SEARCH

SEARCH BY CITATION

Keywords:

  • bone fractures;
  • epidemiology;
  • fat mass;
  • obesity;
  • osteoporosis;
  • quantitative ultrasonography

Abstract

  1. Top of page
  2. Abstract
  3. Introduction
  4. Methods
  5. Results
  6. Discussion
  7. Conflict of interest statement
  8. Acknowledgements
  9. References

Abstract.  Moayyeri A, Luben RN, Wareham NJ, Khaw K-T (University of Cambridge, Cambridge; Institute of Metabolic Science; Cambridge, UK). Body fat mass is a predictor of risk of osteoporotic fractures in women but not in men: a prospective population study. J Intern Med 2012; 271: 472–480.

Objectives.  Obesity has generally been associated with higher bone density and lower fracture risk. However, weight-related indices of obesity may be related differently to health end-points, compared with fat-related indices (such as body fat distribution and fat mass), as they may capture different dimensions of obesity and the associated biological effects. The aim of this study was to examine the association between percentage body fat (%BF) and prospective risk of fracture.

Methods.  The European Prospective Investigation into Cancer (EPIC) in Norfolk was a population-based prospective study. A total of 14 789 participants (6470 men, aged 42–82 years at baseline) were included. The main outcome measures were quantitative ultrasound of the heel and incident hip and any osteoporotic fractures.

Results.  A total of 556 participants suffered a fracture (184 hip fractures) during 8.7 ± 0.8 years of follow-up. Risk of hip fracture decreased linearly with increasing %BF amongst women but not men. After adjustment for age, history of fracture, height, smoking, alcohol intake and heel broadband ultrasound attenuation (BUA), the hazard ratio (95% CI) for a 10% higher %BF on risk of hip fracture was 0.56 (0.39–0.79) in women and 0.92 (0.39–2.21) in men. The effect size in women was approximately equivalent to a difference of 5 years in age or 1 standard deviation (17 dB MHz−1) increased BUA. A nonlinear negative association was also observed between %BF and risk of ‘any type of fracture’ amongst women but not men.

Conclusions.  The %BF appears to predict hip fracture risk in women with an effect size comparable to that of bone density as measured by heel ultrasound. This effect was not observed in men. Understanding the differences in relationships between different indices of obesity as well as sex differences may help to elucidate the metabolic and other underlying mechanisms involved in bone health and fracture risk.


Introduction

  1. Top of page
  2. Abstract
  3. Introduction
  4. Methods
  5. Results
  6. Discussion
  7. Conflict of interest statement
  8. Acknowledgements
  9. References

Obesity and osteoporosis are two major epidemics in the modern world. It is estimated that globally there are more than 1 billion overweight adults of whom at least 400 million are obese [defined as a body mass index (BMI) >30 kg m−2] [1]. National surveys have shown that, for instance, more than 22% of the population of UK and 30% of US citizens are obese [2, 3]. Osteoporosis is another important public health problem, characterized by excessive skeletal fragility and susceptibility to low-trauma fractures amongst the elderly. Globally, 30–50% of women and 13–30% of men will suffer from a fracture related to osteoporosis in their lifetime [4]. Recent studies suggest that there may be relationships between obesity and osteoporosis at the molecular and clinical levels [5].

Fat mass is a component of total body weight and one of the indices of obesity. Body fat mass and bone mineral density (BMD) are known to be under strong genetic regulation, and an association between fat mass and fracture susceptibility is plausible from a genetic point of view [6]. Several lines of clinical evidence support a beneficial effect of fat mass on BMD, hence reducing the risk of osteoporosis [7–11]. By contrast, the findings of others have suggested that excessive fat mass may not protect against osteoporosis [12–15]. There is compelling evidence for both viewpoints from in vitro and in vivo studies, and several potential biological mechanisms have been proposed [16, 17]. These inconsistent findings reflect the inherently complicated nature of the relationship and call for new approaches and strategies to explore the potential effects of fat mass on bone [16].

