Several studies have modeled fractures over time using clinical risk factors with or without bone mineral density (BMD).1–6 The c-statistic (c index in a Cox model) is a measure of model discrimination. A value of <0.6 is thought to have little or no clinical value, 0.6 to 0.7 limited value, 0.7 to 0.8 modest value, and >0.8 genuine clinical utility (0.5 is equivalent to a random coin toss).7 Reported c-statistics in hip fracture models are generally at least 0.7 to 0.8, but fall to <0.7 in models combining hip fracture with other bones.
We hypothesized that the superior model discrimination associated with a hip fracture model compared to a model that combines hip with other fractures, implies that risk factors differ for different bones, or their associations with different bones differ (or some combination of the two). In a review paper, Kelsey and Samelson8 identified substantial differences in the associations between age, falling, type of fall, and frailty with hip, spine, wrist, and upper arm fractures. They stated, “We strongly recommend that studies identify risk factors on a site-specific basis.”
The Global Longitudinal Study of Osteoporosis in Women (GLOW) has collected sufficient data on fractures of specific bones over a 3-year period to model 10 bone sites individually, using risk factors at study baseline. Some models are, to the best of our knowledge, the first to be presented. Results should help to clarify why model discrimination does not necessarily improve when various fractures are combined, as in the example of major fracture, and enable an assessment of whether various bone sites seem similar.
In addition to any fracture, “first-time” bone fracture results are presented. If models for women without a prior fracture differ from those that include prior fracture as a risk factor, this should enhance understanding of how a specific bone becomes damaged in an earlier stage of disease, and may have implications for fracture prevention. In a site where future fracture risk increases greatly with a prior fracture, first-fracture prevention may be particularly important.
Materials and Methods
The GLOW study methods have been described elsewhere.9 In brief, women aged ≥55 years who were under the regular care of a physician in the previous year were asked to participate, with 60,393 (43% of those contacted by their physician) returning the baseline survey. The women are from 10 countries (47% from the United States) and 17 study sites. The participants were surveyed annually for 3 years after baseline. All data were self-reported and completed surveys were scanned into a central database. Data on BMD were not collected: 33% of the study population had not undergone a BMD examination, while others did not know their BMD value.
At baseline, each woman was asked if she had broken any of the following 10 bones since the age of 45 years: hip, pelvis, spine, upper arm/shoulder (UAS), wrist, rib, upper leg (UL), clavicle, lower leg (LL), and ankle. Month and year of fracture were recorded for follow-up fractures, to ensure they were incident. In the small number of cases where the date of fracture was missing, if no prior fracture of that same bone had been indicated in a prior survey, it was assumed the fracture was new and had occurred halfway through the interval. Otherwise, it was assumed the fracture was old and was therefore not included. Some women appended handwritten notes beside survey questions. In cases where we were uncertain if a woman incurred a legitimate postbaseline fracture (eg, a sprain or joint replacement was indicated), we took what we consider a conservative approach and did not count it as an incident fracture.
Major fracture (3-year fracture of any of hip, spine, UAS, or wrist) is included to illustrate that modeling a composite outcome may mask risk profile differences in individual bones. “First-time” fracture indicates an incident fracture for a woman who had no prior baseline fracture since the age of 45 years of any of the 10 bones. “Any” fracture indicates either a first-time incident fracture, or an incident fracture for a woman who had already broken at least one of the 10 bones since the age of 45 years by study baseline.
A detailed description of the statistical methods is given in Supplemental Information A. In all analyses, the outcome was time to first fracture after baseline; if no fracture occurred, the outcome was censored at the last observed time (last survey). Specific fracture types were treated as independent outcomes.
Because not all women returned all three follow-up surveys, the counting process (CP) approach to modeling10 was used in all main analyses, to include as much follow-up survey information as possible. For example, if a woman returned follow-up surveys 1 and 3, but not 2, she would contribute data to the fracture risk set from baseline to the end of year 1, would disappear from the risk set in year 2, and would return in year 3. Three-year cumulative fracture incidence was estimated using the Kaplan-Meier method. Unadjusted hazard ratios (HRs) and p values were obtained by univariable Cox regression.
