LAUSR

Abstract Outcome data are routinely collected in healthcare practices and used for quality of care assessment and improvement. Logistic regression trees are a popular method for subgroup identification for binary outcome data. Outliers often exist in healthcare data, and many studies have addressed this problem with respect to model fitting in logistic regression. However, outlier problems are more complex in the context of tree models, as they involve subgroup identification in addition to model fitting. This study considers the outlier problem in logistic regression tree modeling of outcome data. It reveals the effects of outliers on split variable selection in identifying subgroups and proposes a method to construct logistic regression trees that are robust to outliers. The effectiveness of the proposed method and its advantages over alternatives are demonstrated in a simulation study and case studies.