The evidence question the statute has to answer

The Affordable Care Act asked the federal government to do something no prior American statute had attempted at the same scale: to move the uninsured rate of a nation of more than 300 million people through a single package of insurance provisions, the principal ones taking effect on January 1, 2014. The five machines that made up that package, the coverage provisions whose effects this article measures, are mapped in detail in the companion treatment of the statute’s key provisions. This article carries a publication date of 2011-06-15, which places it before any of those provisions operated. It is drafted under a standing series authorization to describe measured outcomes from January 2014 through about 2022, and nothing later. Every number below therefore carries a named source and a period. Findings that are close to settled are presented as settled. Findings that remain contested among health economists are labeled contested, with the methodological reason for the disagreement stated in plain terms.

Two symmetrical overreaches have to be cleared away before the evidence can be read straight. The first says the law failed because premiums in the individual market rose sharply in its early years. The second says the law succeeded because the number of people with insurance rose. Both pick one outcome and treat it as the whole story. A statute this large never produces a single outcome. It moved the uninsured rate, it moved the composition of who holds insurance, it moved medical debt, it moved measurable health in some dimensions and not in others, and it moved premiums in different market segments in different directions. This article, with its companion sections, holds all of those outcomes at once and names which tradeoffs the statute deliberately made and which ones arrived as surprises.

The brief for this article organizes the record into five findings. The first is about the number of people carrying insurance: how far the uninsured rate fell between 2013 and 2016, why it drifted upward in the years after, why it later reached record lows, and which program delivered most of the gains. The second is about the natural experiment created when the Supreme Court’s 2012 decision made the Medicaid enlargement optional, and what researchers learned by comparing the states that adopted it with the states that did not. Those two findings are developed in the sections that follow. Three further findings, on financial protection, on health outcomes beyond mortality, and on premiums and the individual market, are taken up in the later sections of this article, and the closing passage below hands the inquiry to them.

The series thesis that runs underneath this article is that a statute should be assessed against its own aims while separating what the text did from what implementation and litigation allowed it to do. The coverage title of the Affordable Care Act aimed to reduce the number of uninsured Americans, with the Congressional Budget Office and the Joint Committee on Taxation projecting at passage that the enacted legislation would leave about 31 million fewer nonelderly people uninsured by 2019. Against that aim, the 2013 to 2016 movement is the central fact, and it is a large one. But the text as implemented was not the text as written. The Court’s 2012 decision rewrote the Medicaid title from mandatory to optional, the 2017 tax legislation zeroed the coverage requirement’s penalty, and administrative choices in 2017 and 2018 throttled the enrollment machinery. Each of those interventions changed the measured outcome, and none of them was in the original design. The evidence article therefore does double duty. It reports what happened, and it apportions the happening among the statute, the states, the courts, and the administrators. That apportionment is where the Medicaid first finding earns its place as this article’s namable claim: the law’s largest coverage effect ran through the channel the Court made optional, so the national average is always, in part, a measure of state decisions rather than of the statute alone.

A final word on posture. The measurement of a law’s effects is not the same as advocacy for the law. Where analysts disagree, this article gives the strongest version of each side, names the authors, and explains why their designs produce different answers. Where the evidence converges, it says so. Nothing here is extrapolated to any proposal under debate in any later year, and no sentence reaches past about 2022. The question is narrower and harder than advocacy: what happened, how do we know, and what remains genuinely in dispute.

Finding one: how the Affordable Care Act moved the uninsured rate

The headline series for American insurance coverage is the Census Bureau’s Current Population Survey, released each September in the Income and Poverty report. In 2013, the last full year before the principal provisions took effect, the survey put the share of the population without health insurance at 13.3 percent. In 2016, the third year of full operation, it put that share at 8.8 percent. A fall of four and a half percentage points in three years is, by the standards of American social policy, an unusually fast movement in a national rate, and it represented tens of millions of people moving from uninsured to insured status. The survey’s numbers for the intervening years trace the slope: the rate fell in 2014, fell again in 2015, and reached 8.8 percent in 2016.

That national average conceals the mechanism, and the mechanism matters more than the average. The statute attacked the uninsured population through three channels at once: an enlargement of Medicaid to adults with incomes up to 138 percent of the federal poverty level, subsidized private plans sold through new exchanges, and insurance market rules that barred denial or surcharges for preexisting conditions. Of the three, the first did the most work. The public program accounted for a clear majority of new insurance between 2013 and 2016, a decomposition confirmed by the Urban Institute and the Robert Wood Johnson Foundation in their analysis of the period: about 18.5 million people gained coverage in their accounting, and the Medicaid expansion supplied a majority of that gain, about 10.9 million. The exchanges served a smaller population, largely because the pool of people eligible for subsidies and inclined to buy was smaller than the pool made newly eligible for Medicaid, and because take up among the eligible ran well below one hundred percent. This is the finding that cuts against the popular framing of the law. In public discussion the exchanges are the visible face of the statute, the part with the website and the enrollment deadlines. In the enrollment data they are the smaller channel. The law’s largest coverage effect ran through a program expansion that the Supreme Court later made optional, which means the measured national performance of the statute is partly a measure of state decisions rather than of the statute itself. The program that carried those gains, a joint federal state entitlement whose structure predates the statute by decades, is explained in the series account of how the Medicaid statute is built.

The divergence between the states that adopted the enlargement and the states that did not is visible in every survey series. Analyses by the Urban Institute’s Health Policy Center and by KFF found that uninsured rates fell substantially further in adopting states than in holdout states over the 2013 to 2016 window. In the Urban Institute and Robert Wood Johnson Foundation analysis, the uninsured rate in expansion states fell from 15.3 percent to 7.0 percent across those years, a decline far larger than in the states that declined the expansion. The composition of the remaining uninsured population shifted as well. In states that adopted the enlargement, the people who remained without insurance were disproportionately those ineligible for any assistance: undocumented immigrants, people caught in administrative churn, and those who declined to enroll. In states that did not adopt it, a large group sat in the statute’s coverage gap, earning too little to qualify for exchange subsidies, which began at the poverty line, and too much to qualify for Medicaid under the state’s old rules. The gap was a direct consequence of the statute having been written for a mandatory enlargement. When the mandate became optional, the assumption underneath the subsidy schedule broke, and the people below the poverty line in holdout states fell through. The later sections of this article take up the gap and its size; what matters for the coverage finding is that the national averages mix two different policy regimes, and the average therefore understates what the law achieved where it was fully implemented and overstates what it achieved where it was not.

Before the main provisions operated, one early piece of the statute was already moving the numbers. The provision allowing young adults to remain on a parent’s plan until age 26 took effect in late 2010, and the Department of Health and Human Services reported that about 3.1 million young adults had gained coverage through the provision by December 2011. This matters for two reasons. It shows that the law’s coverage machinery began working years before the headline date, and it means the 2013 baseline against which the main provisions are measured already contains some of the statute’s own effect. A comparison of 2016 with 2010 would show a larger total movement; a comparison of 2016 with 2013 shows the movement attributable to the 2014 provisions specifically. The honest presentation uses the 2013 baseline and says why.

Why does the 2013 baseline already contain some of the law’s own effect?

The dependent coverage provision took effect in September 2010, well before the main provisions, and about 3.1 million young adults had gained coverage through it by December 2011, according to the Department of Health and Human Services. Measuring from 2013 therefore excludes that earlier gain and understates the statute’s total coverage effect.

The dependent coverage provision took effect in September 2010, more than three years before the main provisions, and the Department of Health and Human Services reported that about 3.1 million young adults had gained coverage through it by December 2011. Any measurement that starts in 2013 therefore excludes this earlier gain, which means the 2013 to 2016 comparison understates the statute’s total coverage effect.

They were concentrated at the bottom of the income distribution. Census data for 2013 show the uninsured were disproportionately low income adults, young adults, and Latino residents, with the highest uninsured rates in the South. A large share of uninsured adults had incomes below 138 percent of the poverty line, the group the Medicaid enlargement was designed to reach.

The composition of the uninsured population in 2013 explains why the statute was built the way it was. Employer sponsored insurance covered roughly half the nonelderly population and was never the target of the reform; the law’s drafters assumed, correctly, that most workers with job based plans would keep them. The uninsured were instead concentrated among people the existing system served worst: low wage workers whose employers did not offer insurance, part time and seasonal workers, young adults aging out of parental or student plans, and adults without dependent children, who were excluded from Medicaid in nearly every state regardless of how poor they were. That last exclusion is the key to the whole architecture. Before 2014, Medicaid was a categorical program. It covered low income children, pregnant women, some parents at very low income thresholds that varied wildly by state, and elderly and disabled people qualifying through Supplemental Security Income. A childless adult working full time at the minimum wage qualified in essentially no state. The enlargement to 138 percent of the poverty level for all adults was therefore not an incremental tweak. It created an entirely new eligible population numbering in the tens of millions, and it did so in the part of the income distribution where the uninsured were most densely packed.

What did the individual market look like for someone with a chronic condition before 2014?

In most states, insurers selling directly to individuals could deny applications, exclude preexisting conditions, or charge higher premiums based on health history. The Commonwealth Fund’s 2010 survey found that of 26 million adults who had tried to buy individual coverage in the prior three years, about 16 million found affordable coverage very difficult or impossible to obtain.

In most states, insurers selling directly to individuals could deny applications, exclude preexisting conditions, or charge higher premiums based on health history, a practice called medical underwriting. The Commonwealth Fund’s 2010 Biennial Health Insurance Survey documented the consequences: of an estimated 26 million adults who had bought or tried to buy individual market coverage in the prior three years, about 16 million found it very difficult or impossible to find an affordable plan, and an estimated 9 million were turned down, charged a higher price because of a health problem, or had a preexisting condition excluded from their coverage.

The result was a market that worked tolerably for the young and healthy and failed everyone else. The statute’s market rules, guaranteed issue and community rating, which barred those practices, were aimed squarely at this failure. They were also the provisions that made the coverage requirement and the subsidies necessary companions: bar insurers from pricing on health status, and healthy people will wait until they are sick to buy, unless something pulls them in. The mandate, the subsidies, and the market rules were a tripod, and the coverage finding cannot be understood without seeing all three legs.

Geography concentrated the problem further. The South held a disproportionate share of the uninsured in 2013, a pattern driven by lower median incomes, lower rates of employer sponsored insurance, and state Medicaid programs with some of the stingiest eligibility thresholds in the country. This geographic pattern would later map almost exactly onto the map of states that declined the enlargement, which is why the state by state comparison at the heart of the second finding is so stark. The states where the need was greatest were, with important exceptions, the states where the policy response was withheld.

The public program carried most of the coverage gains

The arithmetic favored the public program. The newly eligible Medicaid population was large, the benefit was free or nearly free to the enrollee, and enrollment could happen year round rather than only during an open enrollment window. The exchange population was smaller by construction: subsidies phased in above the poverty line and phased out at higher incomes, the unsubsidized middle of the individual market faced the full premium, and many eligible people never completed the application.

The reason is arithmetic rather than ideology. The Urban Institute and Robert Wood Johnson Foundation decomposition of the 2013 to 2016 period assigned about 10.9 million of the roughly 18.5 million coverage gains to the Medicaid expansion, a clear majority. The administrative data tell the same story in absolute terms. The Centers for Medicare and Medicaid Services reported total Medicaid and Children’s Health Insurance Program enrollment of about 61.2 million in December 2013, rising steadily through the window; by December 2022 the same series reached 92,340,585. Exchange plan selections for the 2016 open enrollment period ran to about 12.7 million according to the Centers for Medicare and Medicaid Services, with effectuated enrollment, the number of people actually paying premiums and holding coverage in a given month, running lower. Even taking the most generous exchange measure and the most conservative Medicaid measure, the public program’s gains were larger. The Medicaid number is also the cleaner gain, since nearly all of it represents people who were previously uninsured, whereas some marketplace enrollees had prior coverage. The states with the largest drops in uninsurance were overwhelmingly the states that adopted the enlargement.

