Our project will look at rates of ischemic heart disease (IHD) among Danish employees as a function of weekly working hours and nighttime work.

From the viewpoint of cardiovascular risk factors, there appear to be both advantages and disadvantages of long working hours and nighttime work.

One of the advantages of nighttime work is that it usually eliminates exposure to rush hour commuting stress, both to and from work; such stress has been associated with psychological strain [1,2], increased blood pressure [3], and increased rates of acute myocardial infarction [4,5]. Moreover, a survey of more than 400 people in United Kingdom indicated that 44% of people believed that rush hour traffic was the single most stressful part of their life [6]. Daytime workers may reduce their exposure to rush hour traffic, at least in one direction, by choosing to work long hours.

Another potential advantage of nighttime work and long working hours is that they can generate extra income, compared with ordinary daytime work, and thereby reduce the risk or intensity of financial strain. An increased income has been associated with a decreased risk of IHD [7], while financial strain has been associated with an increased risk of hypertension [8] and acute myocardial infarction [9].

The disadvantage of long working hours and nighttime or shift work is that they usually are associated with short sleep duration, mismatch of circadian rhythm, social disruption, and behavioral changes [10,11], which in turn are associated with an increased risk of IHD [12]. The mechanism may be related to disturbed regulation of inflammatory, metabolic, and cardiovascular processes. Thus, short-term sleep restriction increases heart rate [13] and blood pressure [14], changes immune function [13], increases C-reactive protein concentration [13], and decreases glucose tolerance and sympathetic nervous system activity [15].

The evidence of an association between night or shift work and IHD has been reviewed by Frost et al [16], Ha et al [17], and Vyas et al [18]. Frost et al [16] reviewed 14 research papers and concluded that “the available evidence concerning the influence of the type and duration of shift work, as well as sex, on the risk of IHD is too limited to permit any conclusions on these issues." In a meta-analysis that combined the results from 8 studies, Ha et al estimated the effect of night or shift work versus ordinary daytime work, first at a rate ratio (RR) equal to 1.17 (95% CI 1.00–1.37) and then, after adjustment for publication bias, at RR=1.12 (95% CI 0.94–1.33) [17].

Vyas et al [18] reviewed 35 research papers and used the results of the papers in meta-analyses, which estimated the RR of night or shift work versus ordinary daytime work at 1.24 (95% CI 1.10–1.39) for coronary events, 1.23 (95% CI 1.15–1.31) for acute myocardial infarction, and 1.05 (95% CI 1.01–1.09) for ischemic stroke. However, 87% (1,750,565/2,011,935) of the participants in the meta-analyses came from ecologic studies, which of course makes the estimates less certain than their confidence intervals imply.

Virtanen et al [19] investigated the relationship between long working hours and coronary heart disease in a systematic review and meta-analysis, which included results from 12 research papers. The RR for coronary heart disease among workers with “long" versus “normal" working hours was estimated at 1.80 (95% CI 1.42–2.29) when all studies were included and 1.39 (95% CI 1.12–1.72) when restricted to prospective studies. The cut points for long weekly working hours ranged from >40 to >65 hours.

The reviews by Vyas et al [18] and by Virtanen et al [19] suggested that the disadvantages of nighttime work and long working hours tend to outweigh the advantages of such work time arrangements. However, since the meta-analyses only included published studies, we cannot rule out publication bias. The risk of publication bias is especially high if the included studies are too small to be of interest unless the results are statistically significant [20], and this is definitely the case with the studies that Virtanen et al [19] included in their meta-analysis.

Kivimaki et al [21] conducted a more reliable meta-analysis of the relationship between long working hours and IHD, in which they mitigated the risk of publication bias by conducting and inserting a series of new and hitherto unpublished studies. By combining the results from the unpublished studies with corresponding results from published studies, they obtained the following RRs for long versus normal (35–40) working hours a week: 1.02 (95% CI 0.91–1.15) for 41–48 hours, 1.07 (95% CI 0.92–1.24) for 49–54 hours, and 1.08 (95% CI 0.92–1.24) for >54 working hours. These RRs are most probably the best estimates ever of the strength of association between long working hours and coronary heart disease. We should, however, keep in mind that meta-analyses that include published results never can be done blinded. The researchers often know the literature before they formulate inclusion criteria and, however systematic and objective they attempt to be, it can never be ruled out that the selection of papers as well as the selection of estimates within the papers might be influenced by their a priori expectations. From this viewpoint, it would be good to know whether the null finding by Kivimaki et al would be reproduced if it were tested in an equally sized blinded study, that is, a study in which the inclusion criteria, hypotheses, and statistical models were completely specified and published before any results were available.