Final models were obtained using multiple Cox regression. All baseline variables reported in Table 1 were considered. Selected backward elimination was used to remove factors in clusters (round one removes factors with p > 0.5, round two removes those with p > 0.3, etc.), until a relatively parsimonious model is obtained with factors with p ≤ 0.05.
Table 1. Unadjusted Numbers of Fractures and HRs for Fractures 0 to 3 Years After Study Baseline by Certain Baseline Characteristics and Selected Univariate Information
Cumulative 3-year incidence, %
Cumulative incidence of first-time fracture
Risk factors at study baseline
Median age, years
67 (n = 54,229)
Median weight, kg
68 (n = 52,994)
Median BMI, kg/m2
26 (n = 51,867)
Age ≥80 years, %
12 (n = 6592)
Women with any follow-up after baseline, n = 54,229. The following were missing information for ≥1000 women (3%–10%): paternal hip fracture (n = 5332), rheumatoid arthritis (n = 4332), maternal hip fracture (n = 3275), body mass index (n = 2363), weight (n = 1236), emphysema (n = 1269), told have high cholesterol (n = 1171), and osteoarthritis (n = 1047).
BMI = body mass index; Clav = clavicle; F10 = fractured any of 10 bones listed; HR = hazard ratio; LL = lower leg; Major = fracture of hip, spine, UAS, or wrist; N/A = not applicable; N/E = not estimable (no events among women with the baseline risk factor); UAS = upper arm/shoulder; UL = upper leg.
Model discrimination (the c index) was assessed using the Harrell macro for Cox regression,11 while goodness-of fit (GOF) was assessed by the May and Hosmer12 method. Any p ≤ 0.05 indicates poor model fit in at least one risk group.
We did not attempt to internally validate our results using a split sample technique, because use of the maximum sample size leads to the best unbiased estimates, and is least likely to cause type II error. Also, no study can effectively validate itself: “a true evaluation of generalisability (also called transportability) requires evaluation on data from elsewhere.”13 All analysis was conducted using the SAS software package, version 9.2 (SAS Institute, Cary, NC, USA).
Supplemental Information B shows the pattern of follow-up surveys received between November 2007 and April 2011. Of the 60,393 women who returned a baseline survey, 54,229 (90%) returned at least one follow-up survey. Follow-up rates for years 1, 2, and 3 were 85%, 81%, and 75%, respectively. Most follow-ups (93%) had “pattern 1,” which was used for model assessment. Another 1294 women had follow-ups at years 1 and 3; hence, they may contribute two observations to the CP data.
Table 1 reports univariable (unadjusted) results for all 10 fractures in the 3-year period, and for the combined outcome of “major” or “osteoporotic” fracture (hip, spine, UAS, and wrist).4 Factors missing in >1000 women are reported in a footnote; only parental hip fracture (10% paternal, 6% maternal), rheumatoid arthritis (8%), and body mass index (4%) were missing in >3% of women.
Wrist fractures were common (1036 fractures in 3 years, an estimated 3-year cumulative incidence of 2.2%), whereas clavicle and pelvis fractures were rare (n < 200). Estimated 3-year hip fracture incidence was 0.75%, or about 2.5 fractures per 1000 women per year. By age group, hip fracture estimates per year are 0.09%, 0.34%, and 0.84% for those 60–69, 70–79, and 80–89 years old, respectively.
Different fractures exhibit different patterns of risk factors. Women who fractured a hip, pelvis, or UL after baseline were considerably older than the study-wide median of 67 years (median ages 77, 75, and 74 years, respectively), whereas those who fractured an ankle or LL were about average (median ages 66 and 67 years, respectively). Of the 353 women who fractured a hip, 41% were aged ≥80 years. Women who fractured a hip, pelvis, or clavicle were lighter than the median weight of 68 kg (medians 63, 62, and 64 kg, respectively); whereas those fracturing an ankle or LL were somewhat heavier (70 kg). Baseline falls seemed to increase risk for any fracture, particularly pelvis or clavicle fractures.
