PaperlayerView exact revision on arXiv

Source-linked AI summary

A Human-Centered Review of the Algorithms used within the U.S. Child Welfare System

Devansh Saxena, Karla Badillo-Urquiola, Pamela J. Wisniewski, Shion Guha

arXiv:2003.03541v1cs.CYcs.AIcs.HC

TL;DR

U.S. child-welfare algorithms address risk, placement, and matching amid concerns about bias, opacity, and limited context, while prior surveys leave their construction, predictors, and target outcomes underexamined. This paper reviews 50 publications through a human-centered design lens and finds fragmented empirical knowledge and a predominant focus on risk assessment. It identifies research opportunities for human-centered CWS algorithms and notes that its review is limited to U.S.-based systems, some of which may not be publicly available.

  • Problem

    CWS algorithms address risk, placement, and matching amid concerns about bias, opacity, and limited context, while the review asks how they are built, which predictors are salient, and what outcomes they target.

    Source: INTRODUCTION · p. 1

  • Method

    The authors qualitatively review 50 publications on U.S. CWS algorithms using a human-centered algorithm-design framework.

    Source: INTRODUCTION · pp. 1–2

  • Results

    Most reviewed algorithms are empirically constructed despite fragmented empirical knowledge, predictors used differ from those found salient in child-welfare literature, and the literature focuses on risk assessment over positive outcomes.

    Source: INTRODUCTION · p. 2

  • Takeaways & Limitations

    The review identifies gaps and research opportunities for human-centered algorithm design in CWS and brings child-welfare domain knowledge into SIGCHI.

    Source: INTRODUCTION · p. 2

    The review is limited to the U.S. system and may miss algorithms unavailable for public review, including internally distributed reports.

    Source: LIMITATIONS AND FUTURE WORK · p. 10

Abstract

from arXiv · show

The U.S. Child Welfare System (CWS) is charged with improving outcomes for foster youth; yet, they are overburdened and underfunded. To overcome this limitation, several states have turned towards algorithmic decision-making systems to reduce costs and determine better processes for improving CWS outcomes. Using a human-centered algorithmic design approach, we synthesize 50 peer-reviewed publications on computational systems used in CWS to assess how they were being developed, common characteristics of predictors used, as well as the target outcomes. We found that most of the literature has focused on risk assessment models but does not consider theoretical approaches (e.g., child-foster parent matching) nor the perspectives of caseworkers (e.g., case notes). Therefore, future algorithms should strive to be context-aware and theoretically robust by incorporating salient factors identified by past research. We provide the HCI community with research avenues for developing human-centered algorithms that redirect attention towards more equitable outcomes for CWS.

INTRODUCTION

The review examines how algorithms are developed and used in U.S. child welfare, where systems address risks, placements, and child–foster-parent matching amid concerns about bias, opacity, and limited context. Its synthesis finds a fragmented empirical foundation and a strong focus on risk assessment rather than positive outcomes for foster children.

Motivation

  • Algorithms address future maltreatment risk, placement recommendations, and child–foster-parent matching, but have drawn criticism for bias, opacity, reductiveness, and limited context.The passage also notes concerns that systems do not incorporate factors from relevant social-science research.Source: INTRODUCTION · p. 1

Findings and research questions

  • The review asks which methods build CWS algorithms, which predictors are salient, and which outcomes organizations predict.It also brings child-welfare domain knowledge into SIGCHI to support collaboration between the disciplines.Source: INTRODUCTION · pp. 1–2

Review scope and contribution

  • The review synthesizes 50 publications using a human-centered algorithm-design lens to examine methods, predictors, and predicted outcomes in U.S. child welfare.It also identifies literature gaps and research opportunities for more human-centered algorithm design.Source: INTRODUCTION · pp. 1–2

Findings and research questions

  • Most reviewed algorithms are empirically constructed despite fragmented empirical knowledge, and the predictors used differ from those found salient in child-welfare literature.The literature has focused on risk assessment rather than positive outcomes for foster children.Source: INTRODUCTION · p. 2

A Human-Centered Approach to Algorithm Design

The paper uses Human-Centered Algorithm Design (HCAD) as a theoretical lens, emphasizing human and social interpretations in algorithm design and evaluation. It situates this review in SIGCHI work engaging with child-welfare communities and algorithmic decision-making.

