Identification and Classification of Patterns of Change in Longitudinal Data

摘要

This study proposes several statistical measures for discriminating different patterns of change in longitudinal data. The main statistical methodologies employed in this research are: 1) Developing Statistical Measures, 2) Principal Component Analysis (Factor Analysis) to select a subset of non-redundant measures, and 3) Cluster Analysis based on selected measures, to identify subgroups of individuals with the similar longitudinal trajectory.;This approach is applied to the data set coming from a prospective longitudinal study evaluating bone mineral density (BMD) and body composition changes in women, aged 18-35 years old, using depot medroxy-progesterone acetate (DMPA) for contraception. Participants were recruited from local family practice, family planning, gynecology, and women's health clinics and through announcement in newspapers, workplaces, and university residence halls. The data is obtained from University of Iowa in Iowa City.;The proposed measures discriminate between trajectories that are increasing, decreasing, stable over time, or unstable, and detect those with abrupt changes or short-term ?uctuations. Once groups with similar trajectories are identified, the differences in patients' characteristics or outcomes are investigated. Applying the proposed method to evaluate the longitudinal changes in leptin in our data set allowed us to identify five different patterns of change and showed that the trajectories for the cases and controls were not the same. The proposed method provides efficient and easily implemented tools for describing the different patterns of change that can emerge from longitudinal studies. SAS program was used for developing statistical measures, data simulations, and application of all the aforementioned statistical procedures.