Skip to content
Search resources Open menu

Topics of Intimate Partner Violence

Child Physical Abuse Risk Assessment

Published

Snap Shot

  • The question: What do child-welfare risk tools and parent screening measures actually predict, and how much weight can a “high risk” or “low risk” label bear?
  • What the research shows: In the studies reviewed, actuarial tools outperform unaided clinical judgment, but accuracy is moderate: a meta-analysis of 30 studies covering 27 instruments found an overall AUC of 0.681, meaning substantial overlap between families who do and do not have further involvement. Safety assessment (present danger) and risk assessment (future agency contact, typically over 18 to 24 months) answer different questions.
  • What it does not show: Actuarial child-welfare tools are built to predict future agency events, such as new reports, substantiations, or placements, not abuse itself, and most combine all maltreatment types. A risk score does not by itself show that a parent caused a particular injury. The CAP Inventory is a screening tool that does not determine whether abuse occurred.
  • The law: Risk opinions face Rule 702 reliability review (Daubert, Kumho Tire, Bernstein). Dependency and termination require proof of conduct and present danger under preponderance and clear-and-convincing standards (Santosky; A.R.S. §§ 8-844, 8-863), not actuarial scores.
  • Issues for each side: Prosecutors and agency attorneys can rely on properly scored tools but should not present a risk category as evidence that abuse occurred; defense counsel should ask what the tool predicts and how it was validated, without dismissing structured tools as worthless or claiming bias without local data.

Child-protection agencies make thousands of decisions each day about which reports to investigate, whether a child can safely remain at home, and which families should receive services. Many of those decisions are now guided by structured risk assessment tools, and the resulting scores and classifications increasingly appear in dependency petitions, case plans, custody evaluations, and expert reports. Attorneys need to understand what these tools are designed to predict, how accurate they are, and where their limits lie, because a “high risk” classification can carry more weight in a courtroom than the underlying research supports, and a “low risk” classification can provide false reassurance.

This paper explains the difference between risk assessment and safety assessment, describes the principal tools used in U.S. child welfare and in forensic evaluations of parents, and summarizes the evidence on their predictive validity, including the important point that actuarial child-welfare tools predict future agency contact rather than abuse itself. It then reviews the critiques concerning racial disparities, the use of these tools in family court, and the legal framework for admissibility and due process as of November 2026. (The HCR-20, a general violence risk instrument sometimes used with parents, is addressed in a separate paper in this series.)

Key Concepts

Risk assessment and safety assessment

Child welfare practice distinguishes two questions. Safety assessment asks whether a child faces present or impending danger of serious harm right now, and what must happen immediately, including removal, to protect the child. It is a short-term, yes-or-no judgment focused on current conditions. Risk assessment asks how likely it is that a family will come to the agency’s attention again for maltreatment over a longer period, typically 18 to 24 months, and is used to decide which families should receive ongoing services and how intensive those services should be. A family can be judged safe today and still high risk over the next two years, or unsafe today because of an acute crisis but low risk over time. Confusing the two questions is a common source of error.

Actuarial, consensus, and clinical approaches

An actuarial tool combines a small number of items, chosen because they were statistically associated with the outcome in a development sample, into a score that places a family in a risk category. A consensus-based tool uses items chosen by expert agreement rather than by statistical testing. Unstructured clinical judgment is a caseworker’s or clinician’s overall impression without a formal tool. Structured professional judgment falls in between: the evaluator considers a list of research-based factors but forms a final judgment rather than adding up a score.

What “predictive validity” means

Predictive accuracy is usually reported as the area under the curve (AUC), the probability that a randomly chosen family that went on to have the outcome scored higher than a randomly chosen family that did not. An AUC of 0.50 is chance and 1.0 is perfect; values around 0.70 are conventionally described as moderate. AUC describes ranking ability across a group. It does not tell a court the probability that a specific family will maltreat, and it does not convey how many families classified as “high risk” will in fact have no further involvement.

The Principal Tools

Structured Decision Making and the California Family Risk Assessment

The Structured Decision Making (SDM) system, developed by the Children’s Research Center of the National Council on Crime and Delinquency (now Evident Change), is a widely used family of child-welfare decision tools in the United States. It includes separate instruments for screening reports, assessing immediate safety, assessing future risk, reassessing risk during services, and assessing family strengths and needs. The SDM risk assessment is actuarial; jurisdictions adapt and validate it on local data. The California Family Risk Assessment (CFRA) is California’s version.

