Testing Auto Insurance Rates for Racial Bias Without Race Data

By Joshua Davis and Mitch Paden

Colorado requires personal auto insurers to govern and evaluate their pricing tools for racial bias, with the first compliance report due July 1, 2026. Most insurers do not collect race. Here is how evaluation is possible without it, and what separates a credible analysis from a misleading one.

Colorado Regulation 10-1-1

On October 15, 2025, Colorado’s amended Regulation 10-1-1 expanded to cover private passenger auto insurers. It applies a governance and risk-management framework to insurers that use external consumer data, algorithms, or predictive models in pricing, requiring them to evaluate those tools for bias and report to the Division of Insurance, with the first compliance report due July 1, 2026. A more detailed quantitative testing standard for auto is still being finalized. Colorado is not alone: New York’s Insurance Circular Letter No. 7 sets similar expectations, and other states are moving the same way.

That creates a practical problem: auto insurers do not ask applicants their race, nor should they. So how do you evaluate pricing for racial bias when you do not have anyone’s race?

Inferring what you don’t collect

The standard method for approximating race is Bayesian Improved First Name Surname Geocoding, or BIFSG. Names and neighborhoods carry probabilistic information about race and ethnicity, and that information is public, in U.S. Census data. A surname like Nguyen, a first name like Jorge, and the demographic makeup of a ZIP code each shift the odds. BIFSG combines all three into an estimated probability that a given policyholder belongs to each racial or ethnic group.

The output is a set of probabilities across racial and ethnic groups rather than a label on any individual, and over a book of business those probabilities let you compare premiums, losses, and loss ratios between groups and ask whether the pricing treats similar drivers differently. It builds on the same proxy approach the Consumer Financial Protection Bureau (CFPB) has used in fair-lending analysis and that state regulators have used in their own insurance studies. The method itself is well understood; producing results that hold up to a regulator or a court is a different matter, and it is where the actuarial judgment lives.

What rigorous testing has to handle

Three problems separate a credible test from a misleading one, and each can quietly flip the conclusion.

The first problem is the most familiar: a premium difference is not a disparity. A premium gap between groups is not discrimination unless it is larger than the difference in cost behind it. The District of Columbia’s 2024 DISB study illustrates the complexity when it studied auto premiums: minority drivers paid more premium, but their incurred losses ran higher still. The honest yardstick is the loss ratio, incurred losses measured against earned premium, and a credible test still has to separate a real premium gap from one that only reflects where people live or what they drive.

The second problem is subtler, and it starts with the fact that you never observe race at all, only estimate it, and the errors in that estimate are not random. Research on these methods finds that misclassification tracks socioeconomic status, with minorities in higher-income areas more often read as White and White residents in lower-income areas read as non-White. The Society of Actuaries’ 2024 review of imputation methods adds that they identify American Indian, Alaska Native, and multiracial people poorly, and that the Census suppresses counts for rare surnames in a way that falls hardest on smaller groups. A related problem is incomplete matching: when a policyholder’s surname does not appear in Census data, such as frequently happens with compound surnames common in Latino communities, the record cannot be assigned a race probability so it drops from the analysis. Random error would average out over a large book, but this demonstrates that records that drop are actually concentrated in the very groups the test exists to protect. It does not merely add noise, it bends the result, and an analysis that ignores it can produce a clean-looking number that is wrong exactly where it matters most.

The third problem is that the outcome depends on choices for which there is no agreed standard, so the same book can support different conclusions depending on who runs the test. Even the direction of the distortion is contested. Return to that misclassification: suppose Black drivers are in fact charged ten percent more than White drivers of the same risk. Because the imputation misreads some Black drivers as White and some White drivers as Black, each estimated group becomes a blend of the two, and the measured gap comes out smaller than the true ten percent. The disparity is real, but the proxy dilutes it, and a test taken at face value understates it.

That is the most common way imputation distorts the answer, but not the only one. Several fair-lending studies have found the reverse, that it can overstate disparate impact when the errors fall unevenly and invent gaps that were never there. The field genuinely disagrees about which way the bias runs, and the answer turns on how the proxy is built and used. That is not a reason to distrust the work; it is a reason to fix the methodology in advance, document it, and check whether the conclusion holds when the choices change.

Test for fairness pricing now

The July 1, 2026 Colorado compliance deadline for bias testing in personal auto insurance marks the beginning of documentation standards and annual reporting requirements.

Colorado’s regulation may signal where other jurisdictions are headed to ensure fairness in personal lines pricing. If your company uses external data, insurance credit scores, algorithms, or predictive models, it is worth assessing your program for potential discriminatory impacts across protected classes now.

Perr&Knight has the experience and expertise to provide this bias-testing analysis end to end: the data engineering and the testing that separates a real disparity from a cost-justified difference. Reach out to our consulting actuaries to provide the insights you need to ensure fair pricing in your personal lines insurance programs.

Society of Actuaries, Statistical Methods for Imputing Race and Ethnicity, April 2024. https://www.soa.org/globalassets/assets/files/resources/research-report/2024/stat-methods-imputing-race-ethnicity.pdf

DC Department of Insurance, Securities and Banking, Evaluating Unintentional Bias in Private Passenger Automobile Insurance, November 2024. https://disb.dc.gov