Sequence Analysis of Child Protective Services (CPS) Report and Foster Care Lifetime Trajectories

Background/Significance:

Understanding the “flow” of children through the child welfare system (both CPS reporting and foster care) has been a core goal of child maltreatment researchers for many years. While some local and regional studies exist, national studies are needed. We take advantage of a newly created integrated national dataset, linking the National Child Abuse and Neglect Data System (NCANDS) Child Files and the Adoption and Foster Care Analysis and Reporting System (AFCARS) Foster Care Files from 2006-2021 to map national long-term report/placement trajectory classes for the first time. For this linkage, we have developed the RAPIDS (Report and Placement Integrated Data System) program, which is undergoing final testing. Our research question is “what are the child welfare system trajectory typologies of children who experience foster care”. These children may or may not have a prior or subsequent CPS report, but most children do have a CPS report prior to their placement.

Methods:

We identified a cohort of children born in 2006 who experiences at least one FC placement between 2006-2021 (N=129,419). We tracked their progress over a 15-year period, mapping each child’s trajectory through the CPS system into a sequence of 180 elements long, one for each month. Each element represents the child’s status for that month, taking one of three statuses: “CPS-report”, “in-foster-care”, or “neither” (“CPS-report” took precedence over “in-foster-care”).

Because sequence analysis is not feasible with such a large sample, we draw samples of 20,000, the largest size recommended in sequence analysis literature. We draw multiple samples to ensure cluster stability. Distances between sequences are measured with dynamic Hamming matching as we assume the probability of transitioning between status varies as children age. We employed cluster analysis using Ward’s method, fitting sets of 6, 5, 4 and 3 clusters. We measure cluster fit “quality” via the point biserial correlation (PBC), Hubert’s C (HC), and average silhouette width (ASW) metrics.

Results:

Although a 6-cluster fit yielded the optimal “quality”, it was unstable. Cluster fits only became stable at three clusters – “early-entry/early-exit”, “school-age-entry”, and “early-adolescence-entry”. We find that these clusters differ significantly on 15 measures, including average age at first FC placement, type of report preceding placement, total time in FC, and discharge reason. These distributions held after resampling.

Conclusions/Implications:

Cluster distributions varied significantly by state, suggesting that local analyses of data may be a necessary next step to capture nuances in the trajectories children take through the child welfare system. Despite this, the three identified clusters showed a number of clear differences, in terms of both which children were in the clusters and their likelihood of later outcomes. For example, the “early-adolescence-entry” cluster tended to have longer stays than other groups, suggesting the need for more aggressive reunification services, and they were less likely to have initially been placed in foster care due to child abuse or neglect, suggesting that the child welfare system may need to focus on addressing risk factors for children beyond traditional child maltreatment.