An early comparison found that actuarial risk tools classified families by later outcomes more accurately than consensus-based tools then in use (Baird & Wagner, 2000). A field study of the CFRA followed 7,685 reports in five California counties for two years (Johnson, 2011). The tool showed imperfect but better-than-chance predictive validity for later substantiated maltreatment, while caseworkers’ unaided clinical risk judgments, available in a small subsample, did not predict outcomes. Workers overrode the tool in only 1.5 percent of cases. The study illustrates both the case for actuarial tools, that unstructured judgment performed poorly in that sample, and their limits, that “better than chance” is a modest standard.

Meta-analytic evidence across tools

A meta-analysis of 30 studies (87,329 cases) covering 27 different child maltreatment risk instruments found an overall AUC of 0.681, which the authors characterized as moderate (van der Put et al., 2017). Actuarial instruments outperformed clinical ones, and first occurrence of maltreatment was predicted somewhat better than recurrence. An AUC near 0.68 means substantial overlap between families who do and do not go on to have further involvement; an AUC in this range implies that some families classified as high risk will not have a new substantiated report and some classified as low risk will, though this paper does not report category-level rates.

Predictive risk modeling

Some agencies now use predictive risk models that apply machine-learning methods to large administrative databases (prior CPS contacts, public benefits, criminal justice, and behavioral health records) to score incoming reports automatically. The best-known example is the Allegheny Family Screening Tool in Pittsburgh, which generates a score at the time of referral indicating the relative likelihood of later foster care placement, to help hotline staff decide which reports to investigate (Vaithianathan et al., 2020). Proponents argue that such models can make screening more consistent than unaided judgment (Cuccaro-Alamin et al., 2017). One validation study found that children scored in the highest risk group had higher rates of later hospital encounters for injury, including abuse-related injury, than children scored low risk, and that the scores were unrelated to cancer encounters, which served as a placebo outcome (Vaithianathan et al., 2020). Supporters see this as evidence that the model captures genuine danger and not only system contact; the authors of that study included developers of the model, and independent replication was not located for this paper. Critics raise concerns about transparency, the use of data that disproportionately describe poor families who rely on public systems, and the risk of reinforcing existing patterns of surveillance.

The Child Abuse Potential Inventory

The Child Abuse Potential (CAP) Inventory is a 160-item self-report questionnaire, developed by Joel Milner, that screens parents for characteristics associated with known physical abusers, such as distress, rigidity, and problems with family and others. It includes validity scales intended to detect faking good, faking bad, and random responding. Unlike the CPS tools, it is used mainly by psychologists, including in forensic parenting evaluations. A 2025 meta-analysis of the CAP literature found that CAP abuse scores were related to most of the theory-based risk factors for child physical abuse that the authors examined, supporting its construct validity, meaning that it measures what it is intended to measure (Milner et al., 2025). The developer was an author and receives royalties, which the article disclosed. Construct validity is not the same as accuracy in predicting future abuse by an individual parent, and the CAP’s own validity scales cannot eliminate the motivation of parents in litigation to present themselves favorably. The CAP is a screening tool. It does not determine whether abuse occurred.

Other measures

Evaluators may also encounter general parenting-stress measures, observational ratings of parent-child interaction, and IPV-focused tools that include items about children, such as versions of the Danger Assessment. These measure related constructs but are not validated predictors of child physical abuse and should be described accordingly.

What the Tools Actually Predict

The outcome that actuarial child-welfare tools are built and validated to predict is almost never abuse itself. It is a future agency event: a new report, a new substantiation (a caseworker’s finding that maltreatment occurred), a foster care placement, or a re-entry into services. Those events depend not only on parental behavior but on who is watching the family, who reports, how caseworkers decide, and how often the family has contact with public systems. Families already known to an agency are more visible, a point illustrated in the intergenerational-transmission research, where in one 30-year prospective study much of the apparent continuity in official records disappeared when parents’ and children’s own reports were examined (Widom et al., 2015).

