# The effects of a FRAX^{®} revision for the USA

## Authors

- First Online:

- Received:
- Accepted:

DOI: 10.1007/s00198-009-1033-8

- Cite this article as:
- Kanis, J.A., Johansson, H., Oden, A. et al. Osteoporos Int (2010) 21: 35. doi:10.1007/s00198-009-1033-8

- 44 Citations
- 256 Views

## Abstract

### Summary

A revision (version 3.0) of the fracture risk assessment tool (FRAX^{®}) is developed based on an update of epidemiological information for the USA. With the revised tool, there were strong correlations (*r* > 0.99) between versions 2.0 and 3.0 for FRAX^{®} estimates of fracture probability, but the revised models gave lower probability estimates.

### Introduction

The aim of this study was to determine the effects of a revision of the epidemiological data used to compute fracture probabilities in the USA with FRAX^{®}.

### Methods

Models were constructed to compute fracture probabilities based on updated fracture incidence and mortality rates in the USA. The models comprised the ten-year probability of hip fracture and the ten-year probability of a major osteoporotic fracture, both including femoral neck bone mineral density (BMD). For each model, fracture and death hazards were computed as continuous functions. The effect of the revised rates on fracture probability was examined by piecewise linear regression using multiple combinations of clinical risk factors and BMD.

### Results

At all ages, there was a strong correlation (*r* > 0.99) between version 2.0 and revised FRAX^{®} estimates of fracture probability. For a major osteoporotic fracture, the revised model gave lower median probabilities by 13% to 24% in men, depending on age, and by 19% to 24% in women. For hip fracture probability, the revised model gave lower median fracture probabilities by 40% and 27% at the ages of 50 and 60 years in men and by 43% and 30%, respectively, in women. At the ages of 70 years and older the revised model gave similar hip fracture probabilities as version 2.0 in both men and women.

### Conclusion

The revised FRAX^{®} model for the USA (version 3.0) does not alter the ranking of fracture probabilities but provides lower probability estimates than version 2.0, particularly, in younger women and men.

### Keywords

Clinical risk factorsFracture probabilityFRAX^{®}OsteoporosisOsteoporotic fractureUS

## Introduction

FRAX^{®} is a computer-based algorithm (http://www.shef.ac.uk/FRAX) that provides models for the assessment of fracture risk in men and women using easily obtained clinical risk factors for fracture and death to estimate 10-year fracture probability [1–3]. The tool can be used alone or with femoral neck bone mineral density (BMD) to enhance fracture risk prediction. The output is the 10-year probability of hip, clinical spine, humerus or wrist fracture and the 10-year probability of hip fracture alone. Some of the risk factors affect the risk of death as well as the risk of fracture, e.g., increasing age, low BMD and smoking. Most other risk engines calculate the probability of a clinical event (e.g. a myocardial infarct) without taking into account the possibility of death from other causes.

Probability of fracture is calculated in men or women from age, body mass index (BMI) computed from height and weight and dichotomised risk variables that comprise a prior fragility fracture, parental history of hip fracture, current tobacco smoking, ever long-term use of oral glucocorticoids, rheumatoid arthritis, other causes of secondary osteoporosis and alcohol consumption of three or more units daily. Femoral neck BMD can be entered additionally, either as a densitometer-specific BMD or as a T-score. The transformation of BMD to T-score, and vice versa, is derived from the NHANES III database for female Caucasians aged 20–29 years [4]. When entered, calculations give the 10-year fracture probabilities as defined above with the inclusion of BMD.

FRAX^{®} has been constructed using information derived from the primary data of nine population-based cohorts from around the world, including centres from North America, Europe, Asia and Australia, and has been validated in 11 independent cohorts with a similar geographic distribution within excess of 1 million person-years [5]. The use of primary data for the model construct permits the determination of the predictive importance in a multivariable context of each of the risk factors, as well as interaction between risk factors, thereby optimising the accuracy in computing fracture probability. The use of primary data also eliminates publication bias.

In addition to the clinical risk factors, fracture incidence and mortality rates vary markedly in different regions of the world [6]. Thus, FRAX^{®} models are ideally calibrated to those countries where the epidemiology of fracture and death is known. At present, FRAX^{®} models are available for Austria, China, France, Germany, Italy, Japan, Spain, Sweden, Switzerland, Turkey, the UK and USA. Other models are being developed, but a limitation is the adequacy of epidemiological data on which the FRAX^{®} models are based.

