Using classification and regression trees (CART) to support worker decision making

Abstract

Several approaches can be taken to predict case membership in classes of a dependent variable. Classification and regression trees (CART) analysis has been cited repeatedly as a powerful nonparametric approach in fields where classification or prediction are of concern. To test CART's utility in a social work setting, authors conducted a secondary analysis of data collected in a national study of child protective screening practices to identify factors involved with worker decisions to investigate child maltreatment reports. The CART analysis revealed complex interaction effects previously unobserved in logistic regression. Comparisons of CART with traditional statistical approaches and other tree-based programs are presented. Key words: classification and regression trees; decision trees; decision making; screening; child protective ********** Depending on research question, basic purpose of a classification study is either to produce an accurate classifier or to uncover predictive structure of phenomenon under consideration (Breiman, Friedman, Olshen, & Stone, 1984). For most social work professionals, both objectives are of interest; for example, to target resources, a program planner must be able to identify groups of clients that are likely to benefit from a specific approach and to understand factors that predict likelihood of success given client's presenting conditions. Similarly, when a social worker recommends care alternatives, prediction of outcome given client's condition, available resources, and factors expected to influence rehabilitation are necessary to appropriately assist client and family in their decision making. Yet, many social work professionals are faced with complex decision problems without benefit of a set of rules to organize data. In these situations, most decision makers tend to polarize around only a few variables, potentially missing important aspects of a problem. Although a variety of traditional statistical approaches can be used to predict classification of cases from complex data sets, classification and regression trees (CART) analysis (Breiman et al., 1984) has been cited repeatedly as a powerful nonparametric approach in applied fields where classification or prediction are of concern, such as medicine (for example, Goldman et al., 1998; Mair, Smidt, Lechleitner, Dienstl, & Puschendorf, 1998; Thomssen et al., 1998) and mental health (Barnes, Welte, & Dintcheff, 1991; Boerstler & de Figueiredo, 1991; Craig, Siegel, Hopper, Lin, & Sartorius, 1997). For example, in a study of low-income psychiatric patients, Boerstler and de Figueiredo found client's discharge from inpatient treatment at most recent admission to psychiatric treatment to be the most consistent, most powerful, and only necessary predictor of high use of outpatient psychiatric services (p. 32); an important implication for program administrators. Mair et al. (1995) used CART to develop an algorithm for use in emergency room settings for early diagnosis of heart attack based on clinical symptoms, ECG, and other myocardial measures from 114 patients. The method's ability to predict a diagnosis was as high as that of other statistical methods; however, CART's graphical features, essential for use in clinical training and practice, were cited as a primary advantage over other methods. To demonstrate CART's potential for use in social work settings, this article presents CART technique, its utility in identifying factors involved with decisions to investigate reports of child maitreatment, and comparisons of CART with traditional statistical approaches and other tree-based software programs. BACKGROUND In response to growing discrepancy between number of reports made to child protective (CPS) and number of reports investigated, Children's Bureau funded an on-site study of CPS screening practices in 12 communities from five states to illustrate worker decision-making practices at intake (Wells, Fluke, Downing, & Brown, 1989). …