A Predictive Analysis of Conversion among Normal, MCI and Dementia Lead Investigator: Yan Zhou Institution : UCLA E-Mail : YanZhou@mednet.ucla.edu Proposal ID : 489 Proposal Description: Logistic regression and Cox survival model are most widely used methods for examining diagnosis conversion. These regression based models have several limitations when there are a large number of predictors. If no clear theory or hypothesis is available for testing only a few specific covariates, it seems impractical to include all main effects as well as higher-order interaction terms in the model. Also, these models are not sensitive to non-linear effects unless otherwise pre-specified. The SEARCH program can be informative in this situation, and we are particularly interested in identifying non-linear effects and interaction effects. After the idea of recursive partitioning first being introduced in Morgan and Sonquist (1963), the method was later generalized by Breimen and colleagues (1984), known as Classification and Regression Trees (CART). Today a number of programs have implemented the idea, although specific algorithms differing from one to another. It remains unclear how the SEARCH program performs compared to other recursive partitioning programs. This is also one of our research questions.