The women's health study was a randomised, double blind, placebo controlled 2×2 factorial trial of low dose aspirin (100 mg on alternate days; Bayer HealthCare) and vitamin E supplementation (600 IU on alternate days; Natural Source Vitamin E Association) in the prevention of cardiovascular disease and cancer among US women. Most participants are white (>95%). The study design has been described previously.12
Briefly, women were eligible if they were aged 45 or more; had no history of coronary heart disease, cerebrovascular disease, cancer (except for non-melanoma skin cancer), or other major chronic illnesses; and did not actively use or have a history of side effects from the study drugs.
To identify women who were likely to be highly compliant, we enrolled eligible women in a three month run-in period of placebo administration. A total of 39
876 compliant women were randomised (19
934 to aspirin and 19
942 to placebo) from 1992 to 1995. Randomised assignments were computer generated in blocks of 16 within five year age strata by study programmers, masked to study investigators and enrolling staff. Participants provided written informed consent.
Every 12 months, based on their random assignments, participants were posted a year's supply of monthly calendar packs with identical white pills containing aspirin or placebo. Women were asked to complete annual postal questionnaires to update information on compliance, side effects, health and lifestyle characteristics, and the occurrence of clinical end points. They were instructed not to use vitamin E supplements and any prescribed or over the counter aspirin or aspirin containing drugs or any other non-steroidal anti-inflammatory drugs during the trial. The trial continued through to the scheduled end (31 March 2004), when overall follow-up on mortality exceeded 99%; the main results of the trial have been published and showed a benefit of aspirin in reducing the risk of stroke but no benefit in reducing the risk of myocardial infarction11
Details of side effects are described elsewhere;11
briefly, the aspirin group had excesses of self reported gastrointestinal bleeding, peptic ulcer, haematuria, easy bruising, and epistaxis.
Although the original purpose of the trial was to evaluate cardiovascular and cancer outcomes, this large scale, long term trial provided an opportunity to incorporate cognitive outcomes to study the potential effect of aspirin on delaying cognitive decline. In 1998 we started a study of cognitive function among women aged 65 or more. Eligible women were aged 65 or more and were still active participants of the women's health study (n=7175). Most of the ineligibility (98.5%) was due to the age restriction in the cognitive study protocol; only 1.5% of deaths or losses to follow-up had occurred by the start of the cognitive study, and these women were distributed equally between the aspirin and placebo groups (figure).
Flow chart of participation in cognitive cohort of women's health study
Of the 7175 women selected for the substudy of cognitive function, 296 (4.1%) were unreachable by telephone and 502 (6.9%) did not participate; thus 6377 (88.9%) women completed the initial cognitive assessment by telephone (figure). The initial assessments were carried out an average of 5.6 years after randomisation. The participation rates in the initial interview were virtually identical in the two groups. In addition, no significant differences were found in important baseline characteristics by treatment assignment.
After the initial cognitive assessment, participants underwent two follow-up assessments, at about two year intervals. High follow-up was maintained (figure): 5845 (91.7%) of those who completed the initial assessment also completed at least one follow-up assessment, and 5073 women (79.5%) completed all cognitive assessments. Among the 8% of participants who did not complete any follow-up assessments, 2% died, 1% was unreachable, and 5% refused. Follow-up rates were nearly identical in the two groups.
Although the sample size of the women's health study was not designed for the cognitive outcome, the 6377 women in this substudy provided sufficient power to detect modest differences in cognitive decline between the aspirin and placebo groups (3215 receiving aspirin and 3162 receiving placebo). For example, for the categorical outcome of substantial cognitive decline we had at least 80% power to detect a modest relative risk of 0.76 in the aspirin group compared with the placebo group; moreover, power was considerably greater for continuous analyses of mean cognitive decline in the two groups.
We have previously published results of the vitamin E treatment14
; we found no relation between vitamin E and cognitive function.
Cognitive function assessment
The telephone cognitive battery was administered by trained nurses masked to participants' treatment group. The assessment included five tests measuring general cognition, verbal memory, and category fluency. To assess general cognition we administered the telephone interview of cognitive status,15
an adaptation of the mini-mental state examination for use by telephone. To test verbal memory we administered the immediate and delayed recalls of the east Boston memory test,16
in which a short paragraph is read and 12 key elements must be repeated immediately and again at 15 minutes. To further evaluate delayed verbal memory we administered a delayed recall of the telephone interview of cognitive status 10 word list. Finally, to assess category fluency,17
women were asked to name as many animals as possible in one minute.
Our primary, prespecified outcome was a global composite score averaging performance across all five cognitive tests, using z scores. Because verbal memory is among the strongest predictors of eventual risk of Alzheimer's disease,1
our key, secondary outcome was a composite score of verbal memory, averaging performance on four tests (the immediate and delayed recalls of both the east Boston memory test and the 10 word list).
In a test-retest study of the telephone interview of cognitive status among 35 high functioning, educated older women, we found a correlation of 0.7 (P<0.001) between two assessments administered 31 days apart. In a validation study of our telephone instrument, among 61 high functioning, educated older women, the correlation was 0.81 comparing overall performance on our telephone administered interview and overall performance on an extensive in-person interview, showing high validity. Cognitive impairment determined from our telephone assessment strongly predicted later dementia in educated women; among 88 older women followed over three years, lower scores on both the telephone interview of cognitive status and verbal memory were associated with significant eightfold and 12-fold increased risks of dementia.
We first examined mean performance at each cognitive assessment using repeated measures analysis of means (which permits examination of each time point, taking into account correlation between assessments). Secondly, we examined mean change in cognitive function over the three cognitive assessments. We treated scores and change in scores at each assessment as repeated continuous outcomes and we modelled the treatment effect by a time by treatment interaction. Because the trends for test scores over time were non-linear, in all analyses we used general linear models of response profiles, modelling time nominally rather than linearly.18
This approach imposes minimal structure on outcome trends and permits valid estimation of effects in non-linear data. We fitted all models by maximum likelihood, incorporating the longitudinal correlation within participants, using unstructured covariance structures; for statistical testing, we used Wald tests.18
All linear models were fitted using Proc Mixed in SAS (version 9).
In secondary analyses we examined effect modification by key risk factors for cognitive decline (most measured at baseline): age, baseline score, perceived change in memory, education, cigarette smoking, alcohol consumption, body mass index, physical activity, hormone replacement therapy use, history of diabetes, hypertension, hyperlipidaemia, depression (at four years after randomisation), and cardiovascular disease (all incident occurrences from randomisation to the end of the study). Cardiovascular disease included all medical record confirmed non-fatal myocardial infarction, non-fatal stroke, cardiovascular related deaths, or vascular disease as evidenced by either a coronary artery bypass graft or percutaneous transluminal coronary angioplasty or stenting.11
We carried out all tests of effect modification by evaluating the interaction terms in the models of mean change.
To assess the effect of aspirin treatment on the risk of “substantial cognitive decline,” we used logistic regression models in which substantial cognitive decline was defined as the worst 10% of the distribution of decline from the initial to the final cognitive assessment of the whole cohort. In all models we adjusted for the time between the first and third assessments.