Study Suggests Florida Could Use Less Foster Care, More “Light Touch” Help for Families

The number of Florida youth in foster care fell from 29,229 in 2006 to 18,040 in 2013. But the number is on a steady incline again, up to 24,059 as of this summer.

The state has struggled to recruit enough foster homes to keep up with the increase. But a study out of one of Florida’s most populous counties suggests that much of this new influx could be handled without the use of an out-of-home placement, and in some cases, without much child welfare involvement at all.

Broward County (seat: Fort Lauderdale) tested its current child welfare decision-making process against a predictive analytics approach, which relies on data collection and machine learning to predict likely future behavior. The study, conducted by a group of researchers and supporters of predictive analytics modeling, suggests that 40 percent of cases referred for either a foster care removal, intensive services or both could have been handled with less-intrusive options.

The study also found that when assigned unnecessarily, some of the more intrusive child welfare interventions — such as counseling and parenting classes — are associated with worse outcomes, including more foster care involvement.

“The assumption has always been that services do no harm,” said Ira Schwartz, one of the study’s authors. “That might not always be the case.”

Broward County currently employs the use of an actuarial risk assessment to guide determinations on substantiation of maltreatment. It has also implemented a more subjective tool to guide decisions made after substantiation.

As is the case in other Florida counties, a unit within the county sheriff’s office investigates abuse and neglect reports. For families that are screened in and referred for intervention, there are basically three tiers of services:

  • Cases that involve foster care placements or intensive home services are delivered through a lead nonprofit agency called ChildNet.
  • The Children’s Services Council (CSC), an independent governmental tax authority, handles prevention services for families where a minor case of maltreatment is substantiated, or where maltreatment is strongly suspected but not substantiated.
  • The county also offers “light touch” services for families, where some form of cash or living assistance can alleviate a stressful situation.

The county and ChildNet handed over full records of 78,394 individual children for whom a call had come into the maltreatment hotline between 2010 and 2015. Study authors constructed a predictive analytics (PA) model to gauge whether that technique would be more accurate than the actuarial tool in predicting substantiation.

The study then used propensity score matching — a statistical matching technique that attempts to estimate impact by accounting for variables that predict receiving the treatment — to determine the effect that receiving various services had on the prevention of a future maltreatment report.

The study uses the presence of a subsequent report of maltreatment as a target outcome to determine the validity of substantiation and services. The authors note that while this is an imperfect metric, they believe it to be more consistent across different child welfare systems than the substantiation of a report.

Richard Wexler, executive director of the National Coalition for Child Protection Reform, said in an email to The Chronicle of Social Change that the study’s use of a subsequent report of maltreatment magnifies potential biases that exist in a system.

“If your system confuses poverty with neglect then more poor children will be reported,” said Wexler, whose organization advocates for greater emphasis on family preservation by child welfare systems. “And if the family stays poor, they are more likely to be reported a second time.”

The predictive analytics model’s finding was that 40 percent of the referrals to the court and ChildNet, most of which involve out-of-home placement, were inappropriate. Those cases, according to the model, could have been handled safely through CSC’s prevention services, or with “light touch” assistance.

The predictive model also found that the cases inappropriately assigned to ChildNet had an “extraordinarily high” and “statistically significant” chance of returning to the system.

“On the surface, it appears that there may be something about the nature and/or intensity of the services for these lower risk cases that might be causing more harm than good,” the study said. “This is a troubling finding and clearly an issue that needs to be carefully researched in the future.”

ChildNet CEO Emilio Benitez said that placing unnecessary requirements on parents can exacerbate the situation.

“It might not be that my child was removed because I was bad parent, but that I’m homeless,” Benitez said. “If I lost my job, and I just don’t have stabilized housing, that doesn’t mean I’m a bad parent. But we almost always make them go to parenting classes.”

Wexler said family preservation supporters have long argued that unwarranted services can do harm.

“We understand that if you barge in on a family already under stress due to poverty and force the family into useless counseling and parent education that will only add more stress, while doing nothing about the poverty, then you will indeed make things worse,” he said.

The study found a similar pattern in the referral of cases to CSC’s prevention services. It found that in 90 percent of the cases referred, “there appears to be no significant difference to be gained by making these referrals.”

Those cases could have been met with light-touch options for family assistance, the study said.

“Cases that need a very light touch, that makes up a lot of neglect cases,” said Schwartz. “A lot of them are just poverty cases, they just need some financial assistance.”

While Wexler advocates for greater reliance on family assistance and less on foster care, he said the findings do not convince him that predictive analytics would ultimately accomplish this.

“Even if the model these researchers have come up with were valid, here’s how long it would be used: Until a child known to the system dies and a caseworker tells the [Miami] Herald, ‘I pleaded with my supervisor to take away that child, but she kept telling me about some algorithm that said he was safe.’”

Benitez said he favors a shift toward the predictive approach.

“I’m hoping it gives us the means to open our minds and imaginations to think of better ways of addressing the risk factors affecting children and families, poverty being a huge one,” Benitez said.

The study, “Predictive and Prescriptive Analytics, Machine Learning and Child Welfare Risk Assessment: The Broward County Experience,” was published in October by Children and Youth Services Review.

Print Friendly, PDF & Email

John Kelly, Editor in Chief, The Chronicle of Social Change
About John Kelly, Editor in Chief, The Chronicle of Social Change 1212 Articles
John Kelly is editor-in-chief of The Chronicle of Social Change. Reach him at

1 Comment

  1. This is a very interesting, maybe groundbreaking study, but its methods and conclusions are way too preliminary for making policy recommendations. Among the problems: incomplete followup of cases for only a year to see if there is a recurrence, assuming that there are no unmeasured issues that social workers saw and the algorithms didn’t, and using the child rather than the family as a unit of analysis. Finally it is concerning that the authors reported about the proportions of children that were provided with too-intensive services such as foster care but not the proportion that were provided with services that are not intensive enough. For more information about the research methods and problems, see

Comments are closed.