For more than 70 years, insurers and insurance regulators have been sensitive to the issue of potentially discriminatory or unfair rating factors.
While assuring fairness to everyone’s satisfaction is a laudable objective worthy of pursuit, it is elusive by its very nature. Fairness, or impartiality, can be a matter of perception. Many questions arise when considering personal auto insurance, which is this article’s focus.
Is it fair for senior citizens to get a premium discount due to high credit scores and lower driving frequency, while at the same time their driving ability deteriorates with age? Is it discriminatory to use location as a factor in a densely populated, high-crime area when residents cannot afford to move away? Is it acceptable to consider education as a factor when people, for reasons that can be both beyond and within their power, do not attain it? What about telematics? How much should driver behavior impact rates when other factors influence the cost of claims?
Ultimately, if certain rating factors are banned despite their statistically sound predictability of future claims, how will that shift the perception of unfairness to other groups of drivers? How will eliminating predictive factors impact solvency?
These questions, which have been the subject of public policy debate for more than two decades, are gaining greater traction. During the last two congressional sessions, legislators have introduced bills to eliminate so-called “income proxies” including credit scoring, education level and employment status that could greatly impact how actuaries develop rates. In 2021 three states, Colorado, Michigan and Washington, either enacted legislation or implemented regulation in response to those who insist personal auto insurance rates are unfair or discriminatory.
That’s part of why property-casualty actuaries — who are required to adhere to the highest standards of professional independence — are weighing in on the conversation. In March 2022, the Casualty Actuarial Society published four papers that consider issues including fairness and disparate impact on protected classes in the United States as part of a Research Paper Series on Race and Insurance Pricing.
“There are a lot of discussions happening now to incorporate ideas of social fairness into actuarial fairness,” observed Roosevelt C. Mosley, CAS president-elect, principal for Pinnacle Actuarial Resources and co-author of the recently released CAS Research Paper, “Methods for Quantifying Discriminatory Effects in Insurance.” Mosley said, “Actuarial fairness centers around whether a risk characteristic is actuarially justified, while social fairness is about avoiding bias against certain classes.”
“There is a larger trend that concerns us,” said David Snyder, vice president of international policy for the American Property Casualty Insurance Association (APCIA), “which is the politicization of rate regulation that has increased quite a bit in the past year or so.”
In general, insurance rates are guided by states’ regulations that rates should not be excessive, inadequate or unfairly discriminatory. Ironically, personal auto insurers are generally prohibited from collecting policyholder data regarding race, income and religion, making it difficult to prove directly if insurers are unfairly pricing certain groups. More to the point, Snyder said, insurers do not want to collect such data.
However, to consumer advocates and some regulators and legislators, some variables, such as occupation, education, location and credit-based insurance scores, are unfair to African American, Hispanic and low-income policyholders.
Rep. Bonnie Watson Coleman (D-N.J.) introduced the Prohibit Auto Insurance Discrimination Act (PAID Act) in the United States Congress last year to exclude “income proxies” including credit scores and credit-based insurance scores. Other factors on the chopping block include gender, ZIP code, census tract and marital status.
During the previous session in 2020, Sen. Cory Booker (D-N.J.) introduced a bill with the same name to ban the same variables. Although the PAID Act has a low chance of passage, its introduction appears contrary to the 1945 McCarran-Ferguson Act, which assigns insurance regulation to the states. Insurance regulators had objected to The Dodd-Frank Act of 2010, which was adjusted post enactment in response to their concerns (see, “Demystifying the Regulatory Web: Dodd-Frank and Its Complex Impact,” AR, March/April 2016).
Although the PAID Act has a low chance of passage, its introduction appears contrary to the 1945 McCarran-Ferguson Act, which assigns insurance regulation to the states.
Using factors such as location and credit-based insurance scores can imply that insurers are deliberately discriminating against certain policyholders. “I have not seen anything from a pricing perspective to say insurers are intentionally discriminatory,” observed Mosley, a 28-year auto insurance actuarial veteran. However, he explained that there could be inadvertent discrimination in the insurance system that should be examined.
In Colorado, a bill was enacted in 2021 which, among other measures, requires insurers to provide statistical evidence that the data and predictive analytics for determining premium rates do not cause unfair discrimination based on an individual’s race, color, national or ethnic origin, religion, sex, sexual orientation, disability, gender identity or gender expression in any insurance practice. Insurers are also barred from using any external consumer data and information source, algorithm or predictive model regarding the same. The law will become effective on January 1, 2023.