Kivimaki et al [21] based their RRs on a total of 2481 new cases among 129,301 participants. Our project has access to person-based data on work time arrangements for approximately 160,000 employees with long or normal working hours, which, upon follow-up, we expect to yield approximately 3200 new cases of IHD. The exposure data of the project have not been linked to health data before, and this provides an opportunity to study the relationship between work time arrangements and subsequent IHD, with a remarkably high statistical power, completely blinded.

We want to know whether the incidence of antihypertensive drug usage and the incidence of hospital treatment or death due to IHD are independent of weekly working hours and nighttime work among full-time employees in Denmark, and will address these research questions in a series of nested hypothesis tests (Textbox 1).

We will set the overall significance level for the effect of weekly working hours at .05 and we will set the overall significance level for the effect of nighttime work at .05. We will solve the multiple testing problems by the following strategy:

Hospital treatment or death due to IHD is the primary outcome of the study, and a statically significant association with this outcome would afford direct statistical evidence of an association with IHD.

Hypertension plays an important role in the etiology of IHD [22], and relative rates of antihypertensive drug usage have been shown to be highly correlated with relative rates of IHD among occupational groups in Denmark [23,24]. We will therefore regard results obtained for antihypertensive drug usage as indirect statistical evidence of an association with IHD if they are statistically significant and show a similar pattern to the results obtained for hospital treatment or death due to IHD.

The study will comply with The Act on Processing of Personal Data, Denmark (Act No. 429 of May 31, 2000), which implements the European Union (EU) Directive 95/46/EC on the protection of individuals. The data usage is approved by the Danish Data Protection Agency, file number 2001-54-0180. The ethical aspect of the project was examined and approved by Statistics Denmark, account number 704291.

The data base of the project will consist of interview data from the Danish Labour Force Survey 1999–2013, which are linked to data from the central person register [25], the employment classification module [26], the national patient register [27], the cause of death register [28], and the national prescription register [29]. The linkage will be based on the participants’ personal identification numbers.

The Danish Labour Force Survey has been conducted since 1994, in accordance with EU directives, which apply to all member states of the EU. It is based on random samples of 15- to 74-year-old people in the Danish population. The samples are drawn quarterly and the participants are invited to be interviewed 4 times over a period of one and a half years. The structured interviews, which are done by telephone, cover various aspects of labor market participation, including specifications on working hours and work schedules [30]. The response rate is currently 65% [30].

The central person register contains information on sex, addresses, and dates of birth, death, and migrations for every person who is or has been an inhabitant of Denmark sometime between 1968 and the present. A person’s socioeconomic status (SES), occupation, and industry have been registered annually in the employment classification module since 1975. The national hospital register has existed since 1977 and contains data from all public hospitals in Denmark (>99% of all admissions). From 1977 to 1994, the register only included inpatients, but from 1995 it has also covered outpatients and emergency ward visits. Since 1994, the diagnoses have been coded according to International Classification of Diseases, Tenth Revision (ICD-10) [31]. The national prescription register covers all redeemed prescriptions at pharmacies in Denmark since 1995, and the products are coded in accordance with the Anatomical Therapeutic Chemical Classification System (ATC).

The labor force surveys gather person-based information on weekly working hours, calculated by adding the hours worked in secondary jobs to the ones worked in a primary job. The participants are asked first how many hours they usually work and then how many hours they worked during the reference week (a predetermined work week, which occurred 1–4 weeks prior to the interview). They are also asked whether and to what extent they work at night. The questions used to gather this information have changed slightly with time. Before 2001, there was no mention of whether meal breaks should be counted as working hours. During 2001–2006, all participants were instructed to exclude meal breaks when they counted their work hours. As of 2007, the time used for meal breaks is to be counted if the person was paid while eating and is to be excluded otherwise. Another peculiarity that was introduced in 2007 is that the participants are asked whether the weekly working hours vary a lot or there are other reasons that make it difficult to provide a meaningful estimate of usual weekly working hours. If they answer “yes" to any of these questions, then “average working hours" is to be used as a proxy for “usual working hours."

