Equations and Inequality
Calculating Race: Racial Discrimination in Risk Assessment
By: Benjamin Wiggins. Oxford University Press. 2020.
Review by James Lynch
Around 1880, actuaries at Prudential Life Insurance saw a problem in their book. Although Black and white policyholders paid the same rates for the same coverage, the two groups did not have the same mortality rate.
On reflection, this was no surprise. Back then, most Black people had been enslaved, or their parents had. The resulting hardships shortened their lives.
The next step Prudential took was reprehensible as seen today but made a certain degree of business sense.
That action lies at the heart of Calculating Race: Racial Discrimination in Risk Assessment. Author Benjamin Wiggins has written a lucid description of how quantitative professionals, in doing their job—collecting data, analyzing it and creating race-based decision rules—“played a significant role in perpetuating racial disparities in wealth, incarceration and housing in the United States.”[1]
His short book tells three tales of how actuaries used race to:
- Set the price of a life policy.
- Determine whether a prisoner should be paroled.
- Determine whether to grant a mortgage.
He argues as well that race continues to be used, not explicitly, but through proxies.
Wiggins, I should note, does not define actuary the way most practicing actuaries would. He considers an actuary to be anyone who has tried to mathematically project an uncertain outcome. That sounds to me like a statistician, with an actuary being a person who tries to turn statistical probabilities into financial estimates.
Wiggins writes clearly about a complex, charged topic. Within the stories he tells, a thoughtful actuary can find lessons in how every step in the analytical process can be affected by the preconceptions one brings to the analysis. When they are race-based, the outcomes can skew in ugly ways.
Actuaries have a remarkable skill. They examine a risky world and turn that risk into a price. If you face more risk than the average person, you pay more for insurance. If you face less risk, you pay less.
Past discrimination often puts a Black person at greater risk. Because risk has been shoved their way, Black folks must pay more for insurance—a financial insult on top of a societal injury. In those circumstances, charging an actuarially fair rate can be seen as perpetuating the discrimination.
Prudential and other firms had quantified the difference in risk that discrimination forced onto Black people. The circumstances of turn-of-the-century Black life—under-resourced schools, dangerous jobs with lousy pay, inferior services and housing—actuaries boiled generations of prejudice down a single number: the premium for a life insurance policy. A Black person’s life was 50% riskier than a white person’s.
So, presented in the late 19th century with indisputable evidence of the societal toll of racism, Prudential and other insurers offered a solution.
They raised rates on Black policyholders by 50% (or reduced the death benefit by a third, which of course amounts to the same thing).
Race-based pricing shocks the senses today, and it was dimly thought of back then, too. Massachusetts banned the practice in 1884, and within seven years another 10 states had done so.
Prudential and other insurers responded to the price restrictions by tightening underwriting standards. Wiggins methodically describes the extra steps a Black customer needed to take. Agents didn’t solicit Black customers—commissions were lower, or zero; Black customers got no help filling out the multi-page application; they had to pass a physical not required of white customers.
Wiggins treats these responses as evidence of deep institutional racism, and that doubtless played a part. He dismisses the company’s view: They wanted to avoid unprofitable business. I would have enjoyed a smarter exploration into whether companies were earnest.
Wiggins could have examined how racism can look like indifference, demonstrated in an 1891 Atlanta Constitution article I found. In it, Metropolitan and John Hancock executives told New York legislators that they cared not whether race was banned as a rating variable.
And they specifically described, in an open public hearing, how they would defy the spirit of the law, using the tactics Wiggins so ably described.[2] These methods weren’t secrets. Executives didn’t try to persuade legislators to let them serve the Black market, albeit at higher prices. They just wrote it off.
There are ways insurers could have supported a Black marketplace. They could have used their data to call for an improvement in the lives of Blacks. Today insurers support large, respected institutions like the Insurance Institute for Highway Safety, which point to data to promote safety and reduce losses. The industry supports laws to reduce accidents: distracted driving, blood-alcohol limits, speed limits.
