Cardiovascular disease, primarily coronary artery disease (CAD), is the leading cause of death among women in the United States. Many observational studies and analyses have suggested that HRT protects against CVD. However, in the past 4 years, three secondary prevention trials (2-4) and, most relevantly, the Women's Health Initiative (WHI) primary prevention study (5) have shown no benefit or increased rates of CVD events among women randomly assigned to HRT. These results were surprising to many because of the benefit shown in many observational studies and because estrogen use is associated with many effects that could have favorable effects on CVD, including lower low-density lipoprotein cholesterol levels (6), lower lipoprotein(a) levels (7), and increased high-density lipoprotein cholesterol levels (8). However, HRT is also associated with potentially unfavorable effects, including increased levels of triglycerides (6), factor VII, and C-reactive protein (8) and decreased levels of antithrombin III (6).

Given the mix of intermediate biological outcomes, limitations of observational data, and recent trial results, we conducted this systematic review and meta-analysis to examine the value of HRT for the primary prevention of CVD. We were particularly interested in whether bias might explain discordant results between recent trials and the observational literature. Our review is one of several that will serve as background for the Third U.S. Preventive Services Task Force (USPSTF) recommendations on HRT and the prevention of chronic diseases.

Methods

Top Methods Results Discussion Author & Article Info References

We searched the topic of HRT and CVD and CAD in the MEDLINE and Cochrane databases from 1966 to December 2000. We also used bibliographies of original research and other publications to identify studies for review. Criteria for inclusion in the systematic review were evaluation of HRT and the primary prevention of CVD, CAD, or both among postmenopausal women and availability of an English-language abstract for review. We included randomized, controlled trials, and cohort and case–control studies if they evaluated CVD or CAD incidence or mortality. We did not review cross-sectional studies because they are limited by prevalence bias. Two investigators reviewed all abstracts to identify papers for full-text review.

We abstracted study data and created evidence tables organized by study type. Definitions of current, past, recent, never, and ever use were taken directly from the individual studies (Appendix Tables 1 and 2). Hormone use was classified in each study as unopposed estrogen replacement therapy or estrogen plus progesterone "combined therapy" when it was specified, which was infrequent. When the type of estrogen or progesterone therapy was not specified, the exposure was categorized as HRT.

Two investigators independently rated the quality of each study based on criteria created by the Third USPSTF (9); discrepancies were adjudicated by a third reviewer. These criteria are shown in Table 1. Appendix Tables 1 and 2 show each study's rating by quality. In ranking the quality of observational studies, we gave significant weight to adequate control of potential CVD risk factors because of known differences in risk profiles among women who use HRT and those who do not (10, 11).

After reviewing and rating the studies, we limited our formal review and meta-analyses to only studies rated as fair or good quality and included randomized, controlled trials; population-based, case–control studies; and cohort studies with internal controls and at least 3 years of follow-up. In studies with multiple publications from the same cohort or population, only data from the most recent publication were included in the meta-analyses, with reference in the text to older publications if they presented unique findings. To evaluate each study's control of confounding variables that might explain their results, for each outcome and each study providing data relevant to the outcome, we listed potential CVD risk factors included in the multivariable models determining relative risk. These included age, diabetes, hypertension, smoking, lipid levels, family history, exercise, socioeconomic status, education level, alcohol use, body mass index, and others. Definitions for the presence or absence of these factors, as well as their assessment, are taken directly from the included studies.

This research was funded by the U.S. Agency for Healthcare Research and Quality under a contract to support the work of the USPSTF. Agency for Healthcare Research and Quality staff and USPSTF members participated in the initial design of the study and reviewed interim analyses and the final manuscript. Since our report was prepared for the Third USPSTF, it was distributed for review to 15 content experts and revised accordingly before preparation of this manuscript.

