Chapter 35 Exercises

Work these with the chapter's central discipline: for every rating factor and every model output, ask the four tests — is it a protected class? does it have a real loss rationale? what is its disparate impact? is it closer to a cause or a coincidence? — and refuse to let "the model said so" stand in for an answer. Items marked with a dagger (†) have worked solutions in Appendix: Answers to Selected Exercises; the rest are for discussion or self-test. Section references like (§35.3) point you back to the chapter.

A. Recall and definitions

  1. † State, in one sentence each, the difference between fair discrimination and unfair discrimination as the words are used in insurance, and explain why insurance cannot function without the first. (§35.1)
  2. Define proxy discrimination and give one original example of a facially neutral factor that could function as a proxy for a protected class. (§35.3)
  3. Define disparate impact and explain how it differs from intentional discrimination. (§35.4)
  4. † List the four operational tests the chapter gives for deciding whether a rating factor crosses from fair into unfair discrimination. For each, give a one-line statement of what it checks. (§35.2)
  5. What is redlining, where does the name come from, and why does its legacy make geography a contested rating dimension today? (§35.5)
  6. Distinguish actuarial fairness from social fairness in one sentence each. (§35.7)
  7. Define price optimization and state the single reason most regulators consider it unfairly discriminatory. (§35.7)
  8. † Explain the "GINA gap": what does the Genetic Information Nondiscrimination Act protect, and which lines of insurance does it not reach? (§35.6)

B. The paradox and the line

  1. A consumer advocate argues that "any system that charges different people different prices is unjust." Using adverse selection (Chapter 1), explain what would happen to an insurance market that took this advice literally — and then state the strongest point on the advocate's side that survives your rebuttal. (§35.1, §35.7)
  2. † For each factor, state whether it is (i) clearly permissible risk-based pricing, (ii) clearly prohibited, or (iii) a contested proxy that requires examination — and justify in one sentence: (a) a driver's at-fault accident in the last three years; (b) the applicant's religion; (c) a credit-based insurance score; (d) the fire-protection class of a building; (e) the applicant's ZIP code in a racially segregated metro area. (§35.2, §35.3)
  3. Explain why removing the protected variable (e.g., race) from a model does not guarantee the model stops discriminating by that characteristic. (§35.3, §35.4)
  4. The chapter says proxy effects must be examined, not automatically banned or automatically allowed. Lay out the genuine argument on each side of whether a predictive-but-disparate factor should be used. (§35.3)

C. Algorithmic bias

  1. † Name the three distinct mechanisms by which a machine-learning model can produce biased outcomes, and for each give a one-sentence description. (§35.4)
  2. Explain why a feedback loop is "the most insidious" mode of algorithmic bias. Walk through one full turn of the loop using a neighborhood-pricing example. (§35.4)
  3. † The chapter states that three common fairness metrics — demographic parity, equalized odds, and calibration — cannot all hold at once when base loss rates differ across groups. Explain in your own words why this is a values problem and not a technical one, and what follows for the model-governance committee. (§35.4)
  4. A data scientist proudly reports that the new pricing model "uses no protected attributes and has higher lift and Gini than the old one" (Chapter 32). What is the fairness question that report leaves completely unanswered, and what audit would you ask for before approving the model? (§35.4)

D. Underwrite this submission

  1. † Underwrite the factor. Your pricing team proposes adding "number of prior insurance lapses" as a rating variable. It is predictive of loss. It uses no protected class. In backtesting it correlates with income, and in your markets income correlates with race. Walk the factor through the four tests of §35.2, state what additional analysis you would require before adopting it, and give your provisional recommendation with reasons. (§35.2, §35.3)
  2. Underwrite the model output. A model declines an applicant for homeowners coverage with a score of 8. You notice the applicant lives in a historically redlined neighborhood and that the model leans heavily on a "neighborhood risk" feature built from third-party data. State what you would do before issuing the declination, and why a fairness override here is harder to justify on a single file than Chapter 32's accuracy override. (§35.4, §35.5)