Prior fracture of a bone was associated with a new fracture of the same bone for all 10 bones. The largest effects are seen for prior hip, UL, spine, or rib fractures (HRs from 7.2 for rib to 11.0 for spine). Prior fracture of a different bone was also likely to be associated with a subsequent fracture, with generally weaker associations (HRs from 2.0 to 4.4).
Maternal hip fracture (MHF) or paternal hip fracture, comorbidities, poor health, and diminished activity were all associated with future fractures, whereas moderate alcohol consumption was associated with lower risk for certain fractures.
Unadjusted estimates for the composite outcome of major fracture represent mathematically weighted averages of estimates for its four components. For example, the median age of women with a major fracture (71 years) was between that of women with a hip (77 years), spine (72 years), UAS (72 years), or wrist fracture (69 years).
Many Table 1 results are statistically significant, but because risk factors are interrelated, certain results that appear striking here are absent from adjusted models. Many factors are related to age in particular.
Multivariable Cox models
Table 2 and Supplemental Tables 1 to 6 contain final model-fitting results. Only that portion of a risk factor estimate whose association with fracture was independent of other factors is shown. For example, if both age and falls are in a model, this means falling adds to fracture hazard regardless of a woman's age, and vice versa. In this sense, results are adjusted for all factors in a given model.
Table 2. Summary of Main Cox Model Baseline Predictors of 3-Year Fracture
Clav = clavicle; HR = hazard ratio; LL = lower leg; Major = fracture of hip, spine, UAS, or wrist; N/E = not estimable (<80 outcomes); UAS = upper arm/shoulder; UL = upper leg; x = no significant effect of a baseline risk factor.
Predictor with largest χ2 value in each model.
Estimated main effect in presence of age by prior fracture interaction.
The U-shaped association has a high HR at low weight, declines over an interval of increased weight, then rises again with a further weight increase.
Table 2 summarizes main results for the 22 models, with “x”s indicating no significant effect of a baseline risk factor. Supplemental Tables 1 to 6 show full model results. Figure 1 presents key findings for the four most common risk factors: age, weight, prior fracture at the same or a different bone site, and two or more baseline falls. If a risk factor was not statistically significant, no result is shown.
Fractures are ordered by descending values of c indices for any fracture models (Table 2). Hip, pelvis, UL, spine, and clavicle fractures all have c indices ≥0.75, whereas UAS, LL, wrist, and ankle fractures have c indices ≤0.66. Model GOF is adequate for all models (p ≥ 0.08). Major fracture (hip, spine, upper arm, or wrist) has a c index of 0.67.
In each model, one may tell how strong a factor is by its χ2 index (χ2/number of fractures); the index may also help one compare how strong a particular factor is in different models. Judging by χ2 indices, age is the most important factor in hip, pelvis, UL, spine, and clavicle fractures (HRs from 1.5 to 2.3 for an additional 10 years of age). However, age is relatively unimportant in rib, LL, wrist, and ankle models (HRs from 0.81 to 1.2), and is not statistically significant in the LL model.
HRs for baseline fracture of the same bone since age 45 years are highest for spine, rib, UL, and hip fractures (HR = 6.6, 4.8, 4.0, and 3.5; <3 for remaining bones). Prior fracture of a different bone appears in all models except hip, with HRs <3. A prior hip fracture by age interaction was observed (p < 0.001): women with a prior hip fracture have an estimated 1.2 times higher hazard of a new fracture if they are 85 compared to 55 years old, but an estimated 16 times higher hazard if 85 versus 55 years old with no prior hip fracture. In an unadjusted analysis, 3-year hip fracture rates among women with a prior hip fracture were 3.6% if aged 55 to 69 years, 4.9% if 70 to 84 years, and 5.1% if ≥85 years (data not shown). Corresponding rates among women with no prior hip fracture were 0.2%, 0.9%, and 2.7%, respectively.
