Audit for equitable homelessness risk decisions
How we evaluated the accuracy and fairness of Allegheny County’s homelessness risk prediction system to help strengthen monitoring and staff training.
Credibility chips: Public sector • USA • AIA
The challenge
The Allegheny Housing Assessment (AHA) tool is used to prioritise support for residents at risk of homelessness.
The system was built on a complex risk model whose effectiveness—both in terms of accuracy and potential bias—was unclear. Without strong evidence and governance, there was a risk that some groups could be disadvantaged and that staff might lack clear guidance on when and how to use the tool.
What we did
Eticas conducted an algorithmic impact assessment of the county’s homelessness risk model, which draws on 964 variables from departments and regulatory bodies serving the county. These include data on demographics, mental health, justice involvement, incarceration, hospital visits, and substance use.
The audit focused on assessing the model’s accuracy and evaluating various metrics to measure bias affecting protected and sociodemographic groups such as age, race, gender, disability, and veteran status.
We evaluated the system’s overall performance and potential group-level disparities across protected cohorts.
Finally, Eticas reviewed the system’s governance—how it was assessed, maintained, and implemented by users.
Audit findings
Accuracy: The system was found to be accurate overall but under-represented certain cohorts, particularly minors (ages 0–17). The root cause was traced to the original training data, which contained very few examples from this group.
Bias: No significant disadvantage was detected for disability, age, race, or property-based factors. While there were no clear indicators of discrimination based on gender or veteran status, some inconsistencies were detected, suggesting these factors should be closely monitored.
System governance: Eticas found that there was no formal, ongoing review of the system’s performance. This gap could allow accuracy and bias issues to emerge over time as homelessness demographics evolve. In addition, both tool users and affected citizens had limited awareness and confidence in the system’s operation and impact.
Recommendations
Eticas proposed four actions:
- Revise system scope:
Handle minors (ages 0–17) through a separate business rule rather than through the model itself. - Re-model regularly:
Retrain the model at least annually, with particular attention to minority cohorts such as single veterans and women, where inconsistencies were noted.
Eticas also recommended thorough documentation of the training dataset and modelling process using model cards, and making this documentation public. Researchers should be granted access to anonymised microdata, under appropriate data-sharing agreements, to independently examine the models. - Improve explainability:
Develop a clear communication strategy for stakeholders about the tool’s purpose, development, implementation, and impact. - Build trust and confidence:
To improve acceptance among frontline workers and clients, Eticas recommended that the Department of Human Services (DHS) conduct formal training for all staff using the AHA tool.
The result
Allegheny County fully embraced Eticas’s recommendations, committing to:
- Closely monitor model performance for women and single veterans, where potential disadvantage was identified.
- Conduct regular re-modelling of the Predictive System for Risk of Homelessness.
- Implement ongoing quality assurance and monitoring.
- Develop a clear communication strategy with stakeholders before, during, and after AHA implementation.
- Provide multiple training sessions for all frontline workers involved in the homelessness coordinated entry system.