The most complete models in terms of fracture risk comprise those for Japan, Sweden, Switzerland, the UK and USA, where incidence rates were available for the major osteoporotic fracture sites. Other models are derived from hip fracture rates alone, which in turn were used to estimate the risk of fracture at other sites. In the case of the USA, the model was calibrated to the non-Hispanic white population using national mortality data and fracture incidence rates from the predominantly white population of Olmsted County, MN, in 1989–1991 [7]. For non-whites, estimates were based on race-specific hip fracture incidence rates together with race-specific mortality [1].

Since the construction of the US model, hip fracture incidence appears to be declining in the US white population, and mortality is improving [8–11]. For this reason, the National Osteoporosis Foundation (NOF) concluded that updating the FRAX^{®} model was appropriate and in 2009, a working group made specific recommendations [12], which have now been adopted. This paper reports the effects of these recommendations

## Methods

### Model parameters

The epidemiologic data used to develop version 2.0 and the revised (version 3.0) FRAX^{®} models for the USA have been described elsewhere [1, 12]. In brief, version 2.0 of the FRAX^{®} model for the USA used hip fracture incidence for the predominantly white population of Olmsted County, MN, in 1989–1991 [7], which were similar to later national estimates reported for US whites for the year 2001 from the Healthcare Cost and Utilization Project, Nationwide Inpatient Sample [13]. The revised model updated these hip fracture rates using national hospital discharge data for white non-Hispanic women and men in 2006 from the same source. Incidence in 1-year age intervals was calculated from the US Census projections for 2006.

With regard to other osteoporotic fractures, the version 2.0 FRAX^{®} model for the USA used in the 1989–1991 Olmsted County data for fracture-specific incidence rates [7]. These were used for the revised model with the exception of vertebral fracture. In the case of vertebral fracture, the version 2.0 estimates comprise not only symptomatic (i.e. clinical) vertebral fractures but also included those found incidentally during routine medical care [12]. In the absence of robust empirical data for the incidence of clinically significant vertebral fractures, for the revised model, it has been assumed that the ratio of clinical vertebral fractures to hip fractures in the USA was the same as that from Malmo, Sweden [14], a methodology used for the construction of FRAX^{®} [2]. The removal of incidental or non-clinical vertebral fractures in the revision will reduce the estimated 10-year probability of a major fracture.

With regard to mortality, FRAX^{®} version 2.0 used age-, sex- and race-specific death rates for the US population in 2001 [15]. In the revision, final mortality rates for 2004 were used [11], which for the non-Hispanic white population were 5–6% lower than in 2001.

### Comparison of models

For the purpose of comparing version 2.0 and revised probabilities of a major osteoporotic fracture (hip, clinical spine, forearm and humeral fractures) and of hip fracture alone, probabilities were computed in men and women at ages 50, 60, 70 and 80 years for all possible combinations of clinical risk factors at BMD T-scores between 0 and −3.5 SD in 0.5 SD steps with a BMI set to 26 kg/m^{2}. Thus, we considered all combinations of six risk factors and eight values of BMD giving a total number of combinations of 512. Note that this was not a population simulation, but an array of all possible combinations. The correlation between version 2.0 and revised (version 3.0) fracture probabilities was examined by piecewise linear regression with knots at probabilities of 40% and 70% for the version 2.0 for a major osteoporotic fracture and at 40% and 60% for hip fracture. The reason for using knots at different probabilities for the two outcomes was because of the difference in the distribution of probabilities. Tabular data compared probabilities with the two versions at the 10th, 50th (median) and 90th percentile of the distribution of version 2.0. Differences in version 3.0 at these percentiles were expressed as 95% tolerance intervals (TI).

## Results

^{®}is shown for women aged 60 or 80 years in Fig. 1. At both ages, there was a close correlation between the two estimates (

*r*> 0.99). As expected from the lower incidence, the revised version gave somewhat lower probabilities than version 2.0. The median value was lower by 19% and 24% at ages 60 and 80 years, respectively.

*r*> 0.99) at both ages. The revised version gave lower estimates at younger ages than was the case for a major osteoporotic fracture. At age 60 years, the median estimate was 30% lower with the revised version, compared with a decrement of 19% for a major fracture (Fig. 2). This effect was not apparent at older ages, and hip fracture probabilities with the two versions lay close to the line of identity.

