Chapter 35 Quiz

Twenty questions — fifteen multiple choice and five short answer — covering ethics, bias, and fairness in underwriting. Answers and brief explanations are in the collapsed key at the bottom; try the whole set before opening it.

Multiple Choice

1. In insurance, the word discrimination in its technical sense means:

  • A) treating people worse because of who they are
  • B) distinguishing among risks on the basis of their expected loss
  • C) refusing to insure protected classes
  • D) charging everyone the same price regardless of risk

2. Why must insurance discriminate by risk to function at all?

  • A) because regulators require risk-based pricing in every state
  • B) because flat pricing across unequal risks triggers an adverse-selection death spiral that destroys the pool
  • C) because it maximizes the insurer's profit
  • D) because the law of large numbers requires identical premiums

3. Which of the following is categorically prohibited as a rating factor, with no actuarial exception, in essentially every state and line?

  • A) credit-based insurance score
  • B) the applicant's race
  • C) a driver's at-fault accident history
  • D) a building's fire-protection class

4. Proxy discrimination is best described as:

  • A) deliberately using a protected class as a rating factor
  • B) a facially neutral, permitted factor that functions as a stand-in for a prohibited characteristic
  • C) charging different prices to different states
  • D) any rating factor that lowers the loss ratio

5. An insurer rates by ZIP code and uses no racial variable, yet produces racially disparate prices. This is possible because:

  • A) the insurer is breaking the law on its face
  • B) ZIP code, in a segregated society, carries racial information, so pricing by ZIP prices partly by race
  • C) ZIP code is a protected class
  • D) the model contains a hidden race variable by accident

6. Deleting the protected variable (e.g., race) from a machine-learning model:

  • A) guarantees the model can no longer discriminate by that characteristic
  • B) does not prevent discrimination, because the model can reconstruct the characteristic from correlated features
  • C) is required by GINA for all lines
  • D) automatically satisfies disparate-impact testing

7. Disparate impact refers to:

  • A) the intent to discriminate against a protected group
  • B) a discriminatory effect on a protected group from a facially neutral practice, regardless of intent
  • C) a rating factor that affects all groups equally
  • D) the difference between a model's lift and its Gini

8. Which mechanism of algorithmic bias describes a model whose own pricing decisions distort the data it later retrains on, compounding the bias over time?

  • A) biased training data
  • B) proxy variables
  • C) a feedback loop
  • D) calibration drift

9. The chapter states that demographic parity, equalized odds, and calibration:

  • A) are three names for the same fairness condition
  • B) can all be satisfied simultaneously by any well-built model
  • C) are mutually incompatible when base loss rates genuinely differ across groups
  • D) are required together by Colorado SB21-169

10. The federal Genetic Information Nondiscrimination Act (GINA) prohibits the use of genetic information in:

  • A) life, disability, and long-term-care insurance
  • B) health insurance (and employment), but not life, disability, or long-term-care insurance
  • C) all lines of insurance without exception
  • D) auto and homeowners insurance only

11. Colorado's SB21-169 is significant primarily because it:

  • A) bans all use of predictive models in insurance
  • B) requires insurers to test algorithms and external data for unfairly discriminatory outcomes — focusing on effect, not just intent
  • C) prohibits credit-based insurance scoring nationwide
  • D) repeals the McCarran-Ferguson Act

12. Price optimization is widely deemed unfairly discriminatory because it:

  • A) uses protected classes directly
  • B) sets premiums partly on a customer's price sensitivity, severing price from expected cost
  • C) charges all customers the same price
  • D) is illegal under GINA

13. Which statement best captures actuarial fairness?

  • A) a price is fair when it does not deepen existing inequality
  • B) a price is fair when it accurately reflects the expected cost of the risk
  • C) a price is fair when every group pays the same average premium
  • D) a price is fair only when set by a regulator

14. The historical practice of denying or pricing up insurance for whole neighborhoods based on racial composition, marked on color-coded maps, is called:

  • A) price optimization
  • B) community rating
  • C) redlining
  • D) experience rating

15. A model is calibrated (a 7 means the same expected loss for everyone) but declines Group A at twice the rate of Group B. The correct underwriting conclusion is:

  • A) the model has a bug and must be discarded
  • B) the model is calibrated yet produces a disparate impact, and deciding what to do is a documented values judgment, not a technical fix
  • C) the model is illegal in all states
  • D) calibration proves the model is fair, so no action is needed

Short Answer

16. Explain why "we don't use race in our model" is not a sufficient defense against a charge of racial discrimination in pricing.

