Predicting Harm, Managing Risk: Analytics in High-Stakes Environments Custom Case Solution & Analysis

Evidence Brief: Allegheny Family Screening Tool Analysis

1. Financial Metrics

• Department Budget: Allegheny County Department of Human Services (DHS) manages approximately 1 billion dollars in annual spending across various social services.
• Cost of Failure: High-cost interventions (foster care) vs. low-cost preventative services. Foster care placements represent the highest per-child expenditure.
• Research Funding: Initial development of the Allegheny Family Screening Tool (AFST) was supported by multi-million dollar grants and internal budget reallocations.
• Efficiency Gains: The tool aims to reduce the 20 percent rate of screen-out errors where high-risk cases were dismissed by human call-screeners.

2. Operational Facts

• Process: Call-screeners receive roughly 15,000 allegations of child maltreatment annually.
• Scoring Mechanism: AFST generates a risk score from 1 to 20 based on administrative data including public assistance records, criminal justice history, and prior child welfare contact.
• Mandatory Action: Scores above a specific threshold (e.g., 15 or higher) originally triggered a mandatory investigation, though this was later adjusted to allow human override.
• Data Sources: The tool integrates data from 29 different government databases to create a unified view of the family.
• Decision Speed: The algorithm provides a score in real-time while the intake worker is on the phone or immediately following the call.

3. Stakeholder Positions

• Erin Dalton (Director, DHS): Advocate for data-driven decision making to minimize human bias and improve consistency in child protection.
• Rhema Vaithianathan (Lead Researcher): Focuses on the statistical validity of the model while acknowledging that the tool predicts system involvement rather than actual maltreatment.
• Intake Workers: Expressed concerns regarding the loss of professional autonomy and the tools inability to account for immediate, qualitative context during calls.
• Civil Rights Advocates: Argue that the tool penalizes poverty, as the data sources (SNAP, TANF) primarily track low-income families while wealthier families remain invisible to the algorithm.

4. Information Gaps

• Long-term Outcomes: Lack of randomized controlled trial data comparing tool-assisted decisions to human-only decisions over a 5-year period.
• Private Sector Comparison: No data on how wealthier families using private insurance and private therapy are tracked for similar risks.
• Algorithm Weighting: Specific weights of individual variables (e.g., the exact impact of a single criminal record vs. three years of SNAP usage) are not fully transparent to the public.

Strategic Analysis: Balancing Prediction and Equity

Core Strategic Question

• How can Allegheny County maintain the predictive accuracy of the AFST while mitigating the systemic bias inherent in poverty-indexed administrative data?
• How should the agency define the boundary between algorithmic recommendation and human accountability?

Structural Analysis

The core problem is a proxy variable trap. The AFST predicts future system involvement (re-referral or placement) rather than the latent condition of child harm. Because the data inputs are skewed toward public service users, the tool structurally over-samples poor families. This creates a feedback loop where poverty is treated as a risk factor for neglect, regardless of actual parental behavior. The bargaining power of the community is rising, and the social license to operate this tool depends on solving this equity gap.

Strategic Options

• Option 1: The Human-Centric Refinement. Transition the AFST from a decision-driver to a secondary validation tool. Remove mandatory screen-ins. Invest in intensive worker training to interpret scores as one data point among many.
  • Rationale: Restores agency to professionals and reduces the impact of data bias.
  • Trade-offs: Higher variance in decisions; potentially higher rates of screen-out errors.
• Option 2: Proactive Prevention Pivot. Use the high-risk scores to trigger voluntary support services (housing, food, childcare) instead of mandatory investigations.
  • Rationale: Addresses the root causes of neglect (poverty) without the trauma of family separation.
  • Trade-offs: Requires significant reallocation of budget from enforcement to services; families may still feel coerced.
• Option 3: Algorithmic Decoupling. Rebuild the model using a more limited set of variables that are less correlated with race and income, even if it reduces the overall predictive power (R-squared).
  • Rationale: Prioritizes equity and public trust over raw statistical accuracy.
  • Trade-offs: Lower accuracy may lead to missing children in genuine danger.

Preliminary Recommendation

Allegheny County should pursue Option 2. The current tool is a mirror of past systemic biases. By using the tool to identify families needing resources rather than families needing surveillance, the agency transforms the algorithm from a threat into a support mechanism. This preserves the predictive value of the data while aligning with the mission of human services.

Implementation Roadmap: Transition to Service-First Analytics

Critical Path

• Month 1-2: Audit current AFST variables to identify those with the highest correlation to poverty and lowest correlation to physical harm.
• Month 3: Reconfigure the workflow so that mid-range scores (10-15) automatically trigger a referral to a non-investigative community support team.
• Month 4-6: Launch a pilot in one geographic zone where the tool is used exclusively for resource allocation, not investigation mandates.
• Month 9: Compare safety outcomes in the pilot zone vs. traditional zones.

Key Constraints

• Budget Silos: Shifting funds from child protective services (CPS) to preventative community programs often faces legal and bureaucratic hurdles.
• Worker Skepticism: Staff must be convinced that not investigating a high-risk score is a valid clinical choice when resources are offered instead.
• Public Perception: A single high-profile tragedy in a pilot zone could lead to the immediate termination of the program.

Risk-Adjusted Implementation Strategy

The implementation must include an independent ethics board with the power to pause the tool if racial disparities in screen-ins do not decrease by 15 percent within the first year. Contingency plans must include a return to human-only screening if the integrated data system experiences a breach or significant downtime. Success depends on the community viewing the tool as a gateway to help, not a precursor to removal.

Executive Review and BLUF

BLUF

The Allegheny Family Screening Tool is a technical success but a strategic liability in its current form. It accurately predicts which families the system will target, but it fails to distinguish between the challenges of poverty and the intent of neglect. The county must pivot. The tool should no longer mandate investigations. Instead, it must be used to automate the delivery of preventative resources. If the agency continues to use poverty-proxied data to drive family separation, it will face a total loss of community trust and inevitable legal injunctions. The strategy must move from predictive policing of families to predictive support of families.

Dangerous Assumption

The most dangerous premise is that administrative data is a neutral reflection of reality. It is not. The data reflects the history of who the government watches. Assuming that more data leads to more truth ignores the fact that the data itself is a product of systemic bias.

Unaddressed Risks

• Model Drift: As social workers change their behavior in response to the tool, the data generated for future versions of the model becomes contaminated, leading to unpredictable shifts in scoring logic.
• Legal Liability: Using an algorithm to trigger a mandatory investigation may violate Due Process rights if the family cannot challenge the underlying logic of the score.

Unconsidered Alternative

The team has not fully considered a complete exit from predictive analytics for intake. A return to human-only screening, supported by intensive clinical supervision and a simplified checklist, might achieve similar safety outcomes with significantly lower ethical and legal overhead.

Verdict

APPROVED FOR LEADERSHIP REVIEW