The Push for Racial Equity in Child Welfare: Can Blind Removals Reduce Disproportionality?

Abstract

  • First quantitative analysis of a reform called “blind removals,” designed to reduce Black child disproportionality in foster care by masking demographic information in removal decisions.
  • Key findings:
    • Most disproportionality in foster care systems appears because Black children are much more likely than White children to be investigated for maltreatment initially. Conditional on being investigated, investigators remove White and Black children at similar rates.
    • Blind removals did not meaningfully reduce the small racial disparities in the removal decision, but they substantially increased the time to removal.
  • Policy implication: blind removals are unlikely to meaningfully reduce overall racial disproportionality in most systems, since the main source of disproportionality is at the investigation stage, not at the removal stage.
  • Contextual note: the paper uses Kent County (Michigan) as the primary case study, with supplementary national data from NCANDS to gauge generalizability.

Introduction and Context

  • Racial disproportionality in U.S. foster care is well-documented: Black children are over-represented despite being a minority in the general population.
  • By adulthood, Black children are disproportionately involved with child welfare—prior work notes high representation in maltreatment investigations and foster care entries.
  • Blind removals concept: inspired by blind auditions in orchestras (Goldin & Rouse, 2000). Idea is that removing demographic information from the decision-making process could reduce implicit bias in removal decisions.
  • Prior to blind removals, investigators had broad discretion to file removal petitions after determining imminent risk; under blind removals, clerks redact race and socioeconomic information before the case is reviewed by a committee of professionals.
  • Pioneered in Nassau County, NY (2010), but with data and implementation challenges; limited causal evidence from Nassau; Kent County (MI) is the main focus due to data availability and implementation in Aug 2019.
  • Debate and questions:
    • If disproportionality stems mainly from biases at the removal decision, blind removals could help.
    • If disproportionality stems from external factors (e.g., differential initial investigation rates), blind removals may have limited effects.
  • Data sources and scope:
    • MDHHS dataset: all maltreatment investigations in Michigan (2010–Mar 2021) linked to placement outcomes.
    • National NCANDS data (2018) used to compare Kent County and Michigan to national patterns and to assess generalizability.
  • Summary of contributions:
    • Demonstrates that most disproportionality arises at the investigation stage, not at removal
    • Finds no evidence that blind removals reduced the removal-rate gap by race
    • Documents substantial increase in time to removal under blind removals
    • Provides policy guidance about when (and where) such reforms might be effective

Background: Process and Data (Pre- and Post-Blind Removals)

  • Michigan/Kent County maltreatment process (pre-blind removals):
    • Intake call → screening → assignment to an investigator (rotational basis) → two decisions:
    • substantiation of allegation after investigation
    • assessment of risk under substantiation; if imminent risk, a court petition for removal is pursued (through supervisor approval)
    • Substantiation and removal decisions are discretionary and subject to supervisor/agency review.
  • Under blind removals (Kent County implementation, Aug 2019):
    • After substantiation and finding imminent risk, all relevant paperwork is redacted (race, zip code, income, school district, etc.).
    • Investigator presents a redacted case to a blind-removal meeting (10–12 professionals) including blind-removal supervisors.
    • Committee reaches a consensus on whether to file a court petition to remove; initial investigator remains involved but does not have demographic information available to the committee.
  • Data and samples used:
    • Michigan MDHHS administrative data (investigations, substantiations, removals) 2010–2021
    • Kent County pre/post data (before Aug 2019 vs after) compared to control counties in Michigan (Wayne, Oakland, Macomb, Genesee)
    • NCANDS 2018 county-level data for cross-state comparisons (removal rates by race conditional on investigation)
  • Important note on external shocks: implementation overlapped with the onset of COVID-19, complicating causal attribution; authors address this with caution and multiple inference techniques (robust SEs, Ferman & Pinto, synthetic controls in Appendix B).

