Abstract
Child welfare agencies are tasked with investigating allegations of child maltreatment and intervening when necessary. Researchers are turning to the field of predictive analytics to optimize data analysis and data-driven decision making. To demonstrate the utility of statistical algorithms that preceded the current predictive analytics, we used Model Based (MOB) recursive partitioning, a variant of regression analysis known as decision trees, on a dataset of cases and controls with a binary outcome of serious maltreatment (defined as hospitalization or death). We ran two models, one which split a robust set of variables significantly correlated with the outcome on the partitioning of a proxy variable for environmental poverty, and one which ran the same variable set partitioned on a variable representing confirmed prior maltreatment. Both models found that what most differentiated children was spending greater than 2% of the timeframe of interest in foster care, and that for some children, lack of Medicaid eligibility almost doubled or tripled the odds of serious maltreatment. We find that decision trees such as MOB can augment risk assessment tools and other data analyses, informing data-driven program and policy decision making. We caution that decision trees, as with any other predictive tool, must be evaluated for inherent biases that may be contained in the proxy variables and the results interpreted carefully. Predictive analytics, as a class, should be used to augment, but not replace, critical thinking in child welfare decision making.