Higher weight is associated with reduced fracture risk for hip, pelvis, spine, rib, and wrist fractures (HRs 0.85 to 0.94 per additional 10 kg), with increased ankle fracture risk (HR = 1.1), and is not statistically significant in the remaining models. Baseline falls appear in all but spine fracture models. For pelvis, clavicle, and ankle fractures, even a single fall in the year before baseline is associated with increased 3-year fracture risk (HRs ≥ 1.5). Maternal/paternal hip fracture and current smoking or alcohol use at baseline do not appear in any model. Of the remaining factors, unexplained weight loss of ≥10 lbs, and asthma or emphysema, appear in three models, whereas reduced general health or physical activity appear in 6 of 10 individual fracture models. General health is particularly important for spine fracture (HR = 4.6 for poor compared to excellent health). Other factors appear in only one or two models.
Major osteoporotic fracture risk factor estimates fall between estimates of its constituent fractures. For example, the HR for an added 10 years of age, 1.4, is between those of hip, spine, UAS, and wrist (HR = 2.3, 1.5, 1.4, and 1.1, respectively).
Two models have unusually large χ2 indices due to a single risk factor: age (HR = 2.3 per 10 years) dominates hip (χ2 index 0.61, c index 0.79), whereas prior spine fracture dominates spine fracture (χ2 index 0.49, c index 0.75). Wrist (χ2 index 0.08, c index 0.64), ankle (χ2 index 0.09, c index 0.64), and major (χ2 index 0.07, c index 0.67) fracture models lack a strong risk factor.
Model c indices are ≥0.70 for first hip, pelvis, UL, spine, and clavicle fracture models. Indices are ≤0.63 for remaining first fractures (Table 2). The c index for major osteoporotic fractures is 0.63. All models (for which it could be computed) have adequate GOF (p ≥ 0.06).
HRs for an additional 10 years of age are higher than in “any fracture” models for a first hip, pelvis, UL, spine, or UAS fracture (HR = 2.8, 2.1, 2.0, 1.7, and 1.5, respectively). Advanced age is associated with reduced first ankle fracture hazard (HR = 0.83), and is again not statistically significant for first LL fracture. Added weight is associated with reduced hazard of a first hip or wrist fracture, but increased hazard of a first LL or ankle fracture. First pelvis fracture has a nonlinear association with weight: compared to a 70-kg woman, the estimated hazard is 3.1 times as high for a 40-kg woman, then gradually declines with additional weight to 0.9 for a 90-kg woman, and returns to 1.0 for a 100-kg woman (p = 0.04 versus a model with linear weight; linear weight itself had p = 0.56).
Baseline falls are associated with 8 of 10 first fractures. This association is highest for a first pelvic fracture, for which two or more baseline falls are associated with nearly four times the fracture hazard. Parental hip fracture is associated with first-fracture in four models, poor health or reduced activity appear in five models (HRs >3 for a first hip or spine fracture). General health has a stronger association with first than with any spine fracture (HR = 6.2 for poor versus excellent health). Remaining factors appear in two or fewer models. First major fracture estimates are again weighted averages of its four constituents. The age HR, 1.5, is between estimates for hip, spine, UAS, and wrist fracture (HR = 2.8, 1.7, 1.5, and 1.2, respectively).
First hip and pelvic fractures are both dominated by age more than in any fracture models (χ2 indices 0.84 and 0.49, both c indices 0.78). The c index is considerably lower in the first compared to any rib fracture model (0.63 versus 0.71). First wrist fracture has the lowest indices of any model (χ2 index 0.04, c index 0.59).