Before 2001, the participants were simply asked whether they worked at night, but from 2001 onward the question has been whether they worked at night during the last 4 weeks. Until 2006 the response categories were “yes, regularly," “yes, occasionally," and “no, never". From 2007 onward the response categories were expanded to “yes, regularly" (ie, more than half of the working days in the last 4 weeks), “yes occasionally" (ie, at least once within the last 4 weeks, but less than half of the working days), and “no, not within the last 4 weeks."

We will disregard the changes in the data collection routines in the primary analyses of this project. We will define the exposure variables as follows.

The primary end point is hospital treatment or death, with IHD as the principal diagnosis or cause of death, respectively. The case definition includes the following ICD-10 codes: I20 angina pectoris, I21 acute myocardial infarction, I22 subsequent myocardial infarction, I23 certain current complications following acute myocardial infarction, I24 other acute IHDs, I25 chronic IHD. The secondary end point is redemption of a prescription for antihypertensive drugs. The following ATC codes are included: C02 antihypertensives, C03 diuretics, C07 alpha- and beta-blockers, C08 calcium channel blockers, and C09 angiotensin-converting enzyme inhibitors and angiotensin-II antagonists.

The participants will be followed from the beginning of the calendar year that succeeds that of their baseline interview. The follow-up will end at the time the participant is diagnosed with IHD, emigrates, or dies, or the end of the study period (December 31, 2014), whichever comes first. To be eligible for inclusion, they should be between 21 and 59 years old at the start of the follow-up period and employed with ≥32 weekly working hours at the time of the interview. People who received hospital treatment for IHD during the calendar year of the interview will be excluded from the IHD analysis. People who redeemed a prescription for antihypertensive drugs during the calendar year of the interview will be excluded from the antihypertensive drug analysis.

We will use Poisson regression to analyze incidence rates of hospital treatment or death due to IHD as a function of weekly working hours (32–40, 41–48, or >48 hours/week), nighttime work (yes vs no), sex, age (10-year classes), calendar time (2000–2004, 2005–2009, or 2010–2014), time passed since start of follow-up (0–4, 5–9, or ≥10 years), employment in the health care industry (yes vs no), and SES (low, medium, high, or unknown). Age, calendar time, and time passed since start of follow-up will be treated as dynamic (time-varying) variables. The remaining variables will be fixed at baseline (the calendar year of the interview). The logarithm of person-years at risk will be used as offset. People who participated in more than one interview will be classified in accordance with the responses given in their first interview. Later interviews will be disregarded.

We will retrieve information on occupation and industry from the employment classification module, and refer it to the status during the calendar year of the baseline interview. Industries were coded in accordance with the Statistics Denmark classification DB93 [34] in 1999–2002, DB03 [35] in 2002–2007, and DB07 [36] in the calendar years 2008–2013. Occupations were coded in accordance with DISCO-88 (the Danish version of the International Standard Classification of Occupations, ISCO-88) [37] in the calendar years 1999–2009 and DISCO-08 (the Danish version of ISCO-08) [38] in the calendar years 2010–2013.

We will code the variable “employment in the health care industry" as “yes" if the 3-digit industrial code of DB93 or DB03 equals 851 or the 2-digit code of DB07 equals 86.

We will base SES on the participant’s occupation and will code it as “high," “medium," or “low" in accordance with the 3-class version of the European Socio-economic Classification (ESeC). The coding will be performed in accordance with the SAS (SAS Institute) programming statements shown in Textbox 2.

To estimate the statistical power of the hypothesis tests, we first needed to estimate the expected number of cases in the various exposure categories.

To obtain such estimates, we followed the entire population of 21- to 59-year-old employees in Denmark from January 1, 2000 and onward in the same registers that we will use to follow up the samples of this project. While discounting prevalent cases (those who had experienced the clinical end point of the study during the calendar year 1999), we noted all new cases that occurred and tabulated those against the time that had passed since the start of the follow-up. Table 3 gives the results, where the values for the first 13 years of IHD cases and the first 12 years of antihypertensive drug cases are based on actual data, whereas we extrapolated the values for the remaining years by use of a second-degree Taylor polynomial.

Since the sample of participants who were interviewed in calendar year n will be followed for 2014– n years, the expected percentages of new cases in that sample can be found in the row headed by 2014– n years of follow-up, and so forth.