E. Find the red flag

  1. † Find the red flag. A vendor pitches your company a "next-generation" rating model that, it boasts, "personalizes every premium to exactly what each customer is willing to pay, maximizing retention and margin." Identify the red flag, name the practice, and state which test from §35.2 it violates. (§35.7)
  2. A territory-rating proposal sets premiums at the level of small, hand-drawn geographic zones rather than standard, loss-justified territories. The zones, plotted on a map, closely track the racial composition of the city. Identify the red flag and the historical practice it most resembles, and say what would make a geographic factor defensible instead. (§35.5)
  3. An accelerated-life-underwriting program quietly incorporates a customer's prescription-drug history and a third-party "wellness score." One input turns out to be derived partly from genetic-test data. Identify the red flag and the statute it implicates, noting the line where the statute does and does not protect the applicant. (§35.6)

F. Price / analyze this risk

  1. † A model is calibrated — a score of 7 means the same expected loss for every group — but its decline rate (share of applicants scored ≥ 7) is twice as high for Group A as for Group B. Explain, in plain terms, how both facts can be true at once, which fairness definition each fact corresponds to, and why this is precisely the §35.4 dilemma rather than a bug to be fixed. (§35.4)
  2. Sketch (in words, or as pseudo-code) the disparate-impact audit you would run on a pricing model: what you would group by, what outcomes you would compare, and the one rule you must never violate about how the protected attribute is used. (§35.4)

G. Memo and ethics

  1. † Write the memo. A regulator has asked your company to justify its continued use of a credit-based insurance score after a consumer group alleged disparate impact. Draft a short (200–300 word) internal memo to your CUO laying out (a) the actuarial case for the factor, (b) the disparate-impact concern, (c) what testing you recommend, and (d) your honest recommendation. Do not resolve the tension glibly in either direction. (§35.2, §35.3, §35.6)
  2. Ethics dilemma. Your model under-prices a low-risk, affluent segment and over-prices a higher-risk, lower-income segment — both accurately, because the loss rates really do differ. A colleague says, "It's calibrated, so it's fair, end of story." Write your response: is calibration sufficient for fairness? Argue the strongest version of both the actuarial and the social position before stating where you come down and why. (§35.4, §35.7)
  3. Ethics dilemma. You discover that a profitable rating factor your team uses is, on examination, a near-pure proxy for a protected class with only a weak independent causal loss story. Removing it will raise the company's combined ratio in that segment and the broker channel will complain. Lay out your obligations under the four-part "underwriter's duty" of §35.7 and what you would actually do. (§35.7)
  4. The chapter names two opposite traps on this topic — the technocrat's trap and the advocate's trap. Define each, give a sentence of dialogue that exemplifies it, and explain why both are failures of the same underlying skill. (§35.7)

H. The Underwriting File

  1. † Underwriting-File extension. Apply the four §35.2 tests to the Harbor Steel property price (debit-rated for the aging roof, the two fires, and the coastal location). Show why each driving factor has a clean causal loss story and conclude whether the pricing is defensible on fairness grounds — and explain why commercial pricing like this raises the proxy problem far less than personal-lines pricing does. (The Underwriting File; §35.2, §35.7)
  2. The Harbor Steel decision is actuarially fair, yet it sits inside a social-fairness problem. Name that problem (using the Chapter 30 term), explain why no single underwriter created or can solve it at the desk, and state honestly what belongs in the file about it and what belongs "above the desk." (The Underwriting File; §35.7)
  3. Underwriting-File extension (David Okafor). Okafor's life-insurance file now contains a genetic-test result. Explain what the law (§35.6) permits a life insurer to do with it, why that is actuarially defensible, and the strongest social-fairness argument against using it. Then state, with reasons, what you think a responsible insurer should do — acknowledging this is a contested values question, not a settled one. (The Underwriting File; §35.6, §35.7)