None of this diminishes what the exchanges did. For the population they served, particularly people with incomes between the poverty line and four times that level who had previously been priced out of individual insurance or excluded for preexisting conditions, the exchanges were the only new option the statute created. But anyone who equates the statute’s coverage effect with exchange enrollment is counting the smaller channel and missing the larger one.

Why does eligibility not equal enrollment?

Eligibility is a legal status; enrollment is a completed transaction, and the two diverge because of price, hassle, and inattention. Subsidized premiums still strained tight budgets, applications stalled on documentation failures, and many eligible people never learned the enrollment window existed. Easing any one of those frictions raises take up, which is why outreach funding moved the numbers.

Price, hassle, and inattention explain most of it. Even subsidized premiums strained households with little slack, the application demanded income documentation through systems that did not always communicate, and many eligible people never heard about the enrollment window. Research on take up finds that easing any one friction raises enrollment, which is why outreach funding mattered so much to the numbers.

Eligibility is not enrollment, and the distance between the two is one of the most studied problems in social policy. The statute made tens of millions of people newly eligible for Medicaid or for subsidized exchange plans, but a substantial fraction of them never completed the process, and understanding why is essential to reading the coverage numbers correctly. The frictions fall into three buckets. The first is price. Even with subsidies, exchange premiums required monthly payments from households with little financial slack, and the deductibles on the cheaper plans ran into the thousands of dollars. For a family choosing between a premium payment and groceries, the rational short term choice is often the groceries, particularly when the penalty for going without was small or, after 2018, zero. The second is hassle. The application asked people to document income, household composition, and immigration status through systems that did not always talk to each other, and a single data matching failure could stall an application for weeks. The third is inattention. Surveys of the uninsured repeatedly found that large numbers of eligible people had never heard of the enrollment window, did not know subsidies existed, or believed, incorrectly, that they were ineligible. None of these frictions is mysterious. All of them respond to outreach, which is why the navigator program existed. Rebecca Myerson and Honglin Li, exploiting the more than 80 percent cut in navigator funding over 2017 to 2019 as a natural experiment in the American Journal of Health Economics in 2022, found the cuts significantly reduced coverage among lower income adults, adults under 45, Hispanic adults, and adults speaking a language other than English at home.

Running in the opposite direction was a phenomenon researchers call the welcome mat effect, sometimes the woodwork effect. The publicity surrounding the 2014 launch, combined with streamlined applications and expanded outreach, pulled in people who had already been eligible for Medicaid under the old rules but had never enrolled. Analyses of the enrollment surge found that part of the Medicaid growth in 2014 and 2015 came from this previously eligible population rather than from the newly eligible adults. The welcome mat effect complicates the decomposition of the coverage gains. Some of the Medicaid growth attributed to the statute’s enlargement was really the statute’s publicity and simplified enrollment machinery reaching people the old law had already covered on paper. This does not shrink the total number of newly insured people. It changes the interpretation: part of what the law did was not to create new eligibility but to make existing eligibility function.

The statute’s drafters anticipated some of this and built enrollment machinery beyond the subsidies. Hospitals were given presumptive eligibility authority to enroll patients on the spot. States were encouraged to use existing SNAP and tax data to auto renew beneficiaries rather than demanding fresh paperwork each year. Community health centers received funding to hire application assisters. Each of these mechanisms has its own evaluation literature, and the consistent finding is that administrative simplification moves enrollment nearly as much as financial generosity does. The corollary is that administrative retrenchment moves it in the other direction, which is what the 2017 and 2018 funding cuts demonstrated in reverse. One further friction deserves a sentence because the later changes article of this series treats it at length: the regulatory interpretation known as the family glitch, which measured affordability only against the cost of covering the employee rather than the family, left some dependents of workers with affordable self only offers ineligible for subsidies. The glitch narrowed the law’s reach for a specific population until regulators revised the interpretation years later, and it belongs to the amendments story rather than to the coverage story, but it is a reminder that the distance between eligibility on paper and insurance in hand was partly a product of regulatory choices as well as of human behavior.

Even with imperfect enrollment, the exchanges moved large numbers of people who had previously been uninsurable at any price. For adults with chronic conditions, the guaranteed issue rules were the whole ballgame: no application process, however streamlined, matters if the insurer can legally refuse to sell. Studies of the exchange risk pools found that enrollees were sicker on average than the general population, which is exactly what the law’s design predicted and exactly what made the mandate and the subsidies load bearing. A market that must take all comers needs healthy buyers to balance the sick ones, and the subsidies were the instrument for pulling the healthy in. When enrollment among the young and healthy softened after 2016, the pools skewed sicker and premiums adjusted upward, a dynamic the later sections on premiums take up in detail. For the coverage finding, the point is narrower: the exchanges did not need to match Medicaid’s numbers to justify their existence. They needed to create a functioning market for a population that had never had one, and on that measure the enrollment data, imperfect as the take up was, show a market that operated at national scale within two years of launch.

The 2016 to 2019 drift in the uninsured rate

Nothing in the macroeconomy explains the turn. Unemployment was falling and incomes were rising through 2017 to 2019, conditions that ordinarily push uninsurance down. The uptick reflected policy choices instead: the shared responsibility payment fell to zero for 2019, federal marketplace advertising funding was cut 90 percent, from 100 million dollars to 10 million dollars, and navigator grants fell about 40 percent in 2017 and then to 10 million dollars for 2018 to 2019, an 84 percent reduction from 2016 levels, according to the Kaiser Family Foundation’s tracking, while repeated public debate about the law’s survival discouraged enrollment.

The Census series captures the reversal precisely. After reaching 8.8 percent in 2016, the Current Population Survey put the uninsured rate at 8.8 percent in 2017, the American Community Survey at 8.9 percent in 2018 and 9.2 percent in 2019, the highest reading since 2014. The National Health Interview Survey, which interviews households throughout the year rather than once in the spring, showed the same shape: 9.0 percent uninsured for the full year 2016, rising to 10.3 percent, or 33.2 million people, for the full year 2019.

Three policy changes did most of the work. First, the tax legislation of 2017 reduced the shared responsibility payment, the penalty for going without insurance, to zero effective in 2019. Whether the penalty had ever been the decisive factor in anyone’s enrollment decision is itself a contested question among economists, but its removal coincided with a visible softening of enrollment among healthier, lower cost enrollees, and analysts at the Congressional Budget Office had projected in advance that zeroing the penalty would raise the uninsured count. Second, the administration in office from 2017 cut the advertising budget for open enrollment and reduced funding for navigator organizations that helped people complete applications, with the cuts concentrated in 2017 and 2018. Enrollment is sensitive to outreach; the navigator program existed precisely because the application was difficult. Myerson and Li, comparing counties more and less exposed to the 2017 to 2019 funding cuts in the American Journal of Health Economics in 2022, found the cuts significantly reduced marketplace coverage among lower income adults and total coverage among adults under 45, Hispanic adults, and adults speaking a language other than English at home. Third, the years 2017 through 2019 featured repeated legislative and legal attempts to repeal or invalidate the statute, including the events taken up in the later changes article of this series. Confusion is itself a policy outcome. The enrollment literature reviewed by Myerson and coauthors in JAMA Network Open in 2022 reported that nearly two thirds of uninsured adults in 2014 had been exposed to little or no information about financial assistance for marketplace coverage, and that in 2018 two thirds of uninsured adults had still never visited a marketplace to check their options. Uncertainty of that kind suppresses take up even when the underlying benefits are unchanged.

A fourth factor operated through premiums rather than through statute, and it belongs to the premiums discussion in the later sections: the discontinuation of direct federal payments for cost sharing reductions in late 2017, and the resulting repricing of silver tier plans, scrambled the enrollment incentives for 2018 in ways that are still debated. What matters for the coverage finding is the net result. The gains of 2014 through 2016 were real, and then a portion of them was given back, not because the underlying demand for insurance disappeared but because the policy environment around enrollment deteriorated.

A methodological caution belongs at this point, because the most cited series in this article straddles a survey redesign. In the 2014 round of the Current Population Survey’s Annual Social and Economic Supplement, the Census Bureau replaced its longstanding health insurance questions with a redesigned set intended to capture coverage more accurately. The Census Bureau published both versions for the overlap year and cautioned against treating the old and new series as seamlessly comparable. Researchers working with the long trend handle this in the standard way: they use the redesigned series from 2013 forward, note the break, and check the trend against the National Health Interview Survey, which underwent no comparable redesign and shows the same arc. The redesign affects the level, not the shape. The steep fall from 2013 to 2016, the partial reversal through 2019, and the descent to new lows by 2022 appear in every series regardless of questionnaire wording. The honest use of the Census numbers is to cite the survey, the year, and the question vintage when precision demands it, and the figures quoted in this article follow the redesigned series that the Census Bureau itself treats as authoritative from 2013 onward.

The Medicaid undercount is a second measurement wrinkle that cuts in a consistent direction. Survey respondents underreport Medicaid enrollment relative to administrative records, a pattern documented across multiple validation studies that match survey responses to state enrollment files. The undercount means the Current Population Survey likely understates the public program’s share of the coverage gains described above and overstates the share attributable to private plans. Researchers who adjust for the undercount generally find an even more Medicaid heavy composition than the raw survey numbers suggest. The practical consequence is that the central claim of the first finding, that the public program did the heavy lifting, is conservative as stated. The surveys, if anything, flatter the exchanges relative to Medicaid, and the finding survives that bias rather than depending on it.

What makes the 2022 low point a complicated measurement?

The 2022 low point is real but not a pure measure of the statute’s permanent design. It was produced by two temporary conditions layered on top of the law: enlarged premium tax credits and the pandemic era suspension of Medicaid disenrollments. Both were scheduled to end, so the 7.9 percent figure describes a moment, not a steady state.

The Census Bureau’s Current Population Survey for 2022 reported 7.9 percent of the population without insurance, the lowest rate the survey had recorded. The figure is real, but it was produced by two temporary conditions layered on top of the statute’s permanent structure: enlarged premium tax credits and a pandemic era suspension of Medicaid disenrollments. Reading it as a pure measure of the statute’s design would overstate what the permanent provisions achieved.

The descent from the 2019 high was as policy driven as the earlier ascent. Two forces did the work. The first was the pandemic relief legislation of 2021, which enlarged the premium tax credits for exchange plans, removed the income cliff that had cut off subsidies above four times the poverty line, and capped what households paid as a share of income. The effect on enrollment was immediate and large: plan selections for the 2022 open enrollment period reached more than 14.5 million, a record at the time according to the Centers for Medicare and Medicaid Services, with particularly strong growth among middle income households newly eligible for assistance. The reconciliation legislation of 2022 extended the enlarged credits further, and the full sequence of these statutory changes is taken up in the later changes article of this series; for the coverage finding, the mechanism is what matters, which is that lowering the net price of exchange plans pulled in buyers who had previously sat out.

The second force was administrative rather than legislative in origin. The Families First Coronavirus Response Act of March 2020 barred states from disenrolling Medicaid beneficiaries for the duration of the public health emergency, in exchange for enhanced federal matching funds. The result was a historically unusual period in which the normal churn of Medicaid, the cycle of people losing eligibility and reapplying, was suspended, and enrollment rose steadily toward 92,340,585 people across Medicaid and the Children’s Health Insurance Program in December 2022, according to the Centers for Medicare and Medicaid Services. Some of that growth reflected genuine new need during the economic shock of 2020, but much of it reflected the suspension of redeterminations: people who would ordinarily have cycled off the rolls stayed on. For the uninsured rate, the effect was straightforwardly downward. People who held Medicaid continuously were people who did not appear as uninsured in the surveys.

The 7.9 percent figure for 2022 therefore has to be read with the same care as every other number in this article. It is the lowest rate the Census survey had recorded, and it is a real measurement of a real population. It is also the product of two temporary conditions, enlarged subsidies and suspended redeterminations, layered on top of the statute’s permanent structure. Whether the rate would have held once those conditions expired is a question the data through 2022 cannot answer, and this article does not try to answer it. What the series through 2022 establishes is the shape of the whole arc: a steep fall from 2013 to 2016, a policy driven partial reversal through 2019, and then a second fall to record lows on the strength of temporary supports.

