Improving Prioritization of Housing Services: Implementation of the Allegheny Housing Assessment (AHA) and the Mental Health Allegheny Housing Assessment (MH-AHA)

Agency Deploying Data-Driven Model: Allegheny County DHS
Scope: Provide decision support to inform prioritization of placement in Permanent Supportive Housing (PSH) and Rapid Rehousing (RRH)
Status: In Use
Date of Deployment: August 2020
Last Updated: August 2025

Consult the Frequently Asked Questions About AHA↗ for additional information.

The AHA is a decision support tool that helps DHS prioritize admissions to Rapid Rehousing (RRH) and Permanent Supportive Housing (PSH). The model assists in prioritization by identifying individuals who are at risk of a set of adverse outcomes related to homelessness.

The AHA is comprised of four component models that each predict a different outcome: inpatient mental health stays, four or more emergency department visits, jail bookings, and homelessness within a year of the assessment date. The first three outcomes were used to develop the November 2020 version of the model, while the homelessness component was added to the 2025 update.

The training data for the updated AHA model is comprised of 18,008 RRH/PSH assessments from 2016 through 2024. The component models are random forests tuned using subject-wise cross-validation to estimate the likelihood of their respective outcomes. These probabilities are weighted and combined to generate a 1-10 integer score, where 10 represents highest risk and 1 represents lowest risk.

Read the most recent AHA methodology update here.
Read the full AHA methodology report here.

While the AUC and Positive Predictive Value performance for the first three outcomes are similar in the 2020 and 2025 versions, inclusion of the homelessness component model in the 2025 update significantly improves AHA’s identification of homelessness risk, from an AUC of 0.54 to 0.68.

In addition to pre-deployment performance evaluation, the scores that are generated daily are monitored on an ongoing basis for both drift and unusual distribution.

Read the most recent AHA methodology update here.
Read the full AHA methodology report here.

The August 2020 AHA model was subject to a thorough fairness and equity review by Eticas, which found that there were few concerns regarding the model’s fairness across various groups. This represented an improvement in the racial equity of housing services in the County. In the process of updating AHA, we ensured that resources would continue to be allocated equitably.

The table below compares the gender and race breakdown of AHA 10s between the 2020 and 2025 models. Allocations across racial groups appear similar in the updated model, but there is a shift in allocations between genders. In the new model, more men and fewer women would be assigned housing for both singles and families. This shift is likely due to the large difference in homelessness risk between men (38% one-year homelessness risk) and women (26%). Because the updated model includes the homelessness component, it prioritizes more men for housing than the earlier version.

DHS will continue to advance equity and transparency in our predictive risk models/data-driven models through external audits and release of statements to the public. To learn about what DHS has already done to support equity and transparency, see Etica’s algorithmic impact assessment and DHS’s response to the algorithmic audit.

There has been rich public engagement throughout the development of AHA, including involvement from policy makers, community groups, industry experts and the general public. Specifically, individuals experiencing homelessness, local service provider agencies, national homelessness experts, advisory boards and committees, DHS leadership, local funding agencies and foundations, and representatives from the U.S. Department of Housing and Urban Development (HUD) have all contributed to AHA’s development. Formats for engagement have included focus groups and presentations, where stakeholders could express optimism or share concerns about the model.

1. Community feedback surfaced concerns about the accuracy and completeness of administrative data on the homelessness population. This feedback resulted in the implementation of quality assurance protocols and alternative self-report predictive assessment tools, improvements that aim to support the assessment process but not replace clinical judgement for decision-making.
2. National Feedback sourced from conferences and conventions provided DHS with cross-sector expert input, including professional feedback from leading experts in fields covering data science, analytics and ethics. The insights drawn from these specialists support DHS’s mission to use predictive risk models ethically and responsibly, for the purpose of improving human service systems’ responses to evolving community needs.

