Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Better Together? A Field Experiment on Human-Algorithm Interaction in Child Protection

Marie-Pascale Grimon, Christopher Mills

Abstract

Despite algorithms' potential to improve public services, adoption has been limited by concerns about effectiveness and equity. We conduct a randomized controlled trial ($N=4,681$) providing real-time algorithm support to Child Protective Services (CPS) workers allocating investigations. 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. Workers reallocated investigations toward children at greater likelihood of harm, without mechanically following algorithmic predictions. Discussion notes suggest the algorithm shifted worker attention to complementary information. Counterfactual exercises show that human-algorithm complementarity would outperform algorithmic automation in efficiency and equity.

Better Together? A Field Experiment on Human-Algorithm Interaction in Child Protection

Abstract

Despite algorithms' potential to improve public services, adoption has been limited by concerns about effectiveness and equity. We conduct a randomized controlled trial () providing real-time algorithm support to Child Protective Services (CPS) workers allocating investigations. 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. Workers reallocated investigations toward children at greater likelihood of harm, without mechanically following algorithmic predictions. Discussion notes suggest the algorithm shifted worker attention to complementary information. Counterfactual exercises show that human-algorithm complementarity would outperform algorithmic automation in efficiency and equity.

Paper Structure

This paper contains 32 sections, 13 equations, 29 figures, 17 tables.

Table of Contents

  1. Introduction
  2. Empirical Context
  3. Child Protective Services in the Evaluation County
  4. Algorithm Tool
  5. Validity of the Algorithm Tool
  6. Data, Sample Characteristics, and Experimental Design
  7. Data Sources
  8. Sample Characteristics
  9. Experimental Design
  10. Empirical Strategy
  11. Validation of Algorithm Use and Random Assignment
  12. Outcomes: Proxies for Child Harm
  13. Results
  14. Impact of Algorithm Availability on Child Health and Wellbeing
  15. Exploring Spillovers and Learning.
  16. ...and 17 more sections

Figures (29)

  • Figure 1: Algorithm Tool Reduced Harm the Most for Historically Disadvantaged Groups
  • Figure 2: Algorithm Tool Reduced Racial Disparities in CPS Surveillance for Low-Risk-Score Children
  • Figure 3: Groups with Most Improved Targeting had Greatest Reductions in Child Harm Improvements in Outcomes are Tied to Workers' Screen-In Allocation Decisions
  • Figure 4: Access to the Algorithm Changed Worker Discussion Content
  • Figure 5: Child Harm under Counterfactual Algorithm-Only Decisions
  • ...and 24 more figures