Bone medication (use of any of 11 bone medications or estrogen at baseline) is associated with neutral or elevated fracture risk for most bones, even after adjustment, likely reflecting use of medication in women with a high fracture risk (confounding by indication). We therefore did not consider it as a risk factor in our models. We can, however, conclude that its use does not materially alter estimates in the 20 fracture models (data not shown). The largest changes in model estimates, after adding bone medication as a covariate, were reductions in linear weight estimates in the any pelvis (from 0.90 to 0.95 per 10 kg) and any rib fracture models (from 0.94 to 0.96 per 10 kg).
Predicting cumulative 3-year fracture risk
Supplemental Information A shows how to convert model estimates from Supplemental Tables 1 to 6 to estimated 3-year cumulative fracture incidence for a given model.
To the best of our knowledge, several of these individual fracture prediction models are the first to be presented. Our results show that fractures at individual sites may have substantially different risk factors, estimates, or both. In the particular example of major osteoporotic fracture, model discrimination is reduced when combining hip with other fractures, largely due to risk profile differences. Although the composite outcome may be clinically useful, certain specific differences among the four bones are obscured when they are pooled, because composite risks are weighted averages of estimates for specific bones. For instance, 10 additional years of age more than doubles hip fracture risk, but increases wrist fracture risk by only 10%; baseline falling is associated with increased risk in three bones, but not spine; and weight is not associated with UAS fracture. Composite outcomes have the merit of simplicity and large sample size; this should be weighed against the cost of a possible loss of specificity between predictor and outcome, which in extreme cases results in a loss, not a gain, in statistical power. If a composite outcome is clinically useful, but contains fractures with substantial differences in risk profiles, a better approach than forming a composite outcome might be to fit individual fracture models, combine predicted fracture rates from the individual models, and deduct a constant for those who fracture more than one bone. In this way, each model is tailored to the specific bone, and each bone's individual risk profile is preserved.
Whereas our main finding is that large risk profile differences exist, some bone sites exhibit similar risks. In particular, it appears that hip, pelvis, UL, and spine fractures share certain risk factors, and their models (any and first-fracture) tend to have the highest c indices. Age is a strong risk factor for all four fractures, and, perhaps because these are weight-bearing bones near the body core, lower weight is associated with increased fracture risk, except for UL. Unlike the other three bones, spine fracture is not associated with falling. Lower leg and ankle fracture models are also similar: age is not associated with increased risk, whereas higher weight is, and c indices are much lower. Wrist fractures are associated with reduced weight, but age is relatively unimportant, and c indices are low. UAS fractures have a modest association with increased age, and no association with weight.
First-fractures have been relatively neglected in research, perhaps undeservedly. Prior fracture at the same bone site is associated with a greater than tripling of risk of a new hip, spine, UL, or rib fracture. If we can predict who is at high first-fracture risk and intervene, this could prevent a cycle of further fractures, even if first-fracture is simply a marker for other risk factors, rather than a direct cause. Alternatively, if a prior fracture represents adverse bone geometry or other factors intrinsic in a woman, a first fracture and fracture recurrence may not be preventable.
Reported first-fracture models may raise general awareness of fracture risk, because women and physicians are often unaware that women with no prior fracture are at risk. Factors such as rapid weight loss, a tendency toward falling, poor health or reduced activity, and parental hip fracture, besides aging, may indicate that a woman is at increased first-fracture risk, and may warrant a BMD test or general consultation. Of the first-fracture factors identified, only falling, poor health, low activity, and weight are potentially modifiable.
Several first-fracture risks emerged that were absent in the estimation of any fracture, despite lower statistical power in first-fracture models. Prior fracture can be a strong factor in an any fracture model, but submerged in the prior fracture may be factors that led to that prior fracture and so do not appear in the any fracture model (they are subsumed in the prior fracture). These new factors may help us to better understand why fractures occur. For example, maternal hip fracture (MHF) increases risk of a first (but not any) hip, pelvis, and UAS fracture. Other factors, such as age and general health, are not new, but appear stronger in certain first-fracture models.
