The WHI study design has been described in detail.21–23
In brief, it was a long-term national health study that focused on strategies for preventing heart disease, breast and colorectal cancer, and osteoporosis in postmenopausal women. Women between the ages of 50 and 79 were enrolled for an observational study or randomised controlled trial (RCT) from 1993 to 1998 at 40 clinical centres throughout the USA. The institutional review boards at all participating institutions, including the coordinating centre, subcontractors and clinical centres, approved the study protocols and procedures.
Participants available for analysis included 161 748 WHI participants: 93 651from the observational study, 16 590 from the RCT of oestrogen plus progesterone (E+P), 10 722 from the RCT of oestrogen only (oestrogen-alone), and 40 785 additional women who were in the diet study and not in an RCT of hormone therapy. Baseline information included demographics, general health, clinical and anthropometric, functional status, healthcare behaviours, reproductive, medical history, family history, personal habits, thoughts and feelings, therapeutic class of medication, hormones, supplements and dietary intake. Outcomes were identified primarily through self-report at semiannual contacts for clinical trial participants and annual contacts for observational study participants. Specific details of illnesses and hospitalisations are obtained as needed via a standardised questionnaire administered by phone or in-person interview, or self-completed form. For primary and secondary outcomes, portions of the medical record (discharge summary and results of relevant diagnostic and laboratory tests) are requested and assembled.
We excluded 9584 participants from the observational study because they would have been excluded from RCTs. Exclusion criteria were platelets less than 75 000/mm3, haematocrit less than 32%, oral daily use of a glucocorticosteroid, body mass index (BMI) less than 18, systolic blood pressure greater than 200 mm Hg, diastolic greater than 105 mm Hg, breast cancer ever, other cancers in the last 10 years, or stroke, transient ischaemic attack, or myocardial infarction (MI) in the last 6 months. These exclusion criteria made the study participants more homogenous (since they had already been applied to the RCTs) and they reduced the likelihood that sleep patterns would be influenced by severe health problems. An additional 3285 participants were excluded from the final analyses because they were missing information on one or more of the questions used to define difficulty sleeping. The number of participants analysed was 148 938.
Outcomes and risk factors
The present study assessed the association of hypnotic use with mortality (the primary outcome in this study and the one most often evaluated for an association with sleep medications) and other disease outcomes that may contribute to mortality: MI, stroke, diabetes, breast cancer, colon cancer, lung cancer, colorectal cancer, lymphoma, melanoma and ovarian cancer. If hypnotic use caused specific health problems that led to mortality, then the strongest associations should be between sleep medication use and these health problems. It is certainly possible that the health problems most associated with hypnotic use were not collected by the WHI or tested in the present study. One of the outcomes cited in several studies of hypnotic use that was not collected by the WHI was suicide.17
Several risk factors related to sleep were tested for an association with the primary and secondary outcome variables.
The sleep factor analysed in greatest detail was the response to the following question about sleep medications: ‘Did you take any kind of medication or alcohol at bedtime to help you sleep?’ The answers were on a five-point scale from not in the past 4 weeks to 5 or more times a week. For simplicity the highest point on the scale was referred to as frequent hypnotic use.
Information about sleep duration was from the response to the following question: ‘About how many hours of sleep did you get on a typical night during the past 4 weeks?’ Answers were on a six-point scale that ranged from 5 or less hours to 10 or more hours.
Information was also obtained about the following insomnia-related factors in the past 4 weeks: quality of sleep, trouble falling asleep, wake up several times, wake up earlier than planned, trouble getting back to sleep and the WHI sleep construct that was formed by summing the scores for the five specific types of difficulty sleeping.
These risk factors were tested for an association with the primary and secondary outcomes after adjusting for a set of risk factors highly associated with mortality as described in statistical methods. One of the mortality risk factors was smoking history, which was obtained from three questions:
- A three category smoking status variable (never/past/current).
- On the average, how many cigarettes do you (did you) usually smoke each day?
- How many years have you been (were you) a regular smoker?
The association of risk factors with each outcome was tested using Cox proportional hazard regression. Each of the risk factors related to sleep were evaluated separately. The hazard ratio (HR) and associated χ2 of the risk factor were obtained after adjusting for age and again after adjusting for mortality risk factors. The mortality risk factors were identified using stepwise proportional hazards regression on all the potential risk factors collected at baseline by the WHI. Factors statistically significant at p<0.0001 level in the full dataset minus participants who would not qualify for RCTs were retained in the model.
The sleep variables were included in the regression models as categorical variables. The category of greatest interest for each sleep variable was the most extreme. Therefore, results were provided for the sleep variable as a whole and for the most extreme category of the sleep variable, which was represented by a binary variable.
The three questions about smoking were included as categorical risk factors in the proportional hazard regression equation. A composite smoking variable for a given subject was created by summing the age-adjusted regression coefficients associated with the values of the smoking variable present in that subject. This composite was an ordinal variable that was associated with an identical χ2 value as the χ2 is associated with the three categorical smoking variables. Because the variable was ordinal, it was possible to find its correlation with other risk factors.
The question about sleep medication was worded so that it included subjects who used alcohol to help them sleep. To assess whether the results were altered by subjects using alcohol as a sleep aid, the analyses on sleep medications were repeated for only those subjects who took less than one drink a week.
Every study that has evaluated the association between use of sleep medications and mortality adjusted for different covariables. The covariables that most influence the association between sleep medication use and mortality are those most confounded with sleep medications. The amount of confounding depends on what other variables are included in the regression equation. We created a score that measures the extent that frequent hypnotic use was confounded with another risk factor that was included in the Cox model. This score is analogous to a measure of confounding in linear regression.25
The confounding score was the product of χj
is the square root of the χ2
value for the other risk factor when it is in the Cox model with frequent hypnotic use, and rhu,,j
is the correlation of the other risk factor with frequent hypnotic use. This score provides an intuitive, quantified measure of the confounding associated with a particular covariable in the present study. For the 10 binary or ordinal covariables in the present study, the correlation of this confounding score with the change in the χ2
value of frequent hypnotic use when the covariable was added to the regression equation was 0.99.
To test whether hypnotic use was a special risk factor for the morbidly obese, as has been suggested,26
we used a statistical test for interaction between hypnotic use and BMI of 35 or greater.