He co-authors a “news analysis” for the Los Angeles
Times that sounds like a pitch by a company selling “predictive analytics” software
for child welfare. There’s also another
cheap shot at family preservation and something that sure sounds like a dog-whistle.
Second of two parts. Read
part one here.
Yesterday’s
post to this blog reviewed the bad journalism practiced by former Los
Angeles Times reporter, and current contributor, Garrett Therolf. I noted that other journalists have accused
him of repeated misrepresentations and, on one occasion, making up a
quote. The post talked about Therolf’s
profound discomfort with any suggestion that there is systemic racial bias in
child welfare.
I said I believe this poor journalism contributes to the
fact that Los Angeles County tears apart families at the second-highest
rate among America’s biggest cities and their surrounding counties. And I noted how he stacked the deck in a
story about racial bias in child welfare. As
Therolf portrayed it in a 2017 story, white people “marshal data,” Black
people want to rely on anecdote and “folkways.”
That post discussed what Therolf said during a January 20
“Ask the Reporters” video event presented by the Times. But Therolf and a current Times reporter,
Matt Hamilton, also wrote a “news analysis” for the Times headlined
Anthony, Noah, Gabriel and beyond: How to fix L.A. County DCFS.
Though called a news analysis, it was actually a sales pitch
for a child welfare fad that is much beloved by those wedded to a
take-the-child-and-run approach to child welfare: predictive analytics. This is an approach in which a computer
algorithm mines vast amounts of data – especially data about poor people – and
tells the family policing agency whether a case is high-risk. It amounts to computerized racial profiling.
It was tried first in criminal justice – and proven
to be racially biased. Yet Therolf and Hamilton present it not just as a
solution to the problems plaguing the Los Angeles Department of Children and
Family Services but the only viable solution. They regurgitate
favorite talking points of Emily Putnam-Hornstein, America’s foremost
evangelist for predictive analytics in child welfare; someone whose own
extremism and penchant for deriding the work of Black people is documented here.
Analyzing the “analysis”
The story is a great illustration of Therolf’s general
approach. So in this post, I will go through parts of the story paragraph-by-paragraph,
(excerpts are in italics) starting with that headline, which references the
three horror story cases Therolf focuses on, to the near exclusion of all else:
News Analysis: Anthony, Noah, Gabriel and beyond: How to
fix L.A. County DCFS
One could as easily have done a story headlined: “Los
Angeles tears apart families at one of the highest rates in America’s largest
urban areas – but children keep dying: How to fix L.A. County DCFS.” But Therolf has shown no interest in L.A.’s
outlier status. While he has pointed out
that Los Angeles has failed to prevent horror story cases for decades, he has
not seen fit to remind readers that the rush to remove more children has done
nothing to make them safer.
In the long, troubled history of L.A. County child abuse
cases, certain names stand out as avatars of how the system can go terribly
awry. Anthony Avalos. Gabriel Fernandez. Noah Cuatro.
Viola
Vanclief, Joseph
Chacón and Andreas
F. could just as easily have become such avatars
– but Viola and Joseph
died in foster care. Andreas was
allegedly tortured and beaten into a coma by his foster mother. They didn’t become “avatars” because the way they
were hurt doesn’t support Therolf’s “master narrative.” The
Love family (whose story does not
appear to have been covered in the Times even once) and the mother who
is the subject of this
commentary in WitnessLA also would make excellent avatars of system
failure.
In fact, all of these children
deserve to “stand out as avatars of how the system can go terribly awry.” Therolf and his colleagues have chosen to
emphasize some and pay far less attention to others.
Is the racial
justice movement really too influential?
But since the spring of 2020, another name has wielded
outsize influence over national perspectives and policies related to child
welfare, and energized activists to push for sweeping reforms: George Floyd.
Notice the use of “outsize” influence as opposed to “strong
influence” or just “influence.” Outsize
can simply mean large, but it also can mean
“exaggerated or extravagant in size or degree.” – as in: Beware! People are now
way too concerned about not taking away Black children and that is putting them
in danger.
If anything, the impact of America’s
racial justice reckoning on child welfare has been undersized, consisting largely of foster care agencies slapping Black Lives
Matter statements on their websites while continuing business as usual.
Indeed, it is a testament to the double standards that so
many journalists apply when the issue is child welfare that in the midst of a
racial justice reckoning, so many reporters bought into a racially biased, and
now widely-debunked myth: the
one that claimed that, due to COVID-19, in the absence of all those
overwhelmingly middle-class disproportionately white “mandated reporters”
keeping their “eyes” constantly on children who are neither, their parents
would unleash on their own children a “pandemic of child abuse.” Garrett Therolf, who in his current job
writes for outlets besides the Times, jumped
right onto that bandwagon.
