Combined Analysis of Women's Health Initiative Observational and Clinical Trial Data on Postmenopausal Hormone Treatment and Cardiovascular Disease

Abstract

Circumstances in which both randomized controlled trial and observational study data are available provide an important opportunity to identify biases and improve study design and analysis procedures. In addition, joint analyses of data from the two sources can extend clinical trial findings. The US Women's Health Initiative includes randomized controlled trials of use of estrogen by posthysterectomy women and of estrogen plus progestin by women with a uterus, along with corresponding observational study components. In this paper, for coronary heart disease, stroke, and venous thromboembolism, results are first presented from joint analysis of estrogen clinical trial and observational study data to show that residual bias patterns are similar to those previously reported for estrogen plus progestin. These findings support certain combined analyses of the observational data on estrogen and the estrogen plus progestin clinical trial and observational study data to give adjusted observational study estimates of estrogen treatment effects. The resulting treatment effect estimates are compared with corresponding clinical trial estimates, and parallel analyses are also presented for estrogen plus progestin. An application to postmenopausal hormone treatment effects on coronary heart disease among younger women is also provided.

In 2002, the US Women's Health Initiative (WHI) reported primary results from the randomized controlled trial of daily use of 0.625 mg of conjugated equine estrogen plus 2.5 mg of medroxyprogesterone acetate among 16,608 postmenopausal women with a uterus (1). In 2004, corresponding results were reported from the randomized controlled trial of use of 0.625 mg of conjugated equine estrogen alone among 10,739 women who were posthysterectomy at baseline (2). Both trials were stopped prematurely: the former on the basis of an elevated breast cancer risk and an assessment that overall health risk exceeded benefits over an average 5.6-year follow-up period; the latter because of, in part, an elevated stroke risk over an average 7.1 years of follow-up.

These clinical trials, as well as secondary prevention trials among women with prior cardiovascular disease (3–9), have substantially impacted the usage patterns for postmenopausal hormone treatments. The initial WHI report of the estrogen plus progestin trial (1) was followed by more detailed reports on specific outcomes, including coronary heart disease (CHD) (10), stroke (11), and venous thromboembolism (12), and corresponding detailed reports (13, 14) will appear for the estrogen trial. These papers examine clinical trial results in subsets of the study population, but precision is limited, as is the power for identifying interactions between baseline characteristics and hormone treatment.

The WHI observational study of 93,676 postmenopausal women provides an important source of data to add precision to estimated hormone treatment effects and to address a broader range of postmenopausal hormone treatment topics than the clinical trials were designed to accomplish. Women aged 50–79 years were recruited to the clinical trial and the observational study from the same underlying populations, over essentially the same time period, at the 40 WHI clinical centers. Many elements were common to the protocol and procedures, including much of the baseline questionnaire and interview data collection and major elements of outcome ascertainment.

To understand the apparent differences between estrogen plus progestin (E+P) trial results and corresponding results from observational studies, WHI investigators recently presented an analysis of cardiovascular disease outcomes in the E+P trial and the 53,054 observational study women who had a uterus and were not taking unopposed estrogen at baseline (15). After control for potential confounding factors and for time since initiation of E+P, the WHI clinical trial and observational study results were in reasonable agreement regarding CHD and venous thromboembolism, although less so for stroke.

We begin here by presenting corresponding analyses of data from the estrogen-alone (E-alone) trial and from the 38,313 women in the observational study who were posthysterectomy and were not using E+P at baseline. We found comparative patterns very similar to those from our E+P analyses in spite of rather different hazard ratios for the two treatments. This finding suggests that residual biases in the observational study may be similar for E+P and E-alone in this population. Under this assumption, we use the E+P clinical trial and cohort data to adjust the E-alone observational study analyses, and we compare the results with the E-alone trial results for each of the three cardiovascular disease outcomes. A corresponding analysis is also carried out to adjust the E+P observational study results. These comparisons support the utility of the adjusted observational analyses, so we complete the presentation with adjusted hazard ratio estimates for E-alone and E+P for CHD in a subgroup of particular interest: women in the age range 50–59 years at the time of WHI enrollment.

