Neural Network Model
Risk-based assessment/intervention strategies can be applied within treatment-focused Differential Response (DR) child welfare systems to prevent recurrent maltreatment by targeting service delivery towards modifiable risk factors that drive the likelihood that a family will be re-reported for maltreatment. That said, the risk assessment tools used in child welfare, to include DR systems, explain little variance in the likelihood of future maltreatment (~15%), have high rates of false positives (up to 29%) and false negatives (up to 45%), and typically contain few modifiable risk factors that can be altered via service delivery. Thus, a key first step in improving the effectiveness of DR service delivery rests upon the identification of a compact and powerful set of modifiable risk factors that help workers select the services most capable of decreasing each family’s risk of recurrent maltreatment. To this end, this study uses neural networks to answer the exploratory question “what best predicts recurrent maltreatment” for 8,943 Missouri families (with one child randomly drawn from each family) that were first reported for maltreatment in 1993 or 1994 and whose administrative records were collected across a series of public agencies to include their child welfare records through 2010. Neural network models will be used to identify the factors — with an emphasis on modifiable factors — that most accurately classify families into one of two possible repeat maltreatment groups: those re-reported for maltreatment and those not re-reported for maltreatment. The results from this study can be used to build a risk-informed model of preventive service delivery that fits well within a DR perspective.
Administration of Children and Families, DHHS, National Quality Improvement Center on Differential Response in Child Protective Services