THE IMPACT OF ELIGIBILITY FOR MEDICAID VERSUS SUBSIDIZED PRIVATE HEALTH INSURANCE ON MEDICAL SPENDING, SELF-REPORTED HEALTH, AND PUBLIC PROGRAM PARTICIPATION

Abstract

We use a regression discontinuity design to understand the impact of a sharp change in eligibility for Medicaid versus subsidized marketplace insurance at 138 percent of the federal poverty line on coverage, medical spending, health status, and other public program participation. We find a 5.5 percentage point shift from Medicaid to private insurance, with no net change in coverage. The shift increases individual health spending by $341 or 2 percent of income, with larger increases at higher points in the spending distribution. Two-thirds of the increase is from premiums and one-thirdfrom out-of-pocket medical spending. Self-rated health and other public program participation appear unchanged. We find no evidence of bunching below the eligibility threshold, which suggests either that individuals are willing to pay more for private insurance or that optimization frictions are high.

I. Introduction

Not all health insurance is created equal. While Medicaid may not, by law, require substantial premiums or cost-sharing from beneficiaries, private insurance plans may impose substantial costs on enrollees through premiums or cost-sharing for health care.¹ Premium subsidies for marketplace plans under the Affordable Care Act (ACA) were not designed to fully cover the cost of coverage; low-income enrollees can expect to contribute as much as 3 percent of their income toward premiums—or more, if they choose an expensive plan. Moreover, the limits on cost-sharing for health care in marketplace plans can be quite high; in 2016, caps on enrollees’ out-of-pocket health-care costs were $6,850 for an individual and $13,700 for a family. Thus, even subsidized marketplace plans can leave low-income households exposed to considerable out-of-pocket spending.

At the same time, marketplace and other private non-group plans may afford enrollees better access to care than does Medicaid. Medicaid reimbursement for providers is typically far lower than what is offered by private plans (Zuckerman, Skopec, and Epstein 2017; Lopez et al. 2020),² and administrative burdens for providers are higher (Dunn et al. 2021). As a result, Medicaid patients experience more difficulty than privately insured patients getting appointments and face longer waiting times (Asplin et al. 2005; Rhodes et al. 2014; Polsky et al. 2015; Oostrom, Einav, and Finkelstein 2017; Candon et al. 2018; Alexander and Schnell 2019). These barriers to access may, in turn, lead to worse health outcomes.

These factors suggest there may be a cost/quality trade-off between Medicaid and marketplace coverage. In this work, we evaluate several aspects of this trade-off. Specifically, we estimate how having private non-group coverage rather than Medicaid affects individual spending for health insurance and medical care, self-reported health status, and use of other public programs among low-income families using a regression discontinuity design (RDD). Our RDD exploits the income threshold for Medicaid eligibility versus subsidies for marketplace coverage in states that expanded Medicaid under the ACA. In effect, we compare individuals whose incomes are just below and just above 138 percent of the federal poverty line (FPL) threshold for Medicaid versus premium tax credits for marketplace coverage. We use detailed income and spending data from the Current Population Survey (CPS) to estimatechanges in overallcoverage from any source as well asspecific types of coverage at this threshold. We find that Medicaid coverage decreases by about 5.5 percentage points at this threshold and that this decline is largely offset by an increase in private non-group coverage. We find no change in overall coverage at the threshold, which is consistent with evidence from tax data (see Appendix Figure E.1 in Lurie, Sacks, and Heim 2021).³ As a result, we can use the RDD framework to estimate the impact of coverage source (Medicaid versus private non-group coverage) on both the average and the distribution of health insurance premiums and out-of-pocket health-care spending, together referred to as total personal health spending. We find fairly large effects of crossing the 138 percent income threshold on personal health spending for an individual: $341 in real 2016 dollars, or about 2 percent of income at the eligibility cutoff. Most of this effect comes from higher premiums. Effects at higher points of the distribution are even larger and also come from higher premium spending and, to a lesser extent, higher out-of-pocket spending. For example, we find a $1,054 increase in total personal health spending at the 90th percentile ($697 for premium and $374 for out-of-pocket spending), or 6 percent of income at the cutoff. These are reduced-form effects. Scaled by the change in coverage, total individual medical spending increases by between $5,600 and $6,200 at the mean. These findings suggest that, for low-income individuals, Medicaid does more to reduce the risk of exposure to high levels of personal health spending than private coverage.

Ideally, in order to fully evaluate the trade-offs between marketplace coverage and Medicaid, we would analyze changes in access to care at the 138 percent income cutoff; however, we are precluded from doing so by data limitations. There are no available data that both measure access to care and have a sufficiently large sample of detailed income data to support our RDD design.⁴ Instead, we look for impacts of coverage type on health, testing for discontinuities in self-rated health and work limitations at the 138 percent income threshold. The existing literature suggests that self-rated health may change relatively quickly in response to gaining coverage; Card, Dobkin, and Maestas (2004) document significant improvements in self-rated health at age 65, the age of near-universal eligibility for Medicare, with these improvements concentrated among individuals with low levels of education. The Oregon Health Insurance Experiment found that individuals randomized to Medicaid eligibility had higher self-rated health within the first year of the program than those who were not (Finkelstein et al. 2012). Our “natural experiment” is somewhat different in that we are not estimating the effect of gaining coverage (the extensive margin) but rather the effect of the type of coverage (the intensive margin): Medicaid versus private non-group. This may explain why, in contrast with these existing studies, we find no systematic effects of coverage type on health outcomes. Specifically, above the 138 percent income threshold, we find small, imprecisely estimated improvements in self-reported health as well as marginally significant increases in the probability of reporting a disability that limits work.

Finally, we assess whether receipt of public benefits such as food stamps (Supplemental Nutrition Assistance Program [SNAP]) and public housing is higher when individuals have Medicaid rather than private non-group insurance. While economists and others have long speculated that enrolling in a public program may facilitate enrolling in others, perhaps because of increased familiarity with what is needed to sign up, evidence of this phenomenon has been very limited (see, for example, Yelowitz [1996]). In the Oregon Health Insurance Experiment, Medicaid eligibility increased SNAP receipt but not other public program participation (Baicker et al. 2014).⁵ Exploiting differences in Medicaid eligibility across state borders, Schmidt, Shore-Sheppard, and Watson (2019) find that the ACA Medicaid expansions increased participation in SNAP and Temporary Assistance for Needy Families (TANF) and possibly the Earned Income Tax Credit (EITC). Noting again that our study design estimates a different parameter, we find no significant effect of coverage type (Medicaid versus private non-group) on receipt of SNAP/food stamps, the Special Supplemental Nutrition Program for Women, Infants, and Children (WIC) program, public housing, or energy assistance.

