Osteoarthritis (OA) is a leading cause of pain and physical disability in middle-aged and older individuals. We undertook this study to determine predictors of the development of severe OA, apart from age and overweight.
Osteoarthritis (OA) is a leading cause of pain and physical disability in middle-aged and older individuals. We undertook this study to determine predictors of the development of severe OA, apart from age and overweight.
Joint replacement surgery due to severe hip or knee OA was recorded over a 15-year period in the prospective Bruneck cohort study. Demographic characteristics and lifestyle and biochemical variables, including the level of soluble vascular cell adhesion molecule 1 (VCAM-1), were assessed at the 1990 baseline visit and tested as predictors of joint replacement surgery.
Between 1990 and 2005, hip or knee joint replacement due to OA was performed in 60 subjects. VCAM-1 level emerged as a highly significant predictor of the risk of joint replacement surgery. Intervention rates were 1.9, 4.2, and 10.1 per 1,000 person-years in the first, second, and third tertiles, of the VCAM-1 level, respectively. In multivariable logistic regression analysis, the adjusted relative risk of joint replacement surgery in the highest versus the lowest tertile group of VCAM-1 level was 3.9 (95% confidence interval 1.7–8.7) (P < 0.001). Findings were robust in various sensitivity analyses and were consistent in subgroups. Addition of the VCAM-1 level to a risk model already including age, sex, and body mass index resulted in significant gains in model discrimination (C statistic) and calibration and in more accurate risk classification of individual participants.
The level of soluble VCAM-1 emerged as a strong and independent predictor of the risk of hip and knee joint replacement due to severe OA. If our findings can be reproduced in other epidemiologic cohorts, they will assist in routine risk classification and will contribute to a better understanding of the etiology of OA.
Osteoarthritis (OA) is the most common joint disorder throughout the world and a leading cause of disability. Involvement of the knee and hip joints leads to the most severe disease phenotype. Pain is the initial and core clinical feature, followed by impaired joint mobility, reduction of muscle strength, and loss of joint function (1, 2). Pain and joint failure are the driving forces for joint replacement surgery, the major therapeutic advance in OA management during the last decades (3). Almost 200,000 hip joints are replaced in the US every year, corresponding to annual intervention rates of 50–130 per 100,000 individuals, which highlights the enormous socioeconomic relevance of OA (4).
Prevalence estimates of OA strongly depend on the method of assessment. Histopathologic studies detected cartilage erosions in no less than two-thirds of individuals in the sixth and seventh decades of life (5). Radiographic screenings in population-based surveys yielded prevalence estimates for knee and hip OA in persons older than 60 years of >20% and ∼10%, respectively (6). Radiography is the standard tool for diagnosing OA; however, radiographic abnormalities correlate imperfectly with clinical symptoms, and rates of clinical OA are usually lower than one would assume based on the frequency of radiographic abnormalities (7–9). Accordingly, only a fraction of patients with radiographic signs of OA develop clinical sequelae and joint dysfunction.
Unlike most other common diseases, little is known about the etiology of OA, and predictors of a severe disease course remain to be identified. Manifestations of OA have been linked to older age, overweight, high-impact physical activities, bone mass, and genetic factors (10–14). Validation in longitudinal large-scale studies focusing on the progression of knee and hip OA, however, is only available for the predictors of age (15, 16) and body mass index (BMI) (17, 18). Recently, pathophysiologic concepts of OA have put more emphasis on interactions between cartilage, bone, and the synovium. Such concepts suggest a role not only of inflammatory markers in the pathogenesis of OA, but also of adhesion molecules, which mediate cell–cell and cell–matrix interactions. The current study is the first to establish a laboratory marker for the risk of severe OA in a long-term prospective population-based study. Joint replacement surgery was used as a reliable indicator of advanced symptomatic OA, thereby avoiding the challenges related to the assessment of severe progressive OA on clinical or radiographic grounds. We found that the baseline level of soluble vascular cell adhesion molecule 1 (VCAM-1), a surface sialoglycoprotein inducibly expressed on chondrocytes and synovial fibroblasts, predicts the 15-year probability of knee and hip joint replacement due to OA independently of age, sex, and BMI.