For each outcome, a meta-analytic model was fitted, stratifying by HRT exposure status: current, past, and ever (if the study did not report results for current and past use separately). The model also allowed for a global effect of HRT to be estimated; this is referred to as the effect of any HRT exposure (current, ever, or past). Our primary analysis compared event rates in patients with "any HRT use" (current, past, or ever use) to those who had "never" used HRT. We also compared event rates in "current," "past," or "ever" HRT use groups to "never" use to further evaluate variable findings among the studies. Results by category of use are shown in Table 2. All relative risk estimates are compared with never use. Estimation was made by using the Bayesian data analytic framework, in which the analogue to the confidence interval is the credible interval (CrI) (16). Further details of the meta-analysis are given in the Appendix.

Results

Top Methods Results Discussion Author & Article Info References

Scope of Literature

The MEDLINE search identified 3035 abstracts of papers evaluating primary prevention published between 1966 and December 2000. From them, 24 cohort studies; 18 case–control studies; one very small randomized, controlled trial; and one meta-analysis were reviewed. After the quality of the studies was rated, 20 observational studies; one randomized, controlled trial conducted among institutionalized women; and 1 meta-analysis that combined data from 23 randomized, controlled trials of HRT for outcomes other than CVD (such as bone density or lipid levels) were included in our meta-analyses. Data from studies graded as poor quality were not included in the meta-analyses. Our search results are shown in the Appendix Figure.

View larger version (31K): [in this window] [in a new window]   Appendix Figure. Search methods and results. HRT = hormone replacement therapy; RCT = randomized, controlled trial.

Cardiovascular and Coronary Artery Disease Mortality

All of the studies contributing to our mortality analyses were prospective cohort studies, except for a nested case–control study from the Nurses' Health Cohort. There was little variation among the cohort studies and this study; therefore, these studies were combined. Similarly, when stratified by study quality (fair or good) in the meta-analyses, the findings were similar and the results were combined. More details about the studies are presented in Appendix Tables 4 and 5.

View larger version (21K): [in this window] [in a new window]   Figure 1. Relative risk or odds ratio for cardiovascular disease mortality. Error bars represent 95% CIs. HRT = hormone replacement therapy. "Current," "past," and "ever" use are categories used in the individual studies. "Any use" is a category created for this meta-analysis combining data from studies evaluating ever and never use of HRT with data from studies evaluating current, past, or never use (current + past + ever use).

In our meta-analysis evaluating CVD mortality, the global measure of HRT exposure, "any use," was associated with a summary relative risk of 0.75 (CrI, 0.42 to 1.23) (Table 2 and Figure 1). Five studies specifically evaluated the risk for CAD death (13, 39-42). In our meta-analysis of HRT and CAD mortality, "any use" had a summary relative risk of 0.74 (CrI, 0.36 to 1.45) (Table 2 and Figure 2).

View larger version (18K): [in this window] [in a new window]   Figure 2. Relative risk or odds ratio for coronary artery disease mortality. Error bars represent 95% CIs. HRT = hormone replacement therapy. "Current," "past," and "ever" use are categories used in the individual studies. "Any use" is a category created for this meta-analysis combining data from studies evaluating ever and never use of HRT with data from studies evaluating current, past, or never use (current + past + ever use).

Cardiovascular Disease Incidence

Two cohort studies, one case–control study, and the data from the meta-analysis described earlier contributed to our meta-analysis of CVD incidence. Since only four studies (11, 13, 46, 47) contributed to this analysis and the effect sizes were in the same direction (range, 1.07 to 1.76), the data were pooled. In our meta-analysis, "any use" of HRT was associated with a summary relative risk of 1.28 (CrI, 0.86 to 2.00) for CVD incidence. Results were similar when a previous meta-analysis (47) was used as the prior distribution (Table 2 and Figure 3). More details about the studies are presented in Appendix Table 6.

View larger version (16K): [in this window] [in a new window]   Figure 3. Relative risk or odds ratio for cardiovascular disease incidence. Error bars represent 95% CIs. Reference 47 is pooled data from a previously published meta-analysis. HRT = hormone replacement therapy. "Current," "past," and "ever" use are categories used in the individual studies. "Any use" is a category created for this meta-analysis combining data from studies evaluating ever and never use of HRT with data from studies evaluating current, past, or never use (current + past + ever use).

