The NHS was initiated in 1976 when 121,700 female registered nurses aged 30–55 years completed a mailed questionnaire. The cohort is followed via biennial questionnaires, as described previously [18, 19]. Among 119,142 active participants in 2000, we excluded women with diabetes prior to 2000 (n = 11,729) or missing responses to questions on habitual sleep duration in 1986 (n = 34,479) or 2000 (n = 13,903), leaving 59,031 women. The Institutional Review Boards at the Harvard T.H. Chan School of Public Health and Brigham and Women’s Hospital approved the study protocol.
Type 2 diabetes
Cases were defined as self-reported diabetes confirmed via a validated supplementary questionnaire regarding symptoms, tests and hypoglycaemic therapy. Before 1998, we used the National Diabetes Data Group diagnostic criteria. After 1998, we used the ADA criteria. A previous validation showed self-reported diabetes diagnosis through supplemental questionnaire confirmation was highly accurate: of 62 cases confirmed by questionnaire, 61 (98%) were reconfirmed by medical records .
Nurses reported sleep duration in the 1986 and 2000 questionnaires, corresponding to hours of sleep in a typical 24 h period; response options were ≤5, 6, 7, 8, 9, 10 or ≥11 h, treated continuously with values assigned to the nearest integer (≤5 = 5 h; ≥11 = 11 h). In 1988, participants reported total years of rotating shifts, characterised as ‘at least three nights per month in addition to working days or evenings in that month’. Response options were: never, 1–2, 3–5, 6–9, 10–14, 15–19, 20–29 or ≥30 years. Given prior evidence that diabetes risk increases most after ≥5 years of shift work, we categorised this variable (<5/≥5 years) . Women reported snoring frequency in 1986, 2000, 2002 and 2008 and physician-diagnosed sleep apnoea in 2008.
Information on confounders, e.g. age, race/ethnicity, employment, smoking, menopause, medications (e.g. hormone therapy, antidepressants [yes/no, first measured in 1996] and antihypertensive drugs) and physician diagnoses was collected via biennial questionnaires. Height was self-reported in the first questionnaire and weight in each biennial questionnaire, from which we calculated BMI (kg/m2). A prior validation showed a correlation of 0.96 between self-reported and technician-measured weight . Food and alcohol consumption was collected every 4 years via validated semi-quantitative food frequency questionnaires. Hypertension and hypercholesterolaemia were considered as physician diagnosis or related medication use. Depression was defined as physician-diagnosed depression or antidepressant medication.
Diet quality was assessed by the Alternate Healthy Eating Index (AHEI-2010), based on a high consumption of vegetables, fruit, nuts/legumes, whole grains, long-chain fats, polyunsaturated fatty acids, and low levels of sugar-sweetened beverages and juice, red/processed meat, trans fat and sodium . Higher scores indicated better quality diets. Physical activity was defined from work/leisure activities as weekly energy expenditure in metabolic equivalent (MET)-h (vigorous activity was ≥6 METs) . In a previous validation , the correlation of questionnaire-reported physical activity was 0.79 with prospective 1-week recalls and 0.62 with activity diaries.
Our main exposure was 14 year change in sleep duration: the difference in h/day between 2000 and 1986. This continuous change was categorised into a priori groupings: decreases in sleep duration (≤−2 h/day or >−2 to <0 h/day), no change in sleep between time points (0 h/day, reference), and increases in sleep duration (≥2 h/day or >0 to <2 h/day). We computed age-adjusted descriptive statistics by category of change in habitual sleep duration from 1986 to 2000 (Table 1).
We used multivariable linear regression to examine associations of changes in sleep duration (1986–2000) with concomitant changes in energy balance factors over approximately the same period as change in sleep. Change in diet was represented by AHEI-2010 (change in score 1986 to 1998, the nearest questionnaire to 2000); physical activity by METS (change in h/week from 1986 to 2000) and energy balance by weight (change in kg from 1986 to 2000). We adjusted for race/ethnicity (non-Hispanic white, yes/no) and sleep duration at the start of the change period (1986) and covariates, including age (years), retired (yes/no), menopause (premenopausal or hormone therapy [never, past, or current user]), alcohol grams in quartiles, smoking (never, past, or current smoker [1–14, 15–24 or 25+ cigarettes/week]), diabetes family history (yes/no), snoring frequency (most nights, some or never), sleep apnoea (ever-diagnosis), antidepressant use (yes/no), shift work history (≥5 years), BMI (<20.9, 21–24.9, 25–29.9, 30.0–31.9, 32.0+ kg/m2), and hypertension or hypercholesterolaemia. Subsequent models adjusted for changes in these covariates (1986–2000).
To examine the primary outcome of diabetes, we used Cox regression models jointly stratified by age in months at the start of follow-up and calendar year of the current questionnaire cycle to estimate HRs and 95% CI according to category of change in sleep duration. Individuals contributed person-time from the return of the 2000 questionnaire until the date of diagnosis of diabetes, death, loss to follow-up or end of the follow-up period (30 June 2012). Models adjusted for the variables listed above at the start of the change period in 1986 and for change in covariates (1986–2000).
We varied our modelling approach by using an additional categorical exposure: we cross-classified women according to sleep duration in 1986 and 2000 (short [≤6 h/day], long [≥9 h/day] and normative [7–8 h/day]), with women reporting 7–8 h/day at both time points as the reference for comparison with women reporting consistently short, consistently long or changes in sleep duration (Fig. 1).
To limit potential reverse causation, we considered excluding cases within the first 2 years of follow-up (n = 758) and those with chronic disease in 2000 (ever-diagnosis of cancer and/or cardiovascular disease [n = 11,186]). We considered excluding women >66 years (median age) in 2000 (n = 30,243) or who had conducted shift work for ≥5 years (n = 9,503).
Data were analysed using SAS for UNIX (version 9.3; SAS Institute, Cary, NC, USA). Statistical significance was set at a two-tailed p < 0.05.