A Human-Centered Approach to Algorithm Design

  • HCAD incorporates human and social interpretations in design and evaluation; this review uses its theoretical strategy to frame its questions and qualitative analysis.The framework also names speculative and participatory strategies, while the review draws specifically on the theoretical perspective.Source: A Human-Centered Approach to Algorithm… · p. 2
  • The paper identifies a lack of human perspectives throughout CWS algorithm design as a reason to examine how HCAD strategies can address questions in the domain.The framework describes theoretical foundations, social-science concepts, and stakeholders’ interpretations as relevant to design and evaluation.Source: A Human-Centered Approach to Algorithm… · p. 2

SIGCHI Research to Support the Child Welfare System

  • Prior SIGCHI work with CWS families, workers, and service providers found discomfort with algorithmic systems, including concerns about deficit-based frameworks and biased data.That work recommended researcher–designer collaboration with public-service agencies; this review investigates existing CWS algorithms and research opportunities.Source: SIGCHI Research to Support the Child… · p. 2

Sociotechnical Systems for Child-Welfare

The paper situates CWS algorithms in data systems that support decision-making and actuarial risk assessment, while noting concerns about overriding caseworker judgment. It reviews computational systems and proposals through a systematic search and qualitative analysis.

Sociotechnical Systems for Child-Welfare

  • SACWIS and related federal databases support data collection and curation, forming a foundation for mandated actuarial risk tools whose authority over caseworker judgment remains controversial.Earlier surveys examined reliability and validity but not algorithms’ construction, variables, or predicted outcomes, which matter in contextually different cases.Source: Sociotechnical Systems for Child-Welfare · pp. 2–3

Scoping Criteria: Defining Algorithms

  • The authors use “street-level algorithms” for systems making on-the-ground decisions about human lives and welfare, and include statistical modeling and machine learning.They distinguish statistical modeling from machine learning based on assumptions about the data.Source: Scoping Criteria: Defining Algorithms · p. 3

METHODS

Systematic Literature Search

  • The search covered multiple digital libraries and cross-referenced citations, imposed no publication-period constraint, and identified 50 eligible articles.Eligibility required peer-reviewed work or a government systems or policy report with technical discussion of computational methods, predictors, and outcomes.Source: Systematic Literature Search · p. 3

Data Analysis Approach

The authors coded the corpus using a structured qualitative process and report descriptive characteristics of the reviewed work. The corpus was largely published in social-science venues, and many studies used psychometric scales or reported model performance.

Data Analysis Approach

  • The authors used grounded thematic coding, defined theory by framework use or theoretically grounded factors, and consulted coauthors to resolve ambiguous codes.The coding addressed the review questions and descriptive characteristics of the article set.Source: Data Analysis Approach · p. 3

RESULTS

  • The results section organizes findings by the three research questions and relates computational methods, predictors, and outcomes.The codebook maps its dimensions to the research questions.Source: RESULTS · p. 3, Table 2 · p. 4

Descriptive Characteristics of the Data Set

  • 40 papers (80%) appeared in social-science venues; 30 (60%) used psychometric scales, 27 (54%) discussed an actual system, 23 (46%) proposed one, and 35 (70%) reported model performance.Three papers (6%) appeared in computer-science conferences, all in 2018; four studies proposed theoretically driven solutions.Source: Descriptive Characteristics of the… · pp. 3–4

Computational Methods used to build Algorithms (RQ1)

CWS algorithms have used inferential statistics more often than machine learning, with GLMs especially common. The review also identifies limits involving case context, severe outliers, and transparency.

Inferential Statistics approaches

  • Inferential statistics accounted for 68% of papers, including 28 papers (56%) using a generalized linear model (GLM).The review groups computational methods into inferential statistics and machine learning.Source: Inferential Statistics approaches · p. 4, Computational Methods used to build… · p. 4
  • GLMs built from statistically significant psychometric-scale items can omit case-specific context and contribute to variable omission bias.The CANS Algorithm is one example of a model that uses only the scale’s most statistically significant items.Source: Inferential Statistics approaches · p. 4
  • Severe abuse and neglect are statistical outliers, and models aimed at average outcomes tend to perform poorly on these cases.The review links this poor performance to ethical and accountability concerns.Source: Inferential Statistics approaches · p. 4
  • Four papers (8%) used discriminant analysis; its data were limited and of questionable quality amid a lack of standardized collection.It was popular during 1985–1990, then faded as regression techniques emerged.Source: Inferential Statistics approaches · p. 4

