Domain Atlas / Child welfare & family services
New Zealand MSD Predictive Risk Modelling
New Zealand's Ministry of Social Development commissioned a child-maltreatment risk-modelling tool that, on a 2012 development sample of 57,986 children and 132 selected variables, reported an area under the ROC curve of 76% and a top risk decile in which 47.8% had a substantiated maltreatment finding by age five; those figures come from development data rather than field performance, the tool was never operationally deployed, and a proposed two-year study that would have scored about 60,000 newborns was halted by the incoming Social Development Minister, who annotated the briefing papers 'Not on my watch! These are children not lab rats.'[4]
What happened
New Zealand's Ministry of Social Development commissioned a predictive risk model to estimate, at or shortly after birth, a child's likelihood of a substantiated maltreatment finding by age five, computed from linked benefit and child-protection (Child, Youth and Family) administrative records. The peer-reviewed development work, published in 2013 and led from the University of Auckland with co-authors at Auckland University of Technology and the University of Southern California, reported on a 2012 sample of 57,986 children — with 132 variables selected from 224 constructed — an area under the ROC curve of 76% and a top risk decile in which 47.8% had a substantiated finding by age five; those figures come from the development data and were never field-validated. An independent ethical review by Tim Dare (2013) judged that the remaining concerns "may plausibly be regarded as outweighed by the very considerable potential benefits," alongside a dedicated Maori ethical review and a Privacy Impact Assessment discussed with the Privacy Commissioner. In November 2014 the ministry lodged an ethics-committee application for a two-year prospective observational study that would have scored about 60,000 newborns by linking maternity, benefit and child-protection data and watching whether high-risk predictions came true. The incoming Social Development Minister withdrew it, annotating the briefing papers "Not on my watch! These are children not lab rats," a stance the Children's Commissioner echoed ("no ethics committee in New Zealand is going to wear that"). Testing was narrowed to about 20 historic anonymised cases; the tool was never operationally deployed, and as of 2026 predictive modelling is still not used by Oranga Tamariki frontline workers.
The sociotechnical reading
Almost every other case in this Atlas is governed after it runs — by courts, commissions, audits, or an agency's own belated stop. New Zealand was governed before it ran — and among the Atlas's never-deployed tools, it is the one stopped by an accountable authority rather than by a test result or a missing legal basis. The control that bound this system was ex ante: a layered ethical and privacy review, and finally an accountable authority who declined to sign off. In map terms the decisive event sat on the authority dimension, not the correction loop — a deployment-permission gate exercised at the design stage, the earliest and cheapest place any system can be stopped. Two lessons follow. First, how you validate a predictive tool is itself a governed act with its own ethics: the flashpoint here was not the model but a proposed study that would observe high-risk children without intervening. Second, the accuracy numbers this tool is still cited with are development-sample figures that were never field-validated — precisely because the study that would have tested them was the thing the gate stopped, a caution against reading unearned performance into a system that never ran. That the same approach was later exported and deployed elsewhere, while its country of origin declined to deploy it, is the sharpest reminder that deploy-or-not is a governance choice, not a technical inevitability.
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.