The Causal Effects of Foster Care: Why the Intensive Margins Matter

Background and purpose:

In this paper, we estimate the causal effects of foster care on a range of health and child protective services outcomes. Prior work exclusively estimates the causal effects of foster care on the extensive margin, providing estimates on the causal effect of removal relative to children who remain at home. While the removal decision is irrevocably correlated with an array of positive and negative lifelong outcomes, depending upon the age of removal, the intensive margin of foster care plausibly drives these effects. Yet, little is known about the intensive margin of foster care.

Methods:

Our analysis harnesses a representative sample of Wisconsin infants with Medicaid-funded births between 2010 and 2019. We leverage our rich, historical linked administrative data from the Wisconsin Administrative Data Core to deploy machine learning methods to generate two predictions. First, we predict removal within 7 days, 30 days, 1 year, and 3 years after birth using cross validated gradient boosted machines, then predict outcomes in the subsequent period up to 3 years. Outcomes include objective measures of child wellbeing identified in Medicaid claims, such as unintentional injuries, drowning, poisoning, falls, and confirmed and suspected child maltreatment, as well as later reports to child protective services (CPS) and removals. We deploy double machine learning methods to estimate the causal effect of foster care on the extensive margin. We then estimate the causal effect of foster care on the intensive margin using a dose-response model that compares outcomes across children according to the amount of time spent in care, minimizing confounding due to selection into foster care, for the same set of outcomes. To account for selection into the timing of removal, we identify five categories of children using machine learning methods to make comparisons across both dimensions (timing of removal, length of stay).

Results:

Our results suggest that the causal effects of foster care on the extensive margin are mostly negative, implying that children appear to fare worse relative to those who remain home. However, we find that children removed immediately after birth (within 7 and 30 days) exhibit lower rates of downing and unintentional injuries relative to those who remain at home. On the intensive margin, we find that each additional day spent in foster care is associated with a 0.16 percent decline in maltreatment-related injuries, and a 0.35 percent decline in a subsequent CPS report and 0.12 percent decline in subsequent removal. We investigate mechanisms such as placement type, payment amount, and caregiver fit as well.

Conclusions and implications:

Our estimates imply that foster care is protective particularly when children are removed from their homes within the first 100 days of birth relative to delayed removals. Our results provide novel evidence on the protective nature of foster care and demonstrate that double machine learning methods are an alternative to random assignment to investigators in states where this identification approach is infeasible.