Epidemiological studies have reported a nonlinear relationship between BMI (a combined measure of weight and height) and risk of osteoporotic fractures [18]. The results of a meta-analysis of 12 prospective studies including about 60 000 men and women suggested that most of the effect of BMI on nonhip fractures is probably mediated by the effects of weight on BMD (as adjustment for BMD removed most of the observed association), but at the hip there is a component that is independent of BMD [18]. However, it is not clear what proportion of the association with BMI may be related to the fat component of body weight. Of note, only a limited number of prospective studies have investigated fracture outcomes together with direct assessment of body fat [13, 19, 20]. Most of the previous clinical studies have used dual-energy X-ray absorptiometry (DXA) for assessment of both fat mass and bone mass. Bioelectrical impedance analysis (BIA) is another valid method for evaluation of body fat in obese individuals [21]. Moreover, whereas DXA measures only the density of the bone, other techniques such as bone quantitative ultrasound (QUS) are known to reflect elasticity and micro-architecture of the bone and to predict fractures as effectively as DXA [22, 23]. There are limited data from studies using these bone measurements [24, 25]. In the present study, we evaluated the association between fat mass (as measured by BIA) and bone density (as measured by QUS of the heel) and also the prospective risk of fracture in a European population.

Methods

  1. Top of page
  2. Abstract
  3. Introduction
  4. Methods
  5. Results
  6. Discussion
  7. Conflict of interest statement
  8. Acknowledgements
  9. References

Participants in this population-based, cross-sectional and cohort study were recruited as part of the East Anglian component of the European Prospective Investigation into Cancer in Norfolk (EPIC-Norfolk) study. Full details of participant recruitment and study procedures have been published previously [26]. Briefly, the original cohort comprised 25 639 men and women aged 40–79 years, recruited between 1993 and 1997 from general practice age-sex registers in the county of Norfolk, East Anglia, UK. Of these subjects, 15 515 attended a second health examination, including QUS of the heel and BIA, between January 1998 and October 2000. To date, all participants are being followed for different health end-points, including fractures. Data from the second health examination and prospective follow-up since then are included in this analysis. The study was approved by the ethics committee of the Norwich District Health Authority, and all participants provided signed informed consent.

Height and weight were measured in light clothing without shoes. Height was measured to the nearest millimetre using a free-standing stadiometer (CMS Weighing Equipment Ltd, London, UK). Weight was measured to the nearest 100 g using calibrated digital scales (Salter Industrial Measurement Ltd, West Bromwich, UK). Resistance (Ω) was assessed using a standard bio-impedance technique (Bodystat, Isle of Man, UK). This method has previously been shown to be valid [27] and reliable [28]. Total body water and fat-free mass were calculated using the impedance index (height2 per resistance), body weight and resistance according to published equations (see table 5 of [29]). Fat mass was calculated as body weight minus fat-free mass. Percentage body fat (%BF) was fat mass expressed as percentage of the total weight.

Quantitative ultrasound was used to measure broadband ultrasound attenuation (BUA; db MHz−1) and velocity of sound (VOS; m s−1) of the calcaneus with the use of the CUBA sonometer (McCue Ultrasonics, Winchester, UK), as described previously [22, 30]. The mean of the measures (left and right feet) was used for analysis. Five machines were used, and each was calibrated daily with its physical phantom and monthly with a roving phantom and on the calcaneus of one of the operators. Owing to a high correlation between BUA and VOS (pairwise correlation coefficient = 0.73), only BUA was considered as the outcome for this analysis. Smoking status was derived from questionnaires, and individuals were categorized as current, former or never smokers. Total alcohol consumption was estimated from the same questionnaires as units of alcohol (approximately 8 g alcohol per unit) consumed in a week.

Individuals were flagged for death certification at the UK Office of National Statistics, with vital status ascertained for the whole cohort. Participants who were admitted to hospital were identified using their unique National Health Service number by data linkage with the east Norfolk health authority database ENCORE, which identifies all hospital contacts throughout England and Wales for Norfolk residents. The International Classification of Diseases 9th and 10th revisions (ICD-9 and ICD-10) diagnostic codes were used to ascertain fractures by site in the cohort up to the end of March 2008 for this analyses. Fractures of the skull, face, metacarpals, metatarsal and phalanges were excluded from the analyses.