One puzzle that these results present is why, given the large estimated financial benefit to households of having Medicaid rather than private non-group coverage, we do not observe bunching at the 138 percent threshold. This lack of bunching, which has also been shown in tax data (see Lurie, Sacks, and Heim [2021], Appendix Figure E.1), suggests either that individuals are largely indifferent between Medicaid and private non-group coverage or that optimization frictions are high. Given existing evidence that consumers are sensitive to premiums in Medicaid (Dague 2014; Cliff et al. 2022) and on the exchanges (Abraham et al. 2017; Drake and Anderson 2020), indifference in the face of higher premium spending implies that private insurance offers some other benefit, real or perceived, to near-poor households. The most likely explanation may be that private insurance affords better access to health care, but as already noted, data limitations meant that we were unable to test this hypothesis using the RDD framework.

II. Background

Federal rules do not allow states to require premium contributions from Medicaid enrollees below 150 percent of poverty and permit only minimal point-of-service cost-sharing (Brooks, Roygardner, and Artiga 2019). Even in the 10 states that have obtained waivers to expand use of these measures, the amounts are minimal: for example, premium contributions of $10 per month and $8 copayments for nonemergency visits to an emergency department (Medicaid and CHIP Payment and Access Commission 2019). Medicaid therefore offers comprehensive coverage that is essentially free for enrollees. In contrast, marketplace plans can be quite expensive for enrollees. While some low-income enrollees in ACA marketplace plans have the option of a “zero premium” plan, the ACA subsidy schedule does not ensure this; rather, the subsidies are designed to limit annual premium costs for enrollees just above the Medicaid eligibility threshold to 3 percent of their income, or about $1,000 for a family of four in 2020. And while ACA cost-sharing subsidies reduce point-of-service costs for enrollees with income below 250 percent of poverty, enrollees can still expect to pay about 10 percent of the cost of their care out-of-pocket, with very high caps on total out-of-pocket spending: $6,850 for an individual and $13,700 for a family in 2016.⁶ As a result, a small increase in family income—from just below to just above 138 percent of the poverty level—creates a substantial notch in what households can expect to spend on health insurance and medical care.

Our analysis complements recent studies by Selden, Lipton, and Decker (2017) and Blavin et al. (2018). Selden, Lipton, and Decker (2017) analyze outcomes for individuals with family incomes between 100 and 138 percent of the FPL, comparing their coverage, access, and spending outcomes in expansion versus non-expansion states before versus after 2014. In non-expansion states, these individuals were eligible for marketplace premium tax credits but not Medicaid. Coverage increased for both groups, but those in expansion states had larger reductions in out-of-pocket medical spending and smaller reductions in reported difficulty accessing care after 2014 than those in non-expansion states. Blavin et al. (2018) take a similar analytic approach, using different data, and also find increases in coverage and reductions in out-of-pocket spending for individuals living between 100 and 138 percent of the FPL in expansion versus non-expansion states. Similar findings emerge from comparisons of individuals in Arkansas, which had a private-insurance-based Medicaid expansion waiver; Kentucky, which had a traditional Medicaid expansion; and Texas, which did not expand Medicaid (Sommers, Blendon, and Orav 2016).

The primary difference between our work and Selden, Lipton, and Decker (2017) or Blavin et al. (2018) is our identification strategy. While both of those papers use a difference-in-differences (DID) strategy to compare individuals between 100 and 138 percent of the FPL in expansion versus non-expansion states, we use an RDD to compare outcomes for those just above and just below the 138 percent of FPL cutoff for Medicaid within expansion states. Thus, we compare those who can sign up for the state’s Medicaid program with those who can access premium tax credits to purchase a private health plan on the state exchange. Methodologically, the main advantage of the RDD relative to the DID is that it is less likely to be confounded by unobserved differences either within or across states, an issue of particular importance here, given the many health-care market changes occurring as a result of both the ACA and the economic recovery. An additional advantage of our identification strategy compared with DID is that it does not require pre-ACA data. Therefore, the revision of the CPS health insurance questions in 2013, which caused a break in trend, does not affect our analysis, nor does the redesign of the income questions in 2014.⁷ Substantively, our approach means that we are estimating a different parameter from papers that rely on DID approaches. The DID analysis estimates the overall impact of expanding Medicaid eligibility, which will combine the effects of (1) an overall increase in coverage, the primary impact of the expansion, and (2) some substitution of Medicaid for private coverage. In contrast, because we find no change in the fraction uninsured at the 138 percent FPL threshold (documented below), our RDD approach focuses on only the second effect: the shift from Medicaid to private non-group coverage. By quantifying the costs and some potential benefits of having private non-group coverage, rather than Medicaid, we shed light on the current policy debate over expanding eligibility for premium tax credits to adults with incomes below the poverty level in states that have not expanded Medicaid (Holahan et al. 2021). Poor adults in the 12 states that have not expanded Medicaid currently have access to neither Medicaid nor premium tax credits; premium tax credits would help them access private non-group coverage, but what trade-offs, relative to having Medicaid, does this entail for beneficiaries? Our work illuminates these trade-offs.

Our work is also related Hinde (2017), which uses 2014 CPS data and an RDD framework to investigate the effect of ACA tax credits and cost-sharing reductions at different income levels (including at 138 percent of the FPL) on private insurance coverage. Similar to our results, that paper finds a 5.4 percentage point increase in private insurance and no change in uninsurance at 138 percent of the FPL. Our contributions beyond what is reported in Hinde (2017) are a much more detailed analysis of how the shift in coverage induced by the discontinuity at 138 percent of the FPL affects the distribution of both premium spending and out-of-pocket health-care spending, as well as an analysis of the effect of this discontinuity on other outcomes, including health and other public program participation.⁸

One drawback of the RDD is its need for large samples of high-quality income data, which limits the range of outcomes we can examine. An important potential advantage of marketplace coverage, which may offset its higher cost to enrollees, is better access to care than Medicaid provides (Asplin et al. 2005; Rhodes et al. 2014; Polsky et al. 2015). In particular, because many state Medicaid programs reimburse providers at rates far below those paid by Medicare or commercial insurers, providers appear to ration slots for Medicaid enrollees more tightly than for privately insured individuals (Polsky et al. 2015). Ideally, like Selden, Lipton, and Decker (2017) and Allen et al. (2021), we would analyze the effect of insurance type on measures of access to care and health-care utilization. However, we were unable to find any data sets that included a large enough sample with both detailed income information and questions on access to medical care and/or health-care utilization.