^{®}by age is given for women in Table 1. For both hip and major fracture outcomes, there was a very close correlation between the two FRAX

^{®}models, and the correlation coefficient exceeded 0.99 at all ages. For a major fracture, probability with the revised version was consistently lower than version 2.0 by about 20% at all ages at the 50th percentile. For example, at the age of 50 years, the probability of a major fracture at the 50th percentile was 16% with version 2.0 and 13% with the revision—a 19% relative decrease. The absolute decrement in probability was 3% (16–13), but ranged from 1–9% at the other ages. In the case of hip fracture probabilities, the median disparity between the two versions of FRAX

^{®}was greater at ages 50 and 60 years (by 43% and 30%, respectively), but there was no difference between versions 2.0 and 3.0 at older ages where hip fractures are more common.

Probability (%) of a major osteoporotic fracture or a hip fracture (with 95% tolerance intervals) in women at the percentiles of the probability distribution (version 2.0) by age

Age | Percentile |
| |||||
---|---|---|---|---|---|---|---|

10 | 50 | 90 | |||||

Version 2.0 | Revision (v3.0) | Version 2.0 | Revision (v3.0) | Version 2.0 | Revision (v3.0) | ||

Major fracture | |||||||

50 | 7 | 6 (4–8) | 16 | 13 (11–14) | 37 | 29 (27–31) | 0.996 |

60 | 12 | 9 (9–10) | 27 | 22 (21–22) | 53 | 44 (44–45) | 0.998 |

70 | 14 | 10 (8–12) | 29 | 22 (20–24) | 61 | 52 (51–54) | 0.995 |

80 | 18 | 12 (9–15) | 38 | 29 (26–32) | 72 | 67 (63–70) | 0.994 |

Hip fracture | |||||||

50 | 0.2 | 0.0 (0.0–0.3) | 2.1 | 1.2 (0.9–1.5) | 18 | 11 (11–12) | 0.998 |

60 | 0.5 | 0.2 (0.0–0.5) | 3.4 | 2.4 (2.2–2.7) | 21 | 16 (16–16) | 0.999 |

70 | 1.6 | 1.5 (1.4–1.6) | 8.4 | 8.0 (7.9–8.1) | 39 | 38 (37–38) | 1.000 |

80 | 4.4 | 4.2 (3.8–4.5) | 21 | 21 (20–21) | 64 | 63 (63–64) | 1.000 |

^{®}models. The disparity between version 2.0 and the revised model was also similar (Fig. 3). For hip fracture probability, the revised model gave lower median fracture probabilities by 40% and 27% at ages 50 and 60 years, respectively. At ages 70 years and older, the revised model gave similar hip fracture probabilities as the version 2.0, as was the case for women.

## Discussion

In this study, we report FRAX^{®} probabilities for the US incorporating revised estimates of fracture and death hazards. The revisions include updated estimates of hip fracture incidence rates based on a national sample rather than a regional sample and more recent mortality data for the USA. These more recent hip fracture rates are lower than previously estimated, which would be expected to decrease the probability of hip fracture. On the other hand, the decreasing mortality would be expected to increase fracture probability in the elderly. The result of these competing hazards is that estimated hip fracture probability was substantially lower in younger men and women using the revised FRAX^{®} tool, but virtually unchanged at ages 70 years and above. In the case of major fractures, the revised tool gave lower probabilities than previously in part due to the hip fracture and mortality assumptions previously discussed, but also, due to use of a more conservative estimate of the incidence of clinically significant vertebral fractures. It is unlikely that the lower fracture rates are related to increased exposure to treatment, since the population impact will be very small [16]. An additional feature of the revision is the use of hip fracture and mortality rates by 1-year age intervals rather than the 5-year intervals used in the previous version. The changes have been made for the white Caucasian population. The assumptions concerning other ethnic groups in the USA remain unchanged, but the revision will affect fracture probabilities in these populations because the fracture rates are assumed to be a constant ratio of those in the Caucasian population [1].