Contact with CPS is far more common than annual figures suggest. An estimated 37.4 percent of U.S. children experience a CPS investigation by age 18, including 53.0 percent of Black children (Kim et al., 2017), and about 12.5 percent experience a confirmed maltreatment finding (Wildeman et al., 2014). County-level estimates for the 20 most populous U.S. counties put the median investigation rate at 34.5 percent, with high rates of later-stage involvement, such as confirmed maltreatment, foster care, and termination of parental rights, in Maricopa County, Arizona (Edwards et al., 2021). Because these outcomes are common and unevenly distributed, a tool that predicts them may partly reflect exposure to the system.

Most tools also combine all maltreatment types. Neglect is by far the most common substantiated category (in federal fiscal year 2024, 63.3 percent of victims experienced neglect only; HHS, 2026), and the items that predict neglect (poverty-related stressors, housing instability, prior reports) differ from those most relevant to inflicted physical injury. A family’s overall “risk level” therefore says little about the specific likelihood of serious physical abuse.

Finally, an actuarial tool says nothing about whether past abuse occurred. A risk score does not by itself show that a parent caused a particular injury, and using it that way reverses the logic of the instrument.

Validity and Fairness Critiques

Racial disparities

Black and American Indian or Alaska Native children are reported, investigated, and substantiated at higher rates than White children; in federal fiscal year 2024, victimization rates were 13.3 per 1,000 children for American Indian or Alaska Native children and 11.8 per 1,000 for Black children, against 7.2 per 1,000 nationally (HHS, 2026). The reasons are contested. One view holds that the disparity largely reflects real differences in exposure to risk factors linked to poverty and structural disadvantage. A national analysis found that racial differences in CPS victimization rates were consistent with differences in other child outcomes not subject to reporter bias, such as infant mortality, which the authors read as supporting that view over reporter bias as the main explanation (Drake et al., 2011). A study of 1,461 investigations found no Black-White differences in case outcomes after adjusting for case characteristics (Font et al., 2012). The opposing view argues that disparities reflect a combination of bias at multiple decision points and structural inequities that shape the risk factors themselves (Dettlaff & Boyd, 2020). Medical research shows that bias can affect individual decisions: in one hospital, minority toddlers with fractures were more likely than White toddlers to receive skeletal surveys and CPS reports even after controlling for the likelihood of abuse (Lane et al., 2002), and a multisite study found disparities in evaluation and reporting for suspected abusive head trauma concentrated among lower-risk patients at a small number of sites (Hymel et al., 2018).

For risk tools, the concern is that items such as prior reports or prior CPS history may encode past disparities, so a tool can be accurate on average and still produce higher false-positive rates for some groups. Whether a particular tool does so is an empirical question that depends on local validation data. Attorneys can ask whether a tool has been validated in the local population and whether its accuracy and classification rates have been examined by racial and ethnic group.

Other limitations

Tools validated in one jurisdiction may perform worse elsewhere. Scores depend on the accuracy of the underlying information, which may be incomplete early in an investigation. Overrides, in either direction, are common sources of disagreement and should be documented. Risk categories are relative, and a “high” category may correspond to a modest absolute probability of a new substantiated report. A tool’s performance on aggregate data does not establish its accuracy for any particular family.

Use in Family and Dependency Courts

CPS risk scores enter court in several ways: as part of the agency’s narrative in a dependency petition, as support for a case plan, as a basis for recommendations about reunification, and occasionally through records subpoenaed in custody litigation. Forensic psychologists may use measures like the CAP within broader parenting evaluations. In each setting, the court should be told what the instrument predicts, the population on which it was validated, its accuracy, and the difference between a statement about group risk and a finding about a specific parent. Agency risk tools were designed chiefly as decision aids for prioritizing cases and allocating services, and the studies cited here validated them against later agency records, not against findings about past conduct.

Limits of the Evidence

Most validation studies measure agreement between a score and a later agency record, so they cannot separate a family’s behavior from how closely the family was watched. Accuracy figures are averages across tools, jurisdictions, and definitions; several studies cited here come from single states or counties; and the developers of some instruments took part in the research on them. Evidence on how tools perform for particular racial and ethnic groups is mixed and often local.