Machine Learning (ML) approaches

  • Machine-learning methods appeared as early as 1986, resurged from 2015, and were used in 13 papers (26%).The review reports growing computer-science interest in CWS research from 2018; Figure 1 depicts method trends.Source: Machine Learning (ML) approaches · p. 4, Figure 1 · p. 5
  • Machine-learning applications included child-foster-parent matching, while Structured Decision-Making pairs risk tools with clinical assessment and has produced mixed results across race and ethnicity.The review also notes tension between caseworkers’ theoretical assessments and the tool’s empirical judgment.Source: Machine Learning (ML) approaches · p. 5
  • Three papers (6%) used unsupervised methods, including neural networks and natural language processing; neural networks were proposed for modeling nonlinear relationships and interactions.The review notes that social-science neural-network use requires transparency about proposed decisions.Source: Machine Learning (ML) approaches · p. 5
  • Although Naive Bayes and Random Forests achieved higher accuracies, Vaithianathan et al. reverted to probit regression because its outcomes were more explainable.The review describes this as a trade-off between accuracy and explainability.Source: Machine Learning (ML) approaches · p. 5

Predictors used in Algorithms (RQ2)

The review coded predictor variables into seven dimensions. Demographic and systemic factors were comparatively infrequent in the algorithms examined.

Predictors used in Algorithms (RQ2)

  • Because most algorithms used over a hundred predictors, the review systematically coded them and grouped emergent codes into seven dimensions.The dimensions were developed through the review’s coding process.Source: Predictors used in Algorithms (RQ2) · p. 5

Demographics and Systemic Factors

  • Child demographics appeared in 20 papers (40%), while biological-parent demographics appeared in 10 (20%); more than half omitted each type.Figure 2 shows a decline in demographic-variable use after 1990.Source: Demographics and Systemic Factors · p. 5
  • Only two papers modeled agency characteristics and one modeled caseworker characteristics, despite caseworker turnover being directly associated with placement instability.The review identifies systemic factors as an important contributor to multiple placement moves.Source: Demographics and Systemic Factors · p. 5

Foster-child related factors

The review describes child needs and risks, biological-parent factors, and foster-parent factors as predictors used in CWS algorithms, while noting gaps in coverage.

Foster-child related factors

  • Child strengths appeared in 11 papers (22%), emotional and behavioral needs in 26 (52%), and day-to-day well-being and functioning in 15 (30%).These measures included skills and optimism, behavioral needs, school attendance, hygiene, and communication.Source: Foster-child related factors · p. 5
  • Risk-assessment systems focused heavily on child risks, whereas placement-recommendation systems focused on child needs.Figure 2 depicts increased use of child strengths, risks, and needs since 1995.Source: Foster-child related factors · pp. 5–6, Figure 2 · p. 6
  • Only one study accounted for children’s interactions with siblings, relatives, agencies, or caseworkers, although systemic or policy factors accounted for 70% of placement moves in one San Diego County study.Moore et al. included factors such as sibling placement and proximity to home or relatives.Source: Foster-child related factors · p. 6

Biological parents related factors

  • Biological-parent risks and needs appeared in 26 papers (52%); risk models used these factors to assess future maltreatment, while placement models used them to gauge trauma and need.Figure 2 shows a decline in biological-parent factors during 2005–2010 and a rise in foster-child factors during that period.Source: Biological parents related factors · p. 6

Foster parents related factors

  • Four papers (8%) modeled foster-parent characteristics, and the review found very little algorithmic research on factors linked to stability and smoother transitions.The cited literature connects trained, prepared, culturally matched foster parents with better placement outcomes.Source: Foster parents related factors · p. 6

Target Outcomes of Algorithms (RQ3)

  • The target-outcomes section uses Figure 3 to depict trends in the outcomes CWS algorithms have sought to model.The supplied passage identifies the figure’s scope but does not report the trends’ values.Source: Target Outcomes of Algorithms (RQ3) · p. 6

Risk Assessment

Risk assessment is the most-studied CWS outcome, while placement recommendations and child-foster parent matching offer distinct ways to address children’s care needs. The section describes limits in risk tools and a shift toward approaches that also consider placement fit and stability.