Because roughly one-half of all 3-year fractures are first fractures, the two sets of models are not independent. This means that first-fracture models may provide more distinct information from any fracture models than is first apparent. For example, the HR for an additional 10 years of age is 2.8 for a first hip fracture, but 2.3 for any hip fracture. If roughly one-half of any hip fractures are first fractures, this implies an HR of 1.8 for a subsequent hip fracture. Age is therefore a much stronger predictor of a first than of a subsequent hip fracture (HR = 2.8 versus 1.8).
We had a substantial number of specific fractures within 3 years of baseline, with <100 only in first pelvic and clavicle fractures. Also, the study population was fairly diverse: women from seven sites in the United States, eight in Europe, and one each in Canada and Australia. Furthermore, since GLOW is a prospective cohort study, this ensures that a baseline risk factor precedes a 3-year fracture. Problems of recall in annual follow-up are therefore limited, although not eliminated. Last, annual follow-up rates were high: 85%, 81%, and 75%, respectively, over the 3 years, with 90% of baseline women providing at least one follow-up.
If nonrespondents to our survey were sicker and more likely to fracture than respondents, our Kaplan-Meier fracture rate estimates would be low. If absolute rates in our study are low, this should not affect HR estimates, as these represent ratios of rates.
All data were self-reported, causing some inaccuracy in risk factor measurement. Alcohol consumption was likely to have been underreported, whereas parental hip fracture and osteoporosis were often missing. As we lacked outside confirmation, fractures at certain sites may have been confused: hip, pelvis, and UL; clavicle and UAS; and LL and ankle. Shoulder fractures at baseline were not explicitly identified (only upper arm), but were added to later surveys; this may have underestimated the effect of prior baseline fracture in the UAS model. A vertebral fracture model that included incident morphometric spine fractures might have performed better than our model, which was limited to clinical vertebral fractures. A study that compared radiographs to self-report of rib fracture estimated that 23% of self-reported rib fractures were not clinical fractures.14 However, prior self-reported rib fracture, despite possible overreporting, was found predictive of future non-rib fracture in both the GLOW study15 and the National Osteoporosis Risk Assessment (NORA).16
Some unknown degree of misclassification of certain outcomes and risk factors therefore exists in GLOW, making it more difficult to derive accurate models, and biasing results toward the null hypothesis of no association between risk factor and fracture. Some associations may therefore be stronger than reported in this study, and other true associations may not have been found. Despite this, 11 of 20 models showed reasonably good discrimination (c index ≥0.70); none failed the GOF test (although some models were marginal); certain factors were quite strong; and they seem, in general, reasonable. Despite possible misclassification, the models should still provide useful information, and a basis for comparison for future studies.
We did not collect BMD or other physical measures, except age, weight, and height. Physical measures are generally important sources of estimation in other disciplines, being objective measures with a wide range of values over which to discriminate outcomes into groups. Although some models had good discrimination, none were excellent, implying that significant risk factors may not have been identified.
The GLOW study is limited mainly to whites, and a smaller number of black and Hispanic women (ethnicity in sites outside the United States and Canada was not collected). We do not know if risk factors for African and Asian women, for instance, would have important differences from those identified here. Also, as most GLOW women are ambulatory, models for women living in nursing homes may differ from those presented here.
We had missing information on certain variables (Table 1, footnote), and follow-up was good but not perfect. Missing data can be a significant problem if risk factor/fracture associations differ between those with and without missing data. For example, if a large proportion of women with missing information on MHF had mothers who fractured a hip, and MHF was more likely to be associated with a fracture among those missing than among those not missing MHF, then our MHF estimate would be too small. This sort of scenario seems unlikely, however. Few variables were missing in >5% of women, thus potential missing data bias is limited.
Sample sizes for pelvis, UL, and clavicle fractures were small (n = 100–200), so we may have missed moderately important factors due to low statistical power. However, given a general rule of ≥10 outcomes per covariate, our models should not have overfitted the data (ie, strained the capacity of the data to provide estimates likely to generalize beyond our particular data set). First-fracture models, having fewer outcomes than any fracture models, had reduced ability to detect risk factors.
