WITHDRAWN: Using a Predictive Risk Model to Assess Racial Disparities in Service Referrals Following a Child Protection Investigation

Background and Purpose: Child protection system (CPS) agencies often provide services to families through referrals to other providers, rather than directly. Yet, we know little about which families are referred to services because that information is generally recorded in non-CPS data systems. It is well documented that Black children are overrepresented at all levels of child protection system involvement, though research is conflicting about the dominant source of this disproportionality. One area that has been underexplored relates to potential racial disparities in referrals to services that could protect against future CPS involvement. We use a unique source of linked administrative and community-based services data from Los Angeles, California to examine (1) whether children are differentially referred to services by racial group and (2) whether risk explains observed differences.

Methods: Risk was estimated using a predictive risk model developed from more than 300 data features and trained to classify risk of foster care placement within two years using records from 2016-2017 (N = 341,428). Records documenting referrals to services from contracted community providers in LA were extracted and linked to child-referral records using the client ID. Using the predicted probabilities calculated by the predictive risk model, we kept the highest risk score calculated for each child. A logistic regression model was used to estimate the odds of receiving a community service referral as a function of investigation disposition, focal child’s age, race, risk, and race-by-risk interactions.

Results: Overall, following an investigation for alleged maltreatment in which the child was not removed, only 21.2% of families were referred to community-based services (N = 44,889). The median risk score of families referred to services (0.31) was slightly higher than the median risk score of those who were not referred to services (0.23). The proportion of children referred to services who were Black (16%) was similar to the proportion of those not referred to services (15%). The proportion of those referred to services who were White (8%) was lower than the proportion not referred to services that were White (12%). The proportion of children referred to services who were Hispanic (67%) was higher than the proportion not referred (58%). Results of the logistic regression indicated that, compared to White children, the unadjusted odds of receiving a referral were higher for Black (OR = 1.47, p < .01), and Hispanic (OR = 1.68, p <.01) compared to White children. After adjusting for risk, the interaction between risk and race, and other factors, the odds of receiving a referral were higher for Black (OR=1.43, p<.01) and Hispanic children (OR = 2.00, p < .01) compared to White children. There was evidence of a significant interaction between race and risk for both Black and Hispanic children.

Conclusions and Implications: Using predicted probabilities from our model, we found that racial disparities in community-service referrals were greater at low risk than at high risk. Across all risk categories, Hispanic children were more likely to receive a referral than any other racial group, and this variation remains unexplained.