The measurement infrastructure underneath all of these numbers deserves a paragraph of its own, because the surveys do not agree with each other and the disagreements are informative. The Census Current Population Survey asks about insurance over the prior calendar year in interviews conducted each spring. The National Health Interview Survey asks about coverage at the time of interview, continuously through the year. The American Community Survey asks about current coverage in a very large sample that supports state and county estimates. Point in time measures run lower than full year measures, because some people are uninsured for part of a year but not all of it. The Census survey is also known to undercount Medicaid enrollment relative to administrative records, a problem researchers call the Medicaid undercount, which means the survey likely understates the public program’s contribution to the gains described above. None of this invalidates the trend. All three surveys show the same arc: down sharply through 2016, up modestly through 2019, down to new lows by 2022. When independent instruments with different designs agree on the shape, the shape is the finding, and the exact level in any one series is a matter of definition. The honest use of these numbers is to cite the survey, the year, and the definition, and to distrust any single figure quoted without all three.

One more compositional point closes the coverage finding. The people who remained uninsured at the end of the window were not a random slice of the population. Across every survey, the remaining uninsured were disproportionately low income, disproportionately Latino, disproportionately young adults, and disproportionately concentrated in the states that had not adopted the Medicaid enlargement. That concentration is itself evidence for the mechanism. Where the statute was fully implemented, the uninsured rate fell furthest. Where a piece of it was withheld by state decision, the people that piece would have covered remained uninsured. The national numbers are averages over a policy that was never, after 2012, a single national policy at all.

Finding two: the Affordable Care Act’s accidental natural experiment

On June 28, 2012, the Supreme Court decided National Federation of Independent Business v. Sebelius. The decision is taken up in detail in the companion article of this series on the case; what matters for the evidence question is one holding within it. Seven justices concluded that the statute’s threat to withdraw all existing Medicaid funds from states that refused the enlargement was unconstitutionally coercive, and the remedy the Court chose was to make the enlargement optional. The statute had been written on the assumption that every state would extend Medicaid to adults up to 138 percent of the federal poverty level, and the subsidy schedule for the exchanges had been built on top of that assumption, with assistance beginning at the poverty line because everyone below it was supposed to be in Medicaid. Overnight, the law’s largest coverage channel became a state by state choice, and the country divided into adopting states and holdout states on a staggered timetable: a large wave taking effect in January 2014, further adoptions in the years after, and a set of holdouts that had still not adopted by the end of 2022. By that point 38 states plus the District of Columbia had implemented the expansion, while holdouts including Texas and Florida had not.

No researcher could have designed a better experiment, and none would have been permitted to. Randomly assigning states to expand Medicaid or withhold it from their low income residents would be unethical and, in any case, beyond any investigator’s power. The Court’s decision produced the next best thing: a policy change adopted by some jurisdictions and not others, at known dates, with a large affected population and rich survey and administrative data on both sides. Health economists recognized the opportunity immediately. The literature that followed is the closest thing modern American social policy has to a controlled trial of public health insurance for low income adults, and its center of gravity is the work of Benjamin Sommers and his colleagues, with the mortality question sharpened most forcefully by Sarah Miller and her coauthors.

Isolating the policy effect through state comparisons

A simple comparison of adopting and holdout states after 2014 would prove nothing, because the two groups differed before the law ever operated. The difference in differences approach subtracts out those permanent differences by comparing changes rather than levels. If adopting states and holdout states were moving in parallel before 2014 and their paths diverged only after the enlargement took effect, the divergence is attributed to the policy itself.

The logic of the design repays a moment of care, because everything contested in this literature turns on it. The identifying assumption, the thing that has to be true for the attribution to hold, is the parallel trends assumption: that without the enlargement, the two groups of states would have continued on their prior paths. Researchers test it by examining the pre 2014 data for divergence, and the strongest papers in this literature show their work on this point rather than asserting it.

The design has known weaknesses, and the honest literature names them. States chose whether to adopt, and the choice was not random. Adopting states had, on average, more generous safety nets, stronger administrative capacity, and political leadership sympathetic to the law. If those same traits independently improve health outcomes, the comparison overstates the enlargement’s effect. Researchers address the selection problem in several ways: by controlling for state level economic and demographic trends, by using border county comparisons that hold geography constant, by constructing synthetic control states weighted to match the pre 2014 path of adopters, and by exploiting the staggered timing of adoption so that later adopters serve as controls for earlier ones. No single fix is decisive. The credibility of the literature comes from the convergence of results across these different approaches rather than from any one paper’s design.

Sommers sits at the center of this literature because he worked the problem from every available angle before and after 2014. With Katherine Baicker and Arnold Epstein, he studied the state Medicaid expansions of the early 2000s in Arizona, Maine, and New York, comparing mortality in those states with matched control states in a paper published in the New England Journal of Medicine in 2012. The finding was a decline in all cause mortality in the expanding states relative to controls of 19.6 deaths per 100,000 adults per year, a 6.1 percent relative reduction. With Sharon Long and Katherine Baicker, he studied the Massachusetts health reform of 2006, the state level precursor to the federal statute, comparing mortality in Massachusetts counties with similar counties in other states before and after the reform, in a paper published in the Annals of Internal Medicine in 2014. The finding was a decline in all cause mortality of 8.2 deaths per 100,000, a 2.9 percent relative reduction. The two papers established a pattern: Medicaid expansions, studied with comparison group designs, were associated with measurable mortality declines, with the larger and more generous expansions showing larger effects.

After 2014, Sommers turned the same machinery on the federal enlargement itself. With coauthors including Baicker and Robert Blendon, he published a three year assessment in Health Affairs in 2017 comparing low income adults in adopting and holdout states. The paper reported what the design was best powered to detect: large gains in insurance coverage, more primary care visits, fewer people skipping medications because of cost, and improvements in self reported health. With Atul Gawande and Katherine Baicker, he published a review in the New England Journal of Medicine in 2017, titled “Health Insurance Coverage and Health: What the Recent Evidence Tells Us,” which surveyed the full body of work including the randomized Oregon lottery evidence. The review’s conclusion was carefully bounded: insurance improves access to care and reduces financial strain, the evidence supports improvements in self reported health and in some clinical measures, and the mortality evidence, while leaning positive, remained the least settled part of the picture. That 2017 assessment is worth quoting in spirit if not in text, because it set the terms on which the later literature would argue: the question was no longer whether insurance does anything, but which outcomes it moves, by how much, and on what timetable.

Why do late-adopting states strengthen the research design?

Each late adoption supplies a fresh before and after experiment with its own time zero: Pennsylvania in 2015, Montana and Louisiana in 2016, Virginia and Maine in 2019, Idaho, Utah, and Nebraska in 2020, Missouri and Oklahoma in 2021, per the Congressional Research Service. Aligning outcomes around each adoption lets researchers check whether effects appear exactly when the policy does.

Each late adoption is a fresh experiment. Pennsylvania followed in 2015, Montana and Louisiana in 2016, Virginia and Maine in 2019, Idaho, Utah, and Nebraska in 2020, and Missouri and Oklahoma in 2021, each with its own legislative or ballot initiative history, according to the Congressional Research Service. Researchers aligned each state’s adoption date as time zero and traced outcomes in event time, meaning the quarters before and after adoption, which allows the pre adoption path to serve as its own control.

The event study design is demanding in a useful way: if the enlargement is really causing the change, the effect should appear after time zero and not before, and the papers that use this design display the full event time path so readers can see whether the pre trends are flat. The consistent finding across these studies is a break at adoption: coverage jumps, uncompensated care falls, and access measures improve, with no corresponding break in the quarters before. The design cannot fully solve the selection problem, since the states that adopted late chose to do so, often through ballot initiatives that bypassed resistant legislatures, but it substantially weakens the objection that adopters and holdouts were simply on different trajectories all along. When the same break appears at different calendar dates in different states, each aligned to its own adoption, the policy is the most plausible common cause.

An independent corroboration comes from outside the survey data entirely. The American Hospital Association’s annual survey data track uncompensated care, the cost of charity care and bad debt that hospitals absorb when patients cannot pay. Analyses of these data found that uncompensated care costs as a share of hospital expenses fell in expansion states after 2014 while remaining flat or rising in holdout states. Kaufman and colleagues, in a Health Affairs study published in 2017, found that the 2014 expansion was associated with larger Medicaid revenue gains for rural hospitals than for urban ones, while the decline in uncompensated care costs as a share of operating costs was larger for urban hospitals than for rural ones. A later Health Affairs study of Louisiana’s 2016 expansion, summarized by the Center on Budget and Policy Priorities, estimated that expansion was associated with a 33 percent reduction in uncompensated care costs as a share of total operating expenses, with a 55 percent reduction for rural hospitals, as the state’s nonelderly uninsured rate fell from 18.3 percent in 2015 to 11.8 percent in 2018. Hospitals do not answer survey questions about insurance status; their accounting records reflect who showed up without coverage and could not pay. The divergence in the accounting data mirrors the divergence in the survey data, which is what corroboration looks like: two independent instruments, one asking people and one counting dollars, telling the same story about the same policy at the same dates. The uncompensated care finding also hints at the fiscal mechanism that made some states adopt the enlargement over their legislatures’ objections. Hospitals in holdout states continued to absorb the cost of the uninsured, and hospital associations became some of the most persistent lobbyists for adoption, a political dynamic that belongs to the legislative history but helps explain why the map of adoption kept changing through 2021.

What keeps the expansion comparison honest about its limits?

Three limits, all stated openly in the literature. First, the estimates capture the enlargement as implemented, bundled with outreach and simplified applications, not insurance in the abstract. Second, adopting states are not random, so results describe the people who actually gained coverage. Third, the design is largely blind to spillovers onto providers and the uninsured who remained.

Three limits, all stated openly in the literature. First, the design estimates the effect of the enlargement as implemented, not the effect of insurance in the abstract: adopting states bundled eligibility with outreach, simplified applications, and navigator funding, so the measured effect includes all of that machinery. Second, adopting states and their enrollees are not random samples, so the estimates describe the effect on the people who actually gained coverage, not an automatic prediction for holdout states. Third, the design is largely blind to spillovers on providers, clinics, and labor markets.

None of these limits invalidates the findings. They define the findings’ scope, and the scope is substantial: the best available estimate of what the policy did, where it was tried, for the people it reached. The access findings from the natural experiment literature are less contested and provide the mechanism through which the longer run health effects, if present, would operate. The Health Affairs three year assessment and subsequent work found that low income adults in adopting states were more likely to have a usual source of care, more likely to have received preventive services, and less likely to report delaying care because of cost than their counterparts in holdout states. These are the intermediate outcomes that any theory of insurance improving health has to pass through: coverage first, then contact with the health system, then management of chronic conditions, then, over years, mortality. The literature has documented the first three steps with considerable confidence. The fourth step is where the contest lives, and it is the contest described above.

Linked administrative records and the mortality question

Mortality among nonelderly adults is, mercifully, a rare event, which means that detecting a change in the death rate requires either a very large sample or a very long follow up. Sarah Miller, Norman Johnson, and Laura Wherry solved the sample problem by linking American Community Survey responses, which record health insurance status for millions of people, to Social Security Administration death records, creating a dataset large enough to detect mortality effects directly. Their paper, published in the Quarterly Journal of Economics in 2021 and based on a 2019 National Bureau of Economic Research working paper, used the state by state variation in the enlargement in a difference in differences design and reported that annual mortality fell by 0.132 percentage points among near elderly adults, a 9.4 percent reduction relative to the sample mean, with the effect growing over time.

