Panel Paper: Implementation of Predictive Risk Modeling to Improve Child Welfare Call Screening Decisions

Thursday, November 8, 2018
Lincoln 3 - Exhibit Level (Marriott Wardman Park)

*Names in bold indicate Presenter

Erin Dalton and Kyle Jennison, Allegheny County, Department of Human Services

Every year there are more than 3.6 million referrals made to child protection agencies across the US. Recent estimates suggest that 37% of US children are investigated for alleged abuse or neglect at least once during childhood (Kim, 2017). This suggests that far from being a rare occurrence, many more children are being pulled into the child welfare (protection) agencies than previously realized. While increasing amounts of administrative data are available to help guide decision-making, it is difficult for child welfare workers to systematically weight and use historical information for all children and adults in a maltreatment allegation. The availability of data, coupled with heightened pressure to ensure that investigative resources are focused on the highest-risk children, have led to growing interest in the role of predictive analytics as a tool to help child welfare workers more consistently, quickly, and accurately assess each referral.
Predictive Risk Modelling (PRM) uses routinely collected administrative data to predict future adverse outcomes that might be prevented through a more strategic delivery of services. PRM has been used previously in health and hospital settings (Panattoni, Vaithianathan, Ashton, & Lewis, 2011; Billings, Blunt, Steventon, Georghiou, Lewis, & Bardsley, 2012) and has been suggested as a potentially useful tool that could be translated into child protection settings (Vaithianathan, Maloney, Putnam-Hornstein, & Jiang, 2013).
In August 2016, Allegheny County, Pennsylvania implemented a predictive risk model to support child welfare decision-making at the call screening stage. The resulting tool – the Allegheny Family Screening Tool – utilizes a logistic regression model to weight more than 100 predictive factors for each child on the referral. It produces a score based each child’s relative risk of placement within two years if the referral were screened in, and risk of re-referral within two years if the referral were screened out. The panelist will share early results from this implementation as well as next steps in the work.