Risk Assessment

  • 28 papers (56%) focused on risk assessment, which has received more attention than any other outcome since regression models were introduced.These actuarial models select factors associated with abuse or neglect and apply them to new alleged cases; they are probabilistic rather than causal and require follow-up validation, while direct model comparisons depend on predictor context and weighting.Source: Risk Assessment · p. 6, Figure 3 · p. 7
  • Risk tools can predict maltreatment more accurately than unaided judgment but lack utility; SDM combines predictive and contextual assessments, yet tool–caseworker disagreements persist.Caseworkers may detest using the tools as intended, but continue relying on them to standardize decisions in high-uncertainty cases.Source: Risk Assessment · pp. 6–7

Placement Recommendations and Successful Placements

  • 20 papers (40%) discussed placement recommendation systems, while studies of the two precursor outcomes stopped appearing after 1990 until 2005.The paper offers increased attention to risk assessment as a plausible explanation for that gap.Source: Placement Recommendations and… · p. 7
  • CANS recommends one of six increasingly intensive care levels based on a child’s needs, using a multidisciplinary team for lower-uncertainty cases but not aiming to improve quality of life.Six papers (12%) discuss CANS, which is developed using the CANS psychometric scale.Source: Placement Recommendations and… · p. 7

Child-Foster Parent Matching

  • Child-foster parent matching pairs children’s needs with caregiver capabilities and preferences; it has been implemented since 2015 in two studies and is associated with stability and permanency outcomes.A Kansas implementation was validated as resulting in more stable placements, while the paper notes limited guidance for computationally modeling this approach within CWS.Source: Child-Foster Parent Matching · p. 7

Relationship between Methods, Predictors and Outcomes

The reviewed methods and predictors vary by target outcome: risk assessment remains common across computational approaches, while placement recommendation models weigh children’s strengths and needs alongside risks. The predictor sets for the two outcomes differ in emphasis.

Relationship between Algorithms (RQ1) and Outcomes (RQ3)

  • 16 studies used generalized linear models for risk assessment and 8 for placement recommendations; supervised machine learning was used for risk assessment in 9 studies.Two studies each used supervised machine learning for placement recommendation and child-foster parent matching.Source: Relationship between Algorithms (RQ1)… · p. 7

Relationship between Predictors (RQ2) and Outcomes (RQ3)

  • Risk models often used traumatic-experience predictors (15 studies), alongside behavioral/emotional needs and risk behaviors (7 studies each), and sometimes reused predictors for child needs and parent risks.The review criticizes aggregating negative aspects while ignoring positive ones.Source: Relationship between Predictors (RQ2)… · pp. 7–8, Table 4 · p. 8
  • Placement models used predictors of child strengths (5 studies), functioning (10), and behavioral/emotional needs (13), balancing them with risk behaviors (12) and traumatic experiences (10).They are used when a child needs a permanent foster-care placement, rather than at the start of an investigation.Source: Relationship between Predictors (RQ2)… · p. 8, Table 4 · p. 8

Relationship between Methods (RQ1) and Predictors (RQ3)

Across methods, models often draw on a narrow set of child needs, risks, and trauma predictors that may omit systemic factors. The review finds limited theoretically grounded work and argues for context-aware approaches that consider practice and caseworker interpretation.

Relationship between Methods (RQ1) and Predictors (RQ3)

  • Supervised machine learning and generalized linear models often use behavioral/emotional needs, risk behaviors, and traumatic experiences, but this narrow predictor set may omit systemic CWS risk factors.These predictors may become available after caseworker investigation and psychometric assessment.Source: Relationship between Methods (RQ1) and… · p. 8

Algorithms Need to be theoretical & context-aware (RQ1)

