Child injury admissions fall with a Colorado decision aid

In a Colorado county's randomized trial with 4,681 children, those whose referrals were screened with an algorithm's risk score had 21 percent fewer injury admissions.

  • 4 for a person

    Children in the score group had 21 percent fewer injury admissions and were 60 percent less likely to fall in the top one percent of the harm index.

  • 2 for an organization

    Screening teams shifted investigations toward children more likely to be harmed without changing the rate of investigation.

Depth runs from 1, a convenience or time saved, to 5, a life or a livelihood preserved.

Status
Reported benefit.
Evidence

foundstudiedreportedclaimed

Only the evaluators support the benefit.

Compared with

Children in households randomly assigned to screening without the score. Measured.

Without the system
a mean of 0.17 injury admissions per child in the control group
With it
0.03 fewer injury admissions per child, a 21 percent reduction
The system’s part
Households were randomly assigned so that teams saw the score for some referrals and not for others, which isolates the effect of showing it. Workers kept full discretion and recorded the score in at least 73 percent of decisions. The model is a LASSO logistic regression trained on about 500,000 past child referrals.
Later
Expanded, 18 February 2026. After the trial the tool became available for all referred children in the county. The authors write that it has since been scaled to other agencies.
Not yet checked
LimitsCosts and harmsLearns from data

Download data Cite Suggest a correction

Benefit type
Improved Physical Health
System
Child protection decision aid tool (predictive risk score)
Made by
Centre for Social Data Analytics (CSDA), led by Rhema Vaithianathan
Used by
Child Protective Services screening teams in Larimer County, Colorado (named in the trial's registry entry, as the paper does not name the county)
Where
Larimer County, Colorado, United States
Stage
In a pilot
Date
Kind
Aggregate result
Broad benefit
Judging the Qualities of Objects, Services, or People, Making Decisions and Solving Problems
Collections
Cross-border records
Last revised
1 November 2026

Sources

  1. Report, by an evaluator that the makers or users commissioned.

    Grimon M-P, Mills C. Better Together? A Field Experiment on Human-Algorithm Interaction in Child Protection. Working paper, arXiv:2502.08501v3 [econ.GN], 18 February 2026, abstract page. The paper states that the study is part of a research program on risk modeling in human services led by Rhema Vaithianathan and Emily Putnam-Hornstein.

    Algorithm access reduced maltreatment-related hospitalizations, especially among disadvantaged groups, while reducing CPS surveillance of Black children. Notably, child injury admissions decreased by 21 percent.

  2. Other, by an evaluator that the makers or users commissioned. Context or contrary evidence.

    Grimon M-P, Mills C. The Effect of Algorithmic Tools on Child Welfare Decision-Making and Outcomes. AEA RCT Registry, AEARCTR-0006311, first published 21 August 2020, last updated 22 September 2022. The entry lists the region as Larimer County in Colorado and reports no results.

    The tool uses machine learning methods trained using several years of historic administrative data and is highly predictive future home removal (e.g., of any child to foster care), re-referral to child protective services (CPS), and child maltreatment death.

Details

AI system

Workflow

The model was trained on about 500,000 past child referrals from the trial county and other Colorado counties. Teams of about five workers saw each child's score from 1 to 20 in the morning meetings where they decided which referrals to investigate. Workers were trained with examples so that they did not decide on the score alone.

Money

Kind of value

Estimated public savings from about 20 fewer maltreated children, against implementation costs of about $280k over two years

Amount

Estimate: $1.25M

Recipients and payers

Government budgets for special education, criminal justice and child public health

Follow-up

Conditions for replication

Workers keep information that the algorithm does not use, such as the current allegation. Teams use the score within a group discussion and keep the final decision.

Entry details

Sector

Child welfare.

Purpose

Rhema Vaithianathan and colleagues at the Centre for Social Data Analytics built the decision aid to predict a child's likelihood of removal from home within two years of a referral. Screening teams use the score when they decide which referrals to investigate.

References

Other identifiers

Aea_rct: AEARCTR-0006311.

36 more questions not yet checked

model, release, provider, role, configuration, inputs and outputs, autonomy, human oversight, beneficiaries, reach, distribution and access, duration, every measurement question, currency and prices, period and unit, realization, gross or net, included costs, comparison, method, return and payback, ownership and control, dependencies, cross-border connections, geopolitical implications, replication attempts, availability, institutional decision, grounds, action and result, success criterion, every evidence question, observed or intended, alternative assessments, related entries, category mappings.

Cite this record

Peter H. Vartanian. (2026). Child injury admissions fall with a Colorado decision aid (AIBD-1131). Artificial Intelligence Benefits Database, edition of 1 November 2026. https://petervartanian.xyz/aibd/records/AIBD-1131.html (archived at https://doi.org/10.7910/DVN/AADSII, version 7.0)

BibTeX Data Field definitions