MATERIALS AND METHODS

WHI cohorts

Women eligible for the study were 50–79 years of age at screening, were postmenopausal, had no medical condition associated with a predicted survival of less than 3 years, and were likely to continue to reside in the vicinity of a WHI clinical center for at least 3 years (16, 17). For the hormone treatment trials, additional exclusionary criteria involved safety, adherence, and retention concerns (1, 2). Women ineligible for, or not interested in, the hormone treatment trials or participating in an overlapping trial of a low-fat eating pattern were given the opportunity to enroll in the observational study, which was intended to provide new risk factor information on major causes of morbidity or mortality among postmenopausal women. All women provided written informed consent for their respective WHI activities and supplied a baseline fasting blood specimen, a medication and dietary supplements inventory, and common core questionnaires including medical history, reproductive history, family history, personal habits, psychosocial information, and a food frequency questionnaire (16–18).

Baseline exogenous hormone use and clinical trial hormone regimens

At baseline, information on lifetime hormone use was obtained for women in the clinical trial and the observational study by a trained interviewer, assisted by a structured questionnaire and chart displaying colored photographs of various hormone preparations. For postmenopausal hormone treatment, detailed information was obtained on the preparation, estrogen and progestin dose, schedule, and route of administration. Age at starting and stopping such preparations was recorded.

Women interested in participating in the hormone treatment trials who were using postmenopausal hormone treatment at initial screening were required to undergo a 3-month washout period prior to randomization. Women with a uterus were potentially eligible for randomization to daily use of 0.625 mg of conjugated equine estrogen plus 2.5 mg of medroxyprogesterone acetate or placebo, while women with a prior hysterectomy were potentially eligible for randomization to daily use of 0.625 mg of conjugated equine estrogen or placebo. A group of 331 women with a uterus were initially randomized to daily use of 0.625 mg of conjugated equine estrogen. Following the PEPI study results (19), data for these women were unblinded, and the women were reassigned to 0.625 mg of conjugated equine estrogen plus 2.5 mg of medroxyprogesterone acetate; data for them are included here in the combined hormone trial data, as in previous reports (1, 10–12, 15).

Sample for analysis

This paper is based on data from the 10,739 women randomized to the E-alone trial, along with the 38,313 women in the observational study who were posthysterectomy and were using either unopposed estrogen or no hormone treatment at baseline in the WHI. It is also based on data from the 16,608 women randomized to the E+P trial, along with a nonoverlapping 53,054 women in the observational study who had a uterus and were using either combined hormone preparations or no hormones at baseline.

Follow-up and outcome ascertainment

Procedures for follow-up and outcome ascertainment in the clinical trial (1, 2, 10–12, 20) involved semiannual contacts and in-clinic annual visits by participants to collect standardized information related to safety and adherence to study medications, and for structured initial reporting of clinical outcome events. In the observational study, annual mailed follow-up forms updated information on hormone treatment, updated other selected risk factor information, and used the same structured initial reporting of clinical events. CHD was defined (10) as myocardial infarction or death due to coronary disease. Stroke was defined (11) as rapid neurologic deficit requiring hospitalization and attributable to obstruction or rupture of the arterial system or a demonstrable lesion compatible with acute stroke. Venous thromboembolism comprised (12) deep vein thrombosis and pulmonary embolism requiring hospitalization. CHD and stroke in both the clinical trial and the observational study, and venous thromboembolism in the clinical trial, involved physician adjudication at each clinical center following review of pertinent documents. Venous thromboembolism in the observational study was based on self-reports of disease requiring hospitalization. In the clinical trial, confirmation rates for these self-reports of venous thromboembolism were about 80 percent. CHD, stroke, and venous thromboembolism events in the clinical trial were further adjudicated by a central committee, with agreement rates of about 95 percent for each clinical outcome (15).

