In response to NIH announcement PA-98-03 1, we will develop and apply tree- and spline-based statistical methods for the analysis of irregularly structured correlated data on substance use. These statistical methods usually produce data-driven models, but we will study ways to incorporate them with existing theory in medical, behavioral, and social sciences. This new theory-embedded data-driven methodology will enhance our understanding of health-related problems. Specifically, we will address issues of: (a) estimating covariance structures for analyzing irregularly spaced longitudinal data and bridging adaptive splines models with classic effects models. (b) Exploring tree-based methods for discrete longitudinal/genetic data analyses; and, (c) developing frailty models and their combination with tree based technique for segregation analysis, particularly useful for genetic epidemiological studies of psychiatric disorders. The secondary goal of this project is to enrich existing statistical software from the previous work of the Principal Investigator. This free software has already been downloaded by many researchers. The emphasis of this second component is: (a) to simplify the user interface for data input and analysis output; (b) to improve the stability of the programs by performing thorough error-free """""""" checking; and (c) to increase the portability of the software for different platforms. This software will provide a much needed data-mining tool for modeling complicated and correlated (e.g., longitudinal) data in medical, behavioral, and social sciences. In addition, we will apply the statistical methods to important data sets related to substance use and address a variety of problems including: (a) determining the effect of cocaine use by pregnant women on infant growth; (b) understanding the risk factors and the genetic epidemiology of substance use disorders, the impact of daily drug use on later employment, and the connection between drug use and income; (c) determining under what conditions our methods can lead to a deeper understanding of behavioral problems than more traditional analyses.
Showing the most recent 10 out of 28 publications