The Bruneck Study is a prospective population-based survey on the epidemiology and pathogenesis of atherosclerosis and disorders of the brain, bone, and joints (19–22). The study protocol was reviewed and approved by the appropriate ethics committees, and all study subjects gave their written informed consent. At the 1990 baseline visit, the study population was recruited as a random sample (stratified according to age and sex) of all inhabitants of Bruneck, Italy (125 women and 125 men in each of the fifth through eighth decades of life). A total of 93.6% of recruited subjects participated, with data assessment completed for 919 subjects. Followup examinations were performed every 5 years (in 1995, 2000, and 2005), and information on joint replacement surgery was complete for 917 individuals (>99%). Five subjects with joint replacement surgery before 1990 were excluded, leaving a study population of 912 subjects for the current analysis.
Localization, date, and circumstances of hip and knee replacement surgery were carefully recorded using 3 sources of information: subject self-report, medical records of the Bruneck Hospital and general practitioners, and a standardized evaluation of radiographs. The situation in Bruneck is unique in 3 ways: 1) the only radiography and orthopedic surgery facility in the whole district is located at Bruneck Hospital, and all radiographs ever obtained and records of all surgeries ever performed on study subjects are available for review; 2) Bruneck Hospital is convenient for obtaining radiographs in the case of moderate-to-severe or long-lasting knee and hip pain; and 3) population mobility in the area has been low over the past 15 years. Joint replacement following bone fractures (documented by radiography) was not considered in the current analysis. All other subjects with joint replacements had documented OA of the hip or knee on at least 2 sequential radiographs before surgery and met radiographic criteria for the diagnosis of OA (23, 24).
BMI was calculated as the weight (kg) divided by the height (m2). Smoking status and alcohol consumption were recorded as detailed previously (19). The activity score was composed of the scores for work (3 categories) and sports/leisure activities (0, ≤2, or >2 hours per week) (19). Socioeconomic status was defined according to a 3-category scale (low, medium, high) based on information about occupational status and educational level of the person with the highest income in the household (19). Diabetes was diagnosed according to the criteria of the American Diabetes Association (25). The category of cardiovascular disease subsumes fatal and nonfatal strokes (according to the criteria of the National Survey of Stroke ), fatal and nonfatal definite myocardial infarctions (according to the World Health Organization definition ), revascularization procedures, and peripheral artery disease (21).
Baseline blood samples were drawn after an overnight fast and 12 hours of abstinence from smoking. Routine blood parameters were assessed by standard methods. Additional blood samples were immediately frozen and stored at –70°C (without any freeze–thaw cycle). The level of soluble VCAM-1 was measured in duplicate using commercially available enzyme-linked immunosorbent assay kits (Bender MedSystems, Milan, Italy). Intra- and interassay coefficients of variation were 4.8% and 11.2%, respectively (28).
For computation of intervention rates, person-years of followup for each participant were accrued from the 1990 baseline until joint replacement surgery, death, or October 1, 2005, whichever came first. To evaluate the association between potential risk predictors and joint replacement, odds ratios (ORs) and 95% confidence intervals (95% CIs) were calculated by means of logistic regression analyses with the test procedure based on the maximum likelihood estimator.
In a first step, all candidate variables shown in Table 1 were tested by separately entering them in base models including variables for age, sex, and BMI. In this explorative step, the VCAM-1 level emerged as a highly significant predictor of the risk of joint replacement due to severe OA (P < 0.00001), while all other variables clearly showed a lack of significance. In a second step, the association between the soluble VCAM-1 level and the risk of joint replacement was further elaborated. In the absence of a generally accepted cutoff, the VCAM-1 level was treated as a continuous variable. Multivariable analyses were adjusted for age, sex, BMI, social status, lifestyle variables, and measures correlated with VCAM-1 level. All analyses were repeated using ln-transformed levels of soluble VCAM-1. Since findings were essentially the same, we present only data from analyses using nontransformed levels for ease of presentation and interpretation. Nonlinear effects of continuous variables on the log odds of joint replacement were proven by a visual inspection of plots of the variable against the log odds of intervention probability, by adding variable × ln(variable) terms to the regression models (Box-Tidwell transformation) and by using orthogonal polynomials. Neither of these procedures identified significant departures from linearity that were relevant to risk prediction and discrimination. In supplementary analyses, the level of VCAM-1 was subdivided into 3 approximately equally sized groups (tertile groups).