The Consumer Federation of America (CFA) supports the bill. The measure shines “a spotlight on pricing practices — such as the use of credit scores — that tend to result in higher premium(s) for consumers of color,” according to a June 9 CFA news release.
Snyder of APCIA sees the Colorado law differently. Determining how to comply has been challenging, he explained, because “nobody knows what constitutes an acceptable balance of correlation to a protected class versus correlation to a business operation,” he said. “And there isn’t even data for many of the protected classes to even begin the analysis,” he added, observing that the situation “will have a negative impact on all companies and especially smaller companies who would have to comply with the law.”
In 2020 Michigan also took steps to combat suspected unfair discernments, but the impact is unclear. Documentation from the Michigan Department of Insurance and Financial Services says that credit score was eliminated. When asked for clarification, Michigan’s insurance department withdrew from a scheduled interview with Actuarial Review. “The legislation banned insurers from using credit scores used by lenders for setting insurance rates,” Mosley said, “but did not ban insurers from using credit-based insurance scores for setting rates.” Rating factors that some label as discriminatory can be predictive of future costs. Consider gender. Some states, such as California, Hawaii, Massachusetts, Michigan, North Carolina and Pennsylvania, already ban the use of the gender factor.
Gender is considered one of the most unfair rating factors by 66% of respondents in a nationally representative survey of 1,095 U.S. adults called, “Which Data Fairly Differentiate? American Views on the Use of Personal Data in Two Market Settings” (see Figure 1). “People largely make these distinctions according to whether they see data as logically related to the behaviors companies are trying to predict and whether data sort individuals in morally consistent ways,” wrote Barbara Kiviat, the Stanford University professor who published the article in 2021.
Elimination of the gender factor helps to avoid potential discrimination against transgender people and others who do not identify as men or women. According to Pew Research, an estimated 1.4 million adults in the United States identify as transgender.
In states without a ban on gender in pricing, insurers tend to ask people their sex identification at birth, Mosley said. Still, there is discussion to permit policyholders to identify as they are now. He also points out that the gender factor demonstrates that the predictive value of variables can change in relevance. Teenage female drivers are becoming as risky as males in the same age group.
Mosley’s observation also raises the reality that factors can impact each other. Age, which is not on the list of questionable factors, can interact with the gender rating variable.
Other factors that some believe should be eliminated — education, occupation, marital status, location and credit-based insurance scoring — are correlated with riskier driving, according to “Behavioral Validation of Auto Insurance Rating Variables,” published by the APCIA in November 2021. The white paper’s authors are Dr. Robert Hartwig, director of the Risk and Uncertainty Management Center and clinical associate professor in the finance department at the University of South Carolina, and Robert Gordon, APCIA’s senior vice president of policy, research and international.
Drivers with some college education or less engage in hard braking 5% more frequently and generate claim costs about 5% to 10% higher than the study’s population overall, according to the white paper. The analysis reflects telematics data based on billions of miles driven from January 2017 to June 2019 submitted by APCIA member companies.
In contrast, drivers with higher educational attainment hit the brakes 5% less often and are associated with at least 5% to 20% lower claim costs compared to the study’s population overall.
Although the education factor is predictive statistically, 62% of respondents in the Kiviat study considered education an unfair variable for personal auto insurance rating. At the same time, hard braking/sharp turning was the third fairest rating variable (see Figure 1). Based on the consumer responses, the perception of fairness could improve if consumers are better educated about rating factors.
Occupation also correlates to hard braking and claim costs, according to the Hartwig/APCIA study. Depending on the policyholder’s job, claim costs can be 5% to 10% higher or lower than the study’s population of drivers. Hartwig pointed out that the occupational risk groups evaluated are not income-based and represent people in every race and ethnicity. The Hartwig/APCIA study noted that teachers, for instance, do not hit the brakes as often as real estate agents, who are often driving in unfamiliar neighborhoods.
Looking at marital status, the Hartwig/APCIA study shows that married people hit the brakes about 10% to 15% less frequently than the general population, compared to single people who do so 10% more frequently. Being married is associated with 20% lower claim costs compared to the general population, while claim costs were 15% greater for singles.