Coronary Artery Disease Incidence

The association between HRT use and CAD incidence was evaluated in 3 cohort studies (11, 13, 48); 9 case–control studies (12, 14, 15, 49-54); and one small randomized, controlled trial (55). Relative risks ranged from 0.33 to 1.90. Because of the marked range of results, we conducted analyses stratifying by study type and identified little difference in our summary estimates. Therefore, the case–control and cohort studies were combined. Since only 1 trial of HRT in primary prevention has formally published data, we included it in our summary estimate. Past use of HRT was evaluated in 6 studies (13-15, 48, 49, 52), and combined therapy was evaluated in 4 studies (15, 48, 51, 55); these findings varied (Figure 4).

View larger version (29K): [in this window] [in a new window]   Figure 4. Relative risk or odds ratio for coronary artery disease incidence. Error bars represent 95% CIs. HRT = hormone replacement therapy; RCT = randomized, controlled trial; SES = socioeconomic status. "Current," "past," and "ever" use are categories used in the individual studies. "Any use" is a category created for this meta-analysis combining data from studies evaluating ever and never use of HRT with data from studies evaluating current, past, or never use (current + past + ever use). Adjustments for SES were adjustments as assessed by social class (12, 13), education (14, 15), and income.

Table 3 indicates that many studies had point estimates below 1.0 (52). This suggests benefit, although only four studies had statistically significant results. The reasons for disparity among the five good-quality studies are uncertain but may be related to variable assessment of confounders, as shown in Table 3. Appendix Tables 1 and 2 show that all studies included in this systematic review assessed and defined HRT use differently. However, these tables do not show a pattern that helps further explain disparate findings among studies evaluating CAD incidence.

Results of the meta-analysis evaluating the association between CAD incidence and HRT use varied by exposure status (Table 2). For CAD incidence, "any use" was associated with a relative risk of 0.88 (CrI, 0.64 to 1.21). To further explore the differences found among the studies evaluating CAD incidence, we conducted sensitivity analyses stratified by several confounding variables. First, we compared the summary relative risks among the studies with statistical adjustment for socioeconomic status or education with those that did not adjust. Among the studies adjusting for these factors, there was no association between any measure of HRT use and CAD events, with relative risks ranging from 0.97 to 1.11 (Table 2). However, the summary relative risks among studies that did not adjust for socioeconomic status were reduced with current exposure (relative risk, 0.71 [CrI, 0.64 to 0.78]) and past exposure (relative risk, 0.78 [CrI, 0.69 to 0.87]) (Figure 4). Similar results were found when the analysis stratified studies that adjusted for alcohol consumption, exercise, or both, suggesting confounding by these factors. Further stratification did not show any differences between results of case–control and cohort studies or between results of fair- and good-quality studies.

Discussion

Top Methods Results Discussion Author & Article Info References

Earlier data on HRT and the primary prevention of CVD showing benefit to users are not supported by newer studies or by our analysis. The major points from our review and meta-analyses are as follows. First, no significant association was identified between past, ever, or any use of HRT and CVD or CAD death. Second, HRT use did not reduce CVD incidence and, in fact, suggests a small increase in risk. Third, HRT showed no benefit in preventing CAD among the studies that adjusted for major CAD risk factors and socioeconomic status or education and showed reduced risk among studies that did not adjust for these factors, suggesting confounding.

We approached this review differently than other researchers (56-58). First, in an effort to measure a global effect on CVD, as well as a more specific effect on CAD, we evaluated CVD and CAD separately when the data allowed. Second, we conducted separate analyses of incidence and mortality for each outcome. Third, we limited our detailed review and meta-analyses to studies of good or fair quality based on preestablished criteria. Finally, in an effort to better explain variable findings, we evaluated risks in our meta-analyses using different measures of exposure (current, past, ever), as well as a global measure (any use).

