Domain Atlas / Child welfare & family services
The map, not the score: place-based risk terrain and the records it concentrates
In Richmond, Virginia, Predict-Align-Prevent's open-source place-based model (built with Urban Spatial; report byline Ken Steif, Matthew D. Harris, and Sydney Goldstein, 2019) reported its highest risk tier capturing about 70% of held-out child-maltreatment events against about 35% for a kernel-density baseline, while covering about 10% of city land area holding roughly 48,500 residents including about 8,200 children — builder-generated figures from a commissioned report — and the builders' own published fairness audit stated verbatim that the meta-model 'generalizes well across neighborhoods of varying poverty rates, but does not generalize well across neighborhoods of varying race,' despite race and income being excluded from the feature set.[2]
What happened
Predict-Align-Prevent (PAP) is a Texas 501(c)(3) founded by Dr. Dyann Daley, formerly of the Center for Prevention of Child Maltreatment at Cook Children's Health Care System in Fort Worth. Its proposition inverts the child-welfare cases this Atlas already holds: instead of scoring families at intake, it builds risk terrain models over a "fishnet" grid — historical geocoded accepted or substantiated CPS cases per small cell, regressed on 200+ engineered environmental exposure features (crime incidents, blighted buildings, nuisance land uses like bars and laundromats, and protective assets like daycares and churches). The output is a five-tier risk map plus an "Align"-phase gap analysis comparing protective-resource supply against modeled risk (a protective score of 10 against a risk score of 90 reads as a −80 misalignment). Models score places, never households or individuals; PAP frames this as its answer to the bias critiques of person-level predictive risk scoring, and the predictor code and variables are open source. No intake screener, investigator, or caseworker ever receives a score about a family — frontline CPS decision-making is untouched by design.
The published record runs through three engagements of increasing operational depth. The 2016 Fort Worth study (Daley et al., Child Abuse & Neglect), trained on 2013 state child-welfare substantiations and Fort Worth police data, reported that in 2014 the top 10% of cells captured 52% of maltreatment cases against 43% for a conventional hotspot model — figures independently corroborated verbatim by the Marchment and Gill (2021) Crime Science systematic review, which also found Daley's study to be the sole child-maltreatment application of risk terrain modeling in the reviewed literature. Fort Worth-era journalism (KERA News, 2016) relayed further builder claims — children in areas concentrating poverty, domestic violence, and assault about 150 times more likely to experience maltreatment; newer models locating over 80% of cases in about 20% of Fort Worth — builder claims relayed by press, never peer-reviewed. Fort Worth remained a study plus local advocacy: no evidence shows Fort Worth CPS operationally allocating resources by the maps. The 2019 Richmond, Virginia engagement, built with the consultancy Urban Spatial (the report's byline: Ken Steif, Matthew D. Harris, and Sydney Goldstein), was an open-source deployment study on 6,500 accepted CPS cases over four years and a 1,910-cell fishnet, using an ensemble meta-model validated by leave-one-group-out cross-validation across 149 neighborhoods. Its headline: the highest of five risk tiers captured about 70% of held-out maltreatment events, against about 35% for a kernel-density baseline, while covering about 10% of city land area — home to roughly 48,500 residents, including about 8,200 children (21% of Richmond's children). The same report published its own fairness audit, with an adverse finding stated verbatim: the meta-model "generalizes well across neighborhoods of varying poverty rates, but does not generalize well across neighborhoods of varying race" — despite race and income being excluded from the feature set. The environmental proxies carry it. No external body is documented acting on that finding.
New Hampshire is where maps demonstrably moved resources. In September 2018 the state won a federal Children's Bureau Community Collaborations cooperative agreement; PAP's geospatial analysis (funded by Casey Family Programs, with state maternal and child health (MCH) staff and the child-welfare division accessing stewarded databases) mapped Manchester, the Winnipesaukee region, and Coos County — selected partly via the CDC Social Vulnerability Index — using address-level inputs from maltreatment locations and crime down to bars, gas stations, laundromats, and crisis shelters. Mapping was completed for all three service areas before sites received implementation resources. The March 2024 grantee profile from the federal Administration for Children and Families' Office of Planning, Research, and Evaluation (OPRE) documents the planner behavior change: Community Implementation Teams led by family resource centers used PAP-identified areas of need to target family outreach, recruitment, and retention, and issued requests for proposals — with stipends — to community agencies to fill needs the PAP process highlighted. The grant ended in September 2023. Oversight throughout was grant-structural rather than algorithmic: continuous-quality-improvement requirements and cross-site documentation reviewed usage and routed resources; no OIG audit, court filing, or independent outcome evaluation of any PAP deployment was found. PAP's own reach-and-impact page claims further work in Arkansas, Washington, and Oregon, and impact figures for New Hampshire (5,000 lead-poisoning training books delivered in high-risk areas, vaccine administration, WIC enrollment) — vendor claims, with no independent confirmation for the Washington and Oregon engagements.
