Domain Atlas / Housing & homelessness services
Allegheny Housing Assessment
A peer-reviewed 2024 evaluation of the Allegheny Housing Assessment found that although the tool was substantially more accurate than the VI-SPDAT survey it replaced and produced similar risk-score distributions across race, it did not reduce the racial disparity in service rates: white single adults were served at about 23.3% versus 19.5% for Black clients.[2]
What happened
In August 2020 Allegheny County's Department of Human Services replaced the interview-based VI-SPDAT survey in its HUD-required coordinated-entry system with the Allegheny Housing Assessment (AHA), a predictive-risk decision-support tool that scores people experiencing homelessness to prioritize admission to scarce housing — bridge, rapid rehousing, and permanent supportive housing. Developed by the Centre for Social Data Analytics at Auckland University of Technology (the county methodology paper is authored by Rhema Vaithianathan and C.I. Kithulgoda), AHA pulls linked administrative data from the county's integrated Data Warehouse — demographics (age and gender, but explicitly not race), and records from birth, child protective services, homeless services, criminal justice, Medicaid, and assisted housing, plus neighborhood data — to estimate the 12-month likelihood of adverse events if a person remains unhoused. The original model combined three outcome models (a jail booking, a Medicaid-funded behavioral-health inpatient stay, and four-or-more emergency-department visits); an August 2025 update, built by an internal DHS team, added a fourth outcome predicting another homelessness episode, which raised the homelessness-prediction AUC from 0.54 to 0.68. The binned probabilities are summed into a 1-to-10 risk score. Clients cannot opt out; DHS uses AHA wherever sufficient data exist, and for the roughly 5% of cases lacking at least 90 days of warehouse history a self-report "Alt-AHA" questionnaire produces a comparable score, with the higher of the two used. The score is not an automated decision: it ranks clients on the priority list, Allegheny Link staff may use case conferencing to override, and a homeless resource coordinator makes the final referral.
AHA is substantially more accurate than the survey it replaced — for the 2025 model, per-outcome AUC runs roughly 0.66 to 0.72, versus roughly 0.52 to 0.59 for the VI-SPDAT. But accuracy was not the whole story. A peer-reviewed 2024 evaluation (Cheng, Drayton, Chouldechova, and Vaithianathan, published at the AAAI/ACM AIES conference and co-authored by AHA's own developer) analyzed 6,542 single-adult assessments from 2018 to 2022 and found that although AHA produced similar score distributions across race, it did not reduce the racial disparity in service rates: white single adults were served at about 23.3% versus 19.5% for Black clients, a gap that persisted from the pre-AHA rates of 17.6% and 14.5%. The study attributed part of the persistence to a racially skewed reliance on the self-report fallback — white single adults were assessed via Alt-AHA about twice as often as Black clients (24.5% versus 11.8%) — and to eligibility factors that narrow coordinated-entry discretion. The 2025 update also shifted allocation by gender: the male share of assigned housing rose from 62% to 76% (and the female share fell from 34% to 24%) because men showed a higher measured one-year homelessness risk, a clean example of an outcome-selection choice reshaping who receives scarce housing. The county documents layered oversight — a county-commissioned external algorithmic impact assessment by Eticas Research and Consulting (2020, which flagged slight under-protection of women, a race-related concern on the jail outcome, and the ethics of rationing scarce help by risk score), a Continuum of Care governance board and HUD requirements, extensive stakeholder engagement, and, by 2025, daily drift and distribution monitoring with automated gating that blocks anomalous scores from release without human validation. The best-known civil-liberties and scholarly critiques of Allegheny County DHS predictive analytics — the ACLU's 2021 reporting and Virginia Eubanks's "digital poorhouse" work — target the sibling Allegheny Family Screening Tool in child welfare, not AHA directly, but their concerns about warehouse-based, poverty-profiling scoring are extended by critics to the approach AHA shares.
The sociotechnical reading
The Atlas's clearest lesson about the human override lives in the sibling AFST case: overrides were where equity was won. AHA is the counter-lesson. Here the score was made more accurate AND near-parity across race, a human override channel existed, and still the racial gap in who actually got scarce housing did not close — because in a rationing decision the disparity is produced downstream of the score, not inside it. Three structural features locate it. First, the self-report Alt-AHA fallback: routing the low-data cases back to self-report reintroduces exactly the disclosure and access bias the warehouse model was meant to remove, and it is used unevenly by race, so the "fix" for missing data becomes a channel for the disparity. Second, the eligibility machinery: coordinated-entry discretion is documented as limited by strict eligibility rules, so the case-conferencing override that carried the equity effect in the AFST shape has far less room to move a rank here. Third, and most quietly, outcome selection: when the 2025 update chose to add a future-homelessness prediction, the allocation shifted sharply toward men — a reminder that WHAT a score is built to predict is itself an allocation decision, made once, upstream of every rank it produces.
That reframes the governance question the map asks. It is not "is the score accurate?" or even "is the score fair across groups?" — AHA answers both better than its predecessor — but "what does the allocation machinery around the score do?" The productive levers sit there: risk-tiering that respects real scarcity (the county assesses far more households than it can serve), data-minimization on the permanent warehouse features that a person cannot escape, and the audit-and-monitoring rhythm (external impact assessment, independent evaluation on shared data, daily drift gating) that is this deployment's genuine strength. And the memory loop is the Lab's contamination stressor made literal: the models are retrained on cohorts whose outcomes are read from the same warehouse the score reads, and housing is protective of the very harms predicted, so today's allocations quietly become tomorrow's training data. The honest boundary is that none of this measures harm to the unhoused people being ranked; the service-rate disparity, the fallback skew, and the gender shift are documented outside any diagram like this one, and what a rank decides here is a one-time allocation at a scarce gate rather than an error that propagates.
The concepts used in this reading are defined in the Field Guide; the governance responses live in the Practice Library. The model organization for this case can be stress-tested in the PAN Lab.