An investigation of missing data methods for classification trees applied to binary response data

There are many different missing data methods used by classification tree algorithms, but few studies have been done comparing their appropri-ateness and performance. This paper provides both analytic and Monte Carlo evidence regarding the effectiveness of six popular missing data methods for classification trees. We show that in the context of classifi-cation trees, the relationship between the missingness and the dependent variable, rather than the standard missingness classification approach of Little and Rubin (2002) (missing completely at random (MCAR), miss-ing at random (MAR) and not missing at random (NMAR)), is the most helpful criterion to distinguish different missing data methods. We make recommendations as to the best method to use in various situations. The paper concludes with discussion of a real data set related to predicting bankruptcy of a firm