We see nothing like that from major insurers of the era. I contrast that with Black-owned life insurers that developed the marketplace white insurers abandoned and tried to educate it, promoting articles like “Tuberculosis? No, Heart Disease Leads Negro Deaths.”[3]
Another smart business response would have been to look for root causes. What elements of Black experiences were abbreviating those lives—hazardous occupations? Decrepit health care?
Frederick Hoffman, not an actuary but one of Prudential’s most influential statisticians, actually undertook something like this, as Wiggins explains. He gathered tremendous amounts of data by race: disease rates from the National Board of Health; arrest and incarceration rates from the Philadelphia Superintendent of Police and the New Jersey State Prison; financial data from bureaucrats in Georgia and Virginia.
Hoffman didn’t use this information to understand why there was a disparity in risk, or to refine Prudential’s rates. He used it to claim that Blacks in America were headed for extinction.
Wiggins quotes Hoffman’s 1896 work, Race Traits and Tendencies of the American Negro: “Of all races for which statistics are obtainable … the negro shows the least power of resistance in the struggle for life.”
Setting revulsion aside, an actuary can see Hoffman made a fundamental analytical error. He didn’t examine his priors. Social Darwinism explained his world. Every disparity his statistics uncovered convinced his racist mind that Blacks in America would die out in a few generations, and that any assistance would just postpone things.
He couldn’t believe that centuries of institutionalized discrimination could have shifted risk so thoroughly. But it had. Racist stereotypes were so entrenched in his mind that he couldn’t see past them, even as the evidence he collected piled up before him.
At the time, W.E.B. DuBois quickly understood the misfire was not in Hoffman’s statistics but in his analysis. Wiggins notes DuBois “took issue with Hoffman because Hoffman ‘continually forgets’ the ‘unusual disadvantages’ of blacks in the United States.”
Wiggins (dis)credits the insurance industry with creating the use of race as a quantifying factor. He next turns to the Illinois parole system.
In the 1800s, criminal justice systems began to gather identifying information on prisoners. In the days before state ID cards and fingerprints became commonplace, prison officials would gather hundreds of physical measurements of their charges: common markers like hair color, eye color and height; and less common ones, like width of the head or length of the left middle finger. Along with crude prison photographs, these would help identify each prisoner uniquely.
The system wasn’t much as an identifier, but as many actuaries can attest: Data that gets collected gets modeled. Those measurements would serve a purpose.
Illinois revised the system to group and discriminate. In the early 1930s, it created a 27-variable model to predict recidivism. Variables put special emphasis on race and neighborhood (which itself is highly correlated with race). To maintain the database, the state created the perhaps unique job of “parole actuary.”
The model didn’t work too well, but the state used it until 1955, when a former parole board actuary retested the 27-variable formula and slimmed it to seven. No variables resembling race or neighborhood survived.[4]
Illinois’ error took a human toll, and it was exacerbated by poor maintenance of the model. More careful analysis—studying which variables were predictive and which were confounding—and regular re-evaluation of the model would have scrubbed race from the model much sooner. Wiggins doesn’t suggest why it took so long to revisit the model, but I think the lack of a financial imperative was likely one reason.
It costs time and money to evaluate a model, and the prison system was indifferent to whether it worked—just as the big life insurers were indifferent about the toll of their underwriting restrictions.
In insurance, rating variables are regularly scrutinized. If their predictive power ebbs, they are given less weight or dropped. Example: In the 1950s, when men did most of the driving, gender was a good proxy for miles driven, hence highly predictive. Times changed, the predictive value has ebbed, and rates have followed.
For the state, the cost to test the recidivism model was measurable; the financial benefit was not. The victims of the delay: Black prisoners.