Sources of Racial Disproportionality in Foster Care

  • Primary claim: disproportionality largely arises at the initial contact with the system (investigation stage) rather than the removal decision.
  • Michigan and Kent County findings (pre-2019 data):
    • Black children are much more likely to be investigated for maltreatment than White children, both in Michigan overall and in Kent County
    • Conditional on being investigated and substantiated, Black and White children have similar probabilities of removal
    • In Michigan as a whole, Black children are about twice as likely as White children to be the subject of a maltreatment investigation; removal conditional on investigation shows minimal racial gaps
    • In Kent County specifically, the share of investigations for White children is around 64% vs 61% for Black children when considering abuse vs neglect; substantiation rates: White ~25.5% vs Black ~27.2%; removal conditional on substantiation: White ~3.2% vs Black ~3.4%
  • NCANDS nationwide patterns (2018):
    • Across counties, conditional on an investigation, removal rates by race are very similar; the typical White-Black removal-rate differential is close to zero, with some counties showing modest gaps in either direction
  • Implications: if most disproportionality comes from greater likelihood of being investigated (external to the removal decision), then policies that target the removal decision (like blind removals) should have limited impact on overall disproportionality
  • Special counterfactual discussed:
    • Equalizing removal probabilities conditional on an initial investigation yields little change in disproportionality
    • Equalizing the probability of an initial maltreatment investigation by race would nearly eliminate disproportionality

Racial Disparities in Investigators’ Decisions (Evidence on Bias at Later Stages)

  • Investigators’ rate differentials by race analyzed at the individual level:
    • Define for investigator i:
    • Substantiation rate Si and removal rate Ri
    • Si^* = Si^W − Si^B, Ri^* = Ri^W − Ri^B
  • Quasi-random assignment of investigators within Kent County allows cleaner inference about racial prejudice vs. case characteristics:
    • Investigators are rotated, making the assignment approximately random within county/year strata
    • For Michigan and Kent, distributions of Si^* and Ri^* are centered around zero with slight left skew (Black children’s investigations show some higher substantiation rates by Black investigators, but removal rates show Black children are removed at higher rates regardless of investigator race)
  • Investigators by race analysis (Kent County):
    • Black investigators substantiate Black children at a higher rate than White investigators, while White investigators substantiate White children more
    • Both Black and White investigators tend to remove Black children at higher rates than White children, indicating no clear, consistent bias tied to investigator race alone
  • Implication: substantial evidence against a simple prejudice explanation for removal disparities; differences appear more tied to case characteristics and the composition of investigations rather than investigator race per se
  • Additional evidence: pre-blind removals, case characteristics show some systematic differences by race (e.g., age at investigation, likelihood of abuse vs neglect), but these differences are small and unlikely to fully explain removal disparities

Case Characteristics by Race (Systematic Differences in Investigations)

  • Michigan vs Kent County pre-blind removals (Table 4):
    • White investigations slightly older on average; Black investigations tend to involve younger children
    • Proportion of investigations for abuse vs neglect is slightly higher for Black children in both Michigan and Kent County
  • Overall interpretation: while there are some measurable differences in case characteristics by race, the patterns are not large enough to fully account for observed disproportionality at later stages; external factors and initial investigation disparities remain central

Counterfactual Disproportionality Under Equalized Rates

  • Objective: quantify how much disproportionality would change if we equalized removal odds or initial investigation odds by race
  • Disproportionality measure: share of Black children in foster care relative to share of Black children in population (stock measure) – computed as
    • Disproportionality = rac{Share{Black, FosterCare}}{Share{Black, Population}}
  • Michigan (actual) vs counterfactuals (Figure 4, Table A1):
    • If removal rates (conditional on substantiation) are completely equalized across White and Black children, disproportionality would be largely unchanged or slightly higher due to minor biases in other steps; not enough to close the gap
    • If initial maltreatment-investigation probabilities by race are equalized, disproportionality would drop dramatically toward 1 (roughly 1.03 in their counterfactual, i.e., near no disproportionality)
  • Conclusion from counterfactuals: disproportionate outcomes are driven far more by differences in the likelihood of being investigated than by differences in removal probabilities conditional on investigation
  • Additional note: the scenario where equalizing removal rates would be a sign of bias in reporting (if Black children face fewer removals despite similar substantiation), but this scenario is deemed unlikely given similar substantiation and case characteristics across races