III. Data

We use data from the CurrentPopulation Survey’s Annual Social and Economic Supplement (ASEC or “March CPS”) for survey years 2010 through 2017. Here, we provide detail on the key CPS ASEC measures of income, health insurance, health, and other outcomes and covariates used in our analysis, followed by a description of our analytic sample.

A. INCOME

Medicaid eligibility is determined based on modified adjusted gross income (MAGI) relative to the federalpoverty threshold. MAGI is nearly identical toadjusted gross income (AGI) but includes several income categories excluded from AGI, such as tax-exempt interest, foreign income, and the nontaxable portion of Social Security benefits. In practice, the difference between AGI and MAGI is small for low-income families, particularly those headed by working age adults who are not yet eligible for Social Security benefits (Hinde 2017; Brooks 2015). Therefore, we rely on AGI to determine eligibility. The Census Bureau supplements the CPS ASEC with a simulated measure of AGI (O’Hara 2004).⁹ In all cases, we divide AGI by the relevant poverty threshold from the Census Bureau for each family size and year to calculate income relative to poverty. Note that the reference period for income questions in the March CPS is the prior calendar year; for example, the 2014 March ASEC has income information for calendar year 2013. In reporting results, the years correspond to the reference period for the data rather than the year in which the survey was conducted.

B. HEALTH INSURANCE

Our analysis is predicated on the existence of a discontinuity in coverage type at the 138 percent FPL threshold: that is, a “first stage” showing a shift from Medicaid to private non-group coverage. The CPS includes questions about different types of health insurance (e.g. Medicaid, Medicare, employer-sponsored, private non-group) at the time of the survey in March and also about coverage held over the course of the prior calendar year. For consistency with the detailed income data, we construct binary indicators for different types of coverage held at any time during the prior calendar year. Following Hinde (2017), these indicators are mutually exclusive in the following hierarchical order: employer-sponsored coverage > private non-group coverage > public coverage > uninsured. So, for example, our Medicaid indicator is 1 for an individual who reports having had Medicaid coverage at any point in the calendar year prior to the survey and also had no coverage from another source. Our uninsured variable is equal to 1 if the individual had no insurance at any time in the calendar year prior to the survey. As described in the robustness checks below, we show that our results are robust to using health insurance categories that are not mutually exclusive (i.e., such that an individual is coded as having private coverage and Medicaid during the prior year) and, separately, robust to reversing the hierarchy of mutually exclusive categories (i.e., such that public coverage > private non-group coverage > employer-sponsored coverage > uninsured). Note that our measure of private non-group health insurance in the CPS includes coverage that may be purchased independently of the ACA-sponsored health insurance marketplaces, which would not qualify for ACA tax credits; however, for low-income families, the great majority of non-group private coverage is purchased through ACA marketplaces (Fiedler 2021).¹⁰ In some analyses, we add employer-sponsored coverage and private non-group coverage for a measure of total private insurance. We also present analyses that use CPS data on whether working individuals without employer-sponsored coverage were eligible for health insurance offered by their own employer.

C. HEALTH SPENDING

The CPS ASEC began collecting information on health spending in 2014. Respondents are asked to report the total amount spent in the prior calendar year on three types of goods: (1) health insurance premiums; (2) medical care, defined as copayments for doctor and dentist visits, diagnostic tests, prescription medicine, glasses and contacts, and medical supplies; and (3) non-prescription or “over-the-counter” health-care products such as vitamins, allergy and cold medicine, pain relievers, or smoking cessation aids. We adopt terminology similar to the Medical Expenditure Panel Study and refer to (2) medical care spending as out-of-pocket health spending. We refer to the sum of premium spending and out-of-pocket health spending as total personal health spending. Note that our outcome is individual rather than household-level spending because our insurance measures are at the individual level. We adjust reported spending using the medical component of the consumer price index and analyze amounts in real 2016 dollars.

D. OTHER OUTCOMES

Additional outcomes that we analyze include whether there is any household-level use of food stamps/SNAP, the WIC program, public housing, the federal school lunch program, or energy assistance. Like the measures of income, health insurance coverage, and health spending, these measures refer to the entire calendar year prior to the survey. In addition, we analyze responses to the following question: “Would you say (name’s/your) health in general is excellent, very good, good, fair, or poor?”¹¹ If access to care and/or quality of care is higher for those with private insurance, this could result in improvements in health. Although subject to limitations (Crossley and Kennedy 2002), self-rated health is a remarkably reliable indicator of both individual mortality and morbidity (e.g., see Idler and Kasl 1995; McCallum, Shadbolt, and Wang 1994; Okun and George 1984). In addition to analyzing the self-rated health coded on a scale of 1 (poor) to 5 (excellent), we also analyze an indicator of whether individuals rate their health as very good or excellent. Finally, we analyze whether the individual has a disability that limits their ability to work. These measures—self-rated health status and work-limiting disability—are asked with reference to the time of the survey, rather than the prior calendar year. As a result, there is some potential for mismatch between these outcomes and the income measures. To the extent that health represents a stock of human capital that adjusts slowly to changes in insurance coverage, the several-month lag between coverage and health measures may be an asset rather than a limitation.

E. COVARIATES

In order to confirm the validity of our RDD design, we test the smoothness through the 138 percent FPL threshold of other individual and household characteristics available in the CPS ASEC including gender, race/ethnicity, marital status, home ownership, urban/suburban location, nativity (born in US/not), citizenship, and age. In addition, we consider covariates that arguably may respond to the differential incentives created by the notch at 138 percent of the FPL—employment last week and employment in the prior year. Some of these measures (home ownership, urban/rural location, and marital status) vary over time and are reported at the time of the survey, so there is the potential for mismatch with our income, health insurance, and health spending measures.

