Predictive Risk Modeling for First Time Investigations

Background and Purpose: Over three million children each year are referred to child welfare systems in the United States (USDHHS, 2016). Some of these children and their families have had no prior exposure to child protective services (CPS), while others have had repeated exposures. Moreover, while many of these “first-time children” will not be involved in another referral, others may end up in a cycle of referrals and out-of-home placement. Distinguishing which first-time children are likely to be subsequently re-investigated by CPS is therefore crucial in determining how to most appropriately match them with preventive services. By the same token, identifying first-time children at low risk of subsequent re-investigation is important to avoid burdening families with unnecessary services. Despite the prevalence of first-time children in CPS investigations, as well as the recognized importance of making accurate assessments early on (DePanfilis and Zuravin, 2001), few studies have examined how this population differs from those with prior CPS contact.

Methods: We examine over 450,000 investigations including almost 470,000 unique children, comprising detailed records of all investigations conducted by a large urban child welfare agency over a six-year period. We define an investigation of a child to be a “first-time family investigation” (FTF) if that child and any other children in the household have had no prior contacts with CPS, and any alleged perpetrator has no prior recent contact with CPS. We first show how investigations involving first-time children differ from investigations involving children and families with prior CPS exposure. We then apply a random forest algorithm to predict the likelihood that a child will be subject to a subsequent substantiated re-investigation within six months. We compare how predictions vary for several alternative outcomes, as well as between FTF investigations and non-FTF investigations. In addition, we describe the data elements that are predictive of re-report, and examine possible effects of predictive methods on racial disproportionality.

Results: Descriptively, we find that FTF investigations are likely to be less severe (by a range of measures), and involve fewer and younger children. Across different outcomes, children in non-FTF investigations are over twice as likely to have an adverse outcome as children in FTF investigations. We also find that although prediction in FTF investigations is substantially more difficult than among all investigations, we can still identify children at high risk and children at low risk of subsequent substantiated re-report with considerably better accuracy than chance alone. Moreover, using a predictive model to identify FTF children at low risk of re-report would have fewer than ⅓ as many “false negatives” as chance alone.

Conclusions and Implications: Our findings suggest that machine learning methods can be used to efficiently allocate preventive services to first-time families with high need, and to avoid burdening low need first-time families who are unlikely to be involved in a subsequent near-term investigation. Our findings also indicate that first-time families are an important and distinct population served by the child welfare system, and should be a focus of preventive efforts.