Adverse Selection

The death spiral, step by step

An insurance company sells a uniform policy at premium P. The pool contains both high-risk customers (with expected annual claims C_H) and low-risk customers (C_L), with C_H > C_L. The actuarially fair premium is the pool average P̄ = α C_H + (1 − α) C_L, where α is the share of high-risk types.

Each customer decides whether to buy. A risk-neutral type buys if and only if the premium does not exceed their own expected loss, possibly adjusted by a willingness-to-pay multiplier reflecting risk aversion. If P̄ exceeds C_L·(willingness) for some low-risk customers, those customers do not buy. The pool that remains has a higher proportion of high-risk types. The new fair premium P̄' is higher than P̄. More low-risk customers exit. The spiral continues until either the market clears at a high-α pool of high-risk customers (partial collapse) or no one buys at all (full collapse).

Akerlof's 1970 paper showed the limit case: when buyers cannot tell quality, only the worst trade. Rothschild and Stiglitz's 1976 paper showed that even with two types, simple uniform contracts can fail to support a pooling equilibrium, and the only stable arrangement is a screening menu that separates the two types by self-selection.

Worked example: a numerical death spiral

Suppose a market with 1,000 potential buyers: 200 high-risk (expected claims $10,000/year) and 800 low-risk (expected claims $2,000/year). Population average claim cost is 0.2 × $10,000 + 0.8 × $2,000 = $3,600. The insurer prices at $3,600.

1. Round 1. Premium $3,600. Each low-risk type values insurance at their expected cost plus a small risk-aversion premium of $400, so their reservation price is $2,400. They will not buy at $3,600. All 800 low-risk types drop out. The pool is now 200 high-risk customers with total expected cost $2,000,000 ÷ 200 = $10,000 per person.
2. Round 2. Premium repriced at $10,000. High-risk types value insurance at their expected cost plus the same $400 premium = $10,400. They still buy. Market equilibrates at premium $10,000, enrollment 200, total written premium $2,000,000. Total welfare gain compared to no insurance: 200 × $400 = $80,000.
3. Counterfactual under symmetric information. If the insurer could observe type and price separately, low-risk types pay $2,400 (their reservation), high-risk types pay $10,400, all 1,000 buy. Welfare gain: 1,000 × $400 = $400,000. Five times larger.

The information asymmetry has destroyed $320,000 of welfare per year by knocking the low-risk types out of the market. Real-world adverse selection rarely produces such clean numbers — but the empirical literature on insurance markets routinely finds welfare losses of 5-25% of total premium volume from this channel (Einav, Finkelstein, Cullen 2010 on employer health plans).

The Harvard case (1993-1995)

Cutler and Reber's Quarterly Journal of Economics paper (1998) documents one of the cleanest natural experiments in the literature. Harvard University's employee health plan offered both a generous PPO and a cheaper HMO. In 1993, Harvard moved from paying the full premium difference to making employees pay the marginal cost of the more expensive plan.

• Before reform: PPO premium for single coverage $980/year (Harvard absorbed the cost).
• 1995, two years after reform: PPO premium $1,800/year; HMO premium $580.
• 1996: PPO premium $2,200, enrollment cut in half from 1992 baseline.
• 1997: PPO eliminated. Harvard discontinued the plan rather than continue raising premiums.

The reform forced employees to face price differences. Healthy employees moved en masse to the HMO. The PPO retained an older, sicker, more chronic-condition-heavy population. The death spiral was complete within four years. The lesson is not that PPOs are bad insurance — it is that any single insurer offering a generous plan in a market with cheaper options will be selected against unless it can pool risk by some other mechanism.

Adverse selection vs related frictions

	Adverse Selection	Moral Hazard	Signaling	Screening	Lemons	Cream-Skimming
Hidden what?	Hidden type (risk)	Hidden action (effort)	Type known by signaler	Type known by counterparty	Hidden type (quality)	Insurer cherry-picks types
Timing	Before contract	After contract	Before contract	Before contract	Before contract	Before contract
Who suffers?	Uninformed side	Insurer / principal	—	—	Market unravels	Bad types excluded
Headline remedy	Mandates, underwriting	Deductibles, monitoring	Costly credentials	Self-selection menu	Warranties, certification	Regulated rates, mandates
Canonical paper	Akerlof 1970	Holmström 1979	Spence 1973	Rothschild-Stiglitz 1976	Akerlof 1970	Newhouse 1996
Welfare consequence	Death spiral, underinsurance	Excessive risk	Deadweight signal cost	Distorted contracts	Trade collapses	Bad risks have no insurance

Adverse selection and moral hazard are the twin pillars of asymmetric-information economics. Cream-skimming is a sister problem on the supply side: insurers compete to enroll only the cheapest types, leaving expensive types stranded.