Other meta-analyses have had different findings. A meta-analysis in 1991 (56) that included 6 case–control studies, 16 cohort studies, and 3 angiography studies found a relative risk of 0.56 (CI, 0.50 to 0.61) for CAD events (incidence and mortality). A 1992 meta-analysis of HRT and CAD calculated a relative risk of 0.65 (CI, 0.59 to 0.71) for CAD events and of 0.63 (CI, 0.55 to 0.72) for CAD death when comparing ever use with nonuse (57). A more recent meta-analysis calculated a summary risk estimate for CAD of 0.70 (CI, 0.65 to 0.75) for HRT use (58). Each of these meta-analyses included studies that we rated as poor quality and excluded from our analysis. As discussed earlier, poor-quality studies tended to suggest greater protection against CVD and CAD in association with HRT use. In addition, previous meta-analyses included angiography studies that involved symptomatic women and cross-sectional design, which has many shortcomings. These meta-analyses also differ from ours because they combine mortality and incidence relative risks for HRT. These differences may explain why our review, unlike previous analyses, suggests no overall benefit from HRT in preventing CVD or CAD.

One of the difficulties in assessing the literature on HRT and CVD is the large span of years represented in the observational studies, during which dramatic changes in clinical practice and CVD knowledge occurred. There have also been significant secular changes in the use of estrogen, including type, administration, and dose, as well as the relatively recent practice of adding progesterone to estrogen therapy. Complicating this evaluation is that many studies measured estrogen use at only one point in time, asked women if they had "ever" used HRT, or combined past use with non-use, allowing substantial room for misclassification. Such misclassification would dilute any potential association between HRT and CVD (Appendix Tables 1 and 2). The importance of characterizing HRT use is illustrated in the Framingham study, where a change in how HRT use was assessed (single vs. multiple assessments) changed the direction of relative risk estimates for total mortality among younger postmenopausal women (11).

One of the most important findings of our review and analysis is how different evaluation and statistical control of CAD risk factors affect summary estimates. This is highlighted in our meta-analysis, which showed markedly different relative risks depending on the inclusion or exclusion of socioeconomic status or education as predictor variables in a study's multivariable analysis. These findings suggest confounding, which is important because lower socioeconomic status is a strong risk factor for CVD and CAD, as well as for most other poor outcomes (59). In addition, women using HRT tend to have higher socioeconomic status, which may explain their better outcomes (60). Similarly, none of the studies with adjustment for alcohol use or exercise, both known to be more common in women who use HRT, showed benefit with HRT use.

Several biases complicate the interpretation of our results, as well as those of others. The first consideration is selection bias, particularly healthy-user bias. Women who use HRT tend to be more affluent, leaner, and more educated; tend to exercise more often; and tend to drink more alcohol (11, 61). Women who take HRT also have different health characteristics before menopause (10, 11, 62). Researchers can adjust for these protective factors analytically when measuring them; what cannot be adjusted for statistically, however, are lifestyle, environmental exposures, and genetic characteristics that are not measured or may not yet be identified as important etiologically in CVD. This is particularly an issue in CVD, where 50% of CVD incidence is unexplained by traditional risk factors (63).

Women prescribed HRT have access to health care and are therefore more likely to be receiving treatment for other CVD-CAD risk factors, such as high cholesterol levels or high blood pressure, which would lower their risk (60). In contrast to reduced mortality in association with current HRT use, there is a consistent although not statistically significant increase in rates of CVD associated with any HRT exposure. Several possibilities may explain this finding. Women receiving HRT have access to health care and may be more likely to receive diagnoses of CVD, such as myocardial infarction, peripheral vascular disease, stroke, or transient ischemic attack. In addition, after receiving these diagnoses, they may undergo more aggressive risk factor modification or medical management, which may partially explain reduced mortality in the setting of average or slightly increased incidence. Finally, several sources suggest that women using HRT have higher rates of ischemic stroke, which would increase rates of CVD events (48, 64, 65).

Selection for healthy users is also a consequence of secular trends in estrogen use (66). Many studies of estrogen use were conducted when physicians were concerned about the risk for HRT and CVD, based on the Coronary Drug Project findings among men and myocardial infarction rates in women taking oral contraceptives (66). In addition, for many of the years represented in these studies, hypertension, diabetes, and CAD were considered contraindications to HRT (58).

