Research design and methods
Data source and study population
The data source for this study was the 2003 Canadian Community Health Survey (CCHS). Respondents’ responses were linked to the Ontario Health Insurance Plan (OHIP) database covering physician services, as well as the Canadian Institute for Health Information Discharge Abstract Database (CIHI-DAD) for hospital admissions. These data sets were encoded and linked at the Institute for Clinical Evaluative Sciences. Information from the OHIP and CIHI-DAD databases was available up to March 31, 2015. The administrative databases were linked to the survey responses at the individual level, with the accuracy of the linkage verified against the Ontario Registered Persons Database using personal information provided by the respondents, such as health number, given names, surnames, date of birth, age, sex, and postal code.
The CCHS collects information on health conditions, health behaviors, and working conditions from representative cross-sectional samples of the Canadian population. The overall response rate from the respondents from Ontario to the 2003 CCHS was 78.5%.18 Of the 40 507 Ontario respondents to the 2003 survey, 34 950 (86%) gave permission to be linked to administrative healthcare data. A successful linkage was obtained for 33 679 of these respondents (96%). For the purpose of this study, we focused on respondents who were currently employed, working 15 or more hours per week, and aged between 35 and 74 years (n=8895).
Outcome: incident diabetes
Incident diabetes was defined as one hospital admission with a diabetes diagnosis, or two physician service claims with a diabetes diagnosis within a 2-year period. Excellent sensitivity and specificity have been reported for this algorithm, 86% and 97%, respectively.19 Although the administrative data available in the OHIP database cannot distinguish between type 1 and type 2 diabetes, this restriction is unlikely to impact on our results given the high prevalence of type 2 diabetes, in particular among older cohorts, and the fact that the incidence of type 1 diabetes is very rare among adults.20
Since diabetes is known to affect work participation,21 cases occurring within the first 2 years of follow-up were removed to limit the possibility of reverse causation. This choice also resulted from the fact that our definition of diabetes required a 2-year period for physician claims. Respondents who did not develop diabetes within the first 2 years of follow-up were right-censored at the development of the disease, death from causes other than diabetes, or the end of the follow-up period (March 31, 2015).
Primary independent variable: work hours
The primary independent variable was self-reported usual work hours in the respondent’s job per week and included both paid and unpaid hours. Work hours were grouped into the following categories: 15–34 hours, 35–40 hours, 41–44 hours, and 45 or more hours per week. This categorization was performed to evaluate the potential adverse effects of working beyond the legal threshold of overtime (41 hours or more per week) of many countries including Canada, USA, China, and Greece.22
Other independent variables
Several other independent variables were also included in the analysis as covariates. Sociodemographic and health-related covariates measured were age; sex (men/women); marital status (with/without a spouse) and presence of children under 12 in the house (yes/no); if the respondent was born in Canada; their ethnicity (white/other); living location (urban/rural); and self-reported chronic medical conditions that have been diagnosed by a health professional and are expected to last or have lasted more than 6 months. Chronic medical conditions were divided into cardiovascular diseases, high blood pressure, back problems, mood and anxiety disorders, and other chronic conditions. A measure of whether a long-term physical or mental health condition limited the type or amount of activity the respondent could do at work (never, sometimes and often) was also included.
Other working conditions were also measured and were based on self-report and on occupational exposures imputed based on respondents’ occupational title. Self-reported exposures included the number of weeks worked in the previous 12 months (1–26 weeks worked, 27–49 weeks, and 50 or more weeks), current shift schedule (regular, evening or night shift, rotating, or other shift schedules), and skills required to do the job (skills learned at the university, college, high school or learned on the job). Imputed occupational exposures based on occupational title included the primary type of posture or body movement required (primarily sitting; occupations involving primarily standing and/or walking; occupations involving combinations of sitting, standing, and walking; and work that involves other body postures) and the handling of loads 10 kg or greater (binary). Imputed exposures were assigned based on the validated Human Resources and Skills Development Canada’s Career Handbook.23 The Career Handbook assigns various occupational exposures to occupations at the four-digit occupational level, equating to 520 different occupational titles. Each of these exposures was assigned by trained occupational analysts using a modified Delphi procedure.23
Body mass index (BMI) and health behaviors were also accounted for. BMI was based on self-reported height and weight (underweight/normal weight, overweight, obese). Health behaviors available were current smoking status (regular smoker, occasional smoker, non-smoker), alcohol consumption (non-drinker, regular drinker but never having five or more drinks in one sitting, regular drinker who has five or more drinks on an occasional to weekly status), and leisure time physical activity (inactive, moderately active, active).
Analyses
Of the original sample of 8895 respondents, 546 respondents (6%) had pre-existing diabetes, identified either through self-report or through their healthcare record, leaving a sample of diabetes-free respondents of 8349. Of this sample 566 (7%) were missing information on working conditions, with an additional 610 (7%) missing information on sociodemographic characteristics, health conditions or health behaviors, leaving a sample of 7173 (86% of original sample). From this sample we excluded respondents who developed diabetes in the first 2 years of follow-up (n=108), leaving a final analytic sample of 7065 respondents (figure 1). Logistic regression analyses examined the relationship between age, sex and our outcome with the probability of missing information on working conditions. Male respondents were more likely to be missing work information, but no relationship was observed between missing work information and age, or the development of diabetes over the study period. A subsequent logistic regression model examined the probability of missing other study variables. This model included age, sex, diabetes outcome, skill level and work hours. Female respondents in lower skilled occupations and those who developed diabetes over the follow-up period were more likely to be missing information on these characteristics. No relationship was observed between age and work hours and the probability of missing information.
Among the sample of 7065 respondents included in the analyses, we had 78 390 person-years of follow-up, with a median length of follow-up in the sample of 11.7 years. Initial descriptive analyses examined the distribution of diabetes incidence rate and cumulative incidence across weekly work hours categories. Cox proportional hazard regression models were then performed to evaluate the relationship between long work hours and the incidence of diabetes over the 12-year follow-up. The first model examined the effect of long work hours on diabetes after adjusting for age, weeks worked in the previous 12 months and skill level. A second model was additionally adjusted including all other covariates, except health behaviors and BMI, which were adjusted for in the third and fourth models, respectively. Health behaviors and BMI were sequentially adjusted for since it is not clear whether these factors are confounders or mediators in the relationship between work exposures and diabetes. Including potential mediators in the regression models could be considered overadjustment.24 All models were conducted separately for men and women.25 Differences between estimates from men and women were assessed by comparing the point estimates and the associated variances of these estimates for models run separately for men and women.26 27
To account for the complex sample design of the CCHS, in line with the guidelines from Statistics Canada, the CIs around each point estimate, with the exception of incidence rates per 1000 person-years, were adjusted using a bootstrap technique.18 All analyses were weighted to account for the probability of selection into the original sample and non-response. All analyses were conducted in SAS V.9.4.