Similarly, while the Los Angeles
Times did a superb job confronting its own racism in many other fields, it neglected to
scrutinize its child welfare coverage.
Perhaps the Times’ new editor, Kevin Merida, will remedy that.
The murder of the Black Minneapolis resident by a police
officer in May 2020 set off a national soul-searching over the country’s racist
past and the prejudices that still haunt its institutions. In L.A. County, that
process has focused intense scrutiny on what a number of racial justice
advocates and elected officials say is an implicit bias that may make some
Department of Children and Family Services workers more prone to regard poor
families and parents of color as unfit to raise their children.
Notice the “few bad apples”-type
framing here. Some workers may
be biased. As opposed to the idea
that systemic racism poisons the entire system.
Therolf isn’t required to agree with that, but he has an ethical
obligation to note the existence of that critique and explore it. But as we’ve seen, Therolf has a record of dismissing
that idea.
In 2020, three-quarters of children removed from their
homes in L.A. County were Latino or Black, according to a motion — authored by
Supervisor Holly Mitchell and passed in July by the Board of Supervisors — to
begin implementing a controversial pilot project called “blind removal.”
The program, first adopted in Nassau County on Long
Island in New York, redacts all race and race-related factors from the dossiers
used by social workers and supervisors in determining child welfare cases. And
it is gaining popularity, despite critics who say that it has shown
insufficient evidence of its efficacy and that it adds one more task to an
overtaxed workforce.
Therolf is partly right about criticism of the program’s
efficacy – criticism from his pal Putnam-Hornstein. (He does not mention other criticism, that
the change doesn’t go far enough in uprooting institutional racism.) But that hardly makes the whole program
“controversial.” Even the revisionist
critique shows an overall decline in removals of Black children. In fact, the data may indicate that the
process contributes to a decline
in removals of all children – perhaps because the process also eliminates
data suggesting whether or not a family is poor.
But the best way to find answers is to create a pilot
project to test the practice. That is
exactly what the Board of Supervisors has ordered.
But while officials scramble to address these
race-related concerns, other child welfare experts assert that another,
relatively new methodology using machine learning and algorithms is more likely
to yield race-neutral and reliable results that, among other benefits, will
enable social workers to accurately identify incidents of child abuse at an
early stage and move swiftly to intervene.
Notice that this method gets a much friendlier introduction.
A method widely criticized as magnifying racial bias is presented as the magic
cure.
Rather than relying on caseworkers’ limited ability to
weigh a family’s full recorded case history — due to limited time and cumbersome
technology — experts had urged the county to partially automate risk analysis
with a new generation of predictive analytics tools to scan and evaluate
hundreds of known variables regarding families, including prior hotline calls
and the child’s age when the first hotline call was received.
Stop and ask yourself: If you were writing promotional copy
for a company selling predictive analytics software, would it have been any
different from that?
Better solutions
One could, in fact, solve the limited time problem by
curbing the deluge of false reports, trivial cases and poverty cases. This could be done by things like eliminating
mandatory reporting -which has been
found, in studies
Therolf is unlikely to tell you about, to
overload the system and drive families away from seeking help.
The deluge of false reports is due in
part to CYA referrals by mandated reporters terrified of what will happen to
them - including what the Times will do to them – if they don’t report
and something goes wrong. As for
cumbersome technology, how about fixing it, instead of building a whole new
technology on top of it?
The use of such tools to predict which children are at
greatest risk has attracted controversy because of the chance that they might
exacerbate racial disparities in child welfare. That’s because Black and Latino
families tend to interact more frequently with entities such as public
hospitals and mandated reporters who generate the data that are used to train
the algorithm on how to detect risk.
Actually, there’s far more reason than that. The
algorithm in Pittsburgh, for example, which was co-authored by
Putnam-Hornstein and for which Therolf has shown particular fondness, relies
heavily on databases that sweep up data only on poor people. It also relies
heavily on past involvement in the family regulation system itself. (By the
way, Putnam-Hornstein hates the phrase “family regulation system.”) That makes
the computer-generated risk score more like a self-fulfilling prophecy than an
actual prediction. Then there’s the fact
that, as noted earlier, the same approach has been proven
to be biased when applied in criminal justice.
Those pesky civil libertarians!
Pushback from the American Civil Liberties Union and
others had stalled the development of the program for years, contributing to
the decision by Cagle’s predecessor, Philip Browning, to retire in frustration.
