1.0–4.9, 5.0–9.9, and ⩾10 years after hire to shift work, respectively, were 1.46, 0.81, 1.13 and 1.29 after adjusting for all above mentioned risk factors and controlling for selection effects as far as possible. These estimates do, however, suggest a slight shift duration response relation if those who did less than a year of shift work are ignored. None of the OR estimates was significantly greater than one although the unadjusted OR of 1.27 for 10 or more was of borderline significance.

This study used a nested case-referent design, in which all subjects were chosen from an industrial cohort. Office workers and white-collar workers were excluded as far as possible (although 34 of the latter were inadvertently included) in an effort to compare shift workers, who were all blue-collar workers, with men of a similar social class. This is important because the risk of IHD is linked to both job grade and social class.18 Despite this restriction to industrial workers, social class, as inferred from job titles, was different for shift workers and day workers and it was an independent predictor of death from IHD, with estimated OR of 1.16 (90% CI 0.91 to 1.48) and 1.44 (90% CI 1.10 to 1.90) for social class IV and V respectively compared with social class III. Thus, further adjustment for social class was necessary and reduced the OR of death from IHD for shift workers, from 1.11 to 1.04 (90% CI 0.81 to 1.30).

There was no evidence of a healthy, or unhealthy, shift worker hire effect as judged either by the pre-employment risk factors in this study or by restricting analysis to those who survived 10 years after hire. Elsewhere Fox and Collier11 and McNamee et al4 showed that the healthy hire effect on mortality had disappeared after 15 and 10 years of follow-up respectively.

Some evidence for a survivor effect was shown by the OR for those who had worked at least 20 years in shift work, compared with long-term day workers (adjusted OR 0.66, 90% CI 0.15 to 2.89), although the comparison was hampered by small numbers. On the other hand, the OR for 20 or more years of mixed shift and day work was 1.24 (90% CI 0.40 to 3.82) which could be explained if workers left shift work to do day work because of developing cardiac-related problems. Despite this suggestion of a shift worker survivor effect which, if uncontrolled, would tend to attenuate any dose-response relation, attempts to control for it using methods suggested in the literature made little difference to the ORs of death from IHD for different levels of duration of shift work.

In this cohort-based study, work status information was extracted from details held in the personnel records of the workers, blind to the knowledge of whether the subject was a case or referent. Thus, the source of information was identical for both cases and referents. Therefore, shift work measurement was free of recall bias and interviewer bias.

The classification of subjects by duration of shift work should have been relatively free of measurement error because the pay codes on which it was based, for 85% of workers, determined a worker’s wage. However, some clerical errors in transferring information to text files cannot be ruled out and some misclassification for other workers may have occurred. To assess the impact of possible misclassification, the analysis was restricted to 490 pairs of cases and matched referents with complete shift work information from pay codes only. The adjusted OR controlled for all pre-employment risk factors, social class, duration of employment and restricted to survivors 10 years after hire, was 1.03 (90% CI 0.79 to 1.34) which is almost identical to the estimate based on all the data. Therefore, it seems unlikely that error due to using sources other than pay codes contributed to the lack of an association between shift work and IHD.

On the other hand, we did not have any information about work status when working for other employers. It is possible that some day workers might have done shift work in other employment in which case they should really have been classified as day workers. The effect of any such misclassification would be to bias the measured effect of shift work towards the null value. The comparison restricted to workers with at least 20 years employment would have partly addressed this problem but given no evidence of greater effects of shift work, the OR for shift workers compared to day workers was 0.88.

In addition to social class, pre-employment SBP, DBP, BMI, smoking and height were measured and there appeared to be little difference between shift and day workers. However there was no information on cholesterol, diabetes and level of physical activity. However adjustment for social class will, to some extent, have controlled for such differences.

Kristensen1 identified several occupational factors other than shift work linked to cardiovascular disease in the research literature: these include excessive cold or heat, stress and noise. According to the former chief medical officer of the company (personal communication), there were no extremes of heat or cold particularly associated with shift work. Some workers were exposed to heat because of wearing protective clothes, but this was the same for both shift and day workers. Also, employee-employer relations were said to be “good” and work-related stress was unlikely to be an important hazard. Organisation of work was similar for both shift and day workers and they did similar tasks. There was exposure to ionising radiation, but this has never been linked with increased risk of heart disease. Workers were exposed to noise in some degree. In another study at this company,19 mean time-weighted average noise levels across employment for shift and day workers were significantly different (p<0.001) with shift workers having the lower value (83.6 dB (SD 5.1) vs 85.2 dB (SD 5.4)). Inclusion of noise as a covariate in the analysis did not greatly change the OR of death for shift workers in comparison with day workers (OR 1.05, 90% CI 0.84 to 1.32).

The findings in this project agree with those of McNamee et al4 who found an RR of 0.90 for shift workers employed at another site of the same company, and with Boggild et al18 who, in a prospective cohort study, found an adjusted RR of 0.90 (95% CI 0.7 to 1.1) for shift workers. However, our results contradict other literature. Contradictory results might be reconciled by considering possible causal pathways for shift work effects—for example, if shift work exerted an effect only through “stress” (that is, activation of the sympathetic and endocrine systems) then conceivably there might be no effect when workers could freely elect to go into or leave shift work. Another possible explanation is that studies which have found effects may have failed to control adequately for social class.5

CONCLUSION

After taking account of confounding, including by social class and by other work exposures, and of selection effects, we found only an overall 4% excess risk of death from IHD for shift workers. There was no evidence of a dose-response relation with duration of shift work, and no difference between workers with at least 20 years shift work and those with at least 20 years of day work. A weakness of the study is that it could not address lifetime shift work exposure. We recommend that shift work researchers pay close attention to the possibility of confounding by social class-related risk factors.