DHS will continue to promote and engage in two-way communication that both centers community voices and informs staff and partners—whether it be through focus groups with people experiencing homelessness or discussions with providers. To learn about what DHS has already done, see the Focus Group Report and the Overview of Models and Implementation video.

Agency Deploying Data-Driven Model: Allegheny County DHS
Scope: Identify individuals at high risk of negative outcomes to prioritize access to mental health residential programs
Status: In Use
Date of Deployment: February 2023
Last Updated: August 2025

Consult the Frequently Asked Questions About MH-AHA↗ for additional information.

In February 2023, DHS implemented the Mental Health – Allegheny Housing Assessment (MH-AHA), developed by a team led by Dr. Rhema Vaithianathan at the Centre for Social Data Analytics. The MH-AHA is a decision support tool that helps DHS prioritize admissions to mental health residential programs. The model assists in prioritization by identifying individuals who are at risk of a set of adverse outcomes related to mental and physical health.

The training data for the MH-AHA model is comprised of 13,673 RRH/PSH assessments from 2016 through 2024 for people who were enrolled in Medicaid at the time of their assessment. The model is comprised of two component models; each predicts a different outcome: 1) inpatient mental health stays and 2) four or more emergency department visits within one year of assessment. The component models use DHS warehouse data to generate probabilities of their respective outcomes. These probabilities are combined to generate a 1-10 integer score, with 10 indicating highest risk and 1 indicating lowest risk. 

Read the full MH-AHA methodology report here.

The updated MH-AHA model was evaluated along several dimensions of performance and fairness, both in absolute terms and by comparison to the February 2023 version of the model. Model performance for the two outcomes is shown in the table below:

The area under the curve (AUC) and Positive Predictive Value (PPV) performance improved from the 2023 model to the 2025 update for both outcomes. AUCs improved from 0.64 and 0.75 to 0.73 and 0.78 for Mental Health Inpatient and Emergency Room 4+ Visits, respectively, and the PPV improved by 7 to 8 percentage points for each outcome.

In addition to pre-deployment performance evaluation, we monitor the scores that are generated, on a daily basis, for drift and unusual distribution.

Read the full MH-AHA methodology report here.

While different subsets of the training set can present different levels of risk (meaning that the risk for certain subgroups can be under- or over-estimated), it is important that the model selects a treatment group that is equally at-risk across subgroups. For this reason, we performed a full equity analysis looking at the precision and recall across the sensitive categories of race and gender.

We also examined how the model update might change the distribution of resource allocations. The table below compares the gender and race breakdown of MH-AHA 9s and 10s between the 2023 and 2025 models. Allocations across racial and gender groups are nearly the same in the updated model, which indicates that there has not been a shift in equity.

DHS will continue to advance equity and transparency around use of predictive risk models/data-driven models through external audits and by releasing statements to the public. To learn about what DHS has already done to support equity and transparency, see Etica’s algorithmic impact assessment and DHS’s response to the algorithmic audit.

There has been broad public and professional engagement in the development of MH-AHA:

1. Feedback provided from AHA’s engagement process directly informed the development of MH-AHA. The cross-sector input gathered for AHA from experts in multiple fields established the foundation that DHS carried forward to implement MH-AHA. Experts included researchers, machine learning experts, homelessness policy and program administration experts, cyber law experts, ethicists, privacy experts and representatives from HUD. To learn more about the community’s involvement in AHA (and thereby MH-AHA), please review the section on Community Input in AHA Adoption.
2. MH-AHA also included input from providers, DHS staff and local judges, who recognize that the availability—or lack thereof—of residential mental health housing is an important consideration when making crucial service and/or judicial decisions. 

These elements of MH-AHA’s development ensure both system-level and community-based considerations shape MH-AHA’s design and use.

DHS will continue to promote and engage in two-way communication that both centers community voices and informs staff and partners—whether it be through focus groups with people experiencing homelessness or discussions with providers. To learn about what DHS has already done, see the Focus Group Report and the Overview of Models and Implementation video.