Telematics demonstrates that population density is a “highly accurate predictor of insurance costs,” according to the Hartwig/APCIA analysis. Drivers in densely populated areas engage in hard breaking about 10% more frequently than the overall study population, the study noted. In contrast, motorists in lower populated areas hit the brakes 20% less.
“With telematics becoming more popular, marital status will likely become a less important factor,” Mosley said. Telematics data will likely show that married people often travel together, reducing the potential miles of exposure.
Where a policyholder lives, often characterized by ZIP code, can be viewed as discriminatory against low-income policyholders. Location relates to accident exposure potential, Mosley said. “It is related to where you drive most often,” he explained, pointing out that densely populated areas tend to have higher accident rates.
Telematics demonstrates that population density is a “highly accurate predictor of insurance costs,” according to the Hartwig/APCIA analysis. Drivers in densely populated areas engage in hard breaking about 10% more frequently than the overall study population, the study noted. In contrast, motorists in lower populated areas hit the brakes 20% less. Claim costs are 20% higher in densely populated areas and 20% lower in rural areas. The location factor involves myriad risks including weather, repair costs and litigiousness.
To evaluate potential unequitable pricing by location, the Pennsylvania Department of Insurance looked at loss ratios in 2016 by ZIP codes characterized by majority populations of African Americans, Hispanics, low income and less educated citizens. For majority Hispanic ZIP codes, the loss ratio was 71.3% compared to the statewide average of 64%. Those representing low median income and the lowest percentage of college graduates experienced 67.8% and 67.4% loss ratios, respectively. African American jurisdictions’ loss ratios were the closest to the average at 66.9%.
“Many of these studies try to find specific groups that are paying more for insurance than another group,” Hartwig said. “In reality, insurance rating is completely blind to race and ethnicity,” he offered, adding, “These studies fail to provide any proof that insurers are discriminating against anyone.”
Credit-based insurance scoring
FICO reports that 95% of all personal lines insurers use credit-based insurance scores — and for good reason. An insurance score is one of the most statistically predictive factors around.
Attacks on credit-based insurance scores are not about actuarial soundness, Mosley said, and the contention that credit-based insurance scores are harmful to minorities and the poor is “not a universal truth.” Eliminating credit score to avoid bias, he added, “actually could end up causing more harm than good.”
The Washington state experience is a case in point. Last year, the state’s insurance commissioner banned the use of credit-based insurance scoring, but a judge later overturned it. Removing credit scoring reportedly increased premium for more than one million policyholders. Senior citizens, who tend to have favorable credit scores and may not drive as often, were purportedly the hardest hit. Since then, the state’s insurance department attempted to temporarily ban credit scoring through rulemaking. Three insurance industry groups filed a lawsuit against the ban and the matter is on hold.
“People across the socioeconomic spectrum have good credit and bad credit,” Hartwig said. State insurance departments and the federal government have studied the efficacy and fairness of credit-based insurance scores for decades, he explained. They reach the same conclusion that “credit-based insurance scores are highly correlated with loss, and their use in rating models increases rate accuracy, enhances competition and contributes to an overall rate structure that is fair.”
Years before the credit score ban in Washington, the Arkansas Insurance Department found the rating factor was advantageous for the 57.4% of consumers who received a decrease in their auto insurance premium. About 19.2% saw no effect and 23.4% realized an increase, according to a 2017 report.
The cost to insure policyholders with low credit-based insurance scores, as measured by pure premium, is about 28% higher than the driving population overall, according to the Hartwig/APCIA report. Using insurance scoring has also reduced the population of state auto insurance plans of last resort because insurers can take on higher risks, the report noted.
Consumers are mixed on using credit information to rate personal auto coverage. The Kiviat study ranked it fourth in fairness. Thirty-six percent believe its use is fair and 18% were neutral, but 46% saw it as unfair (see Figure 1).
A couple of insurtech companies are looking to get away from credit-based insurance scoring by relying more on telematics data. Root Insurance has committed to eliminating credit-based insurance scoring for rating by 2025 to reduce what it calls “unfair discriminatory biases” in insurance rates.
Lemonade will use “a continuous stream of data” in lieu of proxies, said Yael Wissner-Levy, the insurtech’s vice president of communications, in a November 2021 news release announcing the company’s purchase of Metromile.
