Get Full Government Meeting Transcripts, Videos, & Alerts Forever!
Get email alerts on the Predictive Risk Modeling topic
No spam. Unsubscribe anytime.
Outside expert urges predictive risk models, warns SDM checklists underperform
Summary
UNC researcher Emily Putnam Ornstein briefed the committee on predictive risk modeling as an alternative to checkbox SDM tools, citing Allegheny and Los Angeles County case studies that used models to flag high‑risk investigations for enhanced supervisory support and reported measurable safety improvements in targeted groups.
Emily Putnam Ornstein, a professor in social work who has advised multiple child welfare agencies, told the committee that many current intake and risk‑assessment tools in use are dated operator‑driven checklists that place heavy documentation burdens on workers and can perform poorly in predicting future harm. She argued automated predictive models that pull historic case management data and produce risk stratification for supervisors can be more accurate and act as a workforce intervention.
Putnam Ornstein reviewed two implementation patterns: (1) models that run at intake to assist screening decisions (Allegheny County prototype), and (2) models that run after a referral is accepted and then notify supervisors about newly assigned investigations that appear unusually high risk (Los Angeles County). She described LA's deployment as a daily, automated feed from the case management system (SACWIS/CCWIS) that identified a top‑risk slice (roughly the top 10%) for "enhanced support," prompting extra supervisory teaming, collateral checks and service prioritization.
She said results from a limited pilot in Los Angeles showed a reduction in near‑fatalities and fatalities for the highest‑risk investigations that received enhanced supervisory attention. Putnam Ornstein said mechanisms included slight increases in placements for the highest‑risk group and measurable increases in services delivered to families flagged for enhanced support. She cautioned that predictive models are not a panacea and urged transparency about inputs, public oversight and community engagement to address questions about bias and privacy.
On deployment logistics, Putnam Ornstein said predictive overlays can work with legacy systems by crosswalking available data and that launching a local model can take months, not years, with development costs she has seen in the range of $250,000–$500,000 and modest ongoing maintenance fees. She recommended that when New Mexico's CCWIS comes online the model be revalidated and, if necessary, retrained on the new data elements.
Committee members pressed on data readiness, CCWIS integration, machine learning transparency and whether school attendance or other external records were incorporated; Putnam Ornstein said jurisdictions vary in available features but models have been developed even for smaller counties and that community decisions should guide which external data sources are used. She urged the committee to weigh quicker, incremental deployments to provide supervisors with better information rather than waiting for a full CCWIS build alone.
The hearing ended with requests for CYFD and staff to share the functional design of the CCWIS build and for follow‑up conversations on costs and pilot options for predictive analytics.