F. DEFINING THE ANALYTIC SAMPLE

Our analytic sample is restricted to individuals between the ages of 27 and 64 at the time of the survey and with household income between 88 and 183 percent of the FPL: plus or minus 50 percentage points on either side of the 138 percent FPL cutoff.¹² The 50 percentage point bandwidth was chosen based on Calonico, Cattaneo, and Titiunik’s (2014) optimal bandwidth procedure; below, we demonstrate that the results are not sensitive to a range of alternative bandwidths. The age restriction accounts for access to parental coverage for adults below age 26 and for near-universal access to Medicare for those ages 65 and older. We further restrict the sample to individuals who do not have Supplemental Security Income (SSI) since these individuals are categorically eligible for Medicaid in most states.¹³

Our main analysis focuses on data from states that expanded Medicaid once those expansions and the ACA premium tax credits were in effect, so that the 138 percent FPL threshold separates Medicaid eligibility from premium tax credits. This generally means that our analytic sample contains observations in Medicaid expansion states for 2014 through 2016 (from the CPS conducted in March 2015, 2016, and 2017). States that had Medicaid eligibility thresholds higher than 138 percent of the FPL or that expanded after 2016 are excluded (see Online Appendix Table A1 for details). Within expansion states, we present results for all adults ages 27 through 64 who are within the income band defined above and also for the subset who do not have access to employer-sponsored health insurance (ESI) through either their own job or a spouse’s, since for nearly everyone in this group, Medicaid and private non-group coverage represent the main coverage options available to them.¹⁴

In supplementary analyses, we present results from two placebo groups for which we do not expect to observe discontinuities in outcomes at 138 percent of the FPL: (1) non-expansion states in 2014 through 2016 and (2) both expansion and non-expansion states in 2009 through 2012 (from the CPS conducted in March 2010 through 2013). If we found discontinuities for these groups, it would be a sign that something is wrong with our analytic approach.

IV. Empirical Strategy: Regression Discontinuity Design

To understand the impact of health insurance type on spending, we would, in principle, estimate the following reduced-form equation:

$$Y_i=\alpha +\beta InsTyp{e}_i+X_i\delta +{\epsilon }_i$$ (1)

where $Y_i$ is an outcome measure such as out-of-pocket spending for individual $i$; $X_i$ is a set of demographic characteristics of individual $i$; $InsType$ is the type of health insurance individual $i$ has (e.g., Medicaid or private insurance); and ${\epsilon }_i$ is an unobserved error. A fundamental and well-known problem in interpreting $\beta$ as the causal effect of health insurance type on outcomes is that coverage is endogenous; it both affects and is affected by medical spending and other outcomes of interest, confounding observational comparisons of people by insurance type.

To circumvent this problem, we exploit a discontinuity in eligibility for ACA health insurance programs based on family income. The ACA introduced both an optional expansion of Medicaid eligibility to all adults with incomes up to 138 percent of the FPL, as well as tax credits for the purchase of private insurance through health insurance marketplaces for individuals above the Medicaid threshold but below 400 percent of the FPL. As a result, beginning in 2014, adults in Medicaid expansion states faced a sharp discontinuity in their eligibility for different programs, with those with family income just below 138 percent of the FPL eligible for Medicaid—which, as explained above, imposes minimal costs on beneficiaries—while those with slightly higher incomes qualified instead for tax credits to subsidize private coverage, which might still entail substantial cost to them. If there is no overall change in coverage, as we document in the analysis, then this discontinuity allows us to assess the relative costs/benefits of Medicaid versus subsidized private insurance for low-income individuals.

Formally, we estimate a first stage as follows:

$$\begin{gathered} InsTyp{e}_i={\beta }_0+{\beta }_{1}1\left(\text{INC}:{\text{FPL}}_i>1.38\right)+g\left(\text{INC}:{\text{FPL}}_i\right) \\ +g\left(\text{INC}:{\text{FPL}}_i\right)\times 1\left(\text{INC}:{\text{FPL}}_i>1.38\right)+{\epsilon }_i \end{gathered}$$ (2)

where INC:FPL stands for income relative to the federal poverty line such that ${\text{FPL}}_i>1.38$ is an indicator of whether household income is above 138 percent of the FPL.¹⁵ The function $g(\cdot )$ captures health insurance gradients with respect to INC:FPL that are allowed to vary on both sides of the 138 percent FPL eligibility threshold. In addition to insurance type, we also analyze uninsurance.

Combining equations 2 and 1, the resulting reduced-form model for outcome $y_i$ is

$$\begin{gathered} y_i={\gamma }_0+{\gamma }_{1}1\left(\text{INC}:{\text{FPL}}_i>1.38\right)+f\left(\text{INC}:{\text{FPL}}_i\right) \\ +f\left(\text{INC}:{\text{FPL}}_i\right)\times 1\left(\text{INC}:{\text{FPL}}_i>1.38\right)+v_i \end{gathered}$$ (3)

where $y$ is a measure of medical spending, health, or program participation for individual $i$ and all other variables are defined as above. Our main estimate of interest, ${\gamma }_1$, gives the causal effect of relative access to Medicaid versus subsidized marketplace insurance on spending and health.

In our model, income relative to the federal poverty line ($\text{INC}:{\text{FPL}}_i$) is the “running” or “forcing” variable for the analysis; that is, the variable that, as a result of policy, discontinuously alters the likelihood of Medicaid versus private non-group insurance coverage. In our graphical presentations, we aggregate income into 5 percentage point bins for ease of interpretation. The means of the outcome of interest in each bin are displayed as open circles, along with lines for the parametric fits from our RDD models. While income relative to FPL is binned for the purposes of visual presentation, all regressions use a continuous measure of the running variable. In addition, although the parametric fits exclude data within +/−2 percentage points of 138 percent of the FPL, the means are based on all data from the analytic sample

There are two primary approaches to estimating the RDD: a parametric global polynomial approach and a nonparametric local linear approach. Our main specification takes a local linear approach using an optimal bandwidth around the 138 percent FPL eligibility threshold based on a standard algorithm from Calonico, Cattaneo, and Titiunik (2014). Although in principle the algorithm should be used to generate unique bandwidths for each outcome, we use a uniform bandwidth across all of our outcomes for comparability. The optimal bandwidths varied from 42 to 53 points, so we chose to fix it at 50 percentage points. In sensitivity analyses, we show that the estimates are not sensitive to the choice of bandwidth over a very large range of income relative to the FPL. In addition, because individuals on either side of the 138 percent FPL threshold may be miscategorized, which should bias towards zero any estimated difference in outcomes between the Medicaid eligible versus subsidized private insurance eligible populations, we remove from our main regression analysis individuals within +/−2 percentage points of 138 percent of the FPL. In sensitivity analyses, we show that our main results for coverage and spending are not sensitive to including all the data or excluding smaller or larger “donuts.” Because our running variable, income relative to the FPL, is measured continuously, we report robust standard errors. The standard errors are almost identical when clustered by income (available upon request) to account for possible misspecification (Lee and Card 2008).

