Homework 1 — Adverse Selection in Real Markets

Instructions

This homework covers health insurance purchasing and adverse selection. It is empirical only — the theory for this module is assessed in class (quizzes and the midterm).

You are required to use an AI coding assistant for this assignment. Use GitHub Copilot (free with your Emory GitHub account) or your tool of choice, and let it write the R or Python. Your job is not to produce the code. Your job is to direct the tool, check what it returns, and interpret it as a health economist. The AI will frequently hand you a confident, clean, and wrong answer. Catching that is the assignment.

Submit a rendered notebook (Quarto, R Markdown, or Jupyter) containing your code, its output, and your written answers. Each part is worth 3 points, graded:

  • 3 points: correct, with sound economic reasoning
  • 2 points: close, minor error or thin reasoning
  • 1 point: attempted but the economic judgment is missing or wrong
  • 0 points: no work, or the unexamined AI output pasted in

Verification note (required, embedded): somewhere in your notebook, flag one specific point where the AI’s first output was wrong, incomplete, or misleading, and explain in economic terms how you knew. “It ran without errors” does not count. This is folded into the parts below, not a separate line item.

Homework 1 is due by midnight on Friday, September 25.

Q1 — New Jersey’s individual market, 2000–2008 (15 pts)

Data: hwk1_nj.csv on the shared OneDrive (link on Canvas). Monthly family premiums and year-end “standard” enrollment for three plans in the Individual Health Coverage (IHC) program: Horizon BCBS of NJ Plan D ($1,000 deductible), Aetna Life Plan D ($1,000 deductible), and Celtic Insurance Plan C ($2,500 deductible).

  1. Have the AI build a table of annual premiums for the three plans and total standard enrollment, and plot premiums and enrollment over time.

  2. New Jersey’s market is often called a textbook “death spiral.” Shown rising premiums and falling enrollment, the AI will likely confirm that label. State the precise economic mechanism a death spiral requires — not just the pattern, the mechanism — and explain why the rising-premium / falling- enrollment pattern alone is consistent with but not proof of it.

  3. Two of these plans carry $1,000 deductibles and one carries $2,500, and you are comparing nominal dollars across 2000–2008. Identify how each of these makes a raw premium comparison misleading, and have the AI correct at least one of them. Show what changed.

  4. What would you need to observe to confirm that selection is driving the premium increases, and why is it absent from this dataset?

  5. Give your bottom-line verdict — death spiral, or premium growth that merely looks like one? Defend it in three sentences.

Q2 — Selection across Medicare plan types, 2010–2018 (15 pts)

Data: hwk1_medicare.csv on the shared OneDrive. Medicare Advantage (MA) enrollment and Medicare fee-for-service (FFS) spending by year.

  1. Have the AI plot MA enrollment and FFS spending over 2010–2018 and describe both trends.

  2. The intuitive story is “healthy people pick MA, sicker people stay in FFS, so FFS costs rise.” But the direction of selection here is the opposite of the classic uninsured death-spiral case. Name what kind of selection this is from the plan’s perspective, and explain why the valence flips when everyone is already insured and the choice is across plan types rather than in-or-out.

  3. One of your FFS series is a total and behaves differently from a per-enrollee version. If MA enrollment grows, what happens to total FFS spending mechanically, even with flat per-person costs? Have the AI compute both the total and the per-enrollee series, and show which one supports a selection story and which is just composition.

  4. Risk adjustment pays MA plans more for sicker enrollees. Explain how it should bend the relationship between your two series over time, and whether the data are consistent with that.

  5. Suggest one additional mechanism (besides risk adjustment) that would dampen selection across plan types, grounded in the course model.