  • Only one study built its model on child-welfare literature, and four proposed theory-driven alternatives that have not been consistently implemented.The authors describe empirical child-welfare knowledge as fragmented and decisions as constrained by policies and scarce resources.Source: Algorithms Need to be theoretical &… · p. 8
  • The authors recommend human-centered theoretical approaches grounded in evidence-based social work to incorporate salient factors and understand decision-making pathways.They argue that such approaches can help reveal practitioners’ sense-making and inform feature selection and evaluation.Source: Algorithms Need to be theoretical &… · pp. 8–9
  • Several states’ SDM approaches fell short of integrating predictive and contextual assessments; detailed case notes may provide context-specific signals for methods such as topic modeling.The notes are unstructured text stored in SACWIS systems, and the authors hypothesize they can provide theoretical signals for predictive modeling.Source: Algorithms Need to be theoretical &… · p. 9

Going Beyond what is "Easily Quantifiable" (RQ2)

Current CWS models often rely on easily available child and parent characteristics, including predictors with weak validation and no accounting for risk over time. The paper argues for theoretically grounded, stakeholder-informed approaches that define outcomes beyond risk reduction and attend to foster children’s quality of life.

Going Beyond what is "Easily Quantifiable" (RQ2)

  • Most models use readily available, quantifiable child and parent characteristics, although some common predictors lack predictive validity and others are insufficiently validated.The authors note that unvalidated predictors such as parenting skills and parent conflict can yield unreliable predictions.Source: Going Beyond what is "Easily… · p. 9
  • Risk assessments do not account for temporality; treating post-allegation escalation as the baseline can inflate scores and lead to excessive interventions.Source: Going Beyond what is "Easily… · p. 9
  • Theoretical feature selection can prioritize predictors validated in social sciences, while case-note analysis may capture stakeholders and distinguish empirically similar cases with different outcomes.The authors point to natural language processing as a way to quantify case notes and make holistic inferences about stakeholders.Source: Going Beyond what is "Easily… · p. 9
  • Participatory design can engage domain experts and stakeholders to examine risk accumulation and systemic factors, while new data systems may collect well-being and systemic predictors.Active attention to stakeholder needs and values may help avoid reliability and validity pitfalls in new predictors.Source: Going Beyond what is "Easily… · p. 9

Improve Lives and not just ‘Minimize Risk’(RQ3)

  • Risk models aim to minimize future harm rather than improve foster children’s quality of life, and “risk of maltreatment” is poorly defined across differing state laws and intervention contexts.Training labels based on substantiation reflect judgments shaped by state policies, intervention criteria, funding, and caseloads; human-centered approaches can help define predictors and outcomes with stakeholders.Source: Improve Lives and not just ‘Minimize… · p. 9
  • Deficit-based models emphasize negative inputs and outcomes, while the authors call for strength-based approaches and child-foster parent matching attentive to families’ values and needs.Source: Improve Lives and not just ‘Minimize… · p. 10

Recommendations for Future Research

The paper recommends participatory, theoretically grounded design and a shift from risk mitigation toward equitable outcomes that improve foster children’s lives. It also notes that its review covers U.S. systems and may omit algorithms not publicly available.

Bridging AI and HCI Through Participatory Design

  • Only one reviewed study engaged child-welfare workers; participatory design in design and evaluation can address worker needs and system fit, including legal constraints on data use.The paper gives juvenile criminal records and histories of physical or sexual abuse as data that legal requirements might exclude from modeling.Source: Bridging AI and HCI Through… · p. 10

Algorithmic Decision Making via Speculative Design

  • 56% of studies took a deficit-based approach to mitigating risks, despite child-welfare literature discussing equitable outcomes such as child-foster parent matching.The authors note that matching has been discussed in the field for almost two decades but was only recently explored in an algorithm, following advances in decision-tree learning.Source: Algorithmic Decision Making via… · p. 10
  • The authors recommend speculative design to help stakeholders imagine problems and solutions beyond current technical feasibility and challenge the status quo of risk-focused systems.Source: Algorithmic Decision Making via… · p. 10

LIMITATIONS AND FUTURE WORK

  • The review is limited to U.S. child-welfare systems and may have missed algorithms unavailable for public review, including internal agency or government reports.Source: LIMITATIONS AND FUTURE WORK · p. 10

CONCLUSION

  • The conclusion calls for stakeholder-engaged, theoretically designed algorithms using comprehensive, well-studied and hard-to-quantify predictors to improve foster children’s lives rather than merely mitigate risk.Source: CONCLUSION · p. 10
Loading 2003.03541v1…