Set against these positive findings is the most famous null result in the literature, and it comes from the one true randomized experiment. The Oregon Health Insurance Experiment, which randomly allocated a limited number of Medicaid slots by lottery in 2008, is the only study in which access to Medicaid was assigned by chance rather than by state policy choice. The investigators, led by Katherine Baicker with Amy Finkelstein and others, published two year results in the New England Journal of Medicine in 2013. The lottery winners used more health care, reported better self assessed health, and experienced substantially less financial strain, findings the later sections of this article take up in the financial protection discussion. On measured physical health outcomes and on mortality, the two year follow up found no statistically significant effect. The null was widely reported, and in some public discussion it was treated as proof that Medicaid does not improve health. That reading overreaches. A two year follow up in a relatively small sample has limited power to detect mortality effects, and the absence of a measurable change in blood pressure or cholesterol at two years does not preclude health effects on longer horizons. The honest reading, and the one the investigators themselves offered, is that the Oregon results bound what can be claimed on short timelines: insurance quickly and reliably reduces financial harm and increases the use of care, while physical health effects, if present, take longer to appear and require larger samples to detect.

The spread of mortality findings therefore has a methodological explanation, and the explanation is the reason the finding is labeled contested rather than settled. Studies with large samples and longer follow ups, meaning the Sommers work on the early 2000s expansions and the Miller linked records study, tend to find mortality reductions. Studies with short follow ups and smaller samples, meaning Oregon at two years, tend to find no significant effect. These results are not contradictory. They are what one would expect if insurance reduces mortality gradually, through sustained access to primary care, medications, and early detection, rather than immediately. The contested part is the magnitude: the point estimates differ across designs, the populations differ, and the translation from any single estimate to a national number of lives saved requires assumptions that the studies themselves do not fully support. What can be said without overreaching is that the weight of the comparison group evidence points toward lower mortality where Medicaid was enlarged, that the randomized evidence does not contradict this on any horizon it was powered to test, and that the exact size of the effect remains an active research question rather than a settled fact.

Where the inquiry goes next

The two findings developed here discipline each other, and that mutual discipline is worth stating before the handoff. The coverage finding on its own is a description: the uninsured rate fell, rose, and fell again, and Medicaid carried the largest share. Description alone cannot attribute the movement to the statute rather than to the economy or to background trends. The natural experiment supplies the attribution: the divergence between adopting and holdout states, breaking at each state’s adoption date, is what licenses the claim that the policy caused the movement. Conversely, the experiment on its own is a method in search of a mechanism. The coverage finding supplies it: the reason the adopting states diverged is that millions of people in those states gained insurance through the public program, while their counterparts across the state line did not. Neither finding is complete without the other, and together they form the evidentiary base on which the remaining questions rest.

Insurance expanded substantially after 2014, the public program carried most of the gains, the gains partially reversed under adverse policy conditions and then reached new highs on temporary supports, and the state by state adoption pattern gave researchers the comparison they needed to attribute effects to the policy rather than to background trends. Three findings remain, and they belong to the sections that follow: the financial protection effects, where the evidence is the most consistent in the literature; the broader health outcomes beyond mortality, where the results are genuinely mixed and the honest presentation is the range; and the premiums and individual market story, where the early years brought sharp increases before stabilization, and where the coverage gap in holdout states stands as the statute’s most visible unfinished business. Each of those findings carries the same discipline applied here: named sources, stated periods, contested findings labeled with the reason, and no extrapolation past the data.

Financial Protection: The Most Consistent Measured Effect

Of the five findings this article carries, the third is the one health economists fight about least. Whether the question is put to a randomized lottery, to credit bureau records, to bankruptcy filings, or to survey responses, the answer keeps coming back the same: when people without insurance gained it through the expansion of the joint federal state program that carried most of the law’s enrollment gains, their money problems got smaller. Not their money problems in some abstract sense. Specific, countable things: bills that went unpaid, debts that were handed to collection agencies, medical spending that swallowed a third or more of a household’s income. The evidence here is consistent enough that it reads less like a debate and more like a ledger, which is unusual in a literature where almost everything else is contested.

The reason for that consistency is worth stating before any study is named, because it explains why the finding holds across such different methods. The program in question asks enrollees to pay almost nothing at the point of care. There are no deductibles to meet, no coinsurance percentages to multiply, and in most states no monthly premium worth the name. Take a population that was paying out of pocket for care it could not afford, give it insurance with essentially zero cost sharing, and the financial outcomes have almost nowhere to go but down. That is not a subtle theoretical claim. It is arithmetic. The subtlety lies only in measuring how large the relief was, how long it lasted, and whether it spilled over into parts of a household’s finances that had nothing to do with medicine. On each of those questions there is a named study with a method and a number.

Why do credit bureau records strengthen the medical debt finding?

Credit bureau records remove the recall error and optimism that contaminate survey answers: a debt is on the file or it is not. Comparing expansion against non expansion states, Hu, Kaestner, Mazumder, Miller, and Wong found collection balances about 1,140 dollars lower among those who gained coverage, in the Journal of Public Economics in 2018.

Survey answers suffer from recall error and from the optimism that creeps into self reports. Credit bureau records do not: a debt exists on a credit file or it does not. Luojia Hu, Robert Kaestner, Bhashkar Mazumder, Sarah Miller, and Ashley Wong published their analysis in the Journal of Public Economics in 2018, using credit reporting data for a large national sample and comparing expansion states against non expansion states with a synthetic control method. Their estimates imply a reduction in collection balances of approximately 1,140 dollars among those who gained coverage through the expansions.

That result rests on administrative records rather than on anyone’s memory, and it matters for three reasons that later work has only confirmed. First, collections are a hard administrative outcome. Second, the synthetic control comparison is built for exactly this problem, constructing a counterfactual version of each expansion state from states that did not expand, so the change measured is the change the expansion caused rather than a national trend that happened to coincide with it. Third, the effect concentrates exactly where theory says it should, in the neighborhoods where the most uninsured low income people live, which makes it hard to argue that the result is an artifact of the method. An artifact would not know which zip codes to land in.

Kenneth Brevoort, Daniel Grodzicki, and Martin Hackmann extended the picture with a different data structure and found the same direction. Their analysis, published as National Bureau of Economic Research Working Paper No. 24002 in 2017, found that the expansions were associated with about 3.4 billion dollars less unpaid medical debt sent to collection agencies in the first two years, alongside improved credit outcomes. The bankruptcy channel completes the ledger. Tal Gross and Matthew Notowidigdo, in the Journal of Public Economics in 2011, used pre Affordable Care Act expansions of the same program across 1992 to 2004 and found that a ten percentage point increase in Medicaid eligibility was associated with an 8 percent reduction in personal bankruptcies, with out of pocket medical costs pivotal in about 26 percent of personal bankruptcies among low income households. The statute under review here did not exist when Gross and Notowidigdo wrote, but the mechanism they identified, insurance absorbing a hospital bill that would otherwise have become unpayable debt, is exactly the mechanism the post 2014 studies measured at larger scale.

How does financial relief reach beyond the medical bill itself?

Unpaid medical bills damage credit files, which raises borrowing costs and crowds out other obligations. When coverage removes the medical bill, the rest of the file recovers: Hu and colleagues found the expansions also reduced non medical debt sent to collections, with larger effects among people who had been hospitalized. Freed cash stayed available for rent, utilities, and food.

A medical bill that goes to collections damages a credit file, which raises the cost of borrowing and can trigger demands on other obligations; remove the medical bill and the rest of the file has room to breathe. Hu, Kaestner, Mazumder, Miller, and Wong reported that the expansions also reduced non medical debt sent to collection agencies, with larger effects among people who experienced a hospitalization, which is the spillover pattern one would expect if freed cash stayed available for rent, utilities, and food.

That spillover is what turns a health policy result into a household economics result. The pattern is economically legible. The effect was more pronounced among older adults than among younger ones, and the authors found no significant change in other measures such as credit card balances in the years right after expansion. That last null is equally legible: people who could not pay a hospital bill were not, for the most part, carrying large revolving balances whose minimums they were meeting on time. There was no slack to release there.

The Oregon work says the same thing from a different direction. The Oregon Health Study Group’s own summary of its randomized findings reports that people who won the 2008 lottery and gained coverage were substantially less likely to experience other forms of financial strain, such as borrowing money or delaying payments on other bills because of medical expenses. That is the randomized version of the spillover the credit records show: cash that would have gone to a hospital or a collector instead stayed available for rent, utilities, and food. Two different designs, one built on a lottery and one built on credit files, converge on the proposition that the financial value of the coverage was not confined to the medical bills it paid. It is worth pausing on how rare that convergence is. In this literature, findings that survive the crossing from one method to another are the findings a reader should trust most, and financial protection survives every crossing researchers have attempted.

There is also a quieter channel through which the financial protection operated, and it concerns people who never got sick. Insurance has value even when it is not used, because the insured household no longer needs to hold precautionary savings against a catastrophic bill or to price that risk into every spending decision. That ex ante value, the value of not having to worry, is real even though no study in this literature puts a clean dollar figure on it for the expansion population. What the studies do capture is its shadow: reduced financial strain, fewer skipped bills, better credit standing. Those are the footprints of a risk that stopped hanging over people.

The credit score channel deserves emphasis because it is the mechanism through which a health policy becomes a labor market and housing policy. A damaged credit file raises the price of a car loan, can disqualify a renter from an apartment, and in some states can cost a job applicant an offer, since employers in many jurisdictions check credit histories. When the expansions kept medical debts off those files, they were not merely saving households the face value of a bill. They were preserving the household’s standing in every market that prices risk off a credit report. Brevoort, Grodzicki, and Hackmann’s findings on improved credit outcomes therefore describe a benefit that compounds: cheaper credit means fewer future financial emergencies, which means fewer future debts. The reverse compounding is what the uninsured faced before, where one hospital bill could cascade into a worse credit score, more expensive borrowing, and the next emergency financed at the worst possible terms.

One caveat belongs in this section, and it concerns the difference between the expansion population and everyone else the statute touched. The financial protection findings are overwhelmingly about the program’s enrollees, people with incomes low enough to qualify and cost sharing low enough to be negligible. They say much less about marketplace enrollees in silver plans with several thousand dollars of deductibles, or about workers in high deductible employer plans, who in the 2014 to 2022 window faced growing out of pocket exposure even as they counted as insured. Survey research from Sara Collins, Herman Bhupal, and Michelle Doty at the Commonwealth Fund, reporting the 2018 Biennial Health Insurance Survey in 2019, documented that underinsurance, defined as out of pocket costs reaching 10 percent of income, or 5 percent for low income adults, or deductibles of 5 percent or more of income, remained common among precisely those groups across this period. The honest boundary of the third finding is therefore this: the statute’s most consistent financial effect ran through the part of the law that asked people to pay almost nothing, and it did not extend in the same measure to the parts that asked them to pay more. That boundary is not a flaw in the finding. It is the finding, stated with its edges visible.

What can a two-year lottery result establish, and what can it not?

A two year follow up is well powered to detect changes in financial strain, depression screening, and whether people received a diagnosis. It is not well powered to detect whether those diagnoses translated into controlled blood pressure or blood sugar, because control requires time, adherence, and repeated contact with a system the newly insured were just learning to use.

A two year follow up is well powered to detect changes in financial strain, in depression, and in whether people got a diagnosis. It is not well powered to detect changes in whether those diagnoses then translated into controlled blood pressure, because control takes time, adherence, and repeated contact with a system the newly insured were just learning to use. The Oregon lottery proved the financial mechanism could work at small scale; the expansion studies proved it did work at national scale.

The Oregon experiment is the closest thing health economics has to a laboratory result, and its financial findings are the template against which every later study in this section was read. In 2008, Oregon used a lottery to allocate a limited number of openings in its Medicaid program for low income uninsured adults, which created random assignment to the chance to enroll. Katherine Baicker, Sarah Taubman, Heidi Allen, Mira Bernstein, Jonathan Gruber, Joseph Newhouse, Amy Finkelstein, and their colleagues, writing in the New England Journal of Medicine in 2013, reported that about two years after the lottery, Medicaid coverage had nearly eliminated catastrophic out of pocket medical expenditures. The probability of having to borrow money or skip paying other bills because of medical expenses fell substantially.

