Skip to content
ParamergeParamerge
Stephen Lieberman
Through Paramerge
Contact

PAN Lab example

Allegheny Hello Baby

The help that keeps a file: a birth-risk prevention model

Allegheny County's Hello Baby model scores newborns from county records. Highest-scored families are offered voluntary help, not investigation, by workers who must report suspected abuse.

See more

The Hello Baby predictive risk model is run by the Department of Human Services of Allegheny County, Pennsylvania. Shortly after birth, it scores every newborn's risk of being removed from the home because of maltreatment by age three, using county records. It then sorts families by score into three service tiers, each offered a different level of voluntary support.

How the program works

The county launched Hello Baby in September 2020. The model sorts families by score into three service tiers: Universal, Family Support, and Priority. Families in the top tier, which holds the highest scores, are offered proactive outreach and voluntary prevention services, not an investigation. Universal supports, such as a support phone line open around the clock and a website, are open to every family. The sources do not say what the Family Support tier offers.

Family Center outreach workers and Healthy Start family-engagement specialists deliver the outreach. This network calls them the outreach providers. Families choose whether to enroll. These workers are mandated reporters, whom the law requires to report suspected child abuse or neglect.

What the score is built from

The model uses the county's existing administrative records. They are the same records behind the Allegheny Family Screening Tool, which scores reports of child maltreatment to support the county's call-screening decisions. The records come from many agencies, including birth, child-welfare, jail, court, housing, and behavioral-health records.

They were collected for other purposes, and families never gave them for this one. Prior contact with county systems raises a family's score.

Consent

The program's most contested feature is passive consent: a family is included unless it opts out. Families are told at hospital discharge and again by a mailed postcard. They have 20 days to opt out before the birth record is scored.

Review before launch

Two independent ethics reviews came before the launch. Michael Veale of University College London wrote one, and Deborah Daro of Chapin Hall wrote the other. The Department of Human Services published a point-by-point response. It noted that the county has no institutional review board, the kind of body that usually reviews research involving people.

How well it ranks families

The Centre for Social Data Analytics wrote the model's methodology report for the county in 2020. The research behind the model studied 52,520 births from 2012 to 2015.

On data held back for testing, the model reported an area under the curve of about 0.93. That measure says how often the score ranks a child later removed from the home above one who was not, where 0.5 is a coin flip.

The top roughly 5 percent of newborns by score accounted for about 54 percent of children later removed from the home by age three. Their removal risk was roughly twenty times that of other newborns. The methodology report gives 22.24 times, with a 95 percent confidence interval of 17.50 to 28.25. That interval is the range the true figure very likely falls in.

What the evaluation found

A 2025 external evaluation by the Urban Institute and Chapin Hall covered children born from 2016 to 2024, taking account of the COVID-19 pandemic's effects. It found the program associated with fewer first investigations of child maltreatment. It also found the program associated with fewer first substantiated investigations, those that confirmed the concern. It found no reduction in out-of-home foster-care placements, the outcome the model was built to predict.

The firewall

The county keeps the newborn score out of the screening of hotline calls and out of investigations. The sources call this separation a firewall. It means the score steers an offer of help, not an investigation.

Function creep is a system's slow spread to uses beyond its original purpose. Here it is the standing pressure to let investigators see the scores. A firewall holds only while someone re-approves it each time a new connection is proposed.

What each side says

Mandated reporters deliver the help, and the score is computed from the county's cumulative records. So accepting outreach can lead to a new report that is written into those records. That report can raise the family's future score.

Critics treat this loop as central. The county argues its measured effect is small.

The central dilemma

The offer of help and the child-protection system share the county records and workers who must report suspected abuse. Critics argue that this can make help feel like surveillance. The case file's reading is that the governance work here is not accuracy. It is keeping the boundary: consent, marking where records came from, storing less, and defending the firewall.

What the available tools can and cannot address

A failure pathway is a link between two parts of the network, where a mistake made by one part can be passed on to the other. A tool closes a pathway when mistakes stop passing along it.

This case has a budget of 10 units. Explore (No Targets) sets no targets. There, and under Service Targets Only, one tool costing 2 units is enough to keep mistakes from building on one another across the network. Escalate checks, Keep skills sharp, Mark AI-written records, or Review on schedule each does it alone.

Under Service Targets Only, each of those four also meets the service target. The service target asks that the model stay useful for choosing which families are offered help. Every other offered tool except Require sign-off also meets it alone. In all, 386 distinct sets of tools meet that level within the budget.

Under Service and Safety Targets and All Governance Targets, you must also close every failure pathway. Both levels can be met within the budget. The cheapest way costs 7 units: Escalate checks and Mark AI-written records at 2 units each, and Store less data at 3.

