Investigating the Performance of cart- and Random Forest- Based Procedures for Dealing with Longitudinal Dropout in Small Sample Designs under mnar Missing Data

This chapter describes a simulation designed to assess whether Classification and Regression Trees (CART)-based weights and CART-based multiple imputations can provide relief in small sample studies with outcome-dependent missing data. It provides brief conceptual overview of the aspects of these techniques most relevant to addressing missing data. The chapter also describes how CART and random forests can be utilized to address missing data. The goal of a CART analysis is to use the values of a set of observed predictors to split the dataset into homogeneous subgroups with respect to a single outcome variable. The intuitive appeal of CART-based methods in handling missing data rests on the uncomfortable reality that, in practice, the true causal model predicting missing data is unknown. The chapter investigates the performance of CART and random forest methods for dealing with missing data under small sample-sizes and a variety of Missing Not at Random missing data mechanisms.