MGMT 8825M Week 8 Position Memo Example

Memorandum

To: Director, Calder County Department of Human Services

From: Policy and Ethics Advisor

Date: May 11, 2026

Re: Whether to keep, restrict or withdraw the hotline screening score

Recommendation

Keep the screening score, but restrict how screeners use it. Screeners should record their own decision on each report before the score is shown, the score should never by itself require a report to be investigated, and the department should publish each year how often screeners and the score disagreed and which families were affected.

The Tool as Documented

Calder County began piloting a predictive risk score at its child abuse and neglect hotline in 2025. The score is modeled closely on the Allegheny Family Screening Tool in Pennsylvania, the most fully documented tool of its kind. Chouldechova et al. (2018) described that tool's design and evaluation in a published case study. When a report is received, a model drawing on county administrative records, including child welfare history, public benefits, behavioral health services, and jail and juvenile probation records, estimates the likelihood that the child will be removed from the home within two years. The estimate is shown to the call screener as a score from 1 to 20. The screener and a supervisor decide whether the report is investigated. In Calder's pilot, as in Allegheny's design, the highest scores trigger a default to investigate that a supervisor must override in writing.

The case study reported that the model predicted placement better than screeners' decisions alone had, and it examined how the model's errors were distributed across groups of children.

The Strongest Opposing Argument

Eubanks (2018) made the strongest case against tools of this kind, examining Allegheny's tool directly. Her argument is that the model can only learn from the records the county holds, and the county holds records mainly on families who use public services: those who receive benefits, use public mental health care or have contact with the county jail. A middle-class family that pays for private therapy, addiction treatment or child care leaves no trace in the data. The model therefore sees poor families far more clearly than affluent ones, and it may mistake a family's contact with the state for a risk to its children. Because more contact produces a higher score, and a higher score produces more investigation, the tool can deepen the very pattern it measures.

This argument cannot be dismissed. The county's own data confirm that families receiving public benefits are overrepresented among high scores.

Three Positions Compared

Keeping the score as built preserves its predictive gains and the protection the highest scores may offer children in danger. But it leaves in place the mandatory default for high scores, which gives the model's view of poor families the force of a rule, and it shows the score before the screener has formed a view, which invites screeners to anchor on it.

Withdrawing the score answers Eubanks most completely. It returns the county to decisions made by screeners alone, who also make errors and who, the case study suggests, were less accurate in predicting placement. Withdrawal would remove a tool that may help screeners notice children at serious risk in reports that read as routine.

Restricting the score keeps it as a second opinion. Screeners record their decision first, then see the score; where the two disagree, a supervisor reviews the report. Removing the mandatory default means the model never decides on its own, and publishing disagreements lets the public see whether the score pushes investigations toward poor families. This position keeps most of the tool's value as a check while limiting the risk that it becomes the decision. Keddell (2019) argued that the data behind child protection algorithms do not represent the incidence of abuse across the population, and that tools built on them can create distorted feedback loops; a published count of disagreements is one way to watch for such a loop.

The Cost of the Chosen Position

Restricting the score has costs, and they fall on two groups. Some children whose reports would have been investigated because of a high score will not be, if screeners and supervisors decide otherwise; some of those children may later be harmed, and the department will not always know which ones. The department will also need more supervisor time to review disagreements, about one additional supervisor at the hotline by current estimates. These costs should be measured in the first year's report, alongside how often the score and screeners disagreed by family income. If restriction proves to miss children the score would have protected, the county should reconsider; if it shows the score pushing investigation toward poor families, withdrawal should be reconsidered instead.

References

Chouldechova, A., Benavides-Prado, D., Fialko, O., & Vaithianathan, R. (2018). A case study of algorithm-assisted decision making in child maltreatment hotline screening decisions. In S. A. Friedler & C. Wilson (Eds.), Proceedings of the 1st Conference on Fairness, Accountability and Transparency (Vol. 81, pp. 134-148). PMLR. https://proceedings.mlr.press/v81/chouldechova18a.html

Eubanks, V. (2018). Automating inequality: How high-tech tools profile, police, and punish the poor. St. Martin's Press.

Keddell, E. (2019). Algorithmic justice in child protection: Statistical fairness, social justice and the implications for practice. Social Sciences, 8(10), Article 281. https://doi.org/10.3390/socsci8100281