Analyses of different points in the distribution of spending—for example, spending at the median, 75th, 90th, and 95th percentile—are estimated using a quantile regression version of equation 3. We bootstrap the standard errors for the quantile regressions, calculating the standard errors as the standard deviation of the coefficient estimates from 500 bootstrap samples drawn with replacement. Inference is quite similar if we use an income-based block bootstrap that randomly samples with replacement the data within each percentile of income relative to the FPL and estimates the models on these random samples (Efron and Tibshirani 1994).

A. TESTING THE ASSUMPTIONS UNDERLYING THE RDD

A key identifying assumption of our RDD is that the unobserved determinants of coverage, spending and health—that is, the elements of the error term $v_i$—are a smooth function of income. We partially test this assumption by running specifications similar to equation 3 with predetermined or plausibly unaffected characteristics, such as gender, race/ethnicity, education, and place of birth as the dependent variable.

Online Appendix Figure A1 demonstrates smoothness in several covariates: marital status, citizenship, age, household size, home ownership (including with a mortgage), and living in the suburbs. We also show the fraction employed last year and last week, conditional on working. Online Appendix Table A2 shows the RDD estimates for these and several other covariates, including the fraction female, fraction white, fraction black, fraction Hispanic, fraction living in the principal city in the metropolitan statistical area, fraction foreign-born, and fraction employed last week. For each covariate, we show the RDD estimates using Calonico, Cattaneo, and Titiunik’s (2014) optimal bandwidth of 50. In most cases, the point estimates are small and they are rarely statistically distinguishable from zero. As expected given how many covariates we consider, we do find significant discontinuities in a few instances. The fraction white increases by 4.3 percentage points off a base of about 44 percent, and age increases by 1.4 years off a base of 43.5 years. In sensitivity analyses, we control for these covariates to test the robustness of our results to their inclusion. Overall, however, the general smoothness in our data suggests that our analysis satisfies the continuity assumption of the RDD design.

Another potential threat to the validity of the RDD design is that individuals near the 138 percent FPL cutoff may manipulate their income based on their preferences for either Medicaid or exchange subsidies; for example, individuals may limit their work hours so as not to earn too much to qualify for Medicaid. In this case, the RDD design would fail to provide a valid comparison between individuals who are identical except for their access to Medicaid versus tax credits on either side of the income threshold. In order to test for this possibility, we look for heaping in the distribution of observations in the CPS just above versus below 138 percent of the FPL. Using CPS data on individuals between 88 and 158 percent of the FPL, we find no evidence of such heaping, as shown in Online Appendix Figure A2a (p > 0:901 for the null of no discontinuity).¹⁶ This is consistent with the results reported by Hinde (2017) and supports the validity of the regression discontinuity design.¹⁷

The absence of bunching at this threshold is consistent with the possibilities that (1) individuals are largely indifferent between Medicaid and subsidized private health insurance coverage at this threshold, (2) optimization frictions make it difficult for individuals to strategically change their income to qualify for the preferred option, or (3) some individuals manipulate their income to be above the threshold to qualify for subsidies and an equal number of others do so to be below it to qualify for Medicaid.¹⁸ Online Appendix Figure A2b shows the result of the manipulation test from Cattaneo, Jansson, and Ma (2018), which also shows no evidence of manipulation (p > 0:538) and lends further support to the research design.¹⁹

V. Results

A. INSURANCE COVERAGE AND MEDICAID VERSUS SUBSIDY ELIGIBILITY

Figure 1 shows the fraction of adults who have Medicaid (panel A) and private non-group coverage (panel B) by household income relative to the FPL in expansion states in 2014 through 2016. Panel A shows that, as expected, Medicaid coverage decreases with relative income, with a sharp drop at 138 percent of the FPL. Panel B of Figure 1 also shows a complementary discontinuous increase in private non-group coverage at 138 percent of the FPL. Table 1 confirms these results, reporting estimates of ${\gamma }_1$ from equation 3 above. The drop in Medicaid at 138 percent of the FPL is 5.5 percentage points, or about 18 percent off a base rate of 30.4 percent of individuals below 138 percent of the FPL on Medicaid.^20,21 Private non-group coverage increased by 3.6 percentage points at the 138 percent FPL cutoff or about 27 percent off a base of 13.8 percent below 138 percent of the FPL with private non-group coverage. We also estimate an (imprecise) increase in reports of employer-sponsored insurance (ESI) at 138 percent of the FPL of about 1.3 percentage points with a total increase in private coverage (the sum of private non-group and ESI) of nearly 5 percentage points (see column 4 of Table 1 and Online Appendix Figure A3). Whether the increase in ESI is due to misreports or a true increase in take-up (in response to the mandate, for example) is unclear. Figure 2 shows the small break in the rate of employer-sponsored coverage (panel A) but no change in uninsurance (panel B) at the 138 percent FPL cutoff; results in columns 3 and 5 of Table 1 confirm this result.

FIGURE 1.

Discontinuity in Medicaid and private coverage in expansion states. Data are from the March CPS for reference years 2014–16. Open circles represent the coverage rate for individuals ages 27 to 64 by 5 percentage point bins of household income relative to the federal poverty line (FPL). The dashed vertical line is 138 percent of the FPL, where coverage policies change in Medicaid expansion states. The solid lines are parametric fits from an RDD model that allows profiles of household income relative to the FPL to vary on either side of the 138 percent FPL threshold.

TABLE 1.

Relationship between income threshold and insurance type, 2014–16: Adults ages 27–64 in expansion states

	Medicaid (1)	Private non-group (2)	ESI (3)	Private coverage (4)	Uninsured (5)
	A. Full expansion sample
Above 138% FPL	−0.0546a	0.0361a	0.0128	0.0489b	0.0036
	(0.0162)	(0.0136)	(0.0185)	(0.0191)	(0.0154)
Observations	16,582	16,582	16,582	16,582	16,582
Mean below 138% FPL	0.304	0.138	0.308	0.446	0.204
R 2	0.025	0.001	0.024	0.026	0.001
		B. Expansion sample without ESI access
Above 138% FPL	−0.0689a	0.0792a	N/A	0.0792a	−0.0045
	(0.0266)	(0.0236)		(0.0236)	(0.0261)
Observations	8,017	8,017		8,017	8,017
Mean below 138% FPL	0.428	0.199		0.199	0.303
R 2	0.013	0.010		0.010	0.001

Note: Data are from the March CPS for reference years 2014–16 and include individuals ages 27 to 64. We exclude Massachusetts, the District of Columbia in all years, Minnesota from 2015 on, and New York from 2016 on because these states had Medicaid cutoffs substantially above 138 percent of the federal poverty line (FPL). All regressions include a constant and an indicator for above 138 percent of the FPL and a linear trend in income relative to the FPL that is allowed to vary on either side of 138. All insurance categories are defined to be mutually exclusive. Robust standard errors are reported in parentheses. ESI stands for employer-sponsored insurance.