Rothschild-Stiglitz separating menu

The classic theoretical response to adverse selection is a self-selecting menu of contracts. Suppose two types: low-risk (probability of loss p_L) and high-risk (p_H). An insurer offers two contracts:

• Contract H: full coverage, high premium (priced for high-risk types).
• Contract L: partial coverage, low premium (priced for low-risk types).

Designed correctly, high-risk types prefer Contract H (they value the full coverage at their high loss probability more than the cheap partial), and low-risk types prefer Contract L (cheap, partial coverage is fine because they rarely claim). Each type self-reveals through their contract choice. The friction is that the low-risk contract must offer worse-than-actuarially-fair coverage to prevent high-risk types from mimicking — there is a deadweight loss equal to the under-insurance imposed on low-risk types.

Rothschild and Stiglitz showed something stronger: in a competitive market, a pooling equilibrium (one contract at the average premium) cannot survive. Any insurer can break the pool by offering a partial-coverage contract that attracts only the low-risk types, leaving the original insurer with the high-risk pool at a loss. The only equilibrium, when it exists, is the separating menu.

Variants and extensions

• Wilson anticipatory equilibrium (1977). Allows pooling under stricter notion of stability — predicts pooling equilibria can survive if entrants must consider competitive response.
• Riley reactive equilibrium (1979). Different solution concept that recovers existence in cases where Rothschild-Stiglitz had none.
• Empirical adverse selection in annuities. Finkelstein and Poterba (2002, 2004) measure positive selection: annuity buyers live 2-4 years longer than the general population on average, exactly because mortality-aware buyers self-select in.
• Advantageous selection. Cawley and Philipson (1999) and Hemenway (1990) show life-insurance buyers are sometimes safer than the population, because the people who buy life insurance also drive carefully — preferences and risk are positively correlated rather than independent. The selection direction depends on the joint distribution of preference and risk.
• Health-insurance death spirals post-ACA. When the individual mandate penalty was zeroed in 2019, premium-weighted average marketplace premiums rose 28% in 2018-2019 in states with the steepest healthy-customer attrition (Kaiser Family Foundation 2019).
• Mortgage-backed securities, 2006-2008. Originators retained little skin in the game and sold the worst loans into securitization pools. Akerlof and Shiller (Animal Spirits, 2009) cite this as adverse selection at trillion-dollar scale.

A brief history

Akerlof's lemons paper (1970) named the unraveling mechanism but worked through used cars. Rothschild and Stiglitz (1976) translated the friction into insurance with formal screening contracts; their paper is the technical foundation of every undergraduate adverse-selection lecture today. Wilson (1977), Riley (1979), Hellwig (1987), and many others refined the equilibrium concept. Empirical adverse selection — measuring it from data rather than theorising about it — became its own subfield with Chiappori-Salanié (2000) and the surveys by Einav-Finkelstein (2011). The 2001 Nobel Prize to Akerlof, Spence and Stiglitz cited the entire information-economics tradition; adverse selection sits at the centre.

Policy debates from the 1990s onward routinely invoke adverse selection by name. The Affordable Care Act's individual mandate was sold as a death-spiral preventer. Germany's statutory health insurance system bans risk-based premium variation precisely because it would invite cream-skimming. Medicare's premium-support proposals run head-on into the same problem: if private plans can risk-select, traditional Medicare gets the sickest. The theory is not abstract.

Common pitfalls

• Confusing with moral hazard. Adverse selection is about who joins; moral hazard is about what they do after joining. Both reduce welfare; remedies differ.
• Assuming spirals always run to total collapse. Real markets often equilibrate at a partial pool because of high willingness to pay or sticky participation.
• Treating mandates as the only fix. Underwriting, group plans, and open-enrollment windows are doing most of the work in U.S. insurance markets, mandate or no mandate.
• Ignoring advantageous selection. Sometimes the people who buy a product are the safest users (life insurance, gym memberships). Empirical sign of selection is not assumed; it is measured.
• Forgetting the equilibrium concept matters. Whether a pooling contract exists depends on whether you assume Nash, Wilson, or Riley equilibrium. Theoretical conclusions can flip.
• Believing premiums are just a pass-through. A rising premium does not merely cover claims; it also reshapes who is in the pool, which feeds back into next year's premium.
• Ignoring switching costs. If people are locked into employer plans by tax advantages, the death-spiral feedback is muted. Removing employer tax preference would amplify selection.