Wiggins’s final history of quantitative bias involves the lending practices of the Federal Housing Authority. Today, most discussion focuses on the redlining maps of the early 1930s.[5] Wiggins notes the maps had less impact than most people believe. They were produced by the Home Owners’ Loan Corporation, whose funding ran out by 1936.
The maps, however, vividly convey the attitudes behind the underwriting guidelines the Federal Housing Administration did produce. Those guidelines embedded racism in housing decisions deep into the 1940s. Remnants tattered policy well into the 1970s.
The role of actuaries here is murkier. Wiggins recounts economists and statisticians at work, plus two brothers styling themselves as “real estate actuaries,” but their work only tangentially resembles what actuaries do.
Actuaries take a data set and develop a model that predicts. The price of a policy and the size of a loss reserve are classic examples. The FHA underwriting guidelines don’t operate that way, at least in Wiggins’ description.
Guidelines included a quantitative score, “Rating of Neighborhood,” which definitively included elements with a racist tinge, such as an estimate of whether the neighborhood is likely to be struck with “the ingress of undesirable racial or nationality groups.” If the neighborhood score was too low, no mortgage could be written there.
But the quantification—subject to a single assessor’s whims—was hardly objective. And Wiggins doesn’t discuss the data from which this model was built. (Maybe there wasn’t any.) He shows no evidence the model was validated over time.
It strikes me that the problem here is not excessive application of data but a dearth of any data to apply. It is those circumstances where prejudice can flourish, and for decades of national housing policy, it certainly did.
Wiggins’ final discussion teeters next to a common trap—that quantitative analysis is inherently discriminatory.
“The actuarial assessment of risk,” he writes, promises “health, wealth, well-being, safety, security, and even the extension of life.” A model that fulfills that promise “must necessarily be historical and must necessarily be discriminatory.”
Wiggins writes that he sees the value in modeling, but suggests that society “share risk equitably,” which he suggests would take “a restructuring of society.”
Short of that ambitious goal, he ends up embracing the work of the Fairness, Accountability, and Transparency in Machine Learning workshop (FAT/ML). Six computer scientists formed this group to “use computationally rigorous methods” to ensure non-discrimination in modeling.
Wiggins discusses the activities of FAT/ML at length. You can still access it at https://www.fatml.org/, but the workshop became a standalone conference called FAccT, https://facctconference.org/.
It’s at this point, I think, Wiggins wanders. He asserts that models capture “disparities in morbidity, mortality, education, incarceration and wealth” but ignore “slavery, indentured servitude, unequal laws, unequal care and unequal opportunities that created or exacerbated those disparities.”
I think he has confused what the models intend to do—demonstrate disparities—with what he wants them to do—document what caused the disparities. The models do an excellent job of quantifying the impact of discrimination. They don’t try to isolate whether slavery had greater impact than Jim Crow laws, unequal school systems or any of the hundreds and hundreds of discriminatory behaviors of the past. But the models display, vividly and objectively, the cumulative effect of all that discrimination.
Within that, there is hope. The gap in life expectancy by race, and presumably life insurance costs, has shrunk over the decades—evidence that we have made strides. That a gap still exists indicates that there is a long way to go.
JAMES LYNCH, MAAA, FCAS, is a freelance writer.
Endnotes
[1] I read an electronic copy to review it, but have also heard the audiobook, which is well-presented and costs less. It may be a better way for actuaries to enjoy the work.
[2] “Higher Premiums: Charged Negroes by New York Insurance Men,” The Atlanta Constitution, Feb. 2, 1891, p. 4.
[3] “Tuberculosis? No, Heart Disease Leads Negro Deaths,” The New York Amsterdam News, July 13, 1935, p. A6.
[4] The remaining variables were: Age leaving home; social development pattern; work record; most serious previous sentence; total criminal record; schooling completed; and use of prison time.
[5] If you can stomach them, they can be found online at “Mapping Inequality.” Most of the nation was mapped, so you can likely find your neighborhood.