Ten-year probability of a major fracture and hip fracture (%) in men and women from the USA and different European countries with no clinical risk factors (BMI set to 25 kg/m^{2}), a prior fragility fracture, and a T-score at the femoral neck of −2.5 SD with no other CRFs

Men aged 65 | Women aged 65 | |||||
---|---|---|---|---|---|---|

No CRFs | Prior fracture | T = −2.5 | No CRFs | Prior fracture | T = −2.5 | |

Major fracture | ||||||

USA v 2.0 | 7.5 | 14 | 14 | 12 | 23 | 17 |

USA v 3.0 | 5.6 | 11 | 11 | 9.3 | 18 | 13 |

Sweden | 6.1 | 12 | 13 | 10 | 19 | 15 |

Switzerland | 5.6 | 11 | 11 | 9.5 | 18 | 14 |

Austria | 4.9 | 9.7 | 10 | 8.5 | 17 | 13 |

UK | 4.9 | 9.5 | 9.7 | 8.4 | 16 | 12 |

Italy | 3.7 | 7.5 | 8.1 | 6.5 | 13 | 9.8 |

Germany | 3.3 | 6.6 | 7.1 | 5.8 | 11 | 8.8 |

France | 2.6 | 5.3 | 5.7 | 4.5 | 9.0 | 6.9 |

Spain | 2.1 | 4.2 | 4.6 | 3.6 | 7.3 | 5.6 |

Hip fracture | ||||||

USA v 2.0 | 0.7 | 1.8 | 3.4 | 1.2 | 2.9 | 2.7 |

USA v 3.0 | 0.8 | 1.9 | 3.5 | 1.2 | 3.0 | 2.9 |

Sweden | 1.3 | 3.2 | 6.0 | 2.1 | 5.1 | 4.8 |

Switzerland | 0.9 | 2.2 | 4.1 | 1.5 | 3.6 | 3.4 |

Austria | 1.0 | 2.5 | 4.7 | 1.7 | 4.3 | 4.1 |

UK | 0.8 | 1.9 | 3.5 | 1.3 | 3.1 | 3.0 |

Italy | 0.8 | 1.9 | 3.6 | 1.3 | 3.2 | 3.1 |

Germany | 0.7 | 1.7 | 3.1 | 1.2 | 2.9 | 2.8 |

France | 0.5 | 1.3 | 2.5 | 0.9 | 2.2 | 2.1 |

Spain | 0.4 | 1.1 | 2.1 | 0.7 | 1.8 | 1.7 |

Importantly, the revisions made have little impact on the efficiency with which the FRAX^{®} tool categorizes risk. This is because the revisions do not change the rank order of fracture probability in any population. In the clinical scenarios presented in this paper, the correlation coefficients between versions 2.0 and 3.0 probabilities exceeded 0.99, so that the one can be accurately predicted from the other. In other words, an individual at the 90th percentile of risk would still be at the 90th percentile of risk using the revised FRAX^{®} tool. For example, a white female aged 60 years at the 90th percentile of risk had a 10-year fracture probability of 53% (see Table 1). The revised estimate was 44%, but remained at the 90th percentile of risk. Thus, the consequences of improving accuracy reside in the absolute number generated and not in the rank order of risk. This is of little consequence to the management of patients or the interpretation of clinical studies. There is a useful analogy with the different DXA devices available, where a substantial difference in femoral neck BMD is seen between Hologic and Lunar machines, but the T-score derived from these is more or less identical [18].

An exception arises when fracture probabilities are used in health economic analysis to inform practice guidelines. Several studies have examined intervention thresholds in terms of fracture probability [19–23]. The majority has expressed results as the 10-year probability of hip fracture at which treatment is cost-effective, i.e., using hip fracture as an overall metric to measure the impact of all fracture outcomes. The most recent assessment using this approach was from the National Osteoporosis Foundation (NOF) who updated pre-existing clinical practice guidelines with a further health economic analysis [15, 23]. The analysis, based on a 5-year treatment with bisphosphonates, determined the hip fracture probability at which cost-effectiveness was at or less than $60,000 per quality-adjusted life year gained (the threshold used for cost-effectiveness). In white men and women, treatment became cost-effective at a hip fracture probability of 3.4% and 3.8%, respectively. The corresponding probabilities in blacks were 3.3% and 3.4%. On this basis, the NOF chose a 10-year hip fracture probability of 3% as an intervention threshold. The ‘equivalent’ probability for a major osteoporotic fracture was set at 20%. The conclusions are dependent on the absolute values for probabilities, since for given efficacy, the number of fractures saved is less the lower the fracture probability. The extent by which the current guidance will need modifying is to be determined, but any changes are likely to be relatively modest.

## Acknowledgements

We are grateful to the International Osteoporosis Foundation for their support of the work.

### Conflicts of interest

None.