Escalate checks closes the pathway named Tier picks families for outreach. Mark AI-written records closes County records used for the score. Store less data closes Outreach written to records.

Under Service and Safety Targets, 14 distinct sets of tools meet the targets. Every one includes Escalate checks and Store less data, with Mark AI-written records, Vet connections, or both. Lingering effects is a Dynamics setting in which damage outlasts its cause. It is always off at that level, so Vet connections works fully there.

Under All Governance Targets, 10 distinct sets meet the targets, and every one includes the cheapest three. Lingering effects is always on at that level. With it on, Vet connections needs Understand the system, at 4 units, to close its pathway. That pair with Escalate checks and Store less data costs 12 units, over the budget.

Stylized model of a documented deploymentChild welfare & family services

Open this example in PAN Lab v0.1 to apply pressures and levers and watch what the system does.

What this models

This example runs on the Hello-Baby-class universal-prevention risk model network: 4 components and 9 pathways between them. Every context in the Lab is a stylized model, never a reconstruction of any actual deployment, and each assumption behind it carries a provenance label.

Evidence base: 5 assumed · 1 published baseline. In the Lab, the shaded evidence band behind each headline readout draws its width from the least-established class below.

Show all 6 assumptions
  • assumed

    This network follows the pattern of Allegheny County's Hello Baby program, which scores every newborn to offer prevention services. It is not a reconstruction of the actual program.

  • baseline

    The network draws the link from the county records into each newborn's score from the start. The case file says the score is computed from existing county records, and prior contact with county systems raises it.

  • assumed

    In the county's documented design, the score is kept out of hotline screening and child-protective investigation. The network draws the link that would copy scores there, because this firewall defines the case. Function creep, a system's slow spread to new uses, would open it, for example through an unreviewed data connector.

  • assumed

    Outreach is delivered by mandated reporters, whom the law requires to report suspected abuse or neglect. So contact can itself lead to a new report that is written back into the records. The case file records this risk as contested. Critics treat it as central, while the county argues its measured size is small.

  • assumed

    The network draws links within the outreach teams and within the model. It assumes outreach teams can share ways of working and compare notes on families, which the sources do not document. One model scoring every newborn repeats any blind spot for every newborn it scores.

  • assumed

    The network does not model demographics, or harm that falls differently on different newborns and families. Any such harm is documented in the case file and measured outside any network like this one. The sources give no measured racial disparity figure specific to Hello Baby, so none is stated here.

What this example does not show

Show all 1 limitation
  • This example does not model harm that falls differently on different newborns and families. It traces how mistakes pass between the model, the outreach providers, and the county records. That harm is documented in the case file and measured outside any network like this one. The loop in which accepting help can raise a family's future score is contested. Critics document it as a core risk, while the county argues its measured size is small.

Sources and evidence

What this example rests on, claim by claim. Every entry resolves to the same ledger the Evidence Registry publishes.

  • In Allegheny County's Hello Baby program, the top-tier roughly 5% of newborns by predictive risk score accounted for about 54% of children later removed from the home by age three, at roughly twenty times the removal risk of other newborns (methodology relative risk 22.24, 95% CI 17.50-28.25); the model reported an AUC of about 0.93 on holdout data.

    empirical
    • Government evaluation Centre for Social Data Analytics (AUT) for Allegheny County DHS, Implementing the Hello Baby Prevention Program in Allegheny County: Methodology Report Version I (2020) https://analytics.alleghenycounty.us/wp-content/uploads/2020/12/Hello-Baby-Methodology-v6.pdf
    • Academic Vaithianathan, Benavides-Prado, Rebbe & Putnam-Hornstein, Using a Predictive Risk Model to Prioritize Families for Prevention Services: The Hello Baby Program in Allegheny County, PA, Prevention Science (2025) https://pmc.ncbi.nlm.nih.gov/articles/PMC12064473/

Where this connects

Institutional pressures in this domain

  • Workload surge — Demand outruns staffing; per-case attention shrinks and review becomes triage.
  • Deadline pressure — Statutory or managerial timeliness rules reward fast approval of machine output over slow disagreement.
  • Staff turnover — Experienced skepticism leaves; new staff calibrate their trust on the tool itself.
  • Data & policy drift — The world, the intake process, and the rules change under a system trained on how things used to be — two mechanisms with different remedies: the statistical properties of what the system processes move (concept drift), or the mixture of inputs arriving in deployment differs from the mixture it was trained on (covariate shift).
  • Compliance over substance — Paper controls (sign-offs, checklists) satisfy audits while the behavior they describe erodes.

All of them in context on the Child welfare & family services domain page.

Levers available here and the patterns behind them

Documented case histories