Date of Award
Doctor of Philosophy (PhD)
Chair and Committee
Risk assessment in child welfare has a long tradition of being based on models that assume the likelihood of recurrent maltreatment is a linear function of its various predictors: Gambrill & Shlonsky, 2000). Despite repeated testing of many child, parent, family, maltreatment incident, and service delivery variables, no consistent set of findings have emerged to describe the set of risk and protective factors that best account for increases and decreases in the likelihood of recurrent maltreatment. Shifts in predictors' statistical significance, strength, and direction of effects coupled with evidence of risk assessment models' poor predictive accuracy have led to questions regarding the fit between assumptions of linearity and the true relationship between the likelihood of recurrent maltreatment and its predictors: Gambrill & Shlonsky, 2000, 2001; Knoke & Trocmé, 2005). Hence, this dissertation study uses a distinctly nonlinear approach to modeling the likelihood of recurrent maltreatment by employing a combination of random forest and neural network models to identify the predictors that best explain the risk of recurrent maltreatment.
The risk of recurrent maltreatment was assessed for a cohort of children living in a large Midwestern metropolitan area who were first reported for maltreatment between January 1, 1993 and January 1, 2002. Administrative child welfare records for 6,747 children were merged with administrative records from income maintenance, mental health, special education, juvenile justice, and criminal justice systems in order to identify the effects that various public sector service system contacts have on the risk of recurrent maltreatment. Each child was followed for a period of at least seven years to identify the risk of recurrent maltreatment in relationship to a second report for maltreatment.
Post-hoc analyses comparing the predictive validity of the neural network model and a binary logistic regression model with random intercepts shows that the neural network model was superior in its predictive validity with an area under the ROC curve of 0.7825 in comparison with an area under the ROC curve of 0.7552 for the logistic regression model. Additional post-hoc analyses provided empirical insight into the four prominent risk factors and four risk moderating service variables that best explain variation in the risk of recurrent maltreatment. Specifically, the number of income maintenance spells received, community-level poverty, the child's age at the first maltreatment report, and the parent's status as the perpetrator of the first maltreatment incident defined 21 risk-based groups where the average probability of recurrent maltreatment was dependent upon values for the four primary risk factors, and the risk of maltreatment was moderated by juvenile court involvement, special education eligibility, receipt of CPS family centered services, and the child's receipt of a mental health/substance abuse service in the community. Findings are discussed within a Risk-Need-Responsivity theory of service delivery: Andrews & Bonta, 2006), which links the empiricism of risk assessment with the clinical implementation of a preventive service delivery plan through the identified modifiable risk factors that drive the likelihood of recurrent maltreatment.
Jolley, Jennifer Marie, "Applying Neural Network Models to Predict Recurrent Maltreatment in Child Welfare Cases with Static and Dynamic Risk Factors" (2012). All Theses and Dissertations (ETDs). 1009.