The current program is a protocol, not an operating system. On December 30, 2025 (accepted September 25, 2025), JMIR Research Protocols published PAP's next-generation design (Green, Glass, Purdy, and Daley): gradient-boosted risk terrain modeling against a kernel-density baseline with spatial cross-validation, plus a Stage-3 counterfactual explanation model that searches service-array permutations per household in high-risk cells — voluntary service matching, the authors' framing, though it softens the strict place-only shape of the earlier deployments. The protocol requires jurisdiction-level data-governance agreements and PII masking, promises algorithmic fairness auditing across poverty and race attributes, reports no external funding, and states that "as of September 2025, a participating jurisdiction is under recruitment" — the in-paper status statement; no newer enrollment evidence was found as of mid-2026. Sector reception is cautious rather than adoring: Prevent Child Abuse America's research review of predictive analytics and risk terrain modeling states it "is not intended to be an endorsement of any of these techniques," and 2021 reporting on the ACLU's "Family Surveillance by Algorithm" report (which counted at least 11 states using predictive analytics in child welfare, geographic hot-spot identification among the common uses, and several discontinuations) collected the field's critics — including Khadijah Abdurahman's objection that such tools use "data from the police about families" and no data about police harms.
The sociotechnical reading
Every other child-welfare cell in this Atlas has a score and someone who sees it — a screener anchored by it, a supervisor overriding it, a caseworker never told it exists. This cell deletes that entire seam, and its lesson is about what governance looks like when there is nothing to override. The model's only output is a map; its only channel of influence is persuasion through a two-hop resource-routing chain — planners read the map, coalitions move outreach and grant money where it points. That shape genuinely dissolves the classic harms: no family is mis-scored, no investigation is triggered by a number, and PAP's place-based framing is a considered answer to the person-level scoring critiques this domain's other cases embody. But it also dissolves the classic safeguards. There is no override to log, so the map's real authority is invisible — no record exists of when a planner ignored it. There is no per-case disposition to audit. And the errors that remain are structural: misallocated prevention capacity, and neighborhoods stigmatized wholesale by a high-risk label, with benefit and harm landing on the same cells — the services and the attention arrive together.
The loop is the reading this Atlas adds, and honesty requires saying whose reading it is. The model trains on geocoded CPS and police records; the map concentrates institutional presence in its top tiers; concentrated presence can raise detection and reporting exactly there; and the next refresh trains on the records that attention wrote. That surveillance-amplification loop is the atlas's own structural analysis — not a PAP admission. What PAP's protocol actually acknowledges is narrower: stigmatization risk from "high risk" labels, mitigated in its account because the intervention "consists of providing optional supportive services rather than imposing nonoptional punitive measures"; the "before there is a suspicion of maltreatment" framing is PAP's own site language. No source quantifies whether any targeted area subsequently saw increased reporting — which is precisely the problem. The loop is unmeasured, structural, and slow: by the current protocol's own illustrative design, an outcome signal arrives five years after services begin. A system can therefore run for a full grant cycle — as New Hampshire's did — with its allocation loop live, its evidence loop open, and nothing in between able to tell more-maltreatment from more-looking in the trend data.
The second distinctive structure is a check that fired and found no one listening. The builders ran their own audit — leave-one-group-out validation across 149 neighborhoods, a density-baseline comparison — and published the adverse result: the model fails to generalize across neighborhoods of differing racial composition even with race excluded from the features, because the environmental variables proxy it. That is the domain's proxy lesson (Allegheny's call data, Chile's neighborhood registries) surfacing in a place-based system that was designed to avoid it. But where those cells document oversight bodies reacting, here the finding simply sat: the only oversight this system ever had was grant-structural — CQI reviews of how maps were used, a federal profile documenting process after the fact — and every documented institutional response to the maps is a resource-routing action, never a model-governance action. A fired check with no consequence pathway is the mirror image of the domain's dormant checks, and arguably more instructive: the failure was known, published by the builders themselves, and structurally actionless.
The instruments that fit this cell are accordingly not the domain's usual ones. There is no score to calibrate and no override to protect. The governable surfaces are the loop's three segments and the checks around them: gate what the refresh may train on (the protocol's own data-governance agreements point here); mark the records with their detection provenance so trend reads and retraining can weigh how a report came to exist; bound what map-directed activity writes back; govern the two human hand-offs that carry all of the model's authority; and put the fired fairness audit and the never-run outcome evaluation on a standing, consequential cadence rather than leaving them to a grant that ends. The distinct lesson the Atlas draws: a place-based design genuinely escapes person-level scoring harms, but it does not escape the record loop — it relocates it to geography, where it runs slower, less visibly, and with fewer natural checkpoints, because nobody owns an override on a map. The honest boundary throughout: the families and neighborhoods under the map are not modeled in the paired Lab, which reads institutional propagation only; performance figures are builder-generated except Fort Worth's independently corroborated numbers; impact claims are vendor-reported; no deployment has an independent outcome evaluation; and the current program is recruitment-stage research, not an operating system.
The concepts used in this reading are defined in the Field Guide; the governance responses live in the Practice Library.