|No joint replacement surgery (n = 852)||Joint replacement surgery (n = 60)||P†|
|Age, years||58.3 ± 11.5||65.2 ± 8.2||<0.001|
|BMI, kg/m2||24.8 ± 3.7||26.0 ± 4.6||0.022|
|Social status, %||0.962|
|Physical activity score (ref.38)||4.3 ± 1.5||4.3 ± 1.7||0.463|
|Diabetes mellitus, %||6.3||18.3||0.066|
|Fasting glucose, mg/dl||100.6 ± 20.1||107.6 ± 25.1||0.238|
|Waist circumference, cm||98.7 ± 8.2||100.5 ± 8.6||0.329|
|Total cholesterol, mg/dl||222.5 ± 40.5||221.5 ± 45.6||0.383|
|HDL cholesterol, mg/dl||56.4 ± 14.4||56.3 ± 12.9||0.672|
|LDL cholesterol, mg/dl||138.2 ± 37.8||140.8 ± 45.6||0.952|
|Systolic blood pressure, mm Hg||145.1 ± 21.3||152.3 ± 23.9||0.888|
|Diastolic blood pressure, mm Hg||88.8 ± 10.0||91.0 ± 11.2||0.884|
|Creatinine, mg/dl||0.9 ± 0.5||0.9 ± 0.2||0.389|
|Uric acid, mg/dl||5.4 ± 1.5||5.7 ± 1.5||0.834|
|GGT, units/liter||21.6 ± 26.6||20.2 ± 21.5||0.566|
|Carotid atherosclerosis score, mm (ref.39)||2.1 ± 3.9||2.3 ± 2.8||0.336|
|Ferritin, μg/liter||148.0 ± 172.9||173.9 ± 194.1||0.367|
|High-sensitivity CRP, mg/liter||2.7 ± 3.8||2.9 ± 3.0||0.601|
|Fibrinogen, gm/liter||262.3 ± 63.2||269.2 ± 51.7||0.147|
|Soluble ICAM-1, ng/ml||328.7 ± 88.6||335.8 ± 76.6||0.621|
|Soluble VCAM-1, ng/ml||675.8 ± 275.5||943.8 ± 589.9||<0.001|
|E-selectin, ng/ml||53.9 ± 19.9||54.5 ± 19.9||0.629|
|P-selectin, ng/ml||198.1 ± 53.9||191.0 ± 53.4||0.830|
|MMP-9, ng/ml||296.8 ± 157.7||282.9 ± 133.4||0.758|
|Adiponectin, gm/liter||12.2 ± 6.2||13.1 ± 7.0||0.973|
|Leptin, pmoles/liter||9.3 ± 7.1||12.1 ± 12.4||0.267|
|Osteocalcin, ng/ml||27.2 ± 16.4||29.4 ± 15.1||0.776|
|β-CrossLaps, ng/ml||0.5 ± 0.3||0.5 ± 0.3||0.847|
|Soluble RANKL, pmoles/liter||1.2 ± 0.9||1.2 ± 1.1||0.908|
|Osteoprotegerin, pmoles/liter||3.7 ± 1.2||4.1 ± 1.1||0.566|
Logistic regression analyses were supplemented and confirmed by Cox proportional hazard models allowing for censored data, after confirming that the assumption of proportionality was met. When the exact date of joint replacement was not available, we used the date of the first radiograph showing the implant as a surrogate. This may be viewed as an adequate approximation in most individuals and errs, if at all, on the conservative side.
To estimate the discriminative value of prediction models (i.e., the ability to correctly classify subjects into 1 of 2 categories), we calculated the C statistic, which is analogous to the area under the receiver operating characteristic (ROC) curve (larger values indicate better discrimination). Comparison of ROCs based on models including and not including VCAM-1 was performed according to the method of DeLong et al (29). Findings remained robust when applying the algorithms proposed by Hanley and McNeil (30). To assess model calibration (how closely the predicted probabilities reflect actual risk), we computed the Hosmer-Lemeshow calibration statistics comparing observed and predicted risk in decile categories of predicted risk (more significant P values indicate better calibration). Estimates of the probability of joint replacement surgery in individual study participants and subgroups were derived from logistic regression equations using a leave-one-out technique (cross-validation) to reduce the extent of bias (overfitting). All reported P values are 2-sided.
