When Is Maltreatment Confirmed? Using Unsupervised Machine Learning to Advance Understanding of Child Protective Services Decisions

Background and Purpose: CPS data sources include maltreatment reports and report dispositions (i.e., whether the report is determined to be substantiated). Reports are highly heterogeneous given jurisdiction-specific and oftentimes ambiguous definitions of maltreatment and limited knowledge of reporters. Substantiated maltreatment reports, therefore, are commonly used by researchers and federal agencies as a proxy for child maltreatment victimization (i.e., confirmed maltreatment). In doing so, heterogeneity within substantiated reports is ignored, despite critical implications for families, including disparities in surveillance (e.g., inclusion on a central child abuse registry) and potential outcomes (e.g., foster care). In turn, these disparities might entrench racial and socioeconomic inequalities. We aimed to examine heterogeneity within the decision to substantiate and categorize substantiated reports into homogenous clusters based on child, report, and neighborhood characteristics. We also aimed to examine the extent to which this heterogeneity is explained by county-level agency policy.

Methods: We used de-identified administrative records from California’s child welfare data collection system. The population-based dataset was defined as all reports between January 1, 2000 and August 1, 2023, the last date of available data (N = 16,154,235). We further refined the dataset to children with geocoded residential addresses that linked to American Community Survey data (N = 13,458,114). We focused this analysis on age 5 given the unique needs of younger children. The universe of child-reports for children age 5 and under was 2,350,184. The dataset used for clustering was only child-reports with a disposition that resulted in a perpetrator being included on the CACI, according to policy at a given time (N = 667,401). We used model-based clustering to classify substantiated reports for young children into homogenous substantiated child-report typologies. We used regression analyses to determine why some neighborhoods had a higher likelihood of reports from a particular cluster.

Results: A 13-cluster solution was found to be the most effective and parsimonious. Three clusters were broadly defined by infants (infant and family violence; medical reporters + infants of color; medical reporters + White infants). Three clusters were broadly defined by family violence (Hispanic; White; Black). Five clusters were related to neglect (Hispanic + mandated neglect reports; Hispanic + family neglect reports; White + mandated neglect reports; White + family neglect reports; Black + neglect reports). Two clusters were defined by the absence of neglect (Hispanic; Black); these clusters featured relatively high rates of sexual, emotional, and physical abuse. Cluster membership differed across time and geography.

Conclusion and implications: To substantiate a maltreatment allegation, California requires that it be “more likely than not that child abuse or neglect occurred.” Yet, the operationalization of this definition varies in complex ways. Results reiterate key indicators of risk of CPS involvement, such as infancy and neglect. And, they suggest that families with multiple children in a household for which CPS agencies have received allegations of multiple types of maltreatment may be at greater risk of deeper penetration of CPS. We explore how decision-making varies across racial and ethnic groups and implications for practice and policy-making at the agency level.