^asignificant at the 1 percent level

^bsignificant at the 5 percent level.

FIGURE 2.

Discontinuity in ESI coverage and uninsurance in expansion states. Data are from the March CPS for reference years 2014–16. Open circles represent the coverage rate for individuals ages 27 to 64 by 5 percentage point bins of household income relative to the federal poverty line (FPL). ESI stands for employer-sponsored insurance. The dashed vertical line is 138 percent of the FPL, where coverage policies change in Medicaid expansion states. The solid lines are parametric fits from an RDD model that allows profiles of household income relative to the FPL to vary on either side of the 138 percent FPL threshold.

We repeat this analysis for the subset of CPS respondents who do not have access to ESI coverage.²² This sample is likely to be particularly responsive to an expansion of Medicaid eligibility or the offer of subsidized private insurance. Those with access to “affordable” ESI are not even eligible for subsidies, where “affordable” means that the required employee premium contribution for single coverage is less than approximately 10 percent of income. Thus, at the 138 percent FPL threshold, individuals without access to ESI have few options other than Medicaid or private non-group coverage.

Before repeating the analysis, we first confirm that the rate of offer of employer-sponsored coverage is smooth through the 138 percent FPL threshold (Online Appendix Figure A4). This is consistent with other research showing very little impact of the ACA on employer provision of health insurance (Abraham, Royalty, and Drake 2016, 2019). Having confirmed this fact, panel B of Table 1 presents results for the subsample of individuals in expansion states who do not have access to employer coverage. Indeed, the discontinuous decline in Medicaid coverage is larger for this group than for the sample as a whole. This group experiences a 6.9 percentage point decline in Medicaid at 138 percent of the FPL (off a base of 43 percent), compared with the 5.5 percentage point decline for the full sample shown in panel A of Table 1. The decline in Medicaid at 138 percent of the FPL among those without access to employer coverage is slightly more than offset by a 7.9 percentage point increase in private non-group coverage (panel B). A slight decrease in other public coverage (not shown) means that, as in the full sample, uninsurance is unchanged at the 138 percent FPL threshold.

The results in Table 1 are similar to those reported by Hinde (2017). Taken together, Table 1 and Figures 1 and 2 suggest that, although those above and below the 138 percent FPL cutoff in expansion states are equally likely to have any insurance, those below the cutoff are more likely to have Medicaid while those above the cutoff are more likely to have private insurance. Consequently, we can attribute any discrete changes in health spending, self-reported health, and take-up of other public programs at the 138 percent FPL cutoff to the difference in the source of coverage induced by changes in eligibility for Medicaid versus subsidies to purchase private health insurance coverage.

B. INDIVIDUAL HEALTH SPENDING

Next, we analyze changes in individual health spending at the 138 percent FPL threshold in 2014 through 2016. Panel A of Figure 3 and column 1 of Table 2 show significant increases at the 138 percent FPL threshold in total personal medical spending (that is, the sum of the individual’s spending on premiums and out-of-pocket medical care). The average increase is $341, or about a quarter of average total spending by Medicaid beneficiaries with incomes below 138 percent of the FPL. This increase is driven more by premium spending (see panel B of Figure 3, and column 2 of Table 2), which increased by about $233, than by out-of-pocket medical care spending (see panel C of Figure 3, and column 3 of Table 2, which increased by about $100. Spending on over-the-counter health products was largely unchanged (column 4 of Table 2). The estimates for the subsample without access to ESI are quite similar in absolute terms, although larger in a relative sense (see panel B of Table 2 and Online Appendix Figure A6). Total personal spending increases by about $386 or about a 40 percent of average total personal health spending by Medicaid beneficiaries with incomes below 138 percent of the FPL and without access to ESI. Nearly two-thirds of the increase can be attributed to premium spending.

FIGURE 3.

Discontinuity in personal medical spending in expansion states. Data are from the March CPS for reference years 2014–16. Open circles represent average spending of the specified type for individuals ages 27 to 64 by 5 percentage point bins of household income relative to the federal poverty line (FPL). OOP stands for out-of-pocket spending, and OTC, over-the-counter spending. The dashed vertical line is 138 percent of the FPL, where coverage policies change in Medicaid expansion states. The solid lines are parametric fits from an RDD model that allows profiles of household income relative to the FPL to vary on either side of the 138 percent FPL threshold.

TABLE 2.

Relationship between income threshold and individual medical spending, 2014–16: Adults ages 27–64 in expansion states

	Total personal medical spending (1)	Individual premium payments (2)	Out-of-pocket spending on medical care and equipment (3)	Spending on over-the-counter health products (4)
	A. Full expansion sample
Above 138% FPL	340.6a	233.2a	100.2	7.1
	(126.90)	(86.24)	(74.21)	(20.15)
Observations	16,582	16,582	16,582	16,582
Mean below 138% FPL ($)	1,376.72	739.76	508.19	128.77
R 2	0.004	0.006	0.001	0.001
	B. Expansion sample without ESI access
Above 138% FPL	385.5b	243.3b	141.7	0.499
	(194.4)	(109.3)	(129.1)	(35.14)
Observations	8,017	8,017	8,017	8,017
Mean below 138% FPL ($)	930.75	392.19	417.29	121.28
R 2	0.003	0.004	0.001	0.001

Note: Data are from the March CPS for reference years 2014–16 and include individuals ages 27 to 64 in Medicaid expansion states. We exclude Massachusetts and the District of Columbia in all years, Minnesota from 2015 on, and New York from 2016 on because these states had Medicaid cutoffs substantially above 138 percent of the federal poverty line (FPL). Total personal medical spending is the sum of individual premium payments and out-of-pocket medical spending. All regressions include a constant and an indicator for above 138 percent of the FPL and a linear trend in income relative to the FPL that is allowed to vary on either side of 138. Spending is in 2016 dollars and was adjusted using the medical component of the consumer price index. Robust standard errors are reported in parentheses. ESI stands for employer-sponsored insurance.

^asignificant at the 1 percent level

^bsignificant at the 5 percent level.