During followup, 60 individuals of the Bruneck cohort had joint replacement surgery due to severe OA (38 with hip replacements, 22 with knee replacements), corresponding to an intervention rate of 5.1 per 1,000 person-years or 6.0 per 1,000 person-years after considering bilateral joint replacement. As expected, the frequency of joint replacement surgery was high in obese subjects (10.7 per 1,000 person-years versus 4.5 per 1,000 person-years in nonobese subjects) and increased with advancing age (7.7 versus 2.1 per 1,000 person-years in subjects ages ≥60 years and ages <60 years at baseline), whereas men and women showed no significant difference (4.6 and 5.6 per 1,000 person-years).
Population characteristics in subjects with and in those without joint replacement are shown in Table 1. In addition to age and BMI, the level of VCAM-1 was substantially elevated in affected individuals (P < 0.00001). However, the values of all other lifestyle and laboratory parameters, including adhesion molecules other than VCAM-1, were distributed similarly in both groups. Of note, the baseline VCAM-1 level was highest among individuals with bilateral joint replacement (data are available online at http://www.medizin3.uk-erlangen.de/e1846/e473). The level of VCAM-1 did not differ between sexes and increased with age (r = 0.211). Moreover, the VCAM-1 level showed modest but significant associations with levels of β-CrossLaps, a marker of bone resorption (partial correlation corrected for age and sex, r = 0.106), total cholesterol (r = −0.183), adiponectin (r = 0.108), E-selectin (r = 0.106), and intercellular adhesion molecule 1 (r = 0.130) (P ≤ 0.001 for all comparisons).
The level of VCAM-1 emerged as a significant predictor of the risk of joint replacement due to severe OA, equaling or even surpassing the effects of age (variable likelihood ratio χ2 = 28.35 and 20.76 for inclusion of VCAM-1 level and age, respectively, in a model including sex) (Table 2). Table 3 shows estimates of the relative risk of joint replacement associated with a 1-SD unit increase in VCAM-1 level and comparisons between the bottom tertile group and the middle and top tertile groups for VCAM-1 level. Results did not change when the analysis was additionally corrected for social status, lifestyle variables, and measures correlated with VCAM-1 level (Table 3).
|Model, variable||Variable LR chi-square||Calibration P value||Model discrimination|
|C statistic (95% CI)||Δ C statistic||P|
|Sex only||NC||0.531 (0.456–0.606)|
|+ BMI||4.75||0.990||0.566 (0.486–0.646)||0.035||0.338|
|+ age||20.76||0.051||0.683 (0.627–0.738)||0.152||<0.001|
|+ soluble VCAM-1 level||28.35||0.836||0.671 (0.601–0.742)||0.140||0.002|
|Age, sex, and BMI||0.055||0.694 (0.637–0.751)|
|+ soluble VCAM-1 level||21.06||0.365||0.734 (0.676–0.792)||0.040||0.047|
|Type of replacement, type of analysis||Risk vs. first tertile of VCAM-1 level||Risk per 1-SD unit increase in VCAM-1 level||Risk vs. first tertile of VCAM-1 level||Risk per 1-SD unit increase in VCAM-1 level|
|OR (95% CI)||P for trend†||Second tertile||HR (95%) CI)|
|Second tertile||Third tertile||OR (95% CI)‡||P||Third tertile||P for trend†||HR (95% CI)‡||P|
|Hip/knee joint replacement|
|Unadjusted analysis||2.0 (0.9–4.9)||4.9 (2.2–10.7)||<0.001||1.7 (1.4–2.1)||<0.001||2.1 (0.9–5.0)||5.2 (2.4–11.2)||<0.001||1.8 (1.5–2.0)||<0.001|
|Multivariable analysis§||1.6 (0.7–3.9)||3.6 (1.6–8.1)||<0.001||1.6 (1.3–2.0)||<0.001||1.6 (0.7–3.7)||3.5 (1.6–7.6)||<0.001||1.7 (1.4–2.0)||<0.001|
|Multivariable analysis¶||1.7 (0.7–4.1)||3.9 (1.7–8.7)||<0.001||1.7 (1.4–2.2)||<0.001||1.6 (0.7–3.7)||3.6 (1.6–8.1)||<0.001||1.8 (1.5–2.1)||<0.001|
|Hip joint replacement|
|Unadjusted analysis||2.0 (0.7–6.1)||5.0 (1.9–13.3)||<0.001||1.6 (1.3–2.1)||<0.001||2.1 (0.7–6.3)||5.4 (2.1–14.3)||<0.001||1.6 (1.3–2.0)||<0.001|