The associations between fat and bone measures as well as fracture risk are suggested not to follow a linear trend. Moreover, methods such as categorization of patients according to arbitrary cut-points or percentiles have limited statistical power to detect associations. We conducted regression analysis using fractional polynomial modelling to explore the association between %BF and BUA. Fractional polynomial modelling proposed by Royston and Sauerbrei [31] is a systematic approach to investigate possible nonlinear functional relationships of continuous variables. This method compares models with different combinations of linear and nonlinear transformations of continuous variables (first- and second-degree transformations) and selects the best fitting model with backward elimination. In case of no significant difference between models, the model with the lowest degrees of freedom (linear rather than first- and second-degree models) will be selected as the best fitting model. Cox proportional-hazards regression analysis with fractional polynomial modelling was used to assess the associations between %BF and prospective risk of fractures. All regression models were adjusted for age, history of fracture, height, smoking status and alcohol consumption. Cox models were additionally adjusted for BUA. Hip fracture was considered as a separate outcome for survival analysis. All database management and statistical analyses were performed using stata software, version 10.0 (StataCorp LP., College Station, TX, USA).

Results

  1. Top of page
  2. Abstract
  3. Introduction
  4. Methods
  5. Results
  6. Discussion
  7. Conflict of interest statement
  8. Acknowledgements
  9. References

After exclusion of participants with incomplete data, 14 789 participants were entered into the analysis. The mean age at baseline was 63.0 years amongst men and 61.8 years amongst women. The %BF was significantly higher amongst women than men (39.7 ± 9.1% vs. 23.5 ± 6.1%; < 0.001). Comparison of baseline characteristics of participants according to quartiles of %BF (Table 1) showed that, apart from history of fracture (and age for men), known risk factors of fracture were significantly different amongst participants with different levels of fat mass. Crude mean BUA generally increased with higher levels of fat mass amongst men and women.

Table 1. Characteristics of participants (8319 women and 6470 men) according to percentage body fat quartiles in the EPIC-Norfolk study
 Percentage body fatP- value
Quartile 1Quartile 2Quartile 3Quartile 4
  1. Data are mean (standard deviation) or number of participants (percentage).

  2. BMI, body mass index; BUA, broadband ultrasound attenuation; VOS, velocity of sound.

  3. aValues are median (interquartile range); Kruskal–Wallis test for P values.

Women<34%34–39%39.5–45%>45% 
n = 2198n = 2116n = 1964n = 2041
 Age (years)60.6 (9.7)61.3 (9.1)62.1 (8.7)62.1 (8.3)<0.001
 Height (cm)161.9 (6)161.3 (6)160.5 (6)159.9 (6)<0.001
 Weight (kg)58.6 (6.6)65.4 (7.0)70.9 (8.1)80.7 (11.4)<0.001
 BMI (kg m−2)22.3 (2.0)25.1 (2.0)27.5 (2.4)31.5 (3.9)<0.001
 Current smoking222 (10.1%)173 (8.2%)128 (6.5%)139 (6.8%)0.001
 Alcohol intake (units per week)a2.5 (6)2.5 (6)2 (5.5)2 (4.5)0.004
 Past history of fracture171 (7.8%)152 (7.2%)133 (6.8%)156 (7.6%)0.6
 BUA (dB MHz−1)68.6 (17.1)71.5 (16.3)72.9 (15.3)76.0 (16.1)<0.001
 VOS (m s−1)1621.2 (43.5)1624.2 (40.3)1625.1 (38.7)1629.2 (37.1)<0.001
 Incident hip fracture53 (2.4%)37 (1.8%)24 (1.2%)20 (1.0%)<0.001
 Any incident fracture122 (5.6%)96 (4.5%)83 (4.2%)92 (4.5%)0.2
Men<20%20–23%23.5–27%>27% 
n = 1724n = 1572n = 1596n = 1578
 Age (years)63.0 (9.5)62.9 (8.9)62.9 (8.9)62.8 (8.7)0.9
 Height (cm)174.5 (6.8)173.9 (6.5)173.9 (6.4)173.4 (6.6)<0.001
 Weight (kg)71.5 (7.4)78.5 (7.1)83.9 (8.1)92.4 (11.0)<0.001
 BMI (kg m−2)23.5 (1.8)25.9 (1.5)27.7 (1.7)30.7 (2.9)<0.001
 Current smoking175 (10.2%)119 (7.6%)105 (6.6%)115 (7.3%)<0.001
 Alcohol intake (units per week)a5.5 (10.5)6 (11.5)6.5 (12.5)6.5 (13.5)0.004
 Past history of fracture94 (5.5%)88 (5.6%)78 (4.9%)105 (6.7%)0.2
 UA (dB MHz−1)88.3 (18.5)90.0 (17.5)91.3 (16.9)90.8 (16.8)<0.001
 VOS (m s−1)1648.2 (41.6)1647.5 (40.0)1644.8 (39.0)1640.4 (37.7)<0.001
 Incident hip fracture17 (1.0%)8 (0.5%)10 (0.6%)15 (1.0%)0.3
 Any incident fracture40 (2.3%)36 (2.3%)44 (2.8%)43 (2.7%)0.7