Statistical analysis

For primary analyses, we used time-to-event methods based on the Cox regression procedure (21), with time from randomization in the clinical trial and time from enrollment in the observational study as the basic “time” variable. Disease incidence rates during follow-up were stratified on baseline age in 5-year categories and on WHI component (clinical trial or observational study; E-alone or E+P subcohort). Hence, we derived E-alone and E+P hazard ratio estimates from comparisons among women in the same baseline 5-year age group, participating in the same WHI component, and whose length of time since WHI enrollment was the same.

Disease events in the E-alone trial were included through February 29, 2004, when study medications were stopped, giving an average follow-up time of 7.1 years, and, in the corresponding subsample of women in the observational study, through December 31, 2004, also giving an average follow-up time of 7.1 years. Disease events in the E+P trial were included through July 7, 2002, when study medications were stopped, giving an average follow-up time of 5.6 years, and, in the corresponding observational study subsample, through February 28, 2003, giving an average of 5.5 years of follow-up. We used the best available outcome data, comprised of all centrally adjudicated CHD, stroke, and venous thromboembolism events in the clinical trial as well as all locally adjudicated CHD and stroke and self-reported venous thromboembolism events in the observational study.

We included baseline characteristics in the Cox regression model to control for confounding. The same factors were used for all three clinical outcomes and for both the E-alone and E+P analyses. As reported by Prentice et al. (15), these factors included age, body mass index, education, cigarette smoking, age at menopause, and physical functioning (22) variables. For each clinical outcome, confounding factor coefficients were estimated separately for the E-alone and the E+P analyses.

Our previous analyses (15) observed an important role for time from initiation of the current episode of postmenopausal hormone treatment in resolving E+P hazard ratio differences between the clinical trial and the observational study. Hence, the analyses presented here estimate hazard ratios for hormone treatment separately for less than 2 years, 2–5 years, and more than 5 years from initiation of the current hormone treatment episode. At a specific follow-up time in the WHI, the time from initiation of the current episode of hormone treatment in the clinical trial was defined as time from randomization; for the hormone treatment groups in the observational study, it was defined by summing duration of the current hormone treatment episode at baseline and time since enrollment in the observational study. A usage gap of 1 year or longer (prior to WHI enrollment) defined a new hormone treatment episode.

The analyses presented here show that E-alone hazard ratios from the clinical trial and the observational study agree fairly well following control for confounding and years since hormone treatment initiation, which was also the case for E+P (15). However, we also provide for some residual confounding in the observational study in analyses that bring together clinical trial and observational study data. To do so, we include an interaction term between hormone treatment and an indicator variable for observational study (vs. clinical trial) in the Cox regression analyses. This interaction term leads to estimated hazard ratios in the observational study that differ by a multiplicative factor from corresponding clinical trial hazard ratios. This factor is referred to in the text tables as “HR [hazard ratio] in the observational study/HR in the clinical trial.”

As reported by Prentice et al. (15), we examined the sensitivity of hazard ratio estimates to lack of adherence to baseline hormone treatment or control group designations during WHI follow-up by censoring the follow-up period for each woman in the clinical trial and the observational study at 6 months beyond the time that her hormone treatment or control group designation changed. This decision was based on a woman stopping hormone treatment if she was in a hormone treatment group or initiating hormone treatment if she was in a comparison (control) group.

Hazard ratio estimates from Cox regression models, nominal 95 percent confidence intervals, and two-sided significance tests (p values) are presented throughout this paper.

