Wyse and Meyer 2023 — Saved by Medicaid New Evidence on Health Insurance and Mortality from the Universe of Low-Income Adults

medicaidACA-medicaid-expansionhealth-insurancemortalitylow-incomecausal-inferencedifference-in-differencesproportional-hazardsurvival-analysisadministrative-datacost-effectivenessincome-mortality-gradientstaggered-didpreregistrationsubstance-use-disorderyoung-adult-mortalitytreatment-on-the-treatedhealth-economics

Summary

Wyse and Meyer (2023) construct a population-level administrative linkage of 37.5 million non-disabled low-income adults (ages 19–59, income <138% of the federal poverty level [FPL]) using the 2010 Census, IRS Modified Adjusted Gross Income (MAGI) records, Centers for Medicare & Medicaid Services (CMS) Medicaid enrollment data, and Social Security Administration (SSA) Numident mortality records through April 2022. Using staggered difference-in-differences (DiD) on Affordable Care Act (ACA) Medicaid expansion timing across states and a discrete-time proportional hazard model, they find that expansions raised Medicaid enrollment by +11.7 percentage points (pp) and reduced annual mortality hazard by 2.5% (intention-to-treat, ITT) or 21% for new enrollees (treatment on the treated, TOT). The paper estimates that 27,400 lives were saved in expansion states 2010–2022, with 43% of the ~830,890 life-years saved accruing to adults ages 19–39 due to their longer remaining life expectancy. The paper uses a preregistered analysis plan — only the second nonexperimental economics study to do so — and finds that universal Medicaid coverage would close only 5–20% of the mortality gap between the highest and lowest income quintiles.

Key Claims

Concepts Introduced or Extended

Entities Mentioned

Quotes

"Our best estimates suggest that universal Medicaid would close roughly 12% of the gap in mortality rates between those in the highest and lowest income quintiles. Since insurance is just one of many factors that contribute to health and mortality, this finding suggests that health policies that focus solely on expanding health insurance coverage are only one part of the comprehensive strategy needed to close the income–mortality gradient."

"We find consistent, negative point estimates for all age groups, all racial and ethnic groups, both genders, all income levels, all family structures, and both employment statuses... the effect estimates for those in their 30s, 40s, and 50s account for most of the lives saved, but those in their 20s and 30s account for a large share of the life-years saved."

"We estimate that ACA Medicaid expansions led to approximately 27,400 lives saved in expansion states from 2010 to 2022, which equals approximately 3,220 per year."

My Take

This is the strongest paper in the Medicaid-mortality literature by most methodological dimensions: largest sample by far, preregistered analysis plan, best income identification (IRS MAGI vs. survey self-report), appropriate survival model (proportional hazard vs. LPM). The −21% TOT estimate is credible and well-supported. The finding that prior studies' very large estimates (Goldin et al. 2021: ~122%) are excluded by this paper's CI is an important methodological contribution — it identifies overcorrection as a likely source of upward bias in small-sample studies.

The most novel and provocative finding is the 5–20% gradient closure from universal Medicaid. This directly challenges the "healthcare access explains SES health disparities" narrative: even a fully-insured population would retain 80–95% of its income-mortality gradient, driven by other determinants (education, behaviors, built environment, stress). The paper is appropriately humble about this — it uses Chetty et al.'s gradient and applies Wyse-Meyer's TOT to compute implied closure, which involves extrapolation.

The young adult mechanism (SUD/external causes) is the most speculative element. It is plausible and the circumstantial evidence is strong (ACA SUD benefit expansion, overdose-dominated young adult deaths), but the paper has no utilization or cause-of-death data to confirm it. Future work with linked cause-of-death data could resolve this.

One methodological caveat: the staggered DiD design is theoretically sensitive to heterogeneous treatment effect timing of the kind flagged by Callaway-Sant'Anna (2021) and Goodman-Bacon (2021). The paper addresses this through event studies and triple-differences but does not formally implement recent staggered-DiD estimators — an acknowledged limitation. With treatment effects that appear to grow over time, this may matter.