Figure 1 shows the association between %BF and BUA in multivariable-adjusted fractional polynomial models. There was a positive linear association between %BF and BUA amongst women, whereas the association amongst men was nonlinear with a steeper slope amongst participants with low fat mass than those with higher %BF.

image

Figure 1. Association between percentage body fat (%BF) and broadband ultrasound attenuation (BUA) of the heel amongst 14 789 EPIC-Norfolk participants. Fitted lines (solid) and 95% confidence limits (dashed lines) are from fractional polynomial models adjusted for age, history of fracture, height, smoking status and alcohol intake.

Download figure to PowerPoint

During 122 330 person-years of follow-up, 556 fractures of any type (163 in men) occurred in EPIC-Norfolk participants, of which 184 (50 in men) were hip fractures. Time to fracture from baseline assessment was 5.0 ± 2.5 years for all fractures and 5.4 ± 2.5 years for hip fractures. Average follow-up time was 8.3 ± 1.6 years. Table 1 shows that, in univariable analysis, lower levels of fat mass were associated with risk of hip fracture amongst women but not men. Women with hip fracture had significantly lower %BF compared with other women (36.1 ± 8.7% vs. 39.7 ± 8.8%; < 0.001). This association was also evident in the multivariable Cox proportional-hazards regression analysis with different levels of adjustment for clinical variables.

Table 2 shows that, amongst women, crude %BF and age-adjusted %BF were significantly associated with reduced risk of hip fracture. Further adjustment for BUA did not essentially change the results (model 3 in Table 2), indicating that the relationship between fat mass and fracture risk is independent of bone characteristics as measured by heel QUS. Table 3 shows the Cox model with all variables (model 4 in Table 2). The effects of a 10% decrease in %BF (which is approximately 1 standard deviation of %BF in women) on risk of hip fracture amongst women was almost equal to an increase in age of 5 years or 1 standard deviation (17 dB MHz−1) decrease in BUA. There was no significant association between %BF and fracture risk amongst men in univariable (Table 1) and multivariable analyses (Tables 2 and 3). The best fitting models for both men and women were linear, and Fig. 2 depicts the linear decrease in risk of hip fracture attributed to %BF amongst women but not men.

Table 2. Association between percentage body fat and risk of hip fracture with different levels of adjustment for known risk factors
 WomenMen
HR95% CIP- valueHR95% CIP- value
  1. Hazard ratios (HRs) with 95% confidence interval (CI) are estimated for 10% increase in percentage body fat from Cox proportional-hazards regression models. Details of model 4 are shown in Table 3.

  2. BUA, broadband ultrasound attenuation.

  3. aThese clinical factors include: history of fracture, height, smoking status and alcohol intake.

Models with adjustment for
 1Crude0.620.50–0.77<0.0010.960.59–1.540.9
 2Age0.610.49–0.77<0.0011.000.61–1.620.9
 3Age and BUA0.690.55–0.870.0021.090.68–1.760.7
 4Age, BUA and other clinical factorsa0.710.56–0.890.0031.100.68–1.790.6
Table 3. Multivariable Cox proportional-hazards regression model for prediction of prospective risk of hip fracture amongst EPIC-Norfolk participants. Continuous variables are standardized to make appropriate comparisons
 WomenMen
HR95% CIP- valueHR95% CIP- value
  1. HR, hazard ratio; CI, confidence interval; BUA, broadband ultrasound attenuation.

Percentage body fat (per 10%)0.710.56–0.890.0011.100.68–1.790.6
Age (per 5 years)1.771.54–2.04<0.0011.861.49–2.32<0.001
Height (per 6 cm)1.231.02–1.490.031.080.80–1.470.6
History of fracture1.510.95–2.380.071.990.84–4.720.12
Current smoking1.000.45–2.100.90.720.25–2.040.5
Alcohol intake (units per week)0.950.90–0.990.0181.010.98–1.030.5
BUA (per 17 dB MHz−1)0.610.47–0.76<0.0010.630.46–0.850.003
image

Figure 2. Association between percentage body fat (%BF) and risk of hip fracture amongst participants of the EPIC-Norfolk study. Fitted lines (solid) and 95% confidence limits (dashed lines) are from fractional polynomial models in women (left) and men (right). All models are adjusted for age, history of fracture, height, smoking status, alcohol intake and broadband ultrasound attenuation. HR, hazard ratio.

