STATISTICAL ANALYSIS OF BODY COMPOSITION & RISK FACTORS

几十年来，统计方法一直应用于科学研究

For decades, statistical methods have been applied in scientific research to facilitate objective interpretation and judgment of the results. Progress in statistical methods relies on incentives to improve the existing methods and to reduce the limitations of applications. In biomedical, health-related, and epidemiological studies, multiple linear and logistic regressions with least square estimation have been widely adopted, despite the fact that some statistical literature has investigated the limitations of least square estimation when predictor variables are interrelated and influential observations are present. Alternative estimation procedures have been developed, but there is a gap between statistical advances and their application to health-related research. Parametric models have been applied to serial data of growth-related variables to simulate individual growth patterns for a study group. These models assume individual growth pattern for the group can be represented by a family of mathematical functions and the parameters in the models differentiate individual differences. Nonetheless, for some variables such as blood pressures, skinfold thicknesses, and risk factors, the patterns of change vary so much that they cannot be represented by a family of mathematical functions. In this case, nonparametric models prove useful. The planned analyses encompass (1) cross-sectional analyses: the development of predictive equations (2) longitudinal analyses: the establishment of risk functions, and the implementation of nonparametric kernel regression. Least squares will be adopted to formulate equation to predict body composition variables from anthropometry and impedance followed by examination of diagnostic measures, multicollinearity and influential observations and their effects on least square estimation. Ridge and robust regressions will be used as alternatives. The predictability of adulthood values from childhood values will e investigated by logistic regression using iterative weighted least squares or maximum likelihood methods. The effects of multicollinearity and influential observations on the logistic regression estimates will be evaluated. Ridge and robust regressions will be employed as adulthood to values in childhood. A kernel estimator for finite sample investigation will be developed and applied to serial measures of body composition and risk factors to provide smooth reference data. The long-term goal of this First Award proposal are to enhance the application of existing statistical methods, to develop improved methods when limitations of existing methods are encountered, and thus to assist investigators int he development of predictive equations, in making risk analyses, and in analyzing long-term serial data. Although the investigation will be performed on existing data bases from longitudinal and cross-sectional studies of human growth, body composition, and risk factors for major diseases, these methods are applicable to many other variables.