Hormone Replacement Therapy and Prevention of Nonvertebral Fractures: Title and subTitle BreakA Meta-analysis of Randomized Trials FREE

Hormone replacement therapy (HRT) is generally thought to prevent fractures. Indeed, for women older than 60 years, the main reason for using HRT is for fracture prevention.¹ The basis of this belief, however, rests almost entirely on nonrandomized studies.^{2 - 4} Only 1 randomized trial of HRT has been reported with fracture reduction as the primary outcome, but this 5-year interim analysis (with low statistical power) of a 20-year study did not show a statistically significant reduction in fractures.⁵ Furthermore, a widely quoted trial⁶ that showed an apparent reduction of vertebral fracture incidence of 60% counted the number of new fractures not the number of women with new fractures. When an analysis was undertaken that compared the numbers of women in each group that had 1 or more new fractures, the effect of HRT was no longer statistically significant.⁷

In addition, the Heart and Estrogen/progestin Replacement Study (HERS), which is the largest HRT trial to date, also produced no evidence of an antifracture effect of HRT.⁸ On the other hand, a small factorial trial of HRT, vitamin D, and placebo among 464 young (aged 47-56 years) postmenopausal women did demonstrate a statistically significant reduction in nonvertebral fractures in an adjusted analysis when HRT was compared with placebo.⁹ This lack of substantial and consistent antifracture efficacy data on HRT from randomized trials has led the US Food and Drug Administration to withdraw conjugated equine estrogen as an approved treatment for osteoporosis.¹⁰

Although there is substantial evidence from nonrandomized studies for an antifracture effect of HRT,^{2 - 4} these data could be subject to a range of biases, which might make the results unreliable.¹¹ Given that HRT is widely recommended in clinical guidelines to be the first-line preventive therapy for osteoporotic fractures among postmenopausal women,¹² this paucity of randomized evidence is disturbing.¹³ Therefore, we undertook a systematic review of all trials of HRT in which data on nonvertebral fracture outcomes were collected, even if such data were not reported in the original published articles.

A search was undertaken to identify any existing systematic reviews. Recent systematic reviews were used to identify all the HRT literature published before 1997,^{12 ,14} along with a review of HRT's effects on cardiovascular events.¹⁵ For more recent studies (1997 to December 2000), MEDLINE, EMBASE, Science Citation Index, and the Cochrane Controlled Trials Register were searched using a combination of the following key words: HRT, ERT, hormon* replacement, (o)estrogen replacement, (o)estradiol, (o)estrone, and dien(o)estrol. We identified many studies that did not report fracture data either as an outcome or adverse event, despite the high likelihood that fractures did occur in those studies. Therefore, we tried to contact authors of all trials published since 1990 to establish whether fracture data had been collected but not reported. Investigators of earlier trials were not contacted on the basis that it was unlikely that fracture data from so long ago, if not already published, would be still available.

We also spoke to and e-mailed researchers in the field and pharmaceutical companies in an effort to identify any unpublished studies that we could include in our review.

We only included trials in which women had been randomized to at least 12 months of treatment and reported or made available nonvertebral fracture data. Fractures could be at any site (other than the spine) and due to any cause, thus, we may have included fractures that had been caused by high trauma or pathology, such as malignancy, and may not therefore be defined as osteoporotic.

Control treatments had to be either inactive placebo, no treatment, or calcium supplementation with or without vitamin D therapy. We included unblinded trials. Most identified trials only reported surrogate outcomes (eg, changes in bone density) and did not report fracture data. Trials were not included that only reported vertebral fractures.

In dose-ranging trials, all treatment arms were combined and this resulted in some trials with the HRT group being larger than the control arm. Relative risks (RRs) for fracture were estimated with 95% confidence intervals (CIs). Both reviewers extracted data independently and appraised the quality of the trials using the Jadad scale and disagreements were resolved by discussion.¹⁶ This validated quality scale included the following criteria: method of randomization, concealment allocation, and reporting of withdrawal and dropouts.

To assess whether bias could be introduced by either differential compliance or follow-up between treatment arms, we undertook a meta-analysis comparing the proportion of women not complying with the study medication and not being followed up.

A priori we would expect some clinical heterogeneity, as observational data does suggest a treatment interaction with age.^{2 - 4} Furthermore, the identified trials were of varying length ranging from 1 to 10 years of treatment. Thus, there were 2 obvious sources of heterogeneity. However, when we formally tested for statistical evidence of heterogeneity, we found none (q statistic, 26.42; P = .19). Despite this, it has been recommended that since statistical tests for heterogenity are relatively weak, obvious clinical sources of heterogeneity should be investigated.¹⁷ As there was some a priori evidence for clinical heterogeneity, the trials were combined using a random effects model. Therefore, we sought evidence for a treatment interaction with age and length of treatment as follows. First, we examined the effect of HRT on fractures in trials undertaken among women with a mean age that placed them within 10 years of menopause (ie, <60 years) and in trials undertaken among older women (ie, with a mean age >59 years). Second, the effects of therapy among women allocated to less than 36 months of treatment were compared with those of women allocated to longer periods of therapy.