RESULTS

There were 21,902 E-alone users at baseline in the observational study, constituting 57.2 percent of the 38,313 observational study women who had had a hysterectomy and were not using E+P preparations. There were 421 CHD events in this E-alone group compared with 548 in the corresponding control group of 16,411 women. The women in the E-alone group were somewhat younger than the comparison women (mean age, 63.0 vs. 65.3 years). The CHD incidence rate ratio for E-alone users compared with nonusers was 0.68 versus 0.96 for the E-alone trial, following age adjustment (to the 5-year distribution in the E-alone trial).

Similarly, 431 E-alone women experienced a stroke during WHI follow-up in the observational study, and 408 women in the control group experienced a stroke, giving an age-adjusted incidence ratio of 0.95 for estrogen use versus control in the observational study compared with 1.37 for the E-alone trial. A total of 265 E-alone women in the observational study experienced venous thromboembolism compared with 274 women in the control group, for an age-adjusted incidence rate ratio of 0.78 compared with 1.33 from the E-alone trial. In addition, the age-adjusted annualized incidence rates (percent) for the E-alone control group in the observational study were, respectively, 0.43, 0.31, and 0.23 for CHD, stroke, and venous thromboembolism compared with 0.57, 0.33, and 0.22 for the placebo group in the clinical trial.

Confounding in the observational study may account for some of the discrepancy between the E-alone hazard ratios in the observational study compared with the E-alone trial. For example, after we adjusted for age, E-alone women in the observational study were less likely than women in the control group to be of minority race/ethnicity (15.0 percent vs. 25.9 percent); less likely to be obese or extremely obese (23.3 percent vs. 35.2 percent); more likely to have more than a high school education (78.7 percent vs. 71.7 percent); less likely to be a current smoker (5.4 percent vs. 8.2 percent); more likely to have a family income of $75,000 or more (20.4 percent vs. 13.6 percent); less likely to have a personal history of diabetes (3.5 percent vs. 7.6 percent), CHD, stroke, venous thromboembolism, or overall cardiovascular disease (9.7 percent vs. 12.9 percent); more likely to be physically active (29.7 percent vs. 23.6 percent engaging in four or more 20-minute episodes of exercise per week); and slightly more likely to have favorable scores on various quality-of-life scales (e.g., average physical functioning score of 79.8 vs. 76.4).

Table 1 provides information on the prior use of E-alone or E+P at baseline in the E-alone trial and in the corresponding observational study subsample. Corresponding data for the E+P trial and corresponding observational study subsample are given in Prentice et al. (15).

Hazard ratio analyses

Following control for each of the potential confounding factors (age, body mass index, education, cigarette smoking, age at menopause, and physical functioning), the E-alone estimated hazard ratios in the observational study divided by those in the clinical trial were 0.77 (95 percent confidence interval (CI): 0.59, 0.99) for CHD, 0.74 (95 percent CI: 0.55, 0.99) for stroke, and 0.63 (95 percent CI: 0.44, 0.92) for venous thromboembolism. These results indicate a significant remaining discrepancy (p < 0.05) regarding each of the three clinical outcomes.

The four columns on the left side of table 2 show corresponding hazard ratios and 95 percent confidence intervals from the clinical trial and the observational study for E-alone as a function of time since initiation of the current hormone treatment episode (<2, 2–5, >5 years). These analyses also include the same potential confounding factors. As was found in the corresponding E+P analysis (15), there is general agreement between these clinical trial and observational study hazard ratios for CHD and venous thromboembolism, but not for stroke. Note, however, the broad confidence intervals for the observational study for less than 2 years since initiation of the time period for each disease, owing to few recent hormone treatment initiators at the time of enrollment in the observational study. In fact, 5, 31, and 385 E-alone users in the observational study developed CHD in these three respective time periods since hormone treatment initiation. Corresponding numbers were 1, 15, and 414 for stroke and 4, 13, and 244 for venous thromboembolism. Hence, neither these estimators nor their confidence intervals are reliable in the category of less than 2 years since hormone treatment initiation.