Figure 4 plots results from quantile regression versions of equation 3, for the median, 75th, 90th, and 95th percentiles of the distributions of total personal medical spending. The figure shows a small increase in spending at the median with much larger increases at higher points in the distribution. Table 3 shows the estimates for not only total personal medical spending but also premium and out-of-pocket medical spending. (We do not report quantile regressions for over-the-counter medical supplies because the average effect reported in Table 2 was negligible [$7].) At the median, total personal medical spending increases only by $119. In other words, the estimated increase of $341 at the mean is driven by increases at higher moments in the distribution. Total personal medical spending just above the 138 percent FPL threshold increases by $360, $1,054, and $1,214 at the 75th, 90th, and 95th percentiles, respectively. Consistent with what we find at the mean, the increase in total personal health spending is driven primarily by increases in premium spending. At the median, there is no effect on premiums and only a small effect ($38) on out-of-pocket medical spending. But at the 75th percentile, individuals just above the threshold spend $381 more on premiums and about $111 more on out-of-pocket medical spending than those just below 138 percent of the FPL. At the 90th percentile, individuals just above the 138 percent FPL threshold spend $697 more on premiums and $374 more on out-of-pocket medical spending than those just below 138 percent of the FPL. The pattern of results for those without access to ESI is quite similar, albeit larger in magnitude at the 95th percentiles for both total personal and premium spending. For example, at the 138 percent FPL threshold, the 95th percentile of total personal spending increases by about $1,200 for the full sample but almost $2,300 for the sample without access to ESI.

FIGURE 4.

Discontinuity in personal medical spending at other points in the distribution. Data are from the March CPS for reference years 2014–16. Open circles represent spending at the specified quantile for individuals ages 27 to 64 by 5 percentage point bins of household income relative to the federal poverty line (FPL). The dashed vertical line is 138 percent of the FPL, where coverage policies change in Medicaid expansion states. The solid lines are parametric fits from quantile regressions that allow profiles of household income relative to the FPL to vary on either side of the 138 percent FPL threshold.

TABLE 3.

Relationship between income threshold and the distribution of individual medical spending

	Total personal health-care spending				Individual premium payments				Individual out-of-pocket spending on medical care and equipment
	Percentile				Percentile				Percentile
	50th	75th	90th	95th	50th	75th	90th	95th	50th	75th	90th	95th
	A. Full expansion sample
Above 138% FPL	119.0b	360.2c	1,054.0a	1,214.0c	0	381.4b	696.7b	417.3	37.6a	110.6b	374.2b	611.6c
	(50.58)	(192.99)	(334.62)	(682.97)	(0)	(149.02)	(303.87)	(529.30)	(14.50)	(49.19)	(190.40)	(317.98)
Observations	16,582	16,582	16,582	16,582	16,582	16,582	16,582	16,582	16,582	16,582	16,582	16,582
Value below 138% FPL ($)	213	1,328	3,960	6,439	0	213	2,553	4,580	4	320	1,065	2,130
	B. Expansion sample without ESI access
Above 138% FPL	43.7	335.lb	841.5	2,271.0b	0.0	0.0	895.2b	1,521.0c	0.0	98.0b	33.6	394.3
	(36.52)	(145.12)	(581.50)	(1,142.69)	(0.00)	(0.00)	(410.89)	(782.52)	(0.00)	(42.34)	(148.92)	(412.41)
Observations	8,017	8,017	8,017	8,017	8,017	8,017	8,017	8,017	8,017	8,017	8,017	8,017
Value below 138% FPL ($)	104	519	2,231	4,826	0	0	457	2,217	0	208	1,000	2,076

Note: Data are from the March CPS for reference years 2014–16 and include individuals ages 27 to 64 in Medicaid expansion states. We exclude Massachusetts and the District of Columbia in all years, Minnesota from 2015 on, and New York from 2016 on because these states had Medicaid cutoffs substantially above 138 percent of the federal poverty line (FPL). Total personal medical spending is the sum of individual premium payments and out-of-pocket medical spending. All regressions include a constant and an indicator for above 138 percent of the FPL and income relative to the FPL that is allowed to vary on either side of 138. Spending is in 2016 dollars and was adjusted using the medical component of the consumer price index. Bootstrapped standard errors are shown in parentheses. ESI stands for employer-sponsored insurance.

^aSignificant at the 1 percent level

^bsignificant at the 5 percent level

^cSignificant at the 10 percent level.

C. SELF-REPORTED HEALTH AND DISABILITY

Next, we examine self-reported health and disability. Panel A of Figure 5 shows a discontinuity plot for self-reported health status on a scale of 1 to 5, with 1 = poor and 5 = excellent, so that higher scores reflect better health. There is no discontinuity evident at 138 percent of the FPL in the figure or in the corresponding regression estimates reported in column 1 of Table 4, although this effect is not very precisely estimated. Estimates using an indicator for being in very good or excellent health also show no significant change in self-reported health, although again, this result is imprecisely estimated; see column 2 of Table 4. In contrast, in both Figure 5, panel B, and column 3 of Table 4 we find a small, marginally significant increase in self-reported work limitations. The increase is about 2.2 percentage points off a base of about 14 percent. Results for both self-reported health and work limitations are similar for the sample without access to ESI. Thus, we find no evidence that health, as measured by either self-rated health or work limitations, is better for those who have access to subsidized marketplace coverage rather than Medicaid.

FIGURE 5.

Discontinuity in self-rated health and work limitations. Data are from the March CPS for reference years 2014–16. Open circles represent average self-rated health (panel A) or share of individuals ages 27 to 64 with a disability that limits work (panel B) by 5 percentage point bins of household income relative to the federal poverty line (FPL). The dashed line is 138 percent of the FPL, where coverage policies change in Medicaid expansion states. Solid lines are linear fits from an RDD model that allows profiles of household income relative to the FPL to vary on either side of the 138 percent FPL threshold. Self-rated health has been recoded such that 1 = poor and 5 = excellent health.

TABLE 4.

Relationship between income threshold and self-reported health

	Self-rated health (1)	Very good/excellent health (2)	Disability limits work (3)
	A. Full expansion sample
Above 138% FPL	0.0363	0.0186	0.0221c
	(0.0409)	(0.0192)	(0.0122)
Observations	16,582	16,582	16,582
Mean below 138% FPL	3.515	0.514	0.117
R 2	0.001	0.001	0.004
	B. Expansion sample without ESI access
Above 138% FPL	0.0098	0.0140	0.0229
	(0.0607)	(0.0278)	(0.0192)
Observations	8,017	8,017	8,017
Mean below 138% FPL	3.428	0.476	0.138
R 2	0.003	0.002	0.002

Note: Data are from the March CPS for reference years 2014–16 and include individuals ages 27 to 64 in Medicaid expansion states. We exclude Massachusetts and the District of Columbia in all years, Minnesota from 2015 on, and New York from 2016 on because these states had Medicaid cutoffs substantially above 138 percent of the federal poverty line (FPL). All regressions include a constant and an indicator for above 138 percent of the FPL and income relative to the FPL that is allowed to vary on either side of 138. Robust standard errors are in parentheses. Self-rated health has been recoded so that 1 is poor and 5 is excellent. ESI stands for employer-sponsored insurance.

