Actuarial Group Takes Steps to Identify Racial Bias in Insurance Rates

As the insurance industry grapples with potential racial bias in the pricing of insurance products in the age of artificial intelligence and modeling, a group of property casualty actuaries have issued two new research reports to help guide the insurance industry away from the discriminatory effects higher insurance pricing for Black people and others of color.

The Casualty Actuarial Society (CAS) published two new reports March 31 to examine methods to identify and measure racial bias when pricing products like auto and homeowners insurance after a history of redlining in the industry.

“We hope with this series to serve as a thought leader and role model for other insurance organizations and corporations in promoting fairness and progress,” stated Victor Carter-Bey, CEO of  CAS, an international organization for credentialing and professional education for actuaries employed in assessing property and casualty risk.

These four rating factors loom large in pricing

One of the two new papers explores four specific rating factors that policymakers and insurance regulators have been scrutinizing to ferret out bias in setting homeowners and auto insurance rates. These four widely used insurance rating factors are: credit-based insurance scores; geographic location; homeownership; and motor vehicle records.

For example, geographic location can affect loss and severity in homeowners and auto pricing, two risk factors, but geographic location can often be a result of “historical policies and practices primarily directed at Black people.”

Black people and families were explicitly segregated and then redlined by banks and racially restrictive covenants through the first half of the 20th century, according to research cited in the paper. Redlining  was the term used by the  Home Owners’ Loan Corp. in the 1930s for its color-coding maps —neighborhoods deemed “undesirable” colored red, the paper showed. The long and sad history of redlining Black communities is pictured and detailed in the CAS paper.

Even today, societal stereotypes can determine where people live. CAS research cited shows White home seekers are more likely to be favored than minorities in houses for sale and in rentals and  apartments.

Variables might be predictive but also problematic

While actuaries might price higher in cities with older housing stock with property risks such as dated electrical wiring systems and higher crime, these variables are not race neutral, in fact, the paper on factors that lead to bias points out. The higher predicted losses leading to higher premiums are also correlated with race, and be a result of decades of inequities and biased long-standing societal approaches to living areas and home ownership, CAS explains.

There is also a widely demonstrated disproportionate impact of natural disasters on minority groups, as witnessed in recent years by Hurricane Katrina and Harvey.

"Hurricane Katrina had lasting effects on residents whose homes flooded, in the form of lower credit scores and lower rates of homeownership, compared to neighbors who did not experience flooding," CAS stated, citing outside research. 

The CAS researchers noted that all four of these factors have correlated to actual insurance loss over the years, but also can be a strong proxy for race and ethnicity so can have a strong negative impact on low income and minority communities.

As insurance consumer activist Birny Birnbaum from the center for Economic Justice wrote last spring in an address to CAS, there is "recognition that the historical discrimination has long-lasting effects that disadvantage those groups. Stated differently, you can’t enslave a population for two hundred years and then expect the legacy of that enslavement will disappear overnight."

Correlation but not causation

Credit scores are also highly controversial because they are predictive of loss but not associated --a credit loss does not mean a property loss -- with it, and insurers use propriety models off-limit to outsiders. But, again, credit-scoring could correlate to race and also reflect decades of discriminatory banking lending practices to Black individuals, CAS explains.

As a first step, the actuarial group recommends that actuaries and insurance professionals think about “value judgments and stereotypes often assigned to policyholders concentrated in higher insurance risk categories.”

Many other factors could influence insurance risk, the researchers said, and when actuaries combine these dozens of other factors in a rating model, the outcomes for policyholders could be much improved. CAS also called for more analysis on social bias on individual raising factors as well as the complex models used.

What does unfair discrimination mean?

The second report released recently strives to identify what discrimination in insurance is, and teases out  terms such as protected class, unfair discrimination, proxy discrimination, disparate impact, disparate treatment, and disproportionate impact.

While these are insurance industry rating terms that might reveal different pricing for minority groups, the different definitions applied can lead to higher prices for some consumers. Whether this is fair or right is often a matter of debate for insurers, but if it does cause disparate impact, regulators and now a growing number of actuaries are concerned, the research paper shows.