The four columns on the right side of table 2 show corresponding hazard ratios for both E-alone and E+P under a more restrictive hazard ratio model that is common to the clinical trial and the observational study except for an estimated proportionality factor for hormone treatment in the observational study compared with that in the clinical trial. Note that the hazard ratios tend to be somewhat lower in the observational study than in the clinical trial for both E-alone and E+P, especially for stroke, suggesting that the proportionality factor should be retained to control for residual bias. The similarity of these estimated proportionality factors for E-alone and for E+P suggests similar sources of residual bias, supporting the concept that comparison of the observational study with the clinical trial regarding E+P may be useful for adjusting the E-alone hazard ratios from the observational study, and vice versa.

Adjusted hazard ratio estimates for the observational study

The leftmost hazard ratio column in table 3 again shows E-alone hazard ratios as a function of years since initiation of hormone treatment; these data were obtained from analysis of the E-alone subset of the observational study for each of the three clinical outcomes. This information is followed by two versions of corresponding adjusted hazard ratios that arise from joint analyses of the E-alone observational study data and data from the E+P clinical trial and observational study cohorts. Because the E-alone and E+P cohorts were distinguished by the presence or absence of hysterectomy at baseline, the confounding factor coefficients were estimated separately for the E-alone and E+P components, although the two sets of estimates of these coefficients were quite similar and are shown in appendix table 1. The first set of adjusted hazard ratios (adjusted 1) arises from a model in which the same residual-bias proportional hazard ratio factors are assumed to apply to the observational study for E+P and E-alone. With some exceptions, this observational study adjustment moves the hazard ratio estimate from that based on observational study data alone toward the corresponding estimates from the E-alone clinical trial shown on the rightmost side of the table. The corresponding confidence intervals are widened by this adjustment.

The framework of analyzing the three cohorts together allows one to test the suitability of a more restrictive model under which the E-alone and E+P hazard ratios differ only by a proportionality factor across the three categories of years since hormone treatment initiation. There is no evidence against this assumption for CHD (p = 0.77) or venous thromboembolism (p = 0.62) and only modest evidence for stroke (p = 0.13). Hence, the third set of columns in table 3 (adjusted 2) shows hazard ratios under this additional modeling assumption. This adjustment procedure yields hazard ratios that are mostly closer to clinical trial hazard ratios than are those from the first adjustment procedure; differences from clinical trial hazard ratios were evidently attributable to chance. Note that the second adjustment procedure also yields narrower confidence intervals by borrowing more strongly on the E+P data to compensate for the paucity of recent hormone treatment initiators at observational study enrollment. In fact, the precision of the confidence intervals from this second set of analyses adjusted for observational study is similar to that for the E-alone clinical trial, supporting the use of the two independent data sources to more precisely address a range of subset analysis questions. Appendix 2 provides additional detail on the statistical models used in these analyses.

Table 4 presents the corresponding sequence of E+P analysis from the observational study and comparison with the E+P clinical trial. Once again, the adjustment procedure, which here considers E-alone clinical trial and observational study data, leads to reasonable hazard ratio agreement with corresponding hazard ratios from the E+P trial, with differences attributable to chance. For stroke, there is evidence (p = 0.01) against hazard ratio functions that are proportional across categories of years since hormone treatment initiation between E+P and E-alone, as is assumed by the second observational study adjustment procedure (adjusted 2), but not for the other outcomes.