By relating the frequency distribution of the participants stratified by calendar year of interview, sex, and exposure category to the percentages given in Table 3, we finally obtained the expected numbers of cases that were needed to calculate the power (see Table 4).

Since our hypotheses will be evaluated by use of chi-square distributed likelihood ratio tests, we have chosen to depict the statistical power as a function of Cohen w, which is an effect size defined by equation (1) (Figure 1). Cohen classified w=0.1 as a small effect, w=0.3 as a medium effect and w=0.5 as a large effect [39].

Figure 2 shows the statistical power of hypotheses 1.1 and 1.2. Figure 3 shows the power of hypotheses 2.1 and 2.2. The calculations are based on the total number of expected cases and the noncentral chi-square distribution with 12 degrees of freedom for the effects of long working hours (hypotheses 1.1 and 1.2) and 5 degrees for the effects of nighttime work (hypotheses 2.1 and 2.2).

Since the questions used to obtain information about nighttime work and weekly working hours were revised in 2001 and then again in 2007, we will perform a sensitivity analysis with the results stratified by calendar period of interview (1999–2000, 2001–2006, and 2007–2013). The end point, covariates, and statistical model of the sensitivity analysis will be the same as the ones used to test hypotheses 1.1.4 and 2.1.3.

To ascertain that an observed instance of hospital treatment during the follow-up is a new episode rather than a revisit in a course of treatment that was initiated before baseline, the primary analysis will exclude all workers who were treated for IHD sometime during the calendar year preceding baseline. It will, however, not exclude all former cases of IHD, and it is possible that the estimates of the primary analysis will be affected by nonexcluded workers who were treated for IHD more than 1 year earlier than baseline. We will address this issue with a sensitivity analysis, which will exclude all workers who received hospital treatment for IHD one or more times during a 5-year period prior to baseline. The analysis will include only those who were at least 20 years old and lived in Denmark throughout the 5-year period of interest. In all other respects, the design will be the same as the one used to test hypotheses 1.1.4 and 2.1.3.

The actual working hours, that is, the hours worked during the reference week, constitute a well-defined quantity with minimal recall bias. The usual working hours are less well defined, and the way they are understood and remembered might vary between individuals. In spite of this drawback, we chose to base our analysis on the workers’ usual rather than their actual working hours. We did so because some of the participants, by chance, would have worked less than usual during the reference week due to, for example, holidays, vacation, or sickness absence, while others would have worked more than usual due to, for example, a deadline or a temporary staff shortage. Since the usual working hours might be associated with recall bias, we will perform a sensitivity analysis in which we include only participants who belong to the same category according to their actual working hours as they do according to their usual working hours. In all other respects, the design will be the same as the one used to test hypotheses 1.1.4 and 2.1.3. Table 5 gives a -tabulation of the actual and usual working hours.

It is recognized that the risk of IHD depends on a person’s body mass index (BMI) and smoking habits. Among Danish employees, the RR for IHD has been estimated at 1.54 for current versus never smokers [40], at 1.41 for 25≤BMI<30 versus BMI<25, and at 2.69 for BMI≥30 versus BMI<25 kg/m²[41].

Unfortunately, the Danish Labour Force Survey does not contain any information about the worker’s weight and smoking habits, which makes us unable to control for these factors in our analyses. We therefore wanted to know in what direction and to what extent we can expect the estimates of the project to be influenced by differences in smoking habits and BMI. To shed some light on this issue, we compiled some descriptive statistics on the prevalence of smoking and high BMI in relation to long working hours and nighttime work among a random sample of employees in Denmark (Table 6,Table 7). The sample was drawn in 2010 and the data were collected by use of a mailed questionnaire. The response rate was 48% [42].

Table 7 suggests that the prevalence of smoking, overweight, and obesity tend to be higher among employees who work at night than among those who don’t. To get an idea of the extent to which such differences may influence the RR of IHD between employees with and without nighttime work, we calculated an expected rate ratio, E[RR], under the assumption that the groups are equal in all respects other than the smoking and BMI distributions, using equation (2) (Figure 1).

Our calculations imply that a failure to control for smoking, overweight, and obesity (in a study population in which the prevalences are equal to those given in Table 7) would increase E[RR] of IHD between employees with and without nighttime work by a factor of 1.07.