Causal Effects of Blind Removals: Main Identification and Challenges

  • Research design for causal effects on removal decisions:
    • Kent County treated (blind removals implemented August 2019)
    • Control counties: Michigan’s four largest: Wayne, Oakland, Macomb, Genesee
    • Post-period restricted to before March 2020 to avoid confounding COVID-19 effects
    • Within-investigator removal-rate differentials (R_i^*) used to measure racially differential removal tendencies
  • Key identification assumptions:
    • The counterfactual post-change removal-rate differentials in control counties approximate Kent County’s would-be post-change differentials absent blind removals
    • Parallel trends assumption: Kent and controls would have followed similar trajectories absent treatment
  • Methods used to support inference:
    • Standard errors robust to heteroskedasticity; Ferman & Pinto (2019) p-values for small-number-of-clusters situations
    • Appendix B uses synthetic control methods to check robustness when there is a single treated unit
  • Model specification for main analysis (difference-in-differences):
    • Y{it} = eta0 + eta{ ext{Kent}} ext{Kent}i + heta{ ext{Post}} t + oldsymbol{ u} ( ext{Kent}i imes ext{Post}t) + oldsymbol{ ext{controls}} + ext{ε}{it}
    • The parameter of interest is
    • oldsymbol{
      u} = ar{R}^{ ext{Kent, post}} - ar{R}^{ ext{Kent, pre}} - (ar{R}^{ ext{Control, post}} - ar{R}^{ ext{Control, pre}})
    • In words: the treatment effect on the White-Black removal-rate differential across investigators, after accounting for pre-trends
  • Main results on removal-rate differentials:
    • Distributional evidence shows no systematic shift in R_i^* or Ri after blind removals in Kent vs controls
    • DID estimate for the removal-rate differential is small (around 0.14 percentage points) and statistically insignificant across multiple inference methods
    • Point estimates vary slightly by specification (robust SEs, clustering, synthetic controls), but overall conclusion remains: no strong evidence that blind removals reduced the White-Black removal-rate gap
  • Effects on overall removal rates (not just the differential):
    • Kent County post-blind removals shows a decline in overall removal rates compared to controls; DID suggests about a 1 percentage-point decline in Kent vs control, but this result is not statistically significant in most specifications
    • Descriptively, there is suggestive evidence of a decline in removals in Kent after implementation, but the effect is modest and cannot be conclusively attributed to blind removals

Effects on Time to Removal and Placement Outcomes

  • Time to removal (potential costs):
    • Kent County: median time from investigation start to removal increased from roughly 14 days pre-blind removals to about 26.5–27 days post
    • Control counties: median time rose only about 3 days in the same period
    • Difference-in-differences suggests an increase of around 9 days in Kent relative to controls (statistically significant at 5% using Ferman & Pinto-based inference)
    • Interpretation: blind removals add time to the removal process, potentially prolonging risk exposure or enabling better case matching, depending on the view
  • Placement outcomes (quality/appropriateness):
    • Proportion of first placements in a family setting (kinship or unrelated caregiver) did not change meaningfully post-intervention in Kent vs controls
    • Shared finding: no strong evidence that blind removals affected the likelihood of family-based vs congregate care placements for initial removals
  • Overall takeaway on costs/benefits:
    • Blind removals increased time-to-removal (a potential cost), with uncertain or no clear gains in placement quality metrics on the first placement decision

External Evidence and Generalizability (NCANDS-based comparison)

  • Across U.S. counties using NCANDS 2018 data:
    • Conditional on an investigation, the average removal-rate differential between White and Black children is near zero for most counties
    • About 9% of counties have a negative large gap (Black children removed at least 5 percentage points more often than White children, conditional on investigation) and about 5% show the opposite pattern
  • County characteristics associated with larger differentials (in NCANDS sample):
    • Counties with a smaller share of children in poverty, higher median income, and a smaller Black share tend to have larger removal-rate differentials (Black children removed more relative to White children, conditional on investigation)
  • Implication: settings with sizable Black-White removal-rate gaps exist but are relatively rare; policy relevance of blind removals may be context-dependent

Policy Implications and Practical Considerations

  • Primary implication: blind removals do not meaningfully address the main source of disproportionality in most systems, which lies at the initial investigation stage
  • If disproportionality is driven by differential reporting or external risk factors (poverty, education, family structure), then policies targeting the removal decision will have limited effects on overall disproportionality
  • In settings where there is a larger disparity at the removal decision (not just at investigation), blind removals could have more potential, but such settings appear less common in national data
  • Practical costs and trade-offs:
    • Time to removal increases, which could prolong exposure to risk or delay needed supports
    • Additional administrative steps and potential staff turnover costs; mixed evidence on whether team collaboration/how much shared decision-making improves outcomes
    • Some gains in perceived fairness and accountability may exist, but empirical evidence on downstream child outcomes remains limited
  • Recommendations for policymakers:
    • Prioritize policies that reduce disproportionality at the investigation/reporting stage (e.g., bias-reduction in reporting, equitable access to reporting channels, economic supports to reduce maltreatment risk)
    • If considering blind removals, target contexts where disproportionality primarily arises from the removal decision rather than investigation rates, and weigh time-to-removal costs
    • Continue using rigorous evaluation designs (differences-in-differences, synthetic controls) to assess causal impacts in different jurisdictions, given heterogeneous local conditions

