Dear Editor,
I am writing to Actuarial Review with questions and comments on the article “Regulating Fairness” that was published in the May-June 2025 issue.
The article repeatedly used the term “bias” but failed to give the mathematical definition of bias as applied to the cases mentioned in the article. My impression is that the term is not being used in the same sense that modelers would use it, which is that for an unbiased model, the gap between observed and the modeling estimates will tend towards zero, but seeing the formula behind the article’s definition of bias would help confirm that. I believe that the article was, in fact, arguing that a disparate impact analysis should override the results from a model that is unbiased in the statistical sense if the results of an unbiased modeling exercise would, under a disparate impact analysis, disadvantage members of a protected class.
Actuaries have an obligation to protect clients against the potential effects of adverse selection by providing advice to clients on expected loss costs or loss ratios when selecting factors for a given rating plan. Our clients rely on us to call balls and strikes without fear or favor. I believe that if an actuary fails to show the best estimate of the expected loss ratios to management for a given rating plan, that individual would not meet the obligation to provide reliable advice to a client. A separate disparate impact analysis could provide useful information to company management, though.
Sincerely,
Michael R. Larsen, FCAS