Questions to Ask About This Research

When a tool’s accuracy is described as “validated,” ask what outcome and what AUC: across 30 studies and 27 instruments the overall AUC was 0.681 (van der Put et al., 2017), the outcomes were mostly future reports or substantiations, and the figure describes group ranking, not the chance that this family will maltreat. When a field study of one tool is cited, ask how large and local the sample was, such as 7,685 reports in five California counties (Johnson, 2011), and whether the same tool was validated where the case arose. When a risk category is used to suggest abuse occurred, ask what the tool was built to predict: agency contact, which is far more common than annual counts suggest (37.4 percent of children investigated by age 18; Kim et al., 2017). When a CAP score is offered, ask what the meta-analytic evidence supports, which is construct validity rather than prediction for an individual parent (Milner et al., 2025), and how self-presentation in litigation was addressed. These questions are explained in “Reading Research Claims: A Guide for Attorneys” on this site.

Child Physical Abuse Risk Assessment and the Law

Admissibility of risk tools and expert opinion

When a risk score or a risk-based expert opinion is offered at trial or in a contested hearing, the admissibility question is governed in federal court and in Arizona by Rule 702. Federal Rule 702 adopts the reliability inquiry of Daubert v. Merrell Dow Pharmaceuticals, Inc. (1993), which Kumho Tire Co. v. Carmichael (1999) extended to all expert testimony, including the technical and experience-based judgment typical of social work and clinical practice. The 2023 amendment makes explicit that the proponent must establish each requirement by a preponderance of the evidence and that each opinion must stay within what the method reliably supports, a point with special force for testimony translating a group-based risk category into a statement about an individual. Arizona amended its Rule 702 in 2012 to conform to the federal rule and follow Daubert (State v. Salazar-Mercado, 2014), and again effective January 1, 2024 (Order R-23-0004). Under State v. Bernstein (2015), the trial court’s gatekeeping role includes whether a reliable method was reliably applied to the facts, though only errors serious enough to make the results unreliable require exclusion, and lesser flaws go to weight. In a risk-assessment context, that framework invites questions about whether the tool was scored correctly, on accurate information, and on a population like the one in which it was validated.

Many dependency hearings are tried to the bench and, depending on the hearing and jurisdiction, may operate under relaxed evidentiary rules. Even then, the same reliability concerns bear on the weight a court gives to a risk classification, and any party can raise them through cross-examination of the caseworker or evaluator.

Due process in dependency proceedings

Parents have a constitutionally protected interest in the care and custody of their children. The Supreme Court has held that termination of parental rights requires at least clear and convincing evidence (Santosky v. Kramer, 1982), and that due process requires appointed counsel for indigent parents in termination proceedings in some but not all cases, to be decided case by case (Lassiter v. Department of Social Services, 1981). Arizona requires a preponderance of the evidence for dependency adjudication (A.R.S. § 8-844(C)) and clear and convincing evidence for termination grounds (A.R.S. § 8-863(B)). Arizona law permits taking a child into temporary custody without a court order only if clearly necessary because of exigent circumstances, meaning probable cause that the child is likely to suffer serious harm in the time it would take to obtain an order and, in most circumstances, that no less intrusive alternative would reasonably and sufficiently protect the child (A.R.S. § 8-821). These standards turn on present danger and proof of past conduct, which are the province of safety assessment and fact-finding, not of actuarial risk scores.

Due process concerns arise when a risk classification functions as a substitute for evidence: for example, when a “high risk” score is used to justify removal without findings about present danger, or when a parent cannot learn what items drove the score. Predictive risk models raise an additional transparency question, since the variables and weights may not be readily explainable. Courts have not developed a body of law specific to child-welfare risk tools, as far as the research for this paper found, but general principles about notice, the opportunity to be heard, and the reliability of evidence apply.

Arizona definitions

Arizona defines abuse, neglect, and dependency in A.R.S. § 8-201, and criminal child abuse in A.R.S. § 13-3623. Neither definition refers to risk scores; both require proof of conduct, injury, or substantial risk of harm. A risk assessment may inform service planning under these statutes, but the legal elements are established by evidence about what happened and what danger currently exists.