Formulas and Key Variables (Definitions)

  • Disparities by race in decisions:

    • Substantiation rate for race r: S_r
    • Removal rate for race r: R_r
    • White-Black differentials:
    • Substantiation differential: S^* = SW - SB
    • Removal differential: R^* = RW - RB

  • Relative (conditional) probabilities by race (table-style description):

    • Pr(AllegedVictim) by race
    • Pr(Sub | AllegedVictim) by race
    • Pr(Rem | AllegedVictim) by race
    • Pr(Rem | Sub) by race

  • Investigator-level difference measures (within Kent/MI data):

    • For investigator i, substantiation rate Si and removal rate Ri
    • Si^* = Si^W − S_i^B
    • Ri^* = Ri^W − R_i^B

  • Investigator stringency measure (to test quasi-random assignment):

    • For each investigation j assigned to investigator i, define removal tendency ZR_ji as the fraction of all other investigations handled by i that resulted in foster care placement:

    • ZR{ji} = rac{1}{ni - 1}

      big|{k eq j} FC{ki}

  • Causal effect of blind removals via Difference-in-Differences (DID):

    • DID parameter:

    • u = ( ar{R}^{K post} - ar{R}^{K pre} ) - ( ar{R}^{C post} - ar{R}^{C pre} )

  • Outcome regression framework for DID (example):

    • Y{it} = eta0 + eta{ ext{Kent}} ext{Kent}i + heta{ ext{Post}} ext{Post}t +
      u ( ext{Kent}i imes ext{Post}t ) + ext{controls} + ext{ε}_{it}

  • Disproportionality stock measure (summary):

    • Disproportionality = rac{Share{Black, FosterCare}}{Share{Black, Population}}

  • Counterfactual benchmarks (Figure 4 discussion):

    • Equalizing initial investigation rates by race would reduce disproportionality substantially; equalizing removal rates conditional on investigation would have a smaller effect

Limitations and Cautions

  • The Kent County policy evaluation has limitations for causal inference:
    • Only one treated county; post-treatment period is relatively short due to COVID-19, complicating clean causal identification
    • Parallel trends assumption is tested with pre-treatment data but remains a potential concern
    • Data quality limitations in Nassau County (qualitative study) compared with Kent County’s richer administrative datasets
  • Robustness checks performed: several inference methods (robust SEs, Ferman & Pinto p-values, synthetic controls) yield consistent qualitative conclusions
  • National generalizability is limited by heterogeneity in county characteristics; NCANDS results show substantial variation across counties

Conclusion

  • The study provides the first rigorous quantitative analysis of blind removals, revealing that:
    • Most racial disproportionality in foster care arises from unequal initial maltreatment investigations, not from the removal decision itself
    • Blind removals did not meaningfully reduce the removal-rate gap by race; they did increase time to removal
    • Policy relevance is context-specific: blind removals are unlikely to be a universal solution to racial disproportionality
  • Implications for policy and future research:
    • Focus on reducing disparities at the investigation/reporting stage may yield larger reductions in disproportionality
    • If pursuing blind removals, carefully weigh the costs (time to removal, resource demands) against potential but limited benefits, and study diverse jurisdictions to identify where the policy could be more effective
  • Final takeaway: blind removals are not a one-size-fits-all fix for racial disparities in child welfare; more promising avenues involve addressing external factors driving initial disparities and improving equity in reporting and investigation practices

References and Context (selected from the article)

  • Goldin, C., & Rouse, C. (2000). Orchestrating impartiality: The impact of “blind” auditions on female musicians. American Economic Review.
  • Pryce, J., et al. (2019). A case study in public child welfare: County-level practices that address racial disparity in foster care placement. Journal of Public Child Welfare.
  • Reddy, J., Williams-Isom, A., & Putnam-Hornstein, E. (2022). Racial sensitivity training: An inadequate solution to systemic racial disparities in child protection systems. Child Abuse & Neglect.
  • Bartholomew et al. and related literature on disproportionality in foster care (contextual references in JPAM piece).

Notes:

  • The analysis emphasizes that while blind removals may improve perceptions of fairness and provide accountability benefits, their causal impact on reducing racial disproportionality in foster care is limited when disparities largely originate at the investigation stage. The paper underscores the importance of targeting the earliest decision points to achieve substantial reductions in disproportionality, while also acknowledging potential costs and unintended consequences of expanding blind-removal procedures.