To apply the observational study adjustment methods to an important subset analysis question, consider the second adjusted observational study analysis for E-alone and E+P in relation to CHD for women in the age range 50–59 years at WHI enrollment (table 5). Because of small numbers of CHD events in this subset (47 in the E-alone clinical trial, 89 in the E-alone observational study, 54 in the E+P clinical trial, 85 in the E+P observational study), the confidence intervals remain fairly wide. Comparisons of a corresponding averaged hazard ratio over the average 7.1-year follow-up period of the E-alone clinical trial, and over the average 5.6-year follow-up period of the E+P clinical trial, are of particular interest (bottom of table 5). For example, the 0.65 (95 percent CI: 0.37, 1.13) average hazard ratio for E-alone obtained from the adjusted observational study analysis agrees closely with that, 0.70 (95 percent CI: 0.36, 1.34), from a corresponding E-alone clinical trial analysis and with a previously presented (13) summary hazard ratio of 0.63 (95 percent CI: 0.36, 1.08) for this subset from the E-alone clinical trial. As an illustration of the use of clinical trial and observational study data to estimate longer-term effects, the statistical models applied on the left side of table 5 give hazard ratios averaged over a 10-year period from E-alone initiation of 0.69 from the clinical trial analysis and 0.58 from the adjusted observational study analysis. The shorter confidence interval for the latter analysis reflects the more precise hazard ratio estimate for more than 5 years since hormone treatment initiation in the observational study compared with the clinical trial. Appendix 2 provides additional details on calculating these average hazard ratios and their confidence intervals.

DISCUSSION

These analyses indicate that the correspondence between WHI clinical trial and observational study data in relation to cardiovascular disease shows very similar patterns for E-alone, as was the case in previous analysis (15) for E+P. Even though E-alone and E+P preparations tend to be used by different women according to hysterectomy status, this similarity in patterns supports an adjustment to observational study analyses using data on the effects of the other hormone treatment preparation. The adjusted hazard ratio analyses allow the baseline disease rates and confounding factor regression coefficients to vary between women with and without a hysterectomy and, importantly, allow the observational study hazard ratios to include a multiplicative residual bias term assumed to be common to E-alone and E+P. The resulting hazard ratios tend to be imprecisely estimated in the early years following the initiation of hormone treatment. However, the data are mostly consistent with hazard ratios that differ between E-alone and E+P by a proportionality factor that is independent of years since hormone treatment initiation. The precision of the hazard ratios from the observational study adjusted under this assumption is similar to, and hazard ratio estimates are mostly comparable to, those for the hazard ratios arising from the E-alone and E+P clinical trials. Similar to the E-alone clinical trial, these analyses indicate an elevated stroke risk and some early elevation in venous thromboembolism from E-alone and an elevation in stroke risk with a more sustained elevation in venous thromboembolism risk for E+P. CHD hazard ratio patterns are not so evident from these observational study analyses but are consistent with some modest early elevation for both E-alone and E+P.

The E-alone and E+P preparations used by women in the observational study were predominantly 0.625 mg of conjugated equine estrogen daily or 0.625 mg of conjugated equine estrogen plus 2.5 mg of medroxyprogesterone acetate daily (table 6). Analyses restricted to these daily preparations and these doses differed little from those given here. Adherence to baseline hormone treatment preparations in the observational study, or to randomized hormone treatment assignment in the clinical trial, could affect the relative magnitude of hazard ratio estimates from the clinical trial and observational study. Many of the preceding analyses were repeated by restricting the follow-up period for each woman to 6 months beyond the time when her baseline hormone treatment status changed. As expected, hazard ratios as a function of years since hormone treatment initiation tended to be further from 1 than those given above following this restriction, but comparisons between clinical trial and observational study hazard ratios seemed to be little affected.

There are many potential applications of the use of WHI clinical trial data to strengthen corresponding observational study analyses. Included may be analyses that examine the risks and benefits of treatments studied in the clinical trials over time periods longer than that considered in the available trials; analyses that attempt to add precision to clinical trial results in important subsets of a study population; and analyses that examine doses, schedules, routes of administration, or preparations other than those studied in the trials. To illustrate one such application, we used WHI clinical trial and observational study data to adjust observational study analyses of E-alone and CHD among women aged 50–59 years at baseline. The resulting average hazard ratio estimate over a 7.1-year follow-up period agrees well with that from the E-alone clinical trial, and it supports the hypothesis of some cardioprotection from E-alone for younger women. The corresponding E+P average hazard ratio estimate over a 5.6-year follow-up period also agrees with that from the E+P clinical trial.