A more subtle bias, which may be apparent to practicing physicians, is a tendency to offer and prescribe HRT to women who are perceived as being in better overall "health," even in the absence of defined CVD or CAD risk factors. Supporting this contention are empirical data showing that women with medical or psychiatric problems are less likely to receive HRT or treatment of other unrelated problems (67). This type of selection bias is more difficult to measure and could lead to systematic overestimates of the benefit of HRT in CVD.

Another aspect of healthy user bias is that women often discontinue HRT when they become ill (68). This tendency would bias studies that evaluate recent or current use by underestimating use in ill patients. Such underestimation would result in reduced relative risk estimates associated with current or recent exposure, suggesting a protective effect of HRT. This bias may have occurred in studies where only current or recent users have reduced risk for CVD or CAD. It is even more strongly suggested in studies where past users have higher rates of CVD than nonusers or current users, suggesting that they may have stopped HRT because of CVD- or CAD-related illness (32). Healthy user bias may particularly affect mortality associations because most deaths occur among women with comorbid conditions that led them or their clinician to stop HRT, making hormone use, particularly current use, appear protective against CVD death. This is suggested in our meta-analyses evaluating CVD and CAD mortality, where only current use of HRT was associated with decreased risk and "any use" of HRT (past, current, or ever) showed no association. Supporting the concept of healthy user bias among women using HRT are several studies showing reduced all-cause mortality, as well as reduced mortality from accidents and homicides, among women who take HRT. These findings may reflect multiple benefits but more likely indicate systematic differences among users and nonusers (42, 43).

Another consideration in evaluating the relationship between HRT and CVD is the issue of adherence bias. Women who take HRT, especially for long periods, are by definition adherent to therapy. In randomized, controlled trials, good adherence to placebo has been shown to decrease CAD events by 30% to 60%, suggesting that good adherence is a marker for other healthy behaviors (43, 69-71). Adherence bias itself may explain much of the benefit seen in observational studies of HRT and CVD.

In the past 4 years, data from two randomized, controlled trials of HRT in the secondary prevention of CAD have been published (2, 3), and one trial of HRT in primary prevention has released information to the public (5). These findings are important because randomization is the only way to deal with the inherent biases in observational studies and to ensure equal distribution of known and unknown CVD risk factors. The Heart and Estrogen/progestin Replacement Study (HERS), which examined secondary prevention, showed a 52% increased risk for myocardial infarction during the first year of use and suggested increased CAD deaths in the first 3 years of use (72). The Estrogen Replacement and Atherosclerosis Trial of secondary prevention showed no benefit from HRT in reducing angiographic CAD progression (73). Finally, early results from the WHI randomized, controlled trial of HRT and primary prevention, involving 27 348 postmenopausal U.S. women, have shown increased rates of stroke, CAD, and blood clots among women randomly assigned to HRT compared with those receiving placebo (5). These findings have persisted into the third year of the study (5, 73).

How can the results of these trials of secondary prevention, and especially the WHI study of primary prevention, be explained, given the results from many previous observational studies suggesting benefit? As discussed earlier, it is likely that selection and adherence bias play a major role in the findings from the observational studies. What is especially surprising in two of these studies, however, is the suggestion of harm in the first 2 to 3 years (5, 72). For years, it has been thought that the most likely biological explanation for some of the observed reduction in CVD risk among HRT users was an improvement in lipid levels. However, the fact that two randomized, controlled studies have shown that women had higher CVD risks in the first 2 to 3 years of the study, one in the setting of more favorable lipid profiles (72), suggests that other important biological effects occur in women who take HRT. For example, HRT plays a complex role in clotting, thrombolysis, and inflammation, which must be considered because CVD and CAD events are partially mediated through these processes. Biological changes associated with HRT may result in a shift in balance toward increased blood clotting, inflammation, or both. This shift may be a more immediate or acute effect of estrogen and is consistent with the observation of increased risk during the early years of estrogen use. An important interaction or synergism between the multiple, relatively prevalent hypercoagulable states and HRT may also account for early increased risk, as suggested in several recent studies (74-76). More studies are needed in this area.