Seventy trials were identified. Fourteen studies that reported fracture data were found in published literature.^{5 ,8 - 9 ,18 - 28} After excluding duplicate publications, we wrote to the authors of 59 trials and received replies from 37. From these, we identified 5 studies that did not publish fracture data in their original published reports but supplied it for this review.^{29 - 33} Furthermore, 4 unpublished trials were identified. Three trials were identified from abstracts of their bone mineral density (BMD) results and the authors of 2 of these studies supplied their fracture data for this review.^{34 - 35} Subsequently, 1 of these has been accepted for publication.³⁶ The authors of the third study, unfortunately, could not release their fracture data because it might jeopardize publication of their study since fracture is one of the main study outcomes.³⁷ Talking with clinical investigators at 2 conferences (April 2000 in Bath, England, and June 2000 in Chicago, Ill) revealed the existence of another trial,³⁸ from which a subgroup analysis of the study pertaining to the effect of estrogens on breast density has been recently published,³⁹ although as yet, neither the BMD or fracture data are in the public domain (Michael Draper, MD [Eli Lilly], written communication, July 2000). Also in one of the published studies, the Postmenopausal Estrogen/Progestin Interventions (PEPI) trial, the data in the original trial publication were reported as numbers of fractures rather than numbers of women with new fractures.²⁵ The PEPI investigators supplied us with the unpublished data of the numbers of women with an incident nonvertebral fracture (Mark Espeland, PhD, written communication, January 2001).

Table 1 shows the characteristics of the women in the 22 included trials, the number of trial participants, length of follow-up, and whether there was use of both progestins and calcium. Figure 1 and Table 2 show the estimated RR of fracture by treatment group and the RR of failure to comply and failure for follow-up. Combining the data from all the trials suggested a statistically significant reduction by HRT in all fractures of between 20% to 30% (Figure 1; Table 3).

We undertook 3 prespecified subgroup analyses. First, we examined the effect of age on effectiveness. Second, we looked at the effect of treatment length. Third, we examined any differential effect of HRT by fracture type.

Combining the 14 trials with published and unpublished fracture data undertaken among women younger than 60 years showed an approximate 35% reduction in fractures (Table 3). However, pooling the 8 trials undertaken among women older than 60 years showed only a 12% nonsignificant fracture reduction (Table 3). Trials of 3 years in length or more did not have an enhanced effect on fractures compared with trials of less than 3 years in length (RR, 0.75 [95% CI, 0.53-1.05]; P = .09 for long trials and RR, 0.49 [95% CI, 0.25-0.97]; P = .04 for short trials).

In 14 trials^{5 ,8 - 9 ,20 - 21 ,23 ,25 - 26 ,28 - 30 ,33 ,35} (Mark Espeland, PhD [Bowman Gray School of Medicine], written communication, January 2001; David Herrington, MD [Wake Forest University School of Medicine], written communication, October 2000; Michael Draper, MD [Eli Lilly], written communication, July 2000) it was possible to isolate hip and wrist fractures and examine the effects of HRT on just these fractures: the estimate of RR tended to be enhanced (Table 3). Excluding 6 studies^{8 ,20 ,26 ,28 ,30 ,33} of the 14 conducted among women older than 60 years showed a similar effect (Table 3).

We assessed 20 studies for quality, the remainder being unpublished or not accessible in draft format. Trial quality was generally good. Eleven trials reported the methods of randomization and 15 carried out double blinding with 9 of these defining the methods used. Fifteen trials reported on the withdrawals from the trial and gave reasons for this. There was some evidence of publication bias when all the trials were ranked by study weight with unpublished studies tending to be smaller with effect sizes closer to the null (Figure 2). While the RR of fracture for unpublished data seemed to be quite different from published data, this difference was not statistically significant. When only double-blind, placebo-controlled trials were included in the analysis, the point estimate of effect only changed slightly (Table 3). In addition, there appeared to be no overall difference between the HRT and control groups in terms of follow-up or compliance with therapy (Table 2).

To assess the robustness of our findings, we undertook a sensitivity analysis. As Table 3 shows, confining the analysis to double-blind, placebo-controlled studies did not materially affect the estimates of effect. Including only published data did improve the estimate of effect by 9% (ie, from an RR of 0.73 to 0.64), which shows the importance of including unpublished data. Indeed, the unpublished studies did have a markedly reduced effect size compared with the published trials (an RR of 1.09 vs 0.64 for unpublished and published studies, respectively).

Of the 22 trials, only 2 studies, the HERS⁸ and the Eli Lilly trial, had the potential to markedly affect the results by their inclusion or exclusion. The HERS study by its relatively large size dominates the estimate of HRT on fractures among older women. Exclusion of this study does lead to an improvement in effect size attributable to HRT (when HERS is excluded, RR, 0.62 [95% CI, 0.38-1.01]; P = .06), although this estimate is still not quite statistically significant. Similarly, excluding the Eli Lilly study from the analysis of younger women changed the RR of fracture among young women taking HRT from 0.67 to 0.59 (95% CI, 0.44-0.80; P<.001).