Our study encourages additional joint analyses of WHI clinical trial and observational study data on a range of topics pertinent to the benefits and risks of postmenopausal hormone treatment. Our analyses also encourage joint analyses of clinical trial and observational study data on such topics as aspirin and heart disease and stroke; beta-carotene and lung cancer, prostate cancer, and heart disease; and vitamin E and heart disease and cancer, where substantial data currently exist from both clinical trials and observational studies.

APPENDIX 1

Appendix table 1 complements the hazard ratio estimates shown in the second analyses of table 3 by giving the hazard ratio estimates and 95 percent confidence intervals for the potential confounding factors.

APPENDIX 2

Statistical Models in Analyses of Tables 2–5

Tables 2–5 are based on Cox regression models (21) for the hazard rate h(t; Z(t)) at each follow-up time t, where Z(t) comprises hormone treatment and potential confounding factor histories for a women prior to time t. This hazard rate is modeled as h_os(t)exp(x(t)′b), where h_os is a “baseline” hazard rate in stratum s, x(t) is a modeled regression variable, and b is a vector of regression coefficients (log hazard ratios) to be estimated. Let V₁, … V₄ be indicator variables for the E-alone observational study, E-alone clinical trial, E+P observational study, and E+P clinical trial cohorts, respectively. The stratification s is defined by indicator variables for the six baseline 5-year age categories 50–54, 55–59, 60–64, 65–69, 70–74, and 70–79. Analyses are stratified by these six age categories and by the cohorts included. For example, the observational study adjusted 1 analysis in table 3 involves 18 strata, six age strata for each of the E-alone observational study, the E+P clinical trial, and the observational study cohorts.

The modeled covariates in each of the analyses shown in tables 2–5 include time-dependent indicator variables x₁(t) for hormone treatment (vs. control) for less than 2, 2–5, or more than 5 years since initiation of hormone treatment. Analyses on the left side of table 2 include these treatment variables plus the baseline potential confounding factors x₂ listed in appendix table 1, so that x(t)′ = (x₁(t)′, x₂′) in separate analysis of the E-alone observational study and the E-alone clinical trial cohorts. Analyses on the right side of table 2 include these variables plus an indicator variable x₃ for hormone treatment in the observational study, which generates the factor labeled hazard ratio in the observational study/hazard ratio in the clinical trial.

The second analysis of table 3 has covariate vector x(t)′ = (x₁(t)′, x₁(t)′y, x₂(t)′V₁, x₂(t)′(V₃ + V₄), x₃), where y is an indicator variable for E+P hormone treatment that allows hazard ratios to differ between E-alone and E+P. The third table 3 analysis replaces x₁(t)′y with y in the covariate definition. Table 4 definitions are analogous to those for table 3, interchanging the roles of E-alone and E+P. Table 5 analyses correspond to the second and third analyses of tables 3 and 4, with restriction to CHD and to women 50–59 years of age at baseline.

The average hazard ratios over a time period T years in table 5 are estimated by adding the hazard ratio for less than 2 years times 2, the hazard ratio for 2–5 years times 3, and the hazard ratio for more than 5 years times T − 5, and dividing by T. The corresponding asymptotic variance is calculated from the variance estimator for the regression parameter estimate from the partial likelihood (21) by using the delta method.

The WHI program is supported by contracts from the National Heart, Lung, and Blood Institute; National Institutes of Health; and US Department of Health and Human Services and by a grant from the National Cancer Institute.

The authors thank the WHI investigators and staff for their outstanding dedication and commitment.

A list of key investigators involved in this research follows. A full listing of WHI investigators can be found at the following website: http://www.whi.org.