With the publication of the HERS results and the preliminary reporting of increased event rates in the first 3 years of the WHI, one of the most pressing questions facing investigators and clinicians is whether this early increase in events is later offset by a reduction, perhaps related to improved lipid profiles or other physiologic changes. Among the studies of primary prevention included in our meta-analyses, only one large, high-quality, case–control study (15) suggests an early increased risk for CAD events (relative risk for myocardial infarction was 1.5 during the first year) that decreases with time. Four other studies evaluating estrogen use in the secondary prevention of CAD also suggest an increase in early CVD events in association with estrogen, which later decreases (77-80). In addition, increased rates of clotting in the first 2 years of HRT use are supported by a recent meta-analysis of thrombosis and HRT (81). Finally, a recent study of HRT use in the secondary prevention of stroke showed a marked increase in stroke risk during the first 6 months of use (relative risk, 2.3) that decreased to 1.1 after 33 months of use (4). Thus, several forms of evidence suggest increased rates of cardiovascular, coronary, and thrombotic events in the first 2 to 3 years of HRT use.

In summary, on the basis of this review and meta-analysis, and after extrapolation from two trials of secondary prevention, there is good reason to question the results of observational studies supporting the use of HRT in the primary prevention of CVD and CAD. Because of the limitations of observational studies, randomized, controlled trials are the best way to evaluate the relationship between HRT and CVD or CAD. We hope that such trials will yield better information in the near future. Until such information is available, the primary prevention of CVD and CAD in women should focus on proven strategies to reduce CVD and CAD risk. On the basis of current evidence, we do not advise consideration of CVD prevention when discussing HRT use with women.

Addendum: One week before this article went to press, the WHI reported the results of a randomized trial comparing estrogen plus progesterone with placebo (82). The data safety monitoring board halted the study because the risk for breast cancer exceeded a prespecified level. The report included CVD outcomes that are highly pertinent to this review of the cardiovascular effects of HRT. This addendum summarizes the WHI cardiovascular outcomes.

The study recruited 50- to 79-year-old postmenopausal women by direct letter solicitation and a study- awareness media campaign and randomly assigned 16 608 of them to take 0.625 mg of conjugated equine estrogens plus 2.5 mg of progesterone acetate in a single tablet or an identical placebo. Follow-up was by semiannual structured interview and self-administered questionnaire with annual in-clinic visits. Cardiovascular outcomes included coronary heart disease (CHD) (defined as acute myocardial infarction, silent myocardial infarction, or coronary heart disease death), as determined at each study center by a blinded adjudicator who used prespecified diagnostic criteria. The authors expressed the primary outcomes as hazard ratios (HRs) with two types of 95% CIs. Nominal CIs corresponded to the results of a simple trial with one outcome; adjusted CIs corrected for many end point determinations over time and were wider than nominal CIs. The authors based their conclusions on the nominal CIs (shown here).

After a mean of 5.2 years of follow-up, the annual rate of CHD was 37 per 10 000 women in the HRT group and 30 per 10 000 women in the placebo group (HR, 1.29 [CI, 1.02 to 1.63]). The annual stroke rate was 29 per 10 000 women in the HRT group and 21 per 10 000 women in the placebo group (HR, 1.41 [CI, 1.07 to 1.85]). The rates of coronary artery bypass grafting and percutaneous transluminal coronary angioplasty were 42 and 41 per 10 000 women in the HRT and placebo groups, respectively. The curves showing the cumulative hazard of CHD diverged shortly after randomization and showed no signs of converging at any time after follow-up. Annual rates of cardiovascular death were 15 per 10 000 women in the HRT group and 13 per 10 000 women in the placebo group.