Actuaries, as part of their job for property insurers, examine whether the rates they charge homeowners vehicle owners are supported by loss experience for the various rating factors employed. While they can be different depending on risk, most U.S. states regulate their scope and impact so they are  not “unfairly discriminatory” to protected classes. 

Do these individual factors or combinations of factors derive their predictive power in full or in part from their correlation with a prohibited characteristic, or act as a proxy for it, the second paper ponders.

“If so, then it must also be determined whether this results in disproportionately higher or lower rates for certain groups within that protected class,” CAS says. The organization notes it is not up to them to decide how much correlation between a rating characteristic and race should be tolerated but wants to be at the table with policymakers for the discussion.

AI can continue and amplify racial discrimination in pricing

Of great concern and attention among state insurance regulators and CAS is the baking discrimination into artificial intelligence (AI) models through proxy or other characteristics identified and selected with machine learning.

AI can “propagate that bias into the choices made from the models’ predictions,” the CAS paper on discrimination points out, and effect a consumers  insurance rates through statistical bias and mathematical modeling, as a machine does not know what is unfair. Even when subjective human bias is uprooted or not used in any inputs, the inputs themselves can reflect historical bias.

These factors “may be woven into historical data could still make their way through the modeling process and affect the model’s output, the CAS report notes.

For this reason, research is underway among the casualty industry, academics and regulators on model fairness and one approach suggests adding other components to the model to blast out bias, while AI data scientists are working on ways to construct methods for bias detection and mitigation.

The National Association of Insurance Commissioners (NAIC), which is meeting this week in Kansas City, has been exploring bias, AI, machine learning and disparate impact more in depth in the past few years through its Special Committee on Insurance and Race.  The group is studying longstanding and continuing issues of unfair treatment, proxy discrimination, and disparate impact and making recommendations for statutory and regulatory changes, as well as exploring bias in machine learning, in a discussion slated for April 6.  

Some states & companies taking action

At least one state focused on fair outcomes for protected classes has turned down insurers' rate applications because of lack of model information. Connecticut has hired a data scientist. Expect more depending on a states’ resources as the NAIC ramps up detection of bias in machine learning and how to minimize bad outcomes — i.e. unwarranted higher pricing — for consumers. 

And, while California already prevents the use of insurance scores in setting consumer property casualty rates, a few other leading states have legislation prohibiting some factors such as credit-based insurance scores in setting rates, and instead relying on driving record and miles driven and years licensed. These states are Maryland, Washington State and Oregon, the CAS report notes.

While Washington State Insurance Commissioner Mike Kreidler, the longest-serving state insurance commissioner today, issued an emergency order banning credit scores for three years in 2021, it was overturned, also last year. Kreidler, who has served for two decades, is now working on a permanent rule to replace the rejected emergency order, CAS notes.

CAS identified a few companies, like Root Insurance and Loop, an insure tech, who have also been taking action on how auto insurance rates are set. Root Insurance announced back in 2020 that they would be discontinuing the use of credit-based insurance scores by 2025 and have also called on other insurance companies to do the same.

Further reading

CAS released two other research reports on race and insurance earlier in March as it evolves in its ways of thinking about the causes and effects of insurance pricing for consumers. These reports are Methods for Quantifying Discriminatory Effects on Protected Classes in Insurance, and Approaches to Address Racial Bias in Financial Services: Lessons for the Insurance Industry.

The NAIC's committee on Race and Insurance also has much work outlined and ahead, and archives extensive comments from industry and consumer activists on the issue.

The NAIC said it created a team to review rate models that use predictive analytics with currently three expert actuaries, one of whom is also a behavioral data scientist. The standard-setting organization said that while there is no nationwide tool used by regulators to evaluate algorithmic bias in modes, either in the input or the output, such a development is being evaluated by different states and will be discussed by the NAIC’s new Innovation, Cybersecurity, and Technology Committee.

However, “Regulators do look for unfairly discriminatory variables and, where the relationship between the variable is risk is not clear, may ask the company for more information to clarify,” the NAIC told Investopedia.