Darn those pesky civil libertarians! But, wait, what did the ACLU actually
say? You don’t suppose they have an
in-depth critique that they presented at a recent virtual conference, do they?
Isn’t this where real journalists quote the ACLU in order to explain why there
might be a problem? But no; a hallmark
of Therolf’s “reporting” is to minimize dissent from his own point of view, confining
it to a brief paraphrase or not mentioning it at all.
Oh, one other thing Therolf neglects to mention: Los Angeles
County tried a predictive analytics algorithm during Browning’s tenure –
and it failed spectacularly. The
“false positive” rate was more than 95%. Much the same happened in Illinois. What L.A. is
trying now is supposedly a new, improved algorithm. But it’s co-authored by Putnam-Hornstein –
and her Pittsburgh algorithm has much the same problem.
But [Browning’s successor, Bobby] Cagle gained the
majority support of L.A. County’s supervisors that had eluded Browning, and the
tool was piloted last summer to help flag children like Noah who may be at the
highest risk. (Cagle resigned in November, as DCFS faced mounting criticism
over a series of fatalities and abuse of children under the agency’s care.)
A lead designer of predictive analytics, Emily
Putnam-Hornstein at the University of North Carolina, emphasizes that the tool
is designed to be advisory and can be easily set aside by caseworkers if their
investigation verifies that no significant safety threats exist.
Easily set aside? You’re kidding, right? Again, imagine what
would happen to the caseworker who overrides the algorithm and something goes wrong. Here’s what would happen: A Los Angeles
Times headline that would say something
like: “Caseworker ignored ‘high risk’ warning about [name of child] weeks
before he died."
Also, as Therolf describes it, the override would come only
after the incredibly intrusive, traumatic investigation called for by the
algorithm – and algorithm which, when it comes to predicting the kind of horror
stories on which Therolf’s reporting thrives, will almost always be wrong.
Unfortunately, it seems that the chair of the Board of
Supervisors, Holly Mitchell, buys this.
During a Times-sponsored video presentation, she actually said she knows there is
an inherent potential bias in the algorithm but she was counting on caseworkers
to both overcome their own biases and counter any bias in the algorithm.
In a slide show presentation explaining the need for
tools like L.A. County’s, Putnam-Hornstein wrote that her work stems from a
“growing appreciation that current tools are inadequate, clinicians are poor at
weighting factors (and time is scarce!).” An independent evaluation team will ultimately
decide whether the tool helps.
I’ve already addressed the “time is scarce” argument. The “current tools are inadequate” argument
is a Putnam-Hornstein favorite. The
problem isn’t that she’s wrong, it’s that she’s setting up a false choice:
Either use the cruddy system you have now or the cruddy alternative I’m
proposing. That’s something
I’ve addressed before in the context of Los Angeles.
There are other ways to go: Abolish mandatory reporting, allowing
mandatory reporters to become mandatory supporters (Putnam-Hornstein hates that
phrase, too), narrow neglect laws to reduce the confusion of poverty with
neglect, and provide high-quality defense for families, just for starters. All that would allow human caseworkers plenty
of time to learn how to do the job better and with less bias.
As for the claim about an independent evaluation, that needs
to be checked closely, in light of the way a so-called independent ethics
review was handled for Putnam-Hornstein’s Pittsburgh algorithm.
Scapegoating family preservation
And now, behold how, once again as he has so many times
before, Therolf attempts to scapegoat any effort to keep families together:
One near-universal assumption among social workers is
that children belong with their own families in the absence of serious safety
threats.
That is near-universal rhetoric – but the data show
that caseworkers in Los Angeles are far more prone to tear apart families than
their counterparts in other large metropolitan areas.
But critics contend that this can lead caseworkers to
reflexively declare success whenever a child remains with their parents,
without taking full account of whether the child could be at risk.
Which critics – besides you, Garrett? That sounds a lot like Donald Trump, who was
fond of saying “many people are saying” or “a lot of people are saying” when he
meant himself. Also:
● As is so common with Therolf’s stories, this implies that
holding children in foster care is safe and only returning a child home can
place a child “at risk.” That is at odds
with study after study showing high
rates of abuse in foster care.
● And where are the data to support the claim about placing
children at risk? Oh, wait, there’s one
line by one lawyer in one report - from 2012:
“DCFS should change its messaging from Do Not Detain/Keep
The Numbers Down,” wrote the Board of Supervisors’ special counsel in a secret
internal 2012 report.