Those numbers deserve to be read slowly, because they are doing two jobs at once. They establish the scale of the financial harm that being uninsured inflicted on this population: a meaningful share of people near the poverty line were spending catastrophic fractions of household income on medical care out of pocket, which for a household near the poverty line is not a budget item but a catastrophe. And they establish how completely the program removed that harm. The same paper found that coverage increased the use of health care services, including physician visits, prescription drugs, and preventive care, and that it lowered the probability of screening positive for depression by 9.15 percentage points. The financial protection and the mental health result belong together. A bill that cannot arrive is a worry that cannot accumulate, and the depression finding is one of the few places in the literature where the psychological benefit of financial protection is measured directly rather than inferred.

The parallel the brief asked for is therefore not a stretch. It is the same population, the same program design, the same outcomes, measured first with a lottery and then at national scale with administrative data. The Oregon lottery proved the mechanism could work. The expansion studies proved it did work, at scale, under the statute. When the Hu team found collection balances falling and the Brevoort team found credit outcomes improving, they were documenting in credit files what the Oregon researchers had documented in interviews and medical records: that the most reliable thing this kind of coverage does is keep medical bills from becoming financial ruin. The health outcomes literature, which the next section takes up, is where the disagreements live. The financial literature is where they do not.

Health Outcomes: A Range With a Reason, Not a Single Verdict

If the financial findings are the least contested in this literature, the health findings are the most contested, and honesty requires presenting them as a range rather than as a verdict. The range is wide. At one end are studies that find meaningful reductions in mortality from gaining this kind of coverage. At the other end are studies that find no detectable effect on mortality or on measured physical health. Both ends are real research by serious people. The task is not to pick a side and then defend it. The task is to explain why the same question, asked about the same program, produces different answers depending on who is studied, how long they are followed, and what outcome is measured. When that is done carefully, the apparent contradiction mostly dissolves into something a reader can actually use.

The single most important reason for the spread is the difference between studying the people most likely to benefit and studying everyone at once. A mortality effect has to be large among the people who gained coverage to be visible in a population where most people were already insured and never touched the expansion. Several of the studies that examined ACA era expansions at the county or state level estimated that coverage rose by only a few percentage points among all nonelderly adults, because the newly eligible were a minority of the population being averaged. A real reduction in deaths among the newly insured gets diluted by the much larger number of people whose insurance status never changed, and diluted effects are hard to distinguish from noise. Studies that could identify the people most likely to gain coverage, and follow their outcomes specifically, found larger and more consistent effects. That is not a trick of the methods. It is what the word detectable means.

Why do county averages hide the mortality effects that linked records reveal?

Miller, Johnson, and Wherry linked survey records to administrative death data and found a 0.132 percentage point mortality decline, 9.4 percent over the sample mean, in the Quarterly Journal of Economics in 2021. County averages wash that effect out because the newly insured are a small share of everyone measured.

Studies that link individual records and isolate the people most likely to benefit, such as Sarah Miller, Norman Johnson, and Laura Wherry in the Quarterly Journal of Economics in 2021, find that expansion reduced annual mortality by 0.132 percentage points, a 9.4 percent reduction over the sample mean, with the effect growing over time. Studies that average across whole counties or states find smaller effects or none, because the newly insured are a small share of everyone being measured.

The positive end of the range has a history that predates the statute. Benjamin Sommers, Katherine Baicker, and Arnold Epstein published their analysis in the New England Journal of Medicine in 2012, examining three states, Arizona, Maine, and New York, that had expanded their programs to childless adults between 2000 and 2005, and comparing county level mortality trends against neighboring states that did not expand. In their preferred model, annual mortality rates declined by 19.6 deaths per 100,000 adults, a relative reduction of 6.1 percent, which they translated into roughly 2,840 deaths prevented per year for each 500,000 adults gaining coverage. The reductions were largest among older adults, nonwhite residents, and people in poorer counties, and the authors found no effect among people under 35, whose baseline mortality was too low for coverage to move. They also found improvements in coverage, in self reported health, and in the share of people who said they had delayed care because of cost. The study is quasi experimental rather than randomized, and the authors said so plainly, but the pattern of results hangs together clinically as well as statistically: the groups with the most to gain from care gained the most.

The ACA era version of that result comes from Miller, Johnson, and Wherry, and it is stronger in some ways because the data are better. The authors linked large scale federal survey data to administrative death records, which let them identify the near elderly adults most likely to benefit using information on socioeconomic status, citizenship, and program participation, and then compare mortality changes in expansion states against non expansion states. Mortality trends moved in parallel before the expansions, which is the key credibility check for this kind of comparison, and then diverged beginning in the first year of the policy. The estimated 0.132 percentage point decline in annual mortality, 9.4 percent over the sample mean, was driven by reductions in disease related deaths and grew over time. Alternative specifications, placebo tests, and different sample definitions all confirmed the result. That bridging between the strongest observational design in the literature and the administrative data matters: it connects population scale records to the comparison group logic, and finds them telling the same story.

Mark Borgschulte and Jacob Vogler, writing in the Journal of Health Economics in 2020, approached the same question with restricted access microdata covering all deaths in the United States and compared counties in expansion and non expansion states over the first four years after expansion, using propensity score weighting and machine learning techniques to adjust for pre expansion differences. They estimated a reduction in all cause mortality for ages 20 to 64 of 11.36 deaths per 100,000, a 3.6 percent decrease, driven largely by counties with higher pre expansion uninsured rates and by causes of death likely to respond to access to care. Their cost benefit calculation suggested that the welfare improvement from the mortality response could offset the entire net of transfers expenditure associated with the expansion. That last claim is a modeling exercise rather than a measured outcome, and a careful reader should treat it as such, but the mortality estimate itself sits squarely in the middle of the positive range.

A 2021 analysis in The Lancet Public Health by Brian Lee and colleagues, comparing 32 expansion states against 17 non expansion states over four years of implementation, estimated that expansion was associated with 11.8 fewer deaths per 100,000 adults per year, with substantial variation by state, from large reductions in some states to essentially no change in others. States with higher proportions of women and of non Hispanic Black residents saw larger adjusted reductions. The state level variation is itself a finding worth dwelling on: it suggests that how a state implemented the expansion, which populations it reached, and what its baseline health conditions were all shaped the outcome. Expansion was not a single treatment applied uniformly. It was dozens of state level treatments wearing one federal name.

What do condition-specific mortality studies add to the picture?

They shorten the causal chain. A person with kidney failure who gains coverage gets dialysis on schedule, and dialysis keeps people alive; nobody needs a sophisticated identification strategy to see why that would work. These studies function as plausibility checks on the broader mortality findings, because the mechanism from coverage to survival is short and direct.

The range includes studies of specific conditions where the mechanism from coverage to survival is short and direct. Shailender Swaminathan, Benjamin Sommers, Rebecca Thorsness, and colleagues examined one year survival among nonelderly patients with end stage renal disease beginning dialysis from January 2011 through March 2017, and reported in the Journal of the American Medical Association in 2018 that mortality declined by 0.8 percentage points in expansion states against 0.2 percentage points in non expansion states, an adjusted absolute mortality reduction of 0.6 percentage points associated with expansion. The reductions were largest for Black patients, at 1.4 percentage points, and for patients ages 19 to 44, at 1.1 percentage points. Sameed Khatana and colleagues, in JAMA Cardiology in 2019, compared county level cardiovascular mortality among adults ages 45 to 64 from 2010 to 2016 and found that counties in expansion states had 4.3 fewer cardiovascular deaths per 100,000 residents per year than if they had followed the same trends as counties in non expansion states. These condition specific results are easier to believe than population wide ones, precisely because the causal chain is short. Nobody needs a sophisticated identification strategy to see why scheduled dialysis would keep kidney patients alive, which is why these studies function as plausibility checks on the broader mortality findings.

The other end of the range. The null results are real, they are published, and they have explanations that do not require anyone to have made an error. Population level studies with short follow up windows often found no detectable effect on adult mortality, which is what one expects when the coverage gain is a small share of the population being averaged and the follow up window is short. The Oregon experiment itself, in the 2013 New England Journal of Medicine paper by Baicker, Taubman, Allen, Bernstein, Gruber, Newhouse, Finkelstein, and colleagues, found no significant effect of coverage on measured physical health outcomes in the first two years: no significant change in the prevalence or diagnosis of hypertension or high cholesterol, no significant change in average glycated hemoglobin levels among people with diabetes, even as the same study found increased use of services, increased diabetes detection and management, lower depression, and the near elimination of catastrophic out of pocket spending. That is a null result from a randomized trial, the strongest possible design, and it has to be taken seriously.

The two year window and the physical health null

Two years is not long enough for some health measures to move, which is a statement about time, not about insurance. Hypertension, cholesterol, and blood sugar are managed over years, and their complications kill over decades. A study that follows people for two years is well powered to detect changes in financial strain, in depression, and in whether people got a diabetes diagnosis. It is not well powered to detect changes in whether those diagnoses then translated into controlled blood pressure, because control takes time, adherence, and repeated contact with a system the newly insured were just learning to use.

The Oregon authors said exactly this in their conclusion: the study showed no significant improvements in measured physical health outcomes in the first two years, alongside increased service use, better diabetes detection, lower depression, and reduced financial strain. A null on a two year physical health measure is not evidence that coverage does not improve health. It is evidence that two years is not long enough for some health measures to move.

Sommers, Long, and Baicker added a related data point in the Annals of Internal Medicine in 2014, studying mortality after the Massachusetts reform of 2006 and finding a 2.9 percent relative reduction in all cause mortality, or 8.2 deaths per 100,000, associated with the reform. The Massachusetts result belongs in this section because it sits between the pre ACA state expansions and the ACA era expansions in both time and design, and because it shows the mortality finding appearing in a third setting. The consistent thread across the positive studies is that effects appear where the newly insured are concentrated, among older adults, among people with chronic conditions, among populations with high baseline mortality, and that they grow with follow up length. The consistent thread across the nulls is dilution: whole population averages, short windows, outcomes that move slowly.

Self reported health sits in the middle of the range and deserves its own paragraph because it is the outcome most often dismissed and least fairly. Benjamin Sommers, Bethany Maylone, Robert Blendon, E. John Orav, and Arnold Epstein, in a 2017 analysis in Health Affairs comparing low income adults in Arkansas, Kentucky, and Texas from 2013 to 2016, found a 41 percentage point gain in having a usual source of care, 337 dollars lower annual out of pocket spending, and a 23 percentage point gain in excellent self reported health in the expansion states relative to the non expansion state. The Oregon experiment found substantial gains in self reported health in its first year results. Critics sometimes wave these findings away on the grounds that self reports are subjective, but that objection misunderstands what the measure is for. A person who says their health interferes less with their daily activities is reporting a welfare gain that no blood pressure cuff captures, and the depression reductions in Oregon are self reported too, on a validated screening instrument. Subjective does not mean unreal. It means measured at the level where people actually live. The same attribution problem, separating what a health insurance statute did from what background trends would have done anyway, runs through the series account of what Medicare did to elderly poverty, where the evidence discipline is identical even though the population is different.

The intermediate outcomes, the steps between gaining coverage and living longer, point consistently in one direction even where the final mortality outcome is contested. Benjamin Sommers, Atul Gawande, and Katherine Baicker, in a 2017 review in the New England Journal of Medicine, surveyed the recent evidence on coverage and health and found a consistent pattern of improved access to care, increased use of preventive services, and better self reported health among the newly insured, alongside the financial protection findings and the mixed mortality evidence already discussed. People who gained coverage were more likely to have a usual source of care, more likely to receive recommended screenings, and less likely to delay care because of cost. Those are not health outcomes in the strict sense. They are the behaviors through which health outcomes eventually move, and they moved quickly and uniformly. The review’s implicit argument is a temporal one: access changes in year one, diagnoses change in year two, management of chronic conditions changes in years three through five, and mortality, for the conditions where coverage can affect it, changes after that. A literature that finds strong effects on the early steps and mixed effects on the final step is not contradicting itself. It is describing a causal chain at different points along its length, and the honest reader should weight the early steps heavily precisely because they are the mechanism the later steps require.