These proxies, she wrote to Actuarial Review in an email, are age, gender, marital status and profession. “To be clear, Metromile does use those proxies at the moment, but their architecture underweights them already, and over time, should do away with them entirely,” she explained.
However, some telematics measures could be considered discriminatory. Tracking time of day, for example, could be regarded as unfair to blue collar workers who are more likely to be working at night and thus driving at more risky times, Mosley said. When a person drives was considered less fair than credit scoring, according to the Kiviat study.
According to an October 2021 Consumer Reports article, “What You’re Giving Up When You Let Your Car Insurer Track You In Exchange for Discounts,” nine out of the 10 insurers’ telematics programs included in the article, including those of Allstate and USAA, track time of day. All 10 insurers track hard breaking. Six monitor distance and phone use. Five record speeding and accelerating. Insurers also vary on data use.
Telematics-based insurance is still evolving and is relied upon by nine out of the top 10 personal auto carriers for collecting data for their telematics programs, observed Ryan McMahon, vice president of strategy at Cambridge Mobile Telematics. Soon, data “is going to be more contextual with better sensors, revealing better granularity of risk,” he explained. For instance, telematics can now track whether a driver is distracted not from holding a smart device but from touching it.
Despite the presence of telematics for well over a decade, driver adoption has been slow. Depending on the study, many drivers are not aware of it or do not know what it is. A survey by J.D. Power released in June 2021 reports that 16% of respondents are using telematics and 34% of auto insurance customers are willing to try usage-based insurance.
However, telematics can also lead to premium increases, (see “Getting Personal — Can IoT do for Homeowners Insurance What Telematics Did for Auto Coverage?”, AR, May/June 2021). Although the Consumer Reports article is positive about the potential for basing premiums on how people drive instead of “biographical details” such as location, the article also expresses concerns about consumer privacy.
Some consumers and actuaries share concerns about personal data privacy. Sixty percent of the youngest generation of drivers, known as Generation Z, feel some discomfort with sharing location data, and 45% are uncomfortable sharing driving data, according to a survey conducted by The Zebra in 2021. The elder Gen X age group (45 to 54) is more likely to understand what telematics is but are the most unwilling to select a telematics insurance policy. Millennials tend to be more comfortable with sharing data, but not by high percentages.
“Some concerns include unwanted and illegal surveillance by government or corporate entities, use of personal data for predatory practices, and repeated breaches of security of credit card and financial institutions’ financial data,” said Louise Francis, president of Francis Analytics and Actuarial Data Mining. “[This] suggests [that] many collections of personal data are not secure,” she said.
McMahon offered that one way to promote trust in sharing data is for telematics companies to be more transparent. ”We have consistently seen that consumer trust is built by showing drivers what constitutes risk, how that data is used and how it is not used,” he said. “Consumers need to know how to reduce their risk and that data is limited to a specific use.”
Developing fair rates requires a sensitive balance between multiple rating factors to assure fairness to policyholders while helping insurers achieve business goals.
Insurers rely on actuarial soundness to assure rates are fairly discriminatory to all policyholders, regardless of how they are divided into groups. Assigning individuals to clusters of socioeconomic similarity can be fraught with unfair assumptions. Even if it were possible to base rates on a person’s character, the algorithm could be unfair.
Since actuaries are intimately acquainted with rating factors and the data behind them and are required to uphold the highest standards of professional independence, they should have a greater voice in the rating variable conversation.
Annmarie Geddes Baribeau has been covering insurance and actuarial topics for more than 30 years. Find her blog at www.insurancecommunicators.com.
Federal Trade Commission, “Credit-Based Insurance Scores: Impacts on Consumers of Automobile Insurance,” Report To Congress, 2007.
Golden, Linda L., et al., “Empirical Evidence on the Use of Credit Scoring for Predicting Insurance Losses with Psycho-social and Biochemical Explanations,” North American Actuarial Journal, 20:3, pp. 233-251, DOI:10.1080/10920277.2016.1209118.
Kiviat, Barbara, “The Moral Limits of Predictive Practices: The Case of Credit-Based Insurance Scores,” American Sociological Review 2019, Vol. 84:6, pp. 1134–1158.
National Association of Insurance Commissioners, “Milestones in Racial Discrimination within the Insurance Sector,” August 2020.