But then, as now, the numbers weren’t down. Los Angeles was,
and remains, an outlier in child removal.
Therolf provides no link to the “secret internal 2012 report” and no
documentation that any such messaging existed.
If there has ever been such messaging, the data show that the rank-and-file
sure didn’t get the message.
Yet workers in the cases of Noah Cuatro, Gabriel
Fernandez and others went on to finalize decisions to leave children in
dangerous homes without even reading their own agency’s case file.
Note the leap here: The logical assumption when workers make
decisions without reading the file is that they didn’t have time to read the
file – as Therolf himself suggests earlier in this same story. But if you’re pushing a particular
point-of-view, you simply leap to the conclusion that they didn’t read the file
because of some fanatical devotion to keeping families together.
Research has shown that people who are not trained to
assess and investigate child abuse are more successful than DCFS workers at
predicting the most serious harms to children. Researchers have found that the
number of calls made by the public to child protection hotlines was a better
indicator of deadly risk than the conclusions of caseworker investigations.
Inference peddling
Here, as throughout the story, Therolf wants us to take his
word for whatever inferences he chooses to make. There are no links to actual documents, much
less to research. No way to see the
context or to fact-check his claims.
In fact, it appears that Therolf may be making a leap from a
study – by Putnam-Hornstein, of course, showing that children reported to child
abuse hotlines are more likely to die than children not reported to
hotlines.
But, as noted in yesterday’s
post, of all children who are, in fact, the subject of hotline calls
in the course of a year 99.998% do not die.
That makes the horror stories no less horrible and no less a cause for
action. But they are needles in a huge
haystack.
So Putnam-Hornstein’s findings tell us only that while the risk
of child abuse death among children who are subjects of hotline calls is
infinitesimal, the risk to children who are not subjects of such calls is even
smaller. If you then make the number of
hotline calls a huge red flag in an algorithm, you will traumatize vast numbers
of children in innocent families, encourage malicious false reports, and
further inundate caseworkers, leaving them even less time to find children in
real danger and less time to “read their own agency’s case file” – because,
remember, “time is scarce!”
Indeed, it is precisely because the horror stories we think
of when we hear the words “child abuse” are so rare that predictive analytics
experiments crashed and burned in Illinois – and Los Angeles.
Over the last decade of relative paralysis by the Board
of Supervisors to implement effective reforms, hundreds of children whose cases
cried out for help have died at the hands of their caregivers.
Actually, it wasn’t paralysis. The constant outcry from the Board of
Supervisors – and Therolf’s stories -- led Los Angeles County to take far too
many children and overload the system – so of course child abuse deaths did not
stop.
Can’t you hear the whistle blowing?
Therolf goes on to list a series of factors he says explain
why child abuse deaths occur disproportionately in one part of the county, the
Antelope Valley. Then comes something
that sure sounds like a dog-whistle:
Complicating matters further is that the region has one
of the highest shares of Black residents, so any misstep exacerbates concerns
about racial justice.
That sounds like a more genteel version of what you’d hear
on Fox News: They won’t take Black kids in danger because the politically
correct woke mob will get them!
Now, as the county prepares to select its next DCFS
director, that person will have to confront a central problem that has
undermined … so many … initiatives of
the past: trust.
At a recent forum held by Fordham’s School of Law to
discuss Los Angeles County’s use of machine learning, Ron Richter, the former
director of New York City’s child welfare system, said that even “when we talk
about a tool that may help reduce disproportionality and family regulation,
sincere issues of trust surface, especially for those of us who have witnessed
firsthand what child welfare looks like on the ground.”
That’s also true, Richter added, for “those who have been
historically judged by this system and feel strongly that many children and
families have been misjudged.”
But Richter, who now runs a large private foster care
agency, wasn’t the only speaker at the forum.
One of the others was Aaron Horowitz, chief data scientist for the
American Civil Liberties Union. So you
see, Therolf knows full well why the ACLU disagrees with Richter and does not
see this as “a tool that may help reduce disproportionality and family
regulation” – but he doesn’t seem to want L.A. Times readers to know
it. He allows no one to actually make the
case that predictive analytics can magnify racial bias. Though Therolf won’t tell you, you can watch
Horowitz’s presentation, and the entire Fordham event here.
Note also the condescension.
What Richter is really saying is much like what Therolf himself suggested
in
that 2017 article: That we white people are using science, and those Black
people are just scared because they rely on anecdote and “folkways.”
As long as this is how the Los Angeles Times
continues to cover child welfare, the commendable reckoning it undertook last
September, when it looked at its own track record covering race, is incomplete.