The honest summary of this section is therefore not a number but a conditional statement. Coverage of this kind reduced mortality in studies that could see the people it reached and followed them long enough for the effect to appear, with the largest effects among the near elderly, the chronically ill, and disadvantaged populations. It showed no detectable effect in studies that averaged over whole populations or followed people for too short a time for physical health to respond. Those two sentences are not in tension. They are the same sentence, stated with the conditions made explicit, and any presentation of this evidence that drops the conditions is doing advocacy rather than analysis.

Premiums, the Coverage Gap, and the 2013 Cancellations

The fifth finding is the one that made the law’s critics, and it has three parts that must be kept distinct even though they are constantly blended in public argument. First, prices in the individual market rose sharply in the early years after the coverage provisions took effect, then stabilized. Second, prices in the employer market, where most insured Americans get their coverage, continued on the trend they had been on for years, neither accelerating nor pausing. Third, a population was left with no affordable option at all, not because the statute forgot about them but because the Supreme Court removed the provision that was supposed to cover them. Blending these three into a single statement about what the law did to prices is the most common analytical error in the whole debate, and it is worth understanding each part on its own terms before anyone is allowed to generalize.

What drove the individual market repricing of 2017 and 2018?

Insurers priced 2014 with almost no information about the new risk pools, and enrollment skewed sicker and older than assumed. Benchmark premiums rose about 22 to 25 percent for 2017 and 37 percent for 2018, the latter partly from the end of direct federal cost sharing reduction payments, before falling 1.5 percent in 2019 and 4 percent in 2020.

Insurers priced 2014 with little information, enrollment skewed sicker and older than the models assumed, and several large carriers raised rates sharply or exited markets entirely. Benchmark premiums then rose about 22 to 25 percent for 2017 and 37 percent for 2018, with the 2018 jump partly reflecting the end of direct federal reimbursement for cost sharing reductions, before falling 1.5 percent in 2019 and 4 percent in 2020.

The increase was steepest in the years when insurers were still learning what the new risk pools looked like. Plans sold to individuals faced guaranteed issue, community rating, and a required set of essential health benefits beginning with the 2014 plan year, which raised the actuarial value of what was being sold relative to the thin pre 2014 individual market. Benchmark premiums rose about 2 percent for 2015 and 7 percent for 2016, then the repricing peaked: roughly 22 to 25 percent for 2017 and 37 percent for 2018, according to Department of Health and Human Services and analyst series, with the 2018 increase partly reflecting insurers loading the cost of cost sharing reductions into silver tier premiums after the federal government stopped reimbursing them directly. Insurers that had underpriced in 2014 and 2015 corrected in 2016, 2017, and 2018, and the corrections were large because the initial mispricing had been large. After that repricing, benchmark premiums stabilized and insurer finances in the marketplaces improved through the end of the decade, with benchmark premiums falling 1.5 percent for 2019 and 4 percent for 2020. Analysts at KFF, the Congressional Budget Office, and the Department of Health and Human Services all documented the arc from turbulence to steadier ground across the 2014 to 2022 window.

The mechanism behind the early spike is worth spelling out because it is routinely misdescribed as proof that the law’s economics were unsound, when it was mostly proof that pricing a new market is hard. An insurer setting a 2014 rate had to guess the health mix of a population that had never been offered guaranteed issue coverage before. The people most motivated to enroll first were the people who knew they needed care, which is the textbook definition of adverse selection, and it arrived exactly on schedule. Claims ran above expectations. The risk corridor and reinsurance programs that were supposed to cushion early losses paid out less than insurers had planned, in part because Congress later restricted the risk corridor payments, which turned an expected shock absorber into an additional loss. The reinsurance program, which reimbursed insurers for a share of high cost claims in 2014, 2015, and 2016, did function as designed and held down premiums in those years relative to what they would otherwise have been; its scheduled expiration after 2016 removed one more support just as the repricing was peaking, which helps explain why the 2017 and 2018 rate filings looked so severe. None of that implies the market could not work. It implies that the first prices were guesses, the guesses were wrong in a predictable direction, and the market then repriced toward reality.

How did subsidies change the meaning of rising sticker premiums?

The premium tax credits cap what subsidized enrollees pay as a share of income, so when benchmark premiums rose, subsidies rose with them and most enrollees’ net payments stayed roughly flat. The sticker price increases were largely absorbed by the federal budget. The pain concentrated among the unsubsidized: higher income individual market buyers and people in the coverage gap.

The premium tax credits were designed to cap what subsidized enrollees paid as a share of income, which meant that when benchmark premiums rose, the subsidy rose with them and the enrollee’s net payment stayed roughly flat. For most marketplace enrollees, who received subsidies in most years of the period, the sticker price increases were largely absorbed by the federal government. The pain concentrated among the unsubsidized: people with incomes above 400 percent of the poverty level who bought in the individual market, and people in the coverage gap states who earned too little for subsidies but lived where the expansion never arrived.

What the repricing felt like depended entirely on whether the buyer received a subsidy. That distributional detail is the difference between saying premiums rose, which is true of sticker prices, and saying the law made coverage less affordable for everyone, which is not. The statute deliberately traded higher sticker prices for guaranteed issue and community rating, and then used subsidies to hold most buyers harmless. Whether that tradeoff was worth it is a value judgment. That the tradeoff was the design, not an accident, is a fact.

The employer market tells a different story, and it is the story that covers most insured Americans. The Kaiser Family Foundation’s annual Employer Health Benefits Survey, the benchmark series for this market, reported average family premiums of 16,834 dollars in 2014, the year the coverage provisions took effect. By 2018 the figure was 19,616 dollars, and the survey’s 2022 edition reported 22,463 dollars for family coverage. Annual increases in most of these years ran in the low to mid single digits, which the survey’s authors repeatedly described as moderate by historical standards. The pattern across the whole 2014 to 2022 window is continuity, not rupture: employer premium growth neither accelerated because of the law nor slowed because of it, but continued the longer trend of rising faster than wages that predated the statute by decades. Anyone who wants to credit or blame the law for employer premium trends has to explain why the trend looks the same on both sides of 2014, and the survey data make that explanation hard to sustain.

Deductibles and the underinsurance problem

Underinsurance persisted alongside the coverage gains, and it complicates any simple insured versus uninsured story. The Commonwealth Fund’s 2018 Biennial Health Insurance Survey, reported by Sara Collins, Herman Bhupal, and Michelle Doty in 2019, found that a substantial share of insured adults remained underinsured on income based definitions, meaning their out of pocket costs were high relative to what they earned, with the problem concentrated among lower income adults. A person counted as insured in the Census statistics could still face a deductible of several thousand dollars before coverage paid for much beyond preventive care. That exposure matters for this article’s financial protection finding, because it sets the boundary of what the finding claims: the expansions reduced catastrophic spending and medical debt among the people who gained coverage, but they did not eliminate the affordability problem for people whose plans left them exposed. The distinction between being insured and being adequately insured is one the survey literature has tracked for years, and the law moved the first measure far more than the second.

The drafting assumption behind the coverage gap

The law was written on the premise that every state would expand the joint federal state program to adults with incomes up to 138 percent of the poverty level. On that premise, the drafters set the minimum income for marketplace subsidies at 100 percent of the poverty line, because anyone below that line was supposed to be in the expanded program. The Supreme Court’s 2012 decision made the expansion optional instead of mandatory, and in the states that chose not to expand, the design assumption collapsed, leaving adults below the poverty line too poor for subsidies and ineligible for their state’s unexpanded program.

The coverage gap is the part of this story that the statute’s authors never intended to exist, and understanding it requires understanding what they assumed. There was no need to subsidize people below the poverty line in the marketplaces, because in the drafters’ design no such people would need marketplace coverage. The gap is what the statute’s arithmetic produces once expansion becomes optional, which is why the Kaiser Family Foundation’s 2022 analysis could still count about 2.2 million people in the gap across the 12 states that had not expanded. The scale of the gap shrank over time as more states adopted the expansion, which is itself evidence that the gap was a product of state decisions rather than of the statute’s economics.

The gap is also the clearest illustration of the brief’s namable claim, which the verdict will state in full. The statute’s largest coverage effect ran through an expansion that the Court made optional, which means the law’s measured performance is partly a measure of state decisions. The uninsured rate fell much further in expansion states than in non expansion states. The financial protection findings in this article come from expansion states. The mortality reductions come from comparing expansion states against non expansion states. And the gap population, the people with the worst outcomes in the entire post 2014 landscape, lived in the states that opted out. A national average across all of this conceals the actual structure of what happened, which is that the statute worked substantially better where it was allowed to operate as written. That is not a partisan observation. It is what the state by state data show, and it is why researchers treated the expansion decision as a natural experiment rather than as background noise.

There is a further subtlety that even careful accounts sometimes miss. The gap population was not merely uninsured. It was uninsured in a way the statute had specifically tried to prevent and that only a judicial intervention made possible, which means the standard conservative critique that the law left millions behind and the standard liberal defense that the law covered millions are both describing real phenomena that the same design feature produced. The law covered millions through the expansion where states adopted it. It left millions uncovered where states did not. Both statements are true, and both are incomplete without the other, because the mechanism connecting them is a single Supreme Court decision about federalism, not a single judgment about the statute’s merits.

Why does a cancellation count differ from a coverage loss count?

The two estimates measured different things at different moments. The Associated Press counted about 4.7 million cancellation notices across 30 states in December 2013; the Urban Institute estimated about 3.4 million cancellations nationwide. Notices went to everyone whose plan was discontinued, but many recipients moved to compliant replacement coverage rather than becoming uninsured.

Notices were sent to everyone whose plan was being discontinued, but many recipients transitioned to compliant replacement coverage, often with subsidies, rather than becoming uninsured. The Associated Press counted about 4.7 million notices across 30 states in December 2013, while the Urban Institute estimated about 3.4 million cancellations nationwide. The two estimates measured different things at different moments, and the number of people who actually lost coverage was smaller than the number of notices sent.

The episode needs a plain telling because it has been told so many ways. In the fall of 2013, as the 2014 plan year approached, insurers sent cancellation notices to millions of people holding individual market policies that did not comply with the law’s new requirements. The plans being discontinued were non grandfathered policies that lacked the essential health benefits the statute required, things like prescription drug coverage, mental health services, and maternity care, or that fell below the minimum actuarial value standards. Insurers were not dropping these customers the way an insurer drops a sick enrollee. They were discontinuing non compliant products and offering the affected policyholders replacement coverage that met the new standards, which is what the Health Insurance Portability and Accountability Act of 1996 had always permitted when an insurer discontinues a particular plan. The cause was regulatory, not personal: the old products could not legally be sold anymore.

The scale was genuinely large and genuinely disputed at the same time, which is why the numbers need their sources attached. The two estimates differ because they measured different things at different moments with different methods, and because the market was moving under their feet: on November 14, 2013, the Obama administration announced a transitional policy permitting insurers, where state regulators agreed, to renew non compliant plans for policy years beginning between January 1, 2014 and October 1, 2014. Some states and insurers took the offer, some refused. Precision beyond that range is not available, and anyone quoting a single exact figure for this episode is overstating what the data support.

The political damage came from the collision between the cancellations and a presidential promise. Barack Obama had repeatedly said that people who liked their health plans could keep them, and the notices made that promise visibly false for millions of households. The administration’s defenders noted that many of the cancelled plans were thin coverage that the law was designed to replace, and that most affected people could buy compliant coverage, often with subsidies that made it cheaper than what they had. The administration’s critics noted that the promise had been unconditional and that the replacement coverage was in many cases more expensive for people who did not qualify for subsidies. Both observations are factual. The episode’s lesson for this article is narrower: it demonstrated, earlier than the premium debates did, that the statute’s strategy of raising minimum standards would create visible losers even as it created larger numbers of winners, and that the losers would be concentrated, identifiable, and angry in a way the winners, dispersed across the expansion population, would not be. That asymmetry in political visibility has shaped every subsequent argument about the law’s costs, and it was baked into the design from the start.