Download figure to PowerPoint

The inverse association was also observed between %BF and risk of ‘any type of fracture’ amongst women (Fig. 3). The association appeared to be best modelled by a second-degree fractional polynomial curve with the lowest hazard seen at approximately a mean %BF of 40% (left graph; P < 0.001). Table 4 shows the hazard ratios for different categories of %BF in comparison with the mean fat category (35–45%). This Table shows that, whereas low values of %BF are accompanied by substantially higher risk of any fracture (e.g. more than twofold risk for women with <20% fat mass compared to women with 40% fat mass), higher values of %BF are also associated with moderately higher risk of fracture (e.g. women with >55% fat mass had about 50% higher risk of fracture compared to the mean). Again there was no significant association amongst men. Furthermore, analysing all nonhip fractures separately showed a similar pattern of association with wider confidence intervals compared to ‘any type of fractures’ (data not shown).

image

Figure 3. Association between percentage body fat (%BF) and risk of any type of fracture amongst participants of the EPIC-Norfolk study. Fitted lines (solid) and 95% confidence limits (dashed lines) are from fractional polynomial models in women (left) and men (right). All models are adjusted for age, history of fracture, height, smoking status, alcohol intake and broadband ultrasound attenuation. HR, hazard ratio.

Download figure to PowerPoint

Table 4. Hazard ratios (HRs) with 95% confidence interval (CI) for different levels of percentage body fat (%BF) compared to the mean %BF (40%) amongst female participants in the EPIC-Norfolk study
%BF rangeHR95% CINo. of womenNo. of fractures
  1. The values are obtained from a Cox proportional-hazards regression model (second-degree fractional polynomial) for prospective risk of any clinical fracture adjusted for age, height, history of fracture, smoking status, alcohol intake and heel broadband ultrasound attenuation.

<20%2.191.34–3.56385
20–25%1.651.20–2.2722923
25–35%1.211.06–1.392511116
35–45%1.003502157
45–55%1.020.90–1.15157168
>55%1.480.99–2.2146824

Discussion

  1. Top of page
  2. Abstract
  3. Introduction
  4. Methods
  5. Results
  6. Discussion
  7. Conflict of interest statement
  8. Acknowledgements
  9. References

The results of this study show that low body fat mass is an independent risk factor for fractures amongst women but not men. Women with, for instance, <20% body fat have a 3.5 times higher risk of hip fracture and 2.2 times increased risk of any type of fracture compared with women with ∼40% body fat. We observed a positive association between fat mass and bone properties (as measured by heel QUS in our study), but the magnitude of this association may not have been large enough to account for the effect of fat mass on fracture risk. Risk of hip fracture amongst women almost halved with each 10% increase in %BF (Table 3), and a nonlinear reduction in risk was also observed for any type of fracture. Differences in findings between studies may reflect a lack of consideration of possible sex difference and nonlinear associations.

We observed the harmful effects of low fat mass on hip and other fractures only amongst women. Although it is possible that the lack of a significant association amongst men in our study was because of low power to detect such an association, there is no obvious trend towards a change in risk with higher values of %BF in men. It is notable that the range of %BF in men was much narrower than in women. This sex-specific association between fat mass and BMD has also been suggested previously [32]. Hormonal differences between sexes are one possible cause of this effect. Oestrogen reduces osteoclast-mediated bone resorption and stimulates osteoblast-mediated bone formation [33]. After cessation of secretion of oestrogen from the ovaries in postmenopausal women, extragonadal oestrogen synthesis in fat tissue (mediated by the enzyme aromatase) [34] becomes the dominant oestrogen source. This may lead to the harmful effects of low fat mass on bone in postmenopausal women. Androgen deficiency as a result of hypogonadism contributes to bone loss in 20–30% of elderly men, but this association is not affected by body fat mass in men [33].

