Actuarial Notes

Unfair Discrimination

Study guide: Exam 6C

Unfair Discrimination in insurance rating is charging different premiums to risks that present the same expected cost, or charging on a basis that is legally or socially prohibited. Insurance is inherently discriminatory in the neutral sense — the whole point is to separate risks — so the regulatory standard is not "no discrimination" but no discrimination unsupported by expected cost and permitted by law.

\text{Fair} \iff \frac{\text{Premium}_A}{\text{Premium}_B} = \frac{E[\text{Loss}_A]}{E[\text{Loss}_B]}

  • Two distinct failures get confused and should be kept apart:
    • Rate differences with no cost basis — price optimisation by elasticity, or a relativity kept because it is profitable rather than because it is indicated. This is unfair discrimination in the classical sense.
    • Cost-based differences on a prohibited basis — sex, race, or another characteristic the law excludes, even where the loss data supports the difference. Here the rate is actuarially sound and still unlawful.
  • The second is a policy override, not an actuarial error, and the actuary's job is to say so plainly: the variable is predictive, its use is prohibited, and the consequence of prohibition is the cross-subsidy quantified in Risk Classification Restrictions.
  • Proxy discrimination is the live issue. A model that omits a prohibited variable can still reproduce its effect through correlated permitted variables, especially in high-dimensional models. Testing for it means measuring the model's output across the protected groups, not only checking that the variable is absent from the input — see Bias in Actuarial Practice.
  • Under the Standards of Practice and the fair treatment of customers framework, the actuary is expected to be able to explain a rate difference in cost terms. A relativity that cannot be explained is a conduct exposure whatever its lift.
  • Human rights legislation supplies the outer boundary; insurance statutes and rate filings supply the operational one, and the two are not the same list in every province.
  • United States: State rating laws require rates that are not excessive, inadequate or unfairly discriminatory, and the CAS Statement of Principles calls a rate that meets all three an actuarially sound estimate of expected future costs. The NAIC Unfair Trade Practices Act (§4H) lists, among unfair discrimination, refusing or limiting coverage because of sex, marital status, race, religion or national origin, and refusing or limiting a property or casualty risk solely for its geographic location unless sound underwriting and actuarial principles support it; states adopt their own versions. Colorado's SB21-169 extends the prohibition to external consumer data and predictive models whose use unfairly discriminates by protected class, which reaches the proxy problem directly (Insurance Data Regulation). Credit-based insurance scores are the leading U.S. case.

Three Rate Differences Classify each as fair, unfairly discriminatory, or prohibited-though-cost-based.

  1. A driver with two at-fault accidents pays 60\% more than an otherwise identical driver with none.

  2. Two identical risks pay different premiums because one was identified by the model as unlikely to shop at renewal.

  3. Male drivers aged 18–21 show materially higher pure premiums than female drivers of the same age, and the insurer wishes to reflect it.

Answer

  1. Fair. Accident history is causally related to expected loss, objective, verifiable, within the driver's control, and permitted everywhere in Canada. This is textbook acceptable classification.
  2. Unfairly discriminatory. The premium difference tracks elasticity, not expected cost. Identical risks are charged differently for the same coverage, which fails the definition directly, and Canadian regulators do not accept it in a filing.
  3. Prohibited though cost-based in provinces that ban sex as a rating variable. The data supports the difference; the law forbids using it. The predictable consequence is a cross-subsidy from young women to young men, and the honest filing states that rather than concealing it in a "years licensed" proxy.

Case 3 is where candidates lose marks by arguing the statistics. The statistics are not in dispute; the question is whether the variable is permitted, and it is not.

How it appears on the exam

  • Exam 6C: The example above is framed in Canadian terms: whether a rate difference is cost-based, and whether a cost-based variable is prohibited by the province.

Exam traps

  • Cost-based and prohibited are different failures: A rate difference can be fully supported by losses and still be unlawful because the variable is prohibited.
  • Omission is not fairness: A model with no protected variable can reproduce its effect through correlated inputs; the test is on outcomes by group.