Special Issues in Legal Practice

The principles described above apply with equal force to every party in a legal proceeding. The science operates the same way regardless of who is asking the questions, and the research summarized here does not by itself favor either side. What differs between the parties is not the science but the vantage point — the particular places where these issues tend to surface, and the particular errors each party is best positioned to guard against. The two summaries below identify those recurring issues. They are offered for orientation, not as litigation strategy; in every case the controlling question is what the evidence in that case will, and will not, support, and the expert’s role is to inform the trier of fact rather than to advocate for either side.

Special Issues for Prosecutors

Prosecutors and agency attorneys most often meet risk tools in the agency’s own file. The research supports using structured tools over unaided judgment: in the studies reviewed, actuarial instruments outperformed clinical impressions, and caseworkers’ unstructured risk judgments, available for a small subsample, did not predict outcomes in the California field study. A properly completed safety assessment can document why immediate protective action was or was not taken, and a risk reassessment can show how a family’s circumstances changed during services. These records are most persuasive when the tool was scored on accurate information and overrides are explained.

The research does not support presenting a risk category as evidence that abuse occurred, describing a “high risk” family as likely to inflict serious injury, or treating a moderate AUC as precise individual prediction. Because these tools predict future agency involvement, which is shaped by surveillance and reporting patterns, a risk score is weakest where the legal stakes are highest, in establishing what a specific parent did or will do. Questions about local validation and performance across racial and ethnic groups are foreseeable and are best addressed by knowing the tool’s documentation in advance.

Special Issues for Defense

Defense counsel and parents’ attorneys meet the same tools from the other side, often as a “high risk” label in a petition or case plan. The research supports asking what outcome the tool predicts, in which population it was validated, how accurate it is, and how many families in the same category had no later substantiated report. It supports examining whether the score rested on accurate and complete information, whether items such as prior reports reflect unsubstantiated contacts, and whether the safety and risk questions have been conflated.

The research does not support dismissing structured tools as worthless: on average they predicted better than unaided judgment in the studies reviewed, and the alternative to a tool is often an undocumented impression. Nor does it support asserting that a tool is racially biased without local data, since national studies reach conflicting conclusions about the sources of disparity. For self-report measures such as the CAP, a low score does not establish that a parent poses no risk, particularly in litigation, where favorable self-presentation is expected. Experts on either side are most useful when they explain these limits precisely rather than characterizing any instrument as categorically reliable or unreliable.

This paper is for educational purposes only and is not legal advice. The author is a psychologist, not an attorney; summaries of law are general and must be verified against current authority or with a licensed attorney.

References

Baird, C., & Wagner, D. (2000). The relative validity of actuarial- and consensus-based risk assessment systems. Children and Youth Services Review, 22(11–12), 839–871. https://doi.org/10.1016/S0190-7409(00)00122-5

Cuccaro-Alamin, S., Foust, R., Vaithianathan, R., & Putnam-Hornstein, E. (2017). Risk assessment and decision making in child protective services: Predictive risk modeling in context. Children and Youth Services Review, 79, 291–298. https://doi.org/10.1016/j.childyouth.2017.06.027

Dettlaff, A. J., & Boyd, R. (2020). Racial disproportionality and disparities in the child welfare system: Why do they exist, and what can be done to address them? The Annals of the American Academy of Political and Social Science, 692(1), 253–274. https://doi.org/10.1177/0002716220980329

Drake, B., Jolley, J. M., Lanier, P., Fluke, J., Barth, R. P., & Jonson-Reid, M. (2011). Racial bias in child protection? A comparison of competing explanations using national data. Pediatrics, 127(3), 471–478. https://doi.org/10.1542/peds.2010-1710

Edwards, F., Wakefield, S., Healy, K., & Wildeman, C. (2021). Contact with Child Protective Services is pervasive but unequally distributed by race and ethnicity in large US counties. Proceedings of the National Academy of Sciences, 118(30), Article e2106272118. https://doi.org/10.1073/pnas.2106272118

Font, S. A., Berger, L. M., & Slack, K. S. (2012). Examining racial disproportionality in child protective services case decisions. Children and Youth Services Review, 34(11), 2188–2200. https://doi.org/10.1016/j.childyouth.2012.07.012