Direct comparison of our 3-year fracture rate estimates to WHO Fracture Risk Assessment Tool (FRAX) estimates is not possible, since FRAX treats death as a competing risk, whereas Kaplan-Meier estimates do not. Hence our estimates are slightly higher than they should be, though the overestimation in 3-year rates should be small. This difference in methods should not affect HRs.
Because bone medication was not randomly assigned, we were unable to assess its direct association with fracture. Finally, we were limited to 3 years of incident fracture data, whereas other studies have had a follow-up period of up to 10 years. Risk factors and the associated fracture risks in 3-year fracture estimation may differ from those in 10-year models, though to the extent the time frames overlap they will not.
In conclusion, fracture sites vary significantly in their risk factor profiles. It is hoped that these results will encourage thoughtful reexamination of whether a composite outcome, or a different approach, is optimal for assessing the combined risk of fracture at several bone sites. We further hope that the models described will lead to a better understanding of why certain bones fracture, and increase awareness of a woman's first-fracture risk.
GF has received research and salary support from the Alliance for Better Bone Health (Procter & Gamble Pharmaceuticals, sanofi-aventis). SB has received research grants from Amgen, Eli Lilly, Novartis, Pfizer, Procter & Gamble, sanofi-aventis, Roche, and GlaxoSmithKline; is on speakers' bureau for Amgen, Eli Lilly, Merck, Novartis, Procter & Gamble, sanofi-aventis, and Servier; has received honoraria from Amgen, Eli Lilly, Merck, Novartis, Procter & Gamble, sanofi-aventis, and Servier; and is a consultant/advisory board member for Amgen, Eli Lilly, Merck, Novartis, Procter & Gamble, sanofi-aventis, and Servier. JEC is a consultant for Servier, Shire, Nycomed, Novartis, Amgen, Procter & Gamble, Wyeth, Pfizer, The Alliance for Better Bone Health, Roche, and GlaxoSmithKline; has had speaking engagements (with reimbursement, travel, and accommodation) for Servier, Procter & Gamble, and Eli Lilly; and has received research grants from Servier R&D and Procter & Gamble. JP has received research grants from AMGEN, Kyphon, Novartis, and Roche; has received equipment from GE LUNAR; is on speakers' bureaus for AMGEN, sanofi-aventis, GlaxoSmithKline, Roche, Lilly Deutschland, Orion Pharma, Merck Sharp and Dohme, Merck, Nycomed, and Procter & Gamble; and is on advisory boards for Novartis, Roche, Procter & Gamble, and TEVA. AZL has received research and salary support from The Alliance for Better Bone Health (Procter & Gamble Pharmaceuticals and sanofi-aventis). DWH states that he has no conflicts of interest. FHH and SHG have received research and salary support from the Alliance for Better Bone Health (Procter & Gamble Pharmaceuticals, sanofi-aventis).
Data for this study would not exist without the help of GLOW site coordinators and Christine Vigeant, Todd Pearson, Joan Lovell, Diane McBride, Themia Pappas-Fillmore, Lucille Fink, Ben Erban, Meg Brueggemann, LiRong Song, and Leigh Emery, all from the Center for Outcomes Research, Worcester, MA, USA. We also thank Sophie Rushton-Smith for help with formatting and styling. Financial support for the GLOW study is provided by Warner Chilcott Company, LLC, and sanofi-aventis to the Center for Outcomes Research, University of Massachusetts Medical School.
Authors' roles: Study design: FHH and SGH. Study conduct: SB, JEC, JP, AZL, FHH, and SHG. Data collection: SB, JP, and AZL. Data analysis: GF. Data interpretation: GF and DWH. Drafting manuscript: GF. Revising manuscript content: All authors. Approving final version of manuscript: All authors. GF takes responsibility for the integrity of the data analysis.
This work is dedicated to Professor Philip Neil Sambrook–colleague and friend–who helped guide and inspire us during our work together on the GLOW Study.