Domain Atlas / Child welfare & family services
US Birth Match
Five US states (Michigan 2001, Minnesota 2001/2006, Maryland 2009, Texas 2013, Missouri 2021) operate 'Birth Match' programs in which an identity record-join between new birth registrations and a registry of parents with prior terminations of parental rights, serious-harm findings, or specified convictions fires a mandatory child-protective response with no risk score, no threshold, and no screening discretion: Michigan's current policy (PSM 712-2, reissued 2026-04-01) permits screen-out only for an inaccurate identity match or an already-open case and otherwise requires the referral to be screened in and assigned for investigation, and Maryland's 2025 response policy limits screen-out to a non-qualifying conviction or an adopted child while requiring a 24-hour intake coded 'Risk of Harm: birth match' — a mandatory response with no screening discretion, formally structured in Maryland as a voluntary non-CPS assessment families may refuse, with tracing and escalation duties attached. The only tunable parameters anywhere are registry scope and lookback: 2 years in Texas, 10 in Maryland and Missouri, unlimited in Michigan (records to 1978) and Minnesota.[4]
What happened
Birth Match began in Michigan in 2001, after two September 2000 Detroit cases — seven-month-old Miracle, who suffocated, and five-month-old Jamar, who was severely beaten and survived — in which the implicated parents had prior terminations of parental rights (TPR). Before automation, caseworkers informally asked hospitals to report new births to specific parents; the program replaced that ad hoc watching with a registry join. The mechanism, now operating in five states, is deterministic or probabilistic identity record-linkage between two authoritative registries: the state vital-records birth registration stream and the child-welfare information system's roster of parents with prior TPRs, child-death or serious-harm findings, or (in two states) specified convictions. There is no risk score, no threshold, no feature model, and no ranking — registry membership is binary, and the match itself is the referral. In Michigan the linkage runs automatically at birth-record creation inside the case-management system, generates an unassigned referral, and emails the Centralized Intake director; under current policy (the Children's Protective Services Manual, PSM 712-2, reissued unchanged April 1, 2026) the only screen-out grounds are an inaccurate identity match or an already-open case — an accurate match "must be screened in and assigned for investigation" under the history-based allegation of "threatened harm." Michigan Legal Help tells parents plainly: give birth in a Michigan hospital after a prior termination and your local MDHHS office will be notified, and CPS can ask a judge to remove the newborn at the hospital. Minnesota runs a daily match against nightly birth uploads; Missouri's registrar pushes qualifying births three times a week into a mandatory Newborn Crisis Assessment with a 3-hour contact standard; Texas matches twice monthly under an unpublished memorandum of understanding; Maryland's health department matches weekly. The only tunable parameters anywhere in the system are registry scope and lookback: 2 years in Texas, 10 in Maryland and Missouri, unlimited in Michigan (records back to 1978) and Minnesota. Notably, neither Michigan's nor Texas's program rests on statute — both run on interagency agreements and agency policy, changeable without legislative process.
The implementation record is thin and almost entirely the work of one researcher: a 2022 American Enterprise Institute report by Marie Cohen, a birth-match proponent, assembled the only cross-state data through public-records requests — figures the agencies generated ad hoc and Cohen herself flags as "limited and of dubious accuracy." What those figures show is a cascade of mandatory responses and vanishing follow-through. Roughly half of Michigan matches already had an open case; 38–51% of complaints were assigned for investigation across FY2009–FY2021, opened cases fell from 99 (9% of matches) in FY2012 to 30 (3%) in FY2020, and removals from 41 to 11. Michigan's one internal audit (2011–12, 105 cases) found 75% of birth-match investigations identified no threatened harm — and that workers were not following the threatened-harm assessment policy, with supervisors approving the deficient work; the response was mandatory training, and the audit initiative died with personnel turnover. Fewer than 10% of matched families in any state with data received services; in Maryland, exactly 5 of 124 matched families (4%) had documented services in the FFY2020/CY2020 pairing the source itself labels across mixed year types. Minnesota collects no outcome data at all. Most strikingly, the automation is unowned: Michigan's match volume fell from 1,186 (FY2019) to 515 (FY2021) and Maryland's from 243 (CY2019) to 124 (CY2020), and in both cases the operating agency was unaware of the collapse or unable to explain it. Maryland's 2018 expansion law — passed unanimously, extending the lookback from 5 to 10 years and adding a court conviction feed after Baltimore's Child Fatality Review Team documented the matcher's misses — also ordered the Department of Human Services (DHS) to contract an independent evaluation of the match's sensitivity, specificity, and predictive value — in plain terms, what share of the families it should have flagged it caught, how many it missed, how many it flagged who were never going to harm a child, and how often a flag turned out to be right. A records request found the requirement was simply never implemented; no responsive documents exist, and no state publishes birth-match data in any regular report.