Verdict: Several Outcomes Held at Once

The evidence assembled in this article refuses to fit inside either of the two sentences that dominate public argument about the statute. The first sentence says the law failed because premiums rose. The second says the law succeeded because coverage rose. Both sentences pick one outcome and treat it as the whole, and both are overreaches of exactly the symmetrical kind the brief asked this article to hold at once. Premiums did rise in the individual market, sharply at first, and the people who paid those increases without subsidies have a legitimate grievance that no coverage statistic erases. Coverage did rise, substantially, and the people whose medical debts vanished and whose depression lifted have a legitimate experience that no premium chart erases. A verdict that honors the evidence has to keep both of those truths in view at the same time and then say what the statute deliberately traded.

The tradeoffs were not hidden. They were the design. Guaranteed issue and community rating mean that insurers cannot charge sick people more or turn them away, which means healthy people pay more than they would in a market priced on individual risk. Essential health benefits mean that every plan covers maternity care, mental health services, and prescription drugs, which means people who would have bought thinner cheaper plans pay for coverage they may not want. Those are costs imposed on identifiable people, and they were imposed on purpose, to fund the cross subsidies that make coverage attainable for the sick, the old, and the poor. The premium tax credits then hold most marketplace buyers harmless from the resulting sticker prices, which means the federal budget absorbs much of the cost that the regulations create. To describe this structure as a failure because some unsubsidized buyers paid more is to mistake the price of the mechanism for evidence that the mechanism malfunctioned. The mechanism did what mechanisms do: it moved money from one group to another. The question is whether the movement was worth it, and that question has an answer only inside a set of values, not inside a dataset.

What the datasets do answer is narrower and more useful. The financial protection finding is close to settled: across lotteries, credit records, bankruptcy filings, and surveys, gaining this kind of coverage reduced medical debt, unpaid bills, and catastrophic out of pocket spending, with spillovers into non medical finances and credit standing. The mortality finding is genuinely contested in its magnitude but not in its direction among the studies best positioned to detect it: reductions appear where the newly insured are concentrated and follow up is long enough, and they do not appear where populations are averaged or windows are short. The premium finding is settled in its shape: early turbulence in the individual market as insurers learned the risk pools, then stabilization, with subsidies shielding most buyers, while the employer market continued its pre existing trend as if the law had never passed. The coverage gap finding is settled in its cause: a statute written for mandatory expansion met a Court decision making expansion optional, and the people below the poverty line in non expansion states were left with neither the program nor the subsidies. Each of those sentences carries its period and its named source in the sections above. None of them needs the others to be true.

The Medicaid first claim is the thread that ties them together and the one this article is meant to leave in the reader’s hands. The statute’s largest coverage effect ran through an expansion of the joint federal state program, the expansion the Supreme Court made optional, which means the law’s measured performance is partly a measure of state decisions rather than of the statute itself. Where states expanded, the uninsured rate fell further, the financial protection materialized, and the mortality studies found their effects. Where states did not, the gap population accumulated, and the national averages that blend the two groups conceal the structure underneath. Any assessment of the statute that does not separate what the text did from what implementation and litigation allowed it to do is assessing a different law than the one Congress wrote. That separation is the series thesis in its applied form: judge the statute against its own aims, and keep the aims distinct from the obstacles.

There is one more tradeoff the statute made that deserves to be named plainly, because it explains the political afterlife of everything above. The law concentrated its visible costs and dispersed its visible benefits. The premium increases fell on identifiable households who opened identifiable bills. The 2013 cancellations arrived as identifiable letters in identifiable mailboxes. The financial relief, by contrast, arrived as debts that never went to collections, bankruptcies that never happened, and depressions that lifted quietly in households nobody surveyed for their gratitude. Costs that arrive as events are politically louder than benefits that arrive as non events, and that asymmetry has structured a decade of argument in which the law’s critics could always point to someone paying more while its defenders had to point to statistical populations. The evidence in this article does not resolve that asymmetry. It explains it, which is the more durable service.

Nothing in this verdict extends to any proposal before any legislature. The findings carry their periods, 2014 through 2022, and their sources, and they describe what happened under one statute in one country during those years. Whether similar mechanisms would produce similar results under different provisions, different courts, and different state governments is a question for separate analysis with separate evidence. What this article establishes is the record itself: a law that reduced financial ruin reliably, reduced mortality where it could be measured cleanly, raised some prices by design while holding most buyers harmless, and left a gap where a court decision, not the text, created one. Hold all of that at once, and the two symmetrical overreaches fall away on their own.

The statutory changes that altered the trends traced above, including the zeroed penalty and the enlarged subsidies, are taken up in the later article in this series on how the law was amended after 2010.

The five-finding evidence table

Each outcome below carries its direction, its period, its principal studies, and a rating of where the literature stands. Readers who want to keep the statute’s provisions, the study citations, and the case chronology behind these findings in one place can keep your statute notes, citations, and case chronologies together free on VaultBook.

Outcome	Direction of finding	Period	Principal studies	Settled or contested
Coverage gains and the Medicaid majority	Uninsured rate fell sharply; the public program supplied most new coverage	2013 to 2022	Census Current Population Survey and American Community Survey; National Health Interview Survey; Urban Institute and Robert Wood Johnson Foundation; Centers for Medicare and Medicaid Services enrollment data	Settled
Expansion versus non-expansion states	Adopting states gained more coverage and saw larger uncompensated care declines	2014 to 2022	Sommers and colleagues; Kaufman and colleagues; Congressional Research Service adoption timeline	Settled for coverage; contested for mortality magnitude
Financial protection	Medical debt, collections, and catastrophic spending fell consistently	2008 to 2022	Hu and colleagues (Journal of Public Economics, 2018); Brevoort, Grodzicki, and Hackmann (NBER Working Paper 24002, 2017); Finkelstein and colleagues (Quarterly Journal of Economics, 2012); Baicker and colleagues (New England Journal of Medicine, 2013)	Settled; the most consistent finding in the literature
Mortality and health outcomes	Reductions in well powered, long horizon designs; nulls in short or diluted ones	2000 to 2022	Sommers, Baicker, and Epstein (NEJM, 2012); Sommers, Long, and Baicker (Annals of Internal Medicine, 2014); Miller, Johnson, and Wherry (Quarterly Journal of Economics, 2021); Borgschulte and Vogler (Journal of Health Economics, 2020)	Contested in magnitude; direction leans positive
Premiums, the coverage gap, and the 2013 cancellations	Individual market turbulence then stabilization; employer trend unchanged; gap caused by optional expansion	2012 to 2022	Department of Health and Human Services benchmark series; Kaiser Family Foundation Employer Health Benefits Survey; Kaiser Family Foundation (2022); Associated Press (2013); Urban Institute (2013)	Settled on shape and cause; contested on exact cancellation count

Frequently Asked Questions

Q: How many people gained coverage under the Affordable Care Act?

The Census Bureau’s Current Population Survey measured 49.9 million uninsured Americans, or 16.3 percent of the population, in 2010, and 28.1 million, or 8.8 percent, in 2016. The Urban Institute and the Robert Wood Johnson Foundation, analyzing 2013 to 2016, estimated that about 18.5 million people gained coverage, with the Medicaid expansion supplying a majority, about 10.9 million. That Medicaid majority is the important correction to the popular impression that the exchanges did most of the work. Marketplace plan selections reached about 12.7 million during the 2016 open enrollment period, according to the Centers for Medicare and Medicaid Services, well below the public program’s gains. The uninsured rate later reached a record low: the Current Population Survey recorded 7.9 percent in 2022, while the National Health Interview Survey recorded 8.4 percent for the full year. The two surveys use different methods, which is why their levels differ while their trend agrees.

Q: Did the Affordable Care Act reduce the uninsured rate?

Yes, substantially, though the path was not a straight line. The Census Bureau measured an uninsured rate of 16.3 percent in 2010, the year the law passed, and 8.8 percent in 2016, after the coverage provisions had been operating for three years. The rate then ticked upward during the late 2010s: the American Community Survey recorded 9.2 percent in 2019, which researchers attributed in part to the zeroed individual mandate penalty, reduced outreach funding, and the expansion of short term plans. Enrollment rebounded afterward. The Current Population Survey recorded an uninsured rate of 7.9 percent in 2022, the lowest the survey had ever measured, while the National Health Interview Survey recorded 8.4 percent for the full year. Enlarged premium subsidies and continuous Medicaid enrollment during the public health emergency both contributed to the decline. The pattern shows the law reduced the uninsured rate, but the size of the reduction responded to later policy choices that encouraged or discouraged enrollment.

Q: Did the Affordable Care Act reduce mortality?

The evidence leans positive but is genuinely contested in its magnitude. Sarah Miller, Norman Johnson, and Laura Wherry, linking survey records to administrative death records in a paper published in the Quarterly Journal of Economics in 2021, found that Medicaid expansion reduced annual mortality among near elderly adults by 0.132 percentage points, a 9.4 percent reduction relative to the sample mean. Benjamin Sommers, Katherine Baicker, and Arnold Epstein found a 6.1 percent relative mortality reduction from earlier state expansions in the New England Journal of Medicine in 2012. But shorter follow ups and population averages found null results, including the Oregon lottery study’s two year follow up. The methodological reason for the spread is straightforward: deaths are rare events, so detecting an effect requires large samples and long observation windows, and studies with shorter panels were underpowered. The fair summary is that the best powered studies point toward mortality reductions, while shorter or smaller studies could not rule out zero.

Q: Did Affordable Care Act premiums go up?

It depends on which market you mean, and the distinction matters. In the individual market, benchmark premiums rose sharply in the early years: roughly 22 to 25 percent for 2017 and 37 percent for 2018, with the 2018 jump partly reflecting the end of direct federal reimbursement for cost sharing reductions, which insurers loaded into silver tier premiums. Benchmark premiums then fell 1.5 percent in 2019 and 4 percent in 2020. In the employer market, which covers most insured Americans, premium growth continued its pre existing trend; the Kaiser Family Foundation’s Employer Health Benefits Survey reported average family premiums of 16,834 dollars in 2014, 19,616 dollars in 2018, and 22,463 dollars in 2022. So the law’s individual market saw genuine premium turbulence in its first years, while employer premiums followed the same trajectory they had before 2014. Most marketplace enrollees received subsidies that held their net payments roughly flat as sticker prices moved.

Q: What is the Affordable Care Act coverage gap?

The coverage gap is a direct consequence of the law’s text meeting the Supreme Court’s 2012 decision. The statute was written assuming every state would expand Medicaid to adults earning up to 138 percent of the federal poverty line, and it offered marketplace subsidies only to people earning at least 100 percent of the poverty line, on the theory that everyone below that threshold would be on Medicaid. When the Court made expansion optional, adults in non expanding states who earned too much for their state’s traditional Medicaid program but less than the poverty line fell into a gap: ineligible for Medicaid and ineligible for subsidies. The Kaiser Family Foundation estimated that about 2.2 million people fell into this gap across the 12 non expansion states in 2022. It is the clearest example of how the law’s measured performance reflects state decisions layered on top of the statute’s design.

Q: Did Medicaid expansion under the Affordable Care Act improve health?

The financial and access results are strong; the clinical results are mixed. Benjamin Sommers, Bethany Maylone, Robert Blendon, E. John Orav, and Arnold Epstein, comparing low income adults in Arkansas, Kentucky, and Texas from 2013 to 2016 in Health Affairs in 2017, found a 41 percentage point gain in having a usual source of care, 337 dollars lower annual out of pocket spending, and a 23 percentage point gain in excellent self reported health in the expansion states. The Oregon Medicaid lottery, studied by Katherine Baicker, Amy Finkelstein, and colleagues, found that coverage increased health care use and reduced the probability of a positive depression screening by 9.15 percentage points in their 2013 New England Journal of Medicine paper. But the same Oregon study found no statistically significant improvement in measured blood pressure, cholesterol, or diabetes control within its roughly two year follow up, a null result the authors reported directly. So expansion clearly improved financial security and access, with downstream gains visible in mental health but not yet in short run clinical measures.

Q: What does the Oregon experiment tell us about Affordable Care Act coverage?

