Domain Atlas / Child welfare & family services
Douglas County Decision Aide
Work with this case in the PAN Lab ↗
The Douglas County Decision Aide, deployed into the county's RED-Team call-screening process in February 2019, scores each referral from 1 to 20 for a child's likelihood of out-of-home removal within two years; an independent Cornell-led randomized controlled trial found it sped up screening decisions without significantly changing child outcomes, and a companion study found workers attended mainly to extreme scores while largely disregarding mid-range ones.[3]
What happened
Douglas County's Department of Human Services commissioned the Centre for Social Data Analytics (Auckland University of Technology) in 2017 to explore predictive risk modeling, with a feasibility stage completed that August. The result, the Douglas County Decision Aide (DCDA), is a LASSO-regularized logistic-regression model that scores each maltreatment referral from 1 to 20 for a child's likelihood of out-of-home removal within 24 months; race-related predictors were tested and deliberately excluded from the fielded model. The score is embedded in the county's RED (Read, Evaluate, Direct) Team consensus screening — a supervisor with at least two caseworkers, used for roughly 85 percent of referrals — and computed from multi-agency state records (child-welfare, public-benefit, and court data). It launched in February 2019 as a year-long randomized controlled trial, randomized at the team level, and was independently evaluated by Cornell researchers Maria Fitzpatrick and Christopher Wildeman (with Katharine Sadowski), using an ethics framework ported from the Allegheny County evaluation. The peer-reviewed trial found the tool sped up screening decisions without significantly changing child outcomes, with COVID limiting the outcome analysis, and a companion study found workers attended mainly to extreme high or low scores while largely disregarding mid-range ones. County officials said they would share a family's score on request. National reporting later grouped Douglas among Allegheny-inspired tools, and Colorado headlines tied it to a U.S. Justice Department inquiry that actually targeted Allegheny County's tool, not Douglas's.
The sociotechnical reading
On paper this is the most carefully governed deployment in the Atlas: an independent randomized trial, an ethics review, race predictors excluded, ongoing drift monitoring, a required multi-person consensus, and families told their scores on request. And still the rigorous evaluation found only a modest effect — because the binding constraint was not the model's accuracy but the human-algorithm interaction around it. Workers consulted the score selectively, reacting to the extremes and rarely moving mid-range decisions, so the one thing left ungoverned — the mapping from a score to a next action — is what determined impact. The lesson is that accuracy and governance quality are necessary but not sufficient: a safeguard as discretionary as "the team may consider the score" can dilute a signal to near-nothing, in either direction. A second, quieter lesson is about evidence hygiene. The dramatic positive results often attributed to this tool belong to a separate sibling deployment in another Colorado county; keeping straight which evidence attaches to which system is itself a governance discipline, and the press record here shows how easily it collapses.
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.