Hormone replacement therapy is usually regarded as the standard therapy for the prevention of postmenopausal osteoporosis. To test this belief, we have identified 22 randomized trials of HRT with fracture data, including a substantial number of unpublished studies. The results suggest that HRT does prevent fractures with an RR reduction in nonvertebral fractures of about 35% when women commence therapy before age 60 years. For wrist and hip fractures in younger women, the reduction is closer to 50%, which is similar to that found in observational studies.^{2 - 4}

On the other hand, the review suggests that the effect of HRT on fractures for women starting therapy when older than age 60 years may be reduced, although this attenuated effect is primarily due to the HERS study. However, an attenuation of the effect of estrogens, by age, on fracture has been noted in observational data. There are 4 observational studies in which it is possible to distinguish between women who started HRT before age 60 years and those who started HRT later in life. These studies all show a large reduction in fractures among women starting HRT before age 60 years,^{2 - 4 ,40} with 3 studies showing statistical significance.^{2 - 4} However, none of these observational studies shows a statistically significant effect of HRT on fractures among women starting treatment when older than age 60 years. These data are also supported by a large, unpublished trial of HRT among elderly women. This factorial trial of 489 elderly women, which is only available in abstract form (data not released for this analysis due to pending publication), noted a nonstatistically significant odds ratio for nonvertebral fracture of 0.94 for HRT users vs nonusers.³⁷

It is unclear why we have observed a reduced effect of HRT in older women. If this is not a true effect, it could be due to chance or possibly due to study inclusion. The largest trial, the HERS study, in older women had a high prevalence of obesity and there is observational evidence that overweight and obese women respond less well to HRT than do normal-weight individuals.⁴¹ Indeed, data from this observational study in Sweden indicates that there is no effect of HRT on hip fracture rates among women with a body mass index of greater than 25 kg/m². Given that more than 55% of the HERS participants had a body mass index greater than 27 kg/m², this may be an explanation of the HERS result.⁸ Exclusion of the HERS data from the meta-analysis seems to improve the results among older women. However, this explanation seems unlikely as the most recent analysis of the HERS data do not show an interaction with BMI, but the data show a slightly enhanced effect of HRT among women who are younger than 70 years, although this was not statistically significant.⁴²

The modifying effect of age cannot be explained by the inclusion of high-risk subjects in trials among younger women, in which one might expect a higher effect size, as the younger women tended to be healthy with normal BMD (Table 1). The differential result is not due to poorer HRT compliance among older women as there seemed to be a lower risk of noncompliance among older women than among younger women (Table 2).

Our review does not prove that HRT is ineffective among older women. We can only report that there is an absence of evidence rather than an absence of effect. However, given that most women starting HRT are older than 60 years and do so for osteoporosis prevention, randomized data are urgently required to sustain this clinical practice.¹ This is important because fracture incidence is so much higher among older women, even if HRT did have an attenuated effect, as long as this was statistically significant, it might still be cost-effective to offer HRT for fracture prevention in older women. Nevertheless, for women who started HRT before the age of 60 years and are still receiving treatment, observational data would suggest it is still effective; it is the age of starting therapy that our data would suggest is important.

While it would have been helpful to assess whether HRT was more or less effective among other subgroups of women, this was not possible. For instance, data on bisphosphonates suggest that that class of drug is most effective among patients with low BMD.^{43 - 44} Unfortunately, however, only 3 trials were undertaken among women with low BMD; thus, such a comparison would lack any power to show a difference in effect size.^{19 - 20 ,33}

We did not note any increase in effect size with increasing length of the trials. Indeed, longer trials tended to have a lower effect size than shorter studies; however, this was partly because the HERS study dominates the longer studies and this will tend to reduce the estimate of effect for trials undertaken for more than 3 years.

The quality of the trial reporting was good and more recently published trials scored highly on the Jadad scale. It is unclear whether all the fractures included were by self-report or confirmed by radiographic examination. However, fracture self-report is accurate.⁴⁵ All studies, except 1, were not explicitly designed to examine fractures as a study outcome. Thus, most studies only collected the data as part of adverse event reports. However, less than assiduous data collection on fracture events would only bias the results of the review if there were differential data collection by treatment arm. Given that most of the studies were placebo-controlled and the data would have been collected blindly, this is unlikely. Although poor fracture identification is unlikely to lead to bias, it will reduce the precision of the point estimate of effect as the power of any analysis is driven, not only by the sample size, but also the event rate. Therefore, underidentification of fractures will lead to a reduction in statistical power rather than bias per se.

We also examined the possibility that the nonuse of placebos in some of the trials might lead to bias; however, we did not find any evidence of bias when the results of placebo trials were examined separately.

In summary, this review suggests that HRT reduces the incidence of nonvertebral fractures, with a possible attenuation of this benefit when it is begun after age 60 years.