17. Insurance must discriminate by risk, yet must never discriminate by protected class. State the four operational tests the chapter offers for telling the legitimate sorting from the forbidden kind.

18. Distinguish actuarial fairness from social fairness, and give one example where they point in opposite directions.

19. Why is a coastal ZIP code's higher property rate easier to defend against a proxy-discrimination challenge than an inner-city ZIP code's higher rate? (Refer to the causal loss story.)

20. State the fourfold "underwriter's duty" the chapter sets out for handling fairness — and explain why it does not require the underwriter to single-handedly resolve the actuarial-versus-social tension.

Answer Key (click to expand) **1. B** — Technically, to discriminate is to *distinguish*; insurance distinguishes among risks by expected loss. The slur meaning (A) is the *unfair* kind the law forbids. **2. B** — A flat price across unequal risks makes coverage a bargain for bad risks and a rip-off for good ones; the good risks leave, losses rise, the pool spirals. Risk classification is the cure for adverse selection (Chapter 1). **3. B** — Race is categorically off-limits regardless of any correlation with loss. The others are either clearly risk-based (C, D) or contested-but-not-categorical (A). **4. B** — Proxy discrimination is the laundering of a prohibited basis through a permitted, neutral factor that is tightly correlated with it. **5. B** — Residential segregation makes ZIP code carry racial information; pricing by ZIP therefore prices partly by race without any racial variable appearing. **6. B** — A capable model reconstructs a hidden variable from its correlates. Colorblindness disables your ability to *see* proxy effects, not the model's ability to produce them. **7. B** — Disparate impact is about *effect*, measured by comparing outcomes across groups, independent of intent. **8. C** — A feedback loop: the model's decisions skew the data it retrains on, ratcheting a small bias into a larger one. **9. C** — It is a theorem: when base rates differ, the three definitions conflict, so an insurer can satisfy at most a subset and must *choose* — a values decision. **10. B** — GINA covers health insurance and employment but leaves a "gap": life, disability, and long-term-care insurance are not protected. **11. B** — SB21-169 requires testing of algorithms and external data for unfairly discriminatory *outcomes*, reaching the disparate-impact gap the old unfair-trade-practices acts left open. **12. B** — Price optimization prices *willingness to pay* rather than expected cost, breaking the §35.2 principle that price differences must reflect cost differences. **13. B** — Actuarial fairness = price reflects expected risk. (A) and (C) are social-fairness ideas. **14. C** — Redlining; the name comes from the 1930s color-coded residential security maps. **15. B** — Both facts can be true at once because calibration and demographic parity diverge when base rates differ; the response is a documented governance judgment, not a code fix (and certainly not "no action"). **16.** Because a model with no race variable can still reconstruct race from correlated features (ZIP code, name, behavior) and price by it through *proxy discrimination*; and because the test of unfair discrimination is increasingly *effect* (disparate impact), not stated intent. The only real defense is to *measure* the protected-group impact, which colorblindness prevents. **17.** (1) *Protected-class test* — is the factor a prohibited characteristic? (categorical, no exception). (2) *Actuarial-justification test* — does the factor have a genuine relationship to expected loss? (3) *Disparate-impact test* — does the factor fall much more heavily on a protected group in effect? (4) *Causation-vs-correlation test* — is the factor closer to a cause of loss or to a mere correlate of a protected characteristic? **18.** *Actuarial fairness* = the price accurately reflects expected cost (you pay for what you bring). *Social fairness* = the price does not deepen inequality or deny essential access (some risks should be protected even at a subsidized price). They conflict, e.g., when charging a working-class coastal community the full actuarial hurricane rate is actuarially fair but renders the community uninsurable — the protection gap (Chapter 30). **19.** A coastal ZIP's rate is driven by a *causal* peril — real, physical hurricane exposure (Chapter 30) — which gives it a strong causation-vs-correlation defense. An inner-city ZIP's higher rate, by contrast, may be driven by a factor that correlates with race without a clean causal loss story, placing it in the proxy zone of §35.3. The strength of the *causal* story is what separates them. **20.** The duty is fourfold: *competence* (understand your factors/models, including proxy and disparate-impact effects), *candor* (defend every price with a real loss rationale; escalate suspected proxies), *compliance-plus* (meet the legal floor and treat disparate-impact testing as a genuine audit), and *humility* (recognize that the biggest fairness questions — community rating, residual markets, the protection gap — are decided above the desk). It does not require resolving the actuarial-versus-social tension because that conflict is a societal values question no single underwriter can or should settle in a rate filing; the duty is to contribute honest analysis, not to smuggle in an answer.