|Multivariable analysis§||1.7 (0.6–5.1)||4.0 (1.5–10.8)||0.002||1.6 (1.2–2.1)||0.001||1.6 (0.5–4.8)||4.0 (1.5–10.6)||0.001||1.5 (1.2–2.0)||<0.001|
|Multivariable analysis¶||1.8 (0.6–5.6)||4.9 (1.7–13.6)||0.001||1.7 (1.3–2.3)||<0.001||1.7 (0.6–5.2)||4.9 (1.8–13.4)||<0.001||1.7 (1.3–2.2)||<0.001|
|Knee joint replacement|
|Unadjusted analysis||2.0 (0.5–8.3)||4.7 (1.3–16.7)||0.007||1.9 (1.5–2.5)||<0.001||2.2 (0.5–8.7)||5.1 (1.5–18.0)||0.004||2.0 (1.6–2.5)||<0.001|
|Multivariable analysis§||1.5 (0.4–6.2)||3.0 (0.8–10.9)||0.054||1.8 (1.3–2.3)||<0.001||1.5 (0.4–5.9)||2.9 (0.8–10.5)||0.050||1.9 (1.5–2.4)||<0.001|
|Multivariable analysis¶||1.4 (0.3–5.9)||2.4 (0.6–9.3)||0.144||1.9 (1.4–2.7)||<0.001||1.4 (0.3–5.5)||2.3 (0.6–8.7)||0.146||2.0 (1.5–2.8)||<0.001|
A number of prespecified sensitivity analyses further substantiated the key finding. First, the significant association was internally consistent and applied to men and women (data are available online at http://www.medizin3.uk-erlangen.de/e1846/e473) and to both hip and knee joint replacement (Table 3). Second, Cox proportional hazard models allowing for censored data yielded highly consistent findings (Table 3). Cumulative hazard curves according to tertile groups for the VCAM-1 level are depicted in Figure 1. Third, predictive significance was demonstrated in a variety of subgroups and settings such as various age strata, subjects with or without diabetes, presence of common diseases (neoplasm, cardiovascular disease, chronic infection, dementia, and autoimmune disease), and incident cardiovascular disease (Figure 2). Moreover, the baseline level of soluble VCAM-1 predicted the risk of joint replacement due to severe OA over a long period of time (from 1990 to 2005) (Figure 2). Fourth, subjects with the most severe disease expression (i.e., those requiring bilateral joint replacement) exhibited the highest baseline levels of VCAM-1, and the predictive significance was most pronounced in this group (risk of bilateral joint replacement conferred by a 1-SD unit increase in the VCAM-1 level, OR 2.5 [95% CI 1.7–3.6]).
In order to assess the potential usefulness of soluble VCAM-1 as a risk predictor in routine clinical practice, we estimated its effect on model discrimination and calibration and on risk classification. Actually, addition of VCAM-1 to the regression equation including age, sex, and BMI resulted in a significant increase in the C statistic (from 0.694 to 0.734, with a difference of 0.040 [95% CI 0.001–0.079]; P = 0.047) (Figure 3 and Table 2), indicating a gain in model discrimination. ROCs for a model focusing on bilateral joint replacement (data are available online at http://www.medizin3.uk-erlangen.de/e1846/e473) again demonstrated that the model including VCAM-1 provided a better fit (increase in the C statistic from 0.731 to 0.874, with a difference of 0.143 [95% CI 0.055–0.231]; P = 0.0015). Similarly, models that included the level of VCAM-1 demonstrated better calibration (more significant P values in the Hosmer-Lemeshow calibration statistic) (Table 2). Finally, consideration of the VCAM-1 level resulted in a more accurate classification of subjects. A total of 10 of the 60 subjects requiring joint replacement were correctly reclassified from the categories of <5.0% and 5.0–14.9% into the category of ≥15.0%, whereas only 1 subject was erroneously downgraded from the high-risk group to the low to medium–risk group. Overall, only 2.7% of the 113 subjects reclassified to a lower risk group underwent joint replacement, while 15.7% of the 76 subjects reclassified to a higher risk group underwent joint replacement (data are available online at http://www.medizin3.uk-erlangen.de/e1846/e473).