Hymel, K. P., Laskey, A. L., Crowell, K. R., Wang, M., Armijo-Garcia, V., Frazier, T. N., Tieves, K. S., Foster, R., Weeks, K., & Pediatric Brain Injury Research Network (PediBIRN) Investigators. (2018). Racial and ethnic disparities and bias in the evaluation and reporting of abusive head trauma. The Journal of Pediatrics, 198, 137–143.e1. https://doi.org/10.1016/j.jpeds.2018.01.048

Johnson, W. L. (2011). The validity and utility of the California Family Risk Assessment under practice conditions in the field: A prospective study. Child Abuse & Neglect, 35(1), 18–28. https://doi.org/10.1016/j.chiabu.2010.08.002

Kim, H., Wildeman, C., Jonson-Reid, M., & Drake, B. (2017). Lifetime prevalence of investigating child maltreatment among US children. American Journal of Public Health, 107(2), 274–280. https://doi.org/10.2105/AJPH.2016.303545

Lane, W. G., Rubin, D. M., Monteith, R., & Christian, C. W. (2002). Racial differences in the evaluation of pediatric fractures for physical abuse. JAMA, 288(13), 1603–1609. https://doi.org/10.1001/jama.288.13.1603

Milner, J. S., Ammar, J., Gibson, C., Menon, S. V., & Crouch, J. L. (2025). Construct validity of the Child Abuse Potential Inventory: A meta-analytic review. Child Abuse & Neglect, 167, Article 107586. https://doi.org/10.1016/j.chiabu.2025.107586

U.S. Department of Health and Human Services, Administration for Children and Families, Children’s Bureau. (2026). Child maltreatment 2024. https://acf.gov/cb/data-research/child-maltreatment

Vaithianathan, R., Putnam-Hornstein, E., Chouldechova, A., Benavides-Prado, D., & Berger, R. (2020). Hospital injury encounters of children identified by a predictive risk model for screening child maltreatment referrals: Evidence from the Allegheny Family Screening Tool. JAMA Pediatrics, 174(11), Article e202770. https://doi.org/10.1001/jamapediatrics.2020.2770

van der Put, C. E., Assink, M., & Boekhout van Solinge, N. F. (2017). Predicting child maltreatment: A meta-analysis of the predictive validity of risk assessment instruments. Child Abuse & Neglect, 73, 71–88. https://doi.org/10.1016/j.chiabu.2017.09.016

Widom, C. S., Czaja, S. J., & DuMont, K. A. (2015). Intergenerational transmission of child abuse and neglect: Real or detection bias? Science, 347(6229), 1480–1485. https://doi.org/10.1126/science.1259917

Wildeman, C., Emanuel, N., Leventhal, J. M., Putnam-Hornstein, E., Waldfogel, J., & Lee, H. (2014). The prevalence of confirmed maltreatment among US children, 2004 to 2011. JAMA Pediatrics, 168(8), 706–713. https://doi.org/10.1001/jamapediatrics.2014.410

Legal Authorities

Daubert v. Merrell Dow Pharmaceuticals, Inc., 509 U.S. 579 (1993).

Kumho Tire Co. v. Carmichael, 526 U.S. 137 (1999).

Lassiter v. Department of Social Services, 452 U.S. 18 (1981).

Santosky v. Kramer, 455 U.S. 745 (1982).

State v. Bernstein, 237 Ariz. 226, 349 P.3d 200 (2015).

State v. Salazar-Mercado, 234 Ariz. 590, 325 P.3d 996 (2014).

Ariz. Rev. Stat. Ann. § 8-201 (definitions; abuse, dependent child, neglect).

Ariz. Rev. Stat. Ann. § 8-821 (temporary custody; exigent circumstances).

Ariz. Rev. Stat. Ann. § 8-844 (dependency adjudication; preponderance of the evidence).

Ariz. Rev. Stat. Ann. § 8-863 (termination; clear and convincing evidence).

Ariz. Rev. Stat. Ann. § 13-3623 (child or vulnerable adult abuse).

Ariz. Sup. Ct. Order R-23-0004 (2023) (amending Ariz. R. Evid. 702, effective Jan. 1, 2024).

Ariz. R. Evid. 702.

Fed. R. Evid. 702 (as amended Dec. 1, 2023).

Related topic papers: Co-Occurrence of IPV and Child Maltreatment · IPV Risk Assessment Instruments · The HCR-20 V3 in Family Violence Cases · Infant Fractures: Abuse and Other Causes