Christopher Wade Flaherty

TL;DR: In this paper, the authors explored the potential of an artificial neural network to improve prediction of recurrences of child physical abuse using the U.S. Air Force's central registry of child abuse reports.

Abstract: All 50 states have passed some form of mandatory reporting law to qualify for funding under the Child Abuse Prevention and Treatment act of 1974 (P. L. 93-247). Consequently, child protective service (CPS) agencies have experienced a dramatic increase in reports of abuse and neglect without corresponding increases in funding over the past several years. In response, many CPS agencies have turned to formal risk assessment systems to aid caseworker in malting various decisions. Various methodological obstacles have impeded efforts to predict child abuse. The present study explored the potential of an artificial neural network to improve prediction of recurrences of child physical abuse. Conducted on electronic data file compiled by the U.S. Air Force's central registry of child abuse reports, selected variables pertaining to all child physical abuse reports received from 1990–2000 (N = 5612) were examined. Thirteen predictor variables and five interaction terms were identified for analysis. It was hypothesized that each of the thirteen predictor variables and live interaction terms would be correlated with abuse recurrence when controlling for all other variables in the model. Using binary logistic regression (BLR) to analyze data, only four of the main effect variables and one interaction term were correlated with abuse recurrence. It was also hypothesized that an artificial neural network model would predict abuse recurrences better than an alternative method (BLR) due to superior ability to model complex interactions and curvilinear relationships among the selected variables. The hypothesis was not confirmed. Although both methods predicted recurrences significantly better than chance, the BLR model produced predictions that were slightly, but significantly better than the ANN Model. The BLR was also advantageous in terms of providing more information regarding contributions of individual predictors. It was concluded that both BLR and ANNs offer powerful tools to be used in future efforts to build abuse prediction models. When applied to the present data, BLR was more useful.