Superiority of age and weight as variables in predicting osteoporosis in postmenopausal white women

Identification of women at risk for osteoporosis is of great importance for the prevention of osteoporotic fractures. Routine BMD measurement of all women is not feasible for most populations, hence identification of a high-risk subset of women is an important element of effective preventive strategies. Methods: We identified 959 postmenopausal non-Hispanic women aged 51 years and above from the NHANES III study to assess the relative contribution of risk predictors for low BMD at the whole proximal femur and the femoral neck regions. Based on recognized risk factors for osteoporosis identified by a systematic literature search, we ran several multiple linear regression models based on the results of preceding bivariate analyses. We show several models based on their explanatory ability assessed by adjusted r ², ROC, and C-value analyses rather than on the coefficients and P values. We furthermore examined the sensitivity, specificity, and predictive values of our preferred models for various cutoff T-scores—the choice of which will vary depending on different study goals and population characteristics. Results: Age and weight were by far the most informative predictors for low bone mineral density out of a list of 20 candidate risk predictors. Our preferred prediction models for the two regions hence contained only two variables: i.e., age and measured weight. The resulting parsimonious model to predict BMD at whole proximal femur had an adjusted r ² of 0.43, an area under the ROC curve of 0.85, and a C-value of 0.70. Similarly, prediction for BMD at the femoral neck had adjusted r ², area under the curve, and C-value of 0.39, 0.83, and 0.66, respectively. Conclusions: The model equations, predicted T-score = −1.332−0.0404 × (age) + 0.0386 × (measured weight) and predicted T-score = −1.318−0.0360 × (age) + 0.0314 × (measured weight) for whole proximal femur and femoral neck, respectively, can be used in field conditions for screening purposes. More complex prediction equations add little explanatory power. Based on the study goals and the population characteristics, specific cutoff T-scores have to be decided before using these equations.