Program Office—National Heart, Lung, and Blood Institute, Bethesda, Maryland: Barbara Alving, Jacques Rossouw, and Linda Pottern. Clinical Coordinating Center—Fred Hutchinson Cancer Research Center, Seattle, Washington: Ross Prentice, Garnet Anderson, Andrea LaCroix, Charles L. Kooperberg, Ruth E. Patterson, and Anne McTiernan; Wake Forest University School of Medicine, Winston-Salem, North Carolina: Sally Shumaker; Medical Research Labs, Highland Heights, Kentucky: Evan Stein; and University of California at San Francisco, San Francisco, California: Steven Cummings. Clinical Centers—Albert Einstein College of Medicine, Bronx, New York: Sylvia Wassertheil-Smoller; Baylor College of Medicine, Houston, Texas: Jennifer Hays; Brigham and Women's Hospital, Harvard Medical School, Boston, Massachusetts: JoAnn Manson; Brown University, Providence, Rhode Island: Annlouise R. Assaf; Emory University, Atlanta, Georgia: Lawrence Phillips; Fred Hutchinson Cancer Research Center, Seattle, Washington: Shirley Beresford; George Washington University Medical Center, Washington, DC: Judith Hsia; Harbor-UCLA Research and Education Institute, Torrance, California: Rowan Chlebowski; Kaiser Permanente Center for Health Research, Portland, Oregon: Evelyn Whitlock; Kaiser Permanente Division of Research, Oakland, California: Bette Caan; Medical College of Wisconsin, Milwaukee, Wisconsin: Jane Morley Kotchen; MedStar Research Institute/Howard University, Washington, DC: Barbara V. Howard; Northwestern University, Chicago/Evanston, Illinois: Linda Van Horn; Rush-Presbyterian St. Luke's Medical Center, Chicago, Illinois: Henry Black; Stanford Prevention Research Center, Stanford, California: Marcia L. Stefanick; State University of New York at Stony Brook, Stony Brook, New York: Dorothy Lane; The Ohio State University, Columbus, Ohio: Rebecca Jackson; University of Alabama at Birmingham, Birmingham, Alabama: Cora E. Lewis; University of Arizona, Tucson/Phoenix, Arizona: Tamsen Bassford; University at Buffalo, Buffalo, New York: Jean Wactawski-Wende; University of California at Davis, Sacramento, California: John Robbins; University of California at Irvine, Orange, California: Allan Hubbell; University of California at Los Angeles, Los Angeles, California: Howard Judd; University of California at San Diego, LaJolla/Chula Vista, California: Robert D. Langer; University of Cincinnati, Cincinnati, Ohio: Margery Gass; University of Florida, Gainesville/Jacksonville, Florida: Marian Limacher; University of Hawaii, Honolulu, Hawaii: David Curb; University of Iowa, Iowa City/Davenport, Iowa: Robert Wallace; University of Massachusetts/Fallon Clinic, Worcester, Massachusetts: Judith Ockene; University of Medicine and Dentistry of New Jersey, Newark, New Jersey: Norman Lasser; University of Miami, Miami, Florida: Mary Jo O'Sullivan; University of Minnesota, Minneapolis, Minneapolis: Karen Margolis; University of Nevada, Reno, Nevada: Robert Brunner; University of North Carolina, Chapel Hill, North Carolina: Gerardo Heiss; University of Pittsburgh, Pittsburgh, Pennsylvania: Lewis Kuller; University of Tennessee, Memphis, Tennessee: Karen C. Johnson; University of Texas Health Science Center, San Antonio, Texas: Robert Brzyski; University of Wisconsin, Madison, Wisconsin: Gloria E. Sarto; Wake Forest University School of Medicine, Winston-Salem, North Carolina: Denise Bonds; and Wayne State University School of Medicine/Hutzel Hospital, Detroit, Michigan: Susan Hendrix.

Wyeth-Ayerst Laboratories (Madison, New Jersey) provided medication tested in this study. Dr. Langer has been a consultant for Wyeth-Ayerst and has received research support from this company within the past 3 years.

References