Built upon two advanced nonparametric statistical techniques, Multivariate Adaptive Regression Splines and Classification and Regression Trees, tree-based methods will be developed and applied to explore the data from the Yale Pregnancy Outcome Study (YPOS), which was designed to examine the relationship between pregnancy outcome and a variety of risk factors, including prescription drug and alcohol use, tobacco smoke, caffeine consumption, and contraceptive practice. Data from about 7,000 subjects will be available from two related YPOS's. Analyses will also be extended to several other important databases including the 1988 National Health Interview Survey on Child Health. In contrast to traditional statistical methods and software, the mechanisms that we will employ and investigate have several advantages: (i) automatically finding the important variables and significant interactions among a large number of variables, making it more likely that new risk factors for pregnancy outcome study (other studies as well) will be discovered; (ii) identifying high risk individuals; (iii) efficiently using data by dealing with missing data and predictors of mixed (ordinal, nominal, and nested) types appropriately. We will study pregnancy outcomes associated with perinatal death such as intrauterine growth retardation, small for gestational age, and preterm delivery, and determine the relationship between these outcomes and putative risk factors. Although the YPOS data base has been extensively analyzed using more traditional methods, the tree-based methods will provide a deeper understanding of risk factors, and will therefore impact on the development of plans for public health programs to prevent birth defects. The emphasis of this project will be on interactive effects among risk factors in connection to the outcome of interest (e.g., miscarriage or birthweight). The methods and software developed by this study will offer researchers the opportunity to perform more flexible, realistic, and efficient analyses in epidemiologic studies.
Showing the most recent 10 out of 11 publications