Our study provides strong evidence that the level of soluble VCAM-1 is a predictor of the long-term risk of joint replacement due to severe OA. The key association remained robust in a variety of sensitivity analyses and showed a high level of internal consistency in subgroups. Remarkably, inclusion of VCAM-1 levels in risk prediction models resulted in a more accurate classification of given individuals as well as in better overall model discrimination and calibration. Predictive accuracy was highest for subjects with the most severe phenotype (bilateral joint replacement) and was not affected by the presence or absence of other common diseases.
VCAM-1 (CD106) is a cell surface sialoglycoprotein that binds α4β1 and α4β7 integrins and mediates heterotypic cellular aggregation (31). VCAM-1 was first identified on the endothelium of blood vessels exposed to cytokines, where it promotes leukocyte adhesion and homing to sites of inflammation. In joints, VCAM-1 is expressed by microvascular endothelial cells, synovial fibroblasts, and chondrocytes (32). Moreover, adipose tissue is an abundant source of VCAM-1 (33, 34). In chondrocytes, expression of VCAM-1 is induced by inflammatory cytokines, such as tumor necrosis factor α and interleukin-1, and by hyaluronic acid, a major component of the cartilage matrix liberated during cartilage damage (35, 36). Accordingly, increased VCAM-1 levels may mirror active cartilage damage or an inflammatory component in OA. Finally, VCAM-1 has been shown to mediate the interaction of chondrocytes with immune cells and could thus by itself contribute to immune-mediated cartilage damage in OA (36).
Establishment of novel laboratory biomarkers and risk algorithms for the prediction of severe OA is of great interest for several reasons. First, the standard risk factors of age and overweight do not allow accurate risk prediction. Second, OA is a highly prevalent disease necessitating accurate identification of subjects at greater risk of rapid disease progression and development of severe phenotypes. Third, early diagnosis of OA is difficult because disease onset is gradual and precedes clinical manifestation. Finally, improved prediction of the risk of severe OA may help to select patients for intensive self-management programs (aerobic exercise, strength training, and weight loss) and to tailor therapeutic measures.
Our study has several strengths. Its findings are representative of the general community, as ascertained by nearly complete participation and followup of subjects and by inclusion of subjects with a broad range of ages (40–94 years). The study also included high-quality assessment of hip and knee joint replacements (end point) and of lifestyle and laboratory parameters. In addition, costs of joint replacement surgery are fully covered by the public health care system in the survey area, and access to surgery is the same for all social groups. Moreover, there was no association between social status and VCAM-1 levels in our study, and adjustment for social status had virtually no effect on the results obtained.
A potential limitation of the study is that radiographic changes in the knee and hip joints were not systematically assessed in all subjects. However, correlation of radiographic changes with clinical symptoms is weak, and nonassessment of minor disease phenotypes may be expected to weaken evident relationships rather than to create spurious ones. Another limitation is that these results need to be confirmed by other independent epidemiologic studies. Recently, for instance, Ioannidis et al addressed the repeatability of microarray gene expression analyses and showed that the rate of repeatability was rather low (37). The reasons for failed repeatability were mainly unavailability of data, incomplete annotation of data, or incomplete specification of data processing. By aiming for high-quality data with almost complete followup of subjects and by carefully adjusting for potential confounders, we sought to optimize the chances that our results could be repeated. Final proof of repeatability, however, will require a longitudinal analysis of a similar epidemiologic cohort.
In summary, the level of soluble VCAM-1 emerged as a highly significant predictor of the risk of hip and knee joint replacement due to severe OA. Further clarification of the mechanisms underlying the association between the VCAM-1 level and OA may well contribute to a better understanding of disease etiology. Moreover, application of our findings in routine clinical practice awaits prior validation in independent population samples showing repeatability of results.
All authors were involved in drafting the article or revising it critically for important intellectual content, and all authors approved the final version to be published. Dr. Schett had full access to all of the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.
Study conception and design. Schett, Kiechl, Bonora, Oberhollenzer, Lorenzini, Willeit.
Acquisition of data. Schett, Kiechl, Bonora, Zwerina, Mayr, Axmann, Weger, Willeit.
Analysis and interpretation of data. Schett, Kiechl, Bonora.
We thank Dr. Kurt Redlich (Medical University of Vienna) for critical discussion of the study design.