SUBGROUP IDENTIFICATION WITH MISSING DATA

Abstract

This thesis discusses subgroup identification with missing data. Various methods for subgroup identification are reviewed, including examples of global outcome modeling, global treatment effect modeling, and local modeling. The model and procedure of the two-stage multiple changepoint detection (TSMCD) are thoroughly reviewed. Moreover, TSMCD and model-based recursive partitioning (MOB) are investigated in detail by numerical experiments. Different missing data mechanisms, including MCAR and non-MCAR, and different samples, are considered. Criteria, including estimated number of subgroups, test MSE, true subgroup rate, and positive predictive values are used to evaluate and compare the capacity of these methods, while MICE is implemented priorly to impute missing data. Furthermore, several methods for missing data are reviewed, including complete case analysis, examples of imputation methods, and missing-indicator methods. Additionally, this thesis discusses other possibilities for handling subgroup identification with missing data, including the combination of GUIDE and TSMCD, and the generalization of MissInspect.