The structural loop is what distinguishes the case: the registry that fires the trigger is populated by the system's own outputs. A birth-match response that ends in a new TPR writes the parent back onto the match list — permanently, in Michigan and Minnesota — so every subsequent birth re-fires with strictly increasing history, and Michigan adds manual list entries for severe cases without any TPR. Maryland law accelerates the loop: a prior involuntary TPR lets the agency skip reasonable reunification efforts for the matched newborn, making the next termination faster and cheaper, which re-arms the trigger. That coupling — not the matcher — is the target of the recurring Maryland legislative fight: the Right to Fight Act (HB 944 of 2025, died without a committee vote; reintroduced as HB 48 of 2026, heard January 29, 2026, and dead in committee without a vote when the session adjourned sine die on April 13, 2026) would have repealed the waiver, and Civil Rights Corps testified that birth match plus the waiver is "functionally state sterilization," coercing parents into "voluntary" relinquishments to avoid arming the trigger. Maryland's rewritten response policy (SSA/CW #25-03, effective September 1, 2025) meanwhile tightened the pipeline — 48-hour verification, 24-hour intake coded "Risk of Harm: birth match," home visit within 5 days, full assessments within 30 — while still labeling the response a voluntary, non-CPS prevention assessment families may refuse, with benefits-record and hospital tracing when families cannot be located and escalation to a CPS report on suspicion. The scholarly record brackets rather than resolves the debate: Shaw, Barth and colleagues (2013, co-authored with Maryland officials) present the program as timely use of administrative data to protect newborns; Lanier and colleagues (2020) analyze it as an early algorithmic child-welfare policy lacking transparency and accountability; Gibbs and colleagues (2024) map the five-state landscape as a fatality-prevention tool. A federal commission recommended birth match nationally in 2016, and Harvard's Data-Smart City Solutions program highlighted the approach — yet only Missouri (2021) has adopted it since, and Missouri's own startup data show two-thirds of its first matched families were already the subject of a birth-time CPS call from a medical professional.
The sociotechnical reading
Every other child-welfare deployment in this atlas puts a model between the record and the decision — a score, a threshold, a ranking, something with a dial. Birth Match deletes that layer, and the deletion is the lesson. The trigger is a store-to-store join: two registries, a binary membership test, and a referral that policy makes mandatory. That shape appears elsewhere in the atlas only in public-benefits entity resolution — and its appearance in child welfare removes every governance surface the domain's debates assume. There is no threshold to tune, no calibration to audit, no accuracy to contest, no score for a caseworker to override: the entire model surface is a lookback window and a list of trigger classes, which means the entire governable system is the couplings (which registries join, how far back), the writes (what arms the trigger), the discretion (deleted at intake by the mandatory screen-in rule, surviving only downstream inside the mandated response, where the one audit found it degraded), and the watching (which nobody did). The record makes the last point almost literal: this is the atlas's strongest documented case of unowned automation — throughput halved in two states and the operating agencies were unaware of the collapse or unable to explain it, because no oversight loop consumes the trigger's own telemetry.
Two structural features do the deepest work. First, the registry never forgets and the loop arms itself: no state retires a parent from the list on a no-harm finding — only the lookback window ages a trigger out, and Michigan and Minnesota set no lookback limit at all, matching records back decades — and a response that ends in a new termination writes the parent back onto the list, so the system's enforcement output is its own future input, with a statutory waiver in one state making each cycle faster than the last. The false-positive structure of the whole apparatus is therefore governed by parental-recovery dynamics the trigger cannot see; in Michigan and Minnesota, families with repeated no-harm findings stay armed forever, which is precisely the tail no oversight body audits. Second, the one oversight loop that works is a one-way ratchet: fatality-review teams audit the matcher's misses — the dead child with no TPR, the non-parent perpetrator — and their two documented wins both widened the trigger, while the legislated evaluation that would have measured the other side (sensitivity, specificity, predictive value) was never performed. An oversight channel that only ever sees false negatives will only ever expand the system it governs. The honest boundary runs through the middle of the case: the newborns and parents on whom all of this lands are not in any model here — under 10% of matched families ever received a service, no birth-match-specific disparity figure has ever been published, and nearly every number in the record flows through one proponent researcher's records requests that she herself flags as of dubious accuracy. A system that mandates its response, permanently remembers its trigger, feeds its own input, and was never once evaluated is not a model problem; it is a governance topology, and this case file maps it.
The concepts used in this reading are defined in the Field Guide; the governance responses live in the Practice Library.