A subgroup analysis of women who met the entry criteria of the HERS study (68) (previous myocardial infarction or revascularization) had the same results as the women who had no previous CHD. The increased risk for CHD was present in all strata (age, ethnicity, antecedent CHD risk). Limitations of the study included a very high rate of discontinuing HRT (which should have biased the results toward no effect of HRT). The study could not apportion the increased disease risk between estrogen and progesterone.

The WHI results underscore the main conclusion of our review and meta-analyses: Reducing CVD or CHD is not a valid reason for using HRT. Indeed, the WHI confirms the unexpected finding of the HERS study by showing that HRT increases the risk for CHD. According to the WHI results, 10 000 women would suffer an additional 7 CHD events per year if they took HRT. Women at higher baseline risk for CHD are likely to suffer even more events if they take HRT.

Appendix: Bayesian Statistical Model

Four outcomes were analyzed separately: the incidence and mortality of CVD and CAD. Studies contributed multiple data points if they reported separate results for current and past users of HRT. Such studies almost never reported results for the combined group of "ever" HRT users. The results from studies that did not distinguish current users from past users were categorized under "ever" users. The measure of "any use" was created for this meta-analysis and includes data from studies measuring only "ever" use with data from studies measuring "current," "past," or "never" use. Any use equals ever + current + past use and is a global measure of use. All studies reported relative risk estimates using either odds ratios or hazard ratios. Adjusted relative risks were used because they represent the original authors' best estimate of the relative risk. The logarithm of the relative risk (logRR) is used as the data point for the effect size since it is assumed to be normally distributed. Standard errors for logRR were calculated from reported CIs or P values.

The mean logRR for study or data point i is defined as µ_i, which has the general form µ_i = ß₀ + ß₁x_i,1 +...+ ß_j x _i,j, where the x _i,j are indicator variables for study-level factors. In the presence of no study-level covariates, µ_i = ß₀. Study-level factors examined in the analysis include study design (cohort; case–control; and randomized, controlled trial); HRT exposure type (unopposed estrogen, combined therapy, and unspecified HRT); study quality (fair and good); HRT exposure status (current, past, and ever use); and whether the study adjusted for socioeconomic status, alcohol use, exercise, or cholesterol level.

Since studies measure and report HRT use differently, we wanted to preserve stratification by exposure status. However, we also wanted to allow for the estimation of a global measure of relative risk associated with any HRT exposure and for variation between status categories. To do this, we created the category of "any HRT use" by specifying a hierarchical model. Under this model, µ_i = ß_current x _i,current + ß_past x _i,past + ß_ever x _i,ever + ß₁ x _i,1 +...+ ß_j x _i,j, and x _i,current, x _i,past, x _i,ever are indicator variables for whether the data point i corresponds to the exposure category. ß_current, ß_past, and ß_ever are assumed to have normal distribution with a common mean, µ_anyHRT, and variance, ² _anyHRT. The global effect of any exposure is represented by µ_anyHRT. Variance among exposure categories is represented by ² _anyHRT. The model allows for further stratification by adding terms to µ_i.

The logarithm of the relative risk is assumed to have the following distribution: log RR _i " Normal(µ_i + z _i², s ² _i), where z _i " Normal(0,1) and s _i is the standard error calculated from reported CIs or P values. ²represents between-study variance and z _i represents the deviation between the logRR of the individual study or data point and the population. Under a fixed-effects model, ² = 0; under a random-effects model, ² > 0. The model is estimated by using a Bayesian data analytic framework ADDIN ENRfu (16). The data were analyzed by using WinBUGS (MRC Biostatistics Unit, Cambridge, United Kingdom) (83), which uses Gibbs sampling to simulate posterior probability distributions. Noninformative (proper) prior probability distributions were used. Specifically, Normal(0, 10⁶) prior distributions were used for ß_j, ß_current, ß_past, and ß_ever, and inverse (0,001, 0.001) prior distributions were used for ²and ² _anyHRT. Five separate Markov chains with various initial values were used to generate draws from posterior distributions. Point estimates (mean) and 95% credible intervals (2.5 and 97.5 percentiles) were derived from the subsequent 5 x 2000 draws after reasonable convergence of the five chains was attained.