The Oregon Health Insurance Experiment, run by Katherine Baicker, Amy Finkelstein, and their coauthors, is the closest thing health economics has to a laboratory test of Medicaid coverage, because Oregon allocated a limited number of Medicaid slots by lottery in 2008. Published in the New England Journal of Medicine in 2013, with the lead economics paper in the Quarterly Journal of Economics in 2012, the results were nuanced. Winning the lottery increased health care use across the board, raised the detection and treatment of diabetes, reduced the likelihood of a positive depression screening by 9.15 percentage points, and nearly eliminated catastrophic out of pocket medical spending. But it also increased emergency department use, reported by Taubman and colleagues in Science in 2014 at about 0.41 additional visits per person, or 40 percent, over roughly 18 months, and produced no statistically significant change in measured blood pressure, cholesterol, or blood sugar control during the study period. For the Affordable Care Act debate, Oregon showed that coverage delivers large financial protection and mental health gains quickly, while physical health effects need longer follow up than a lottery study allows.

Q: Did people lose their health plans because of the Affordable Care Act?

Yes, in the fall of 2013 several million people received notices that their individual market plans would not be renewed for 2014 because the plans did not meet the law’s new benefit and consumer protection standards. The scale is genuinely disputed: the Associated Press, surveying state regulators, counted about 4.7 million notices across 30 states, while the Urban Institute estimated about 3.4 million cancellations nationwide. What is not disputed is the cause: insurers discontinued non grandfathered plans that failed the new requirements and offered compliant replacements. The episode became politically explosive because President Obama had repeatedly said that people who liked their plans could keep them. In November 2013 the administration announced a transitional policy allowing insurers to renew noncompliant plans, which many states adopted, and most of the people whose plans were cancelled transitioned to compliant coverage, often with subsidies. The episode illustrates a deliberate tradeoff: minimum standards that protected consumers from skimpy coverage necessarily meant some existing plans could not continue.

Q: How do expansion and non-expansion states compare on coverage gains?

Expansion states consistently posted larger coverage gains than states that declined the Medicaid expansion, and researchers have treated the difference as the cleanest available test of the law’s effect. The Urban Institute and the Robert Wood Johnson Foundation, analyzing 2013 to 2016, found that the uninsured rate in expansion states fell from 15.3 percent to 7.0 percent, a decline far larger than in non expansion states. The Census Bureau’s state level data show the same pattern: states that expanded saw uninsured rates fall further and faster after 2014. The comparison is not perfectly clean, because states that expanded also tended to run their own outreach and their own exchanges, which complicates isolating Medicaid alone. But the direction is consistent across designs, and the gap between the two groups of states is one of the least contested findings in the literature on the law. By the end of 2022, 38 states plus the District of Columbia had implemented the expansion.

Q: What mechanisms explain the drop in medical debt under the law?

The mechanism runs through Medicaid’s near total elimination of cost sharing for the poorest enrollees and through the law’s subsidies and out of pocket caps in the marketplaces. When people who previously paid cash for care gained coverage that paid providers directly, unpaid bills stopped accumulating. Luojia Hu, Robert Kaestner, Bhashkar Mazumder, Sarah Miller, and Ashley Wong, using credit bureau data in the Journal of Public Economics in 2018, found that the expansions reduced collection balances by about 1,140 dollars among those gaining coverage, concentrated in high uninsured neighborhoods. Kenneth Brevoort, Daniel Grodzicki, and Martin Hackmann, in National Bureau of Economic Research Working Paper No. 24002 from 2017, found about 3.4 billion dollars less unpaid medical debt sent to collection in the first two years of the expansions, alongside improved credit outcomes. The Oregon lottery study found the same pattern experimentally: winning Medicaid coverage reduced medical debt sent to collections and nearly eliminated catastrophic medical expenditures. Financial protection is the outcome where the evidence is most consistent across every research design used.

Q: Which Affordable Care Act findings are settled and which remain contested?

Close to settled: the law reduced the uninsured rate, Medicaid carried most of the coverage gain, expansion states gained more coverage than non expansion states, and coverage reduced medical debt and catastrophic out of pocket spending. The financial protection results replicate across randomized, quasi experimental, and survey designs, which is why economists treat them as the firmest ground. Genuinely contested: the size of the mortality effect, because estimates range from meaningful reductions in well powered studies to null results in shorter follow ups; the law’s effect on overall health care spending growth, where separating the statute from the broader slowdown is methodologically difficult; and whether the individual market’s early premium spikes reflected one time repricing or structural problems, a debate the market’s later stabilization partly resolved. The settled versus contested map matters because single outcome verdicts on the law almost always cherry pick one column.

Q: How did employer-sponsored insurance change after 2014?

Employer coverage, which insures roughly half of Americans, changed far less than the individual market did, which is exactly what the statute’s drafters intended. The law’s employer provisions were modest: firms with at least 50 full time equivalent workers faced penalties if they did not offer affordable coverage, and small business tax credits proved little used. The Kaiser Family Foundation’s annual Employer Health Benefits Survey showed premium growth continuing its pre 2014 trend of low to mid single digits per year, with average family premiums of 16,834 dollars in 2014, 19,616 dollars in 2018, and 22,463 dollars in 2022. The employer mandate’s enforcement was delayed twice before taking effect in 2015 and 2016, and researchers found little evidence of a shift toward part time work attributable to the threshold. The employer market’s stability is one reason national coverage gains came overwhelmingly from Medicaid and the marketplaces instead.

Q: How do researchers isolate the law’s effect from other trends?

They exploit variation in who got treated and when. The workhorse design compares states that expanded Medicaid with states that did not, before and after 2014, a difference in differences approach that strips out national trends affecting both groups. Sarah Miller, Norman Johnson, and Laura Wherry strengthened this by linking survey records to administrative mortality data in the Quarterly Journal of Economics in 2021, which allowed individual level follow up instead of relying on state aggregates. The Oregon lottery supplied the rare randomized design, though only for the pre 2014 Medicaid population in one state. Each design has a weakness: state comparisons assume the two groups would have trended together, administrative linkages depend on match quality, and Oregon’s results do not automatically generalize. Findings that survive across all these designs, like the financial protection results, earn the most confidence.

Q: What does the coverage gap mean for a family below the poverty line?

Consider a family of three earning about 18,000 dollars in a non expansion state, below the federal poverty line. In an expansion state, the parents would qualify for Medicaid because the law set eligibility at 138 percent of the poverty line. In their state, they earn too much for traditional Medicaid, whose thresholds for working parents in some non expansion states sat well below half the poverty line, yet too little for marketplace subsidies, which the statute starts at 100 percent of the poverty line. The Kaiser Family Foundation estimated about 2.2 million Americans were in this position across the 12 non expansion states in 2022. The family pays full price for any private plan or goes without, and studies of the gap population document higher rates of skipped care and medical debt. The gap is not an accident of implementation; it is what the statute’s arithmetic produces once expansion becomes optional.

Q: How much of the coverage gain came through Medicaid versus the exchanges?

Medicaid dominated, and the split is one of the most underappreciated facts about the law. The Urban Institute and the Robert Wood Johnson Foundation, analyzing 2013 to 2016, estimated that about 18.5 million people gained coverage and that the Medicaid expansion supplied a majority, about 10.9 million. Centers for Medicare and Medicaid Services enrollment data show total Medicaid and Children’s Health Insurance Program enrollment rising from about 61.2 million in December 2013 to 92,340,585 in December 2022, though the later figure partly reflects pandemic era continuous enrollment. Marketplace plan selections reached about 12.7 million during the 2016 open enrollment period and more than 14.5 million for 2022, with actual effectuated enrollment lower in both years. The Medicaid number is also the cleaner gain, since nearly all of it represents people who were previously uninsured, whereas some marketplace enrollees had prior coverage. The popular image of the law as an exchange centered program gets the proportions backwards; it was, in coverage terms, mainly a Medicaid expansion with an exchange attached.

Q: How did the young adult provision to age 26 change coverage?

The dependent coverage provision, which took effect in September 2010 and let young adults stay on a parent’s plan until age 26, was the law’s first coverage expansion and its cleanest success. The Department of Health and Human Services reported that 3.1 million young adults had gained coverage through the provision by December 2011. Sommers and coauthors, comparing 19 to 25 year olds with slightly older adults in the New England Journal of Medicine in 2013, found significant coverage gains concentrated in the targeted age group along with improved access to care. Because the provision applied nationwide regardless of state Medicaid decisions, it also provided a useful falsification test: coverage rose for young adults in both expansion and non expansion states, confirming that the provision, not state policy, drove the gain. Few provisions of the law have a less contested evaluation.

Q: Did emergency department use change after Medicaid expansion?

Yes, and the direction surprised many supporters of the law. The Oregon lottery study, reported by Taubman, Allen, Wright, Baicker, and Finkelstein in Science in 2014, found that winning Medicaid coverage increased emergency department visits by about 0.41 visits per person, or 40 percent, relative to the control group over roughly 18 months, including visits for conditions treatable in primary care. The result held across the study’s follow up period. Later quasi experimental studies of the 2014 expansion found more mixed patterns: some detected increases in emergency department use, others found shifts toward outpatient care over longer horizons. The methodological lesson is that insurance lowers the price of every kind of care, including the emergency department, so coverage alone does not redirect patients to cheaper settings without accompanying changes in primary care capacity. The finding is often cited by critics and misread; higher use under coverage is what economic theory predicts, not evidence that coverage failed.

Q: Did the law narrow racial gaps in health coverage?

Yes, measurably, though gaps remained large. Thomas Buchmueller, Zachary Levinson, Helen Levy, and Barbara Wolfe, writing in the American Journal of Public Health in 2016, used American Community Survey data from 2008 to 2014 and found that the share of uninsured adults fell by 7.1 percentage points for Hispanics, 5.1 percentage points for Blacks, and 3 percentage points for Whites after the main provisions took effect in 2014. Coverage gains were greater in states that expanded Medicaid. The mechanism was straightforward, since minority populations were disproportionately represented among the low income uninsured adults targeted by the Medicaid expansion and subsidies. The authors concluded that the law reduced racial and ethnic disparities in coverage, though substantial disparities remained. Census Bureau data showed the same pattern at the national level: the groups with the highest pre 2014 uninsured rates posted the largest percentage point declines. Coverage is only one input to health equity, but the disparity reduction is among the better documented distributional results.

Q: Did the zeroed mandate penalty in 2019 change enrollment?

The 2017 tax legislation reduced the individual mandate penalty to zero effective in 2019, which gave researchers a natural test of how much the mandate had been holding enrollment up. The Congressional Budget Office had projected in November 2017 that repealing the penalty would leave 13 million more people uninsured by 2027. What actually happened was milder: marketplace enrollment dipped only slightly in 2019 and 2020, and the national uninsured rate rose modestly from 8.8 percent in 2016, in the Current Population Survey, to 9.2 percent in 2019, in the American Community Survey, before falling to a record low of 7.9 percent in 2022 as enlarged subsidies took effect. The smaller than projected response is contested in its interpretation: some economists argue the mandate was always less binding than modeled, while others note that the subsidies enacted in 2021 confound any clean read of the post 2019 period. Either way, the enrollment collapse some predicted did not occur.

Q: Did Medicaid expansion slow rural hospital closures?

This is one of the more concrete downstream effects researchers have tested. Kaufman and colleagues, in a Health Affairs study published in 2017, found that the 2014 expansion was associated with larger Medicaid revenue gains for rural hospitals than for urban ones, while declines in uncompensated care costs as a share of operating costs were larger for urban hospitals. A Health Affairs study of Louisiana’s 2016 expansion, summarized by the Center on Budget and Policy Priorities, estimated that expansion was associated with a 33 percent reduction in uncompensated care costs as a share of total operating expenses, with a 55 percent reduction for rural hospitals, as the state’s nonelderly uninsured rate fell from 18.3 percent in 2015 to 11.8 percent in 2018. The mechanism is direct: rural hospitals carry high shares of uninsured patients, and expansion turned much of that uncompensated care into reimbursement. The finding is quasi experimental rather than randomized, so it carries the usual caveats about state differences, and closures continued in both groups of states. But among the law’s second order effects, the hospital finance channel has some of the clearest empirical support.