Several other mechanisms have been proposed to explain how fat mass may relate to bone characteristics: these mechanisms may have both positive and negative impact on bone health. The interplay between these processes in each individual might ultimately determine the net beneficial or detrimental effects of fat mass on bone health. Two mechanical explanations for the effect of fat mass on bone are the cushioning effects of fat pads on bony areas, such as the hip [35], and increased bone strength in response to the greater mechanical loading imposed by a larger body mass [12]. Furthermore, it is known that adipose tissue is not just an inert organ for energy storage [17]. It expresses and secrets a wide variety of biologically active molecules such as oestrogen, leptin [36, 37], adiponectin [36, 38], resistin [39] and interleukin-6 [40]. The secretion of these hormones as well as bone-active hormones from the pancreas (including insulin, amylin and preptin) [41–43] may contribute to the complex relationship between fat mass and bone. Moreover, adipocytes and osteoblasts both originate from a common progenitor, the pluripotential mesenchymal stem cell. These stem cells display an equal propensity for differentiation into adipocytes or osteoblasts, and the balance of differentiation is regulated by several interacting pathways that may contribute to the final effect of fat mass on bone [16].

An important finding of our study is that low fat mass is a risk factor for hip fracture independently of bone density measured using heel QUS. Most cross-sectional studies assessing the relationship between fat and bone have used hip or lumbar DXA with inconsistent findings [7–15]. A few studies have also used heel QUS for bone assessment. Kroke et al. [24] used a skinfold thickness method to estimate fat and lean mass and reported a significant association with heel BUA amongst pre- and postmenopausal women. Assantachai et al. [25] also reported a significant negative association in categorical analysis for BUA and fat mass. All these studies have used bone measures (either DXA or QUS) to estimate the potential impact on fracture risk. However, our results indicate that the relationship between fat mass and prospective fracture risk is largely independent of BMD. In models to predict fracture using %BF, further adjustment for heel QUS did not change the association substantially (Table 2). This suggests that simple extrapolation of the relationship between fat mass and bone density to estimate fracture risk is unlikely to be satisfactory. In other words, a single cross-sectional bone measurement (either QUS or DXA) may not represent bone health in the complicated relationship with fat mass, and therefore, future studies should use prospective designs with fracture outcomes.

We explored nonlinear associations in our study and observed an interesting pattern of association between %BF and risk of ‘any type of fracture’. Fractional polynomial modelling is an easy and widely available method from an array of statistical methods recently developed for investigating nonlinear associations. This or similar methods (e.g. regression splines) merit a greater role in epidemiology, and future population-based studies evaluating the association between fat and bone health should consider use of such methods. The selection of factors to adjust for multivariable models may also have a significant impact on the final models observed in these studies [12].

Our study has several methodological strengths, including its prospective design and population-based sample that makes our results more generalizable. A potential limitation of this study is a low power to detect associations especially for hip fracture and amongst men. Although there was virtually complete follow-up of the cohort using routine record linkage with national hospitalization data, only fractures that required admission to hospital were identified for this study. This might have resulted in underestimation of the rate of fracture in our population. However, hospitalized fractures are arguably the ones with the most clinical impact. We were also not able to exclude high-trauma fractures, for example because of car accidents; however, it is very hard to distinguish between osteoporotic and nonosteoporotic fractures amongst the elderly involved in an accident and some recommend considering all fractures in this population as osteoporotic.

In conclusion, the results of this study indicate that higher body fat mass is associated with lower risk of fracture amongst women but not men. This relationship appeared to be independent of bone characteristics as measured by heel QUS. Clarifying the nature of this relationship may improve the understanding of the different mechanisms involved in fracture risk that can inform preventive strategies in the future.

Conflict of interest statement

  1. Top of page
  2. Abstract
  3. Introduction
  4. Methods
  5. Results
  6. Discussion
  7. Conflict of interest statement
  8. Acknowledgements
  9. References

None of the authors has any financial or personal interests to declare.

Acknowledgements

  1. Top of page
  2. Abstract
  3. Introduction
  4. Methods
  5. Results
  6. Discussion
  7. Conflict of interest statement
  8. Acknowledgements
  9. References

EPIC-Norfolk is supported by programme grants from the Medical Research Council and Cancer Research UK. The sponsors had no role in the design and conduct of the study, collection, management, analysis and interpretation of the data, or preparation, review and approval of the manuscript.

References

  1. Top of page
  2. Abstract
  3. Introduction
  4. Methods
  5. Results
  6. Discussion
  7. Conflict of interest statement
  8. Acknowledgements
  9. References