Summary
Einav, Finkelstein, Ryan, Schrimpf, and Cullen (2013) decompose adverse selection in health insurance into two components: traditional "selection on levels" (expected health risk) and a novel "selection on slopes" (anticipated behavioral/moral hazard response to coverage). Using a structural model estimated on Alcoa Inc. employee-level panel data with quasi-experimental variation from staggered union contract timing, they find substantial heterogeneity in moral hazard (coefficient of variation [CV] > 2) and that selection on moral hazard is roughly as important in plan choice as selection on health risk. A key policy implication: standard estimates of spending reductions from high-deductible plans substantially overestimate the actual savings because low-moral-hazard types self-select into those plans first. Published in American Economic Review 103(1): 178–219.
Key Claims
- Adverse selection has two components: "Selection on levels" (traditional adverse selection on expected health spending) and "selection on slopes" (selection on the incremental spending that arises from having insurance, i.e., moral hazard). The paper is the first empirical analysis of the latter.
- Average moral hazard: arc elasticity ≈ −0.14. Moving employees from old plans (12.6% average out-of-pocket [OOP] share) to new high-deductible plans (28.4% OOP share) reduces spending by ~$600/year (11–17%). Arc elasticity ≈ −0.14, consistent with the RAND experiment (−0.2) and the quasi-experimental literature range (−0.1 to −0.4). Spending reduction arises entirely through doctor/outpatient visits; no discernible effect on inpatient spending.
- Massive heterogeneity in moral hazard (CV > 2): Average moral hazard parameter ω = $1,330 (≈30% of expected health risk), but standard deviation = $3,200 (coefficient of variation > 2). Spending reduction from switching no-deductible→high-deductible: mean $348, median only $48, 90th percentile > $1,000. Standard deviation ($749) is more than twice the mean.
- Selection on moral hazard ≈ selection on health risk: Moving from the 10th to the 90th percentile of the moral hazard distribution reduces demand for the high-deductible plan by 23 percentage points (pp). Moving from the 10th to the 90th percentile of expected health risk reduces demand by 24 pp. Selection on risk aversion is considerably less important (15 pp range). In other words, adverse selection based on the "slope" of spending (moral hazard) is roughly as quantitatively important as adverse selection based on the "level" (expected health).
- High-deductible plan introduction overestimates savings when ignoring selection on moral hazard: When 10% of employees endogenously select the high-deductible plan, average spending reduction for those who select it = $131 — less than 40% of the $348 population average. Selection on moral hazard means that low-behavioral-response types (who benefit least from insurance coverage) are the first to choose the high-deductible plan when premiums rise. Standard quasi-experimental estimates of average moral hazard — which randomize plan assignment — will therefore substantially overestimate the spending reduction from introducing high-deductible options in voluntary-choice settings.
- Welfare decomposition: Selection on moral hazard accounts for $34 of the $52 welfare cost of adverse selection (65%). However, eliminating selection on moral hazard alone (without eliminating the underlying moral hazard distortion) yields only $25/employee welfare gain — 5% of the $490 gain from eliminating moral hazard entirely. This implies that monitoring technologies that reduce moral hazard may have ancillary benefits for adverse selection, but that the primary welfare gains are from reducing the moral hazard distortion itself.
- Data and identification: Alcoa Inc., 2003–2006. ~7,500 employees. Old plan options (2003–2004): three plans with low cost-sharing (13% average OOP share). New plan options (introduced staggered across union groups 2004–2006): five plans with high cost-sharing (28% average OOP share). Quasi-random treatment assignment from union contract timing. Model estimated via Markov Chain Monte Carlo (Gibbs sampler).
- Connection to essential heterogeneity: Individuals select insurance coverage partly based on their anticipated treatment effect (behavioral response to coverage). This is the health insurance analog of Heckman-Urzua-Vytlacil's essential heterogeneity: individuals who select higher coverage (treatment) are precisely those who respond more to it, making the local average treatment effect (LATE) and average treatment effect (ATE) diverge.
Concepts Introduced or Extended
Entities Mentioned
Quotes
"We use a model of plan choice and medical utilization [to] present evidence of heterogeneous moral hazard as well as selection on it... abstracting from selection on moral hazard could lead to overestimates of the spending reduction associated with introducing a high-deductible health insurance option."
"Heterogeneity in moral hazard is roughly as important as heterogeneity in expected health risk in determining whether to buy a higher- or lower-deductible plan."
"When only 10 percent of the employees select the high-deductible coverage, the average per-employee spending decline for those who select the high-deductible plan is just over $130" — vs. $348 for the full population average.
My Take
This is a technically sophisticated structural estimation paper that makes a genuinely novel conceptual contribution: the identification and measurement of selection on the slope of spending (moral hazard) rather than just its level (expected health). The finding that moral hazard heterogeneity is roughly as important as health risk heterogeneity for plan choice is surprising and practically significant. The policy implication — that voluntary high-deductible plan adoption produces far smaller spending reductions than what randomized experiments estimate — is directly relevant to the Affordable Care Act (ACA)/Health Savings Account (HSA) policy debate and Medicare Part D design. The main limitation is external validity: results are from a single large aluminum manufacturer with union-negotiated benefits. The structural model requires strong functional form assumptions, and the authors acknowledge the model fit for the new options is less precise than for the old options. The welfare estimates should be read as illustrative rather than precise.