Evaluating classification and regression trees with clinical trial data
Gregoski, Mathew Jon
MetadataShow full item record
An examination of classification and regression tree (CART) models among African American (AA) adolescents at risk for Essential Hypertension (EH) was conducted. The purpose of Study I was to compare multiple CART model rule creation and cross-validation techniques with each other using intervention data and validate the results with hierarchical regression models. The analyses utilized a data sample obtained from a randomized clinical trial of 181 AA adolescents considered to be at risk for EH. CART models were created using the Gini, Entropy, Class Probability, and Two-ing selection methods combined with the fraction of random cases, and V-fold cross-validation techniques The CART models examined behavioral stress interventions and the influence of underlying anthropometric, psychosocial, and behavioral variable and their impact on resting systolic blood pressure (SBP), diastolic blood pressure (DBP), and heart rate (HR). The findings imply that CART models using the “Gini or Entropy” selection methods combined with V-fold cross-validations were the best methods for use in clinical trial research. The results of the CART models agreed with previous regression analyses and in some circumstances provided additional information not captured by the regression models. Study II utilized the same 181 participants in Study I and examined the same baseline anthropometric, psychosocial, and behavioral characteristics and treatment group effects. In addition, baseline characteristics, changes in anthropometric, psychosocial, and behavioral characteristics that occurred during the intervention period were also examined for the purpose of determining what treatments and characteristic lead to improved cardiovascular function in the natural environment as measured by 24 hour ambulatory SBP, DBP, and HR. Based on study I CART models using “Gini” and “Entropy” selection methods with V-fold cross-validation were constructed for ambulatory SBP, DBP, and HR. Hierarchical regression models were created that included variables and values based on the rules obtained from the CART analyses. Across all regression models significant effects were found for the subgroups formed from CART outputs. The studies show CART models created with “Gini or Entropy” selection methods combined with V-fold cross-validation are a useful method for maximizing clinical trial success rates at the individual level.