^csignificant at the 10 percent level.

D. USE OF OTHER SOCIAL PROGRAMS

A concern, particularly among states that did not expand Medicaid, was that the ACA’s Medicaid expansions would increase take-up of other social programs and thus indirectly create budgetary pressure in expansion states. We test the impact of the expansion relative to private exchange subsidies on take-up of other social programs. Specifically, we consider receipt of food stamps/SNAP, unemployment income, use of public housing, and take-up of free and reduced-price lunch among children. As shown in Figure 6 and Table 5, we generally find small, insignificant effects that are not very precisely estimated. The one exception, free and reduced-price lunch take-up, increases at 138 percent of the FPL by a marginally significant 4 percentage points off a base of 86 percent of families with school-age children enrolling in the program: the opposite of what we would expect if Medicaid use led households to sign up for other programs. None of these findings is consistent with a systematic increase in public program participation as a result of Medicaid eligibility.

FIGURE 6.

Discontinuity in other public benefit program participation. Data are from the March CPS for reference years 2014–16. Open circles represent share of individuals ages 27 to 64 with household members receiving food stamps (panel A) or receiving unemployment income (UI, panel B) by 5 percentage point bins of household income relative to the federal poverty line (FPL). The dashed vertical line is 138 percent of the FPL, where coverage policies change in Medicaid expansion states. The solid lines are parametric fits from an RDD model that allows profiles of household income relative to the FPL to vary on either side of the 138 percent FPL threshold.

TABLE 5.

Relationship between income threshold and other public programs

	Unemployment income	SNAP participation	Public housing	Free lunch
	A. Full expansion sample
Above 138% FPL	−0.0064	−0.0115	0.0093	0.0410c
	(0.0072)	(0.0159)	(0.0136)	(0.0223)
Observations	16,582	16,582	8,320	7,665
Mean below 138% FPL	0.038	0.292	0.078	0.842
R 2	0.000	0.031	0.001	0.051
	B. Expansion sample without ESI access
Above 138% FPL	−0.0094	−0.0291	−0.0090	0.0611c
	(0.0106)	(0.0243)	(0.0194)	(0.0320)
Observations	8,017	8,017	4,302	3,500
Mean below 138% FPL	0.041	0.323	0.081	0.864
R 2	0.000	0.030	0.001	0.051

Note: Data are from the March CPS for reference years 2014–16 and include individuals ages 27 to 64 in Medicaid expansion states. We exclude Massachusetts and the District of Columbia in all years, Minnesota from 2015 on, and New York from 2016 on because these states had Medicaid cutoffs substantially above 138 percent of the federal poverty line (FPL). All regressions include a constant and an indicator for above 138 percent of the FPL and a polynomial in income relative to the FPL that is allowed to vary on either side of 138. Robust standard errors are in parentheses. Note that not all households are asked about public housing, only those deemed poor enough to potentially qualify. ESI stands for employer-sponsored insurance.

^csignificant at the 10 percent level.

E. ROBUSTNESS CHECKS

In this section, we check the credibility and robustness of our results in three main ways. First, we carry out placebo tests in which we reestimate our models for two different samples in which we would not expect, a priori, to observe discontinuities at 138 percent of the FPL: (1) adults ages 27 through 64 in non-expansion states in 2014 through 2016; and (2) adults ages 27 through 64 in (nearly) all states in 2009 through 2012, before Medicaid expansion and health insurance premium tax credits became available. If we observed discontinuities in coverage or spending at the 138 percent FPL cutoff prior to 2014 or in non-expansion states, this would be an indication that our regression discontinuity design is not valid. Consequently, these analyses help test whether our main findings are a result of misspecification errors or of other unobserved determinants of our outcomes that may vary discontinuously at the 138 percent FPL threshold. Second, we experiment with three variations on our model specifications: varying bandwidth; estimating donut RDD models; including covariates that change at the 138 percent FPL threshold (age and share white) in our models; and using a global polynomial approach. Third, we test the robustness of our results to two alternative ways of coding our health insurance outcome variables: a set that is not mutually exclusive; and a set that reverses the hierarchical order of the mutually exclusive categories such that public coverage > private non-group coverage > employer-sponsored coverage > uninsured. We describe each of these in more detail below, but the bottom line is that the results of our main analysis—a 5.5 percentage point shift from Medicaid to private non-group coverage and a $341 increase in individual spending at 138 percent of the FPL—are robust to all of these checks.

E.1. PLACEBO TESTS USING NON-EXPANSION STATES.

Online Appendix Figure A7 plots the fraction of adults ages 27 through 64 with Medicaid (panel A) and private non-group coverage (panel B) in 2014 through 2016 in states that did not expand Medicaid. As expected, and in contrast to the results for Medicaid expansion states shown in Figure 1, there is no evidence of any discontinuity at 138 percent of the FPL. Online Appendix Table A3 contains estimation results for a subset of five key outcomes from Tables 1 and 2—Medicaid, private non-group coverage, uninsured, total personal medical spending, and premium spending. We find no discontinuities in these outcomes in non-expansion states. This lends additional support to our RDD design in expansion states since it suggests there is not something else going on at 138 percent of the FPL (for example, eligibility for some other program) that might be driving the discontinuities we observe in expansion states.

E.2. PLACEBO TESTS USING 2009–12 DATA.

Online Appendix Figure A8 shows RDD plots for Medicaid and Online Appendix Figure A9 for non-group coverage for 2009 through 2012 (that is, prior to Medicaid expansion; refer to Online Appendix Table A1 for a listing of states in each category) in states that subsequently would (panel A) and would not (panel B) expand Medicaid. As expected, no discontinuities are evident in these years in either group of states. Online Appendix Table A4 reports regression results confirming the lack of significant discontinuities in this time period, for Medicaid or any of the other key outcomes, in either expansion states (panel A) or non-expansion states (panel B).

E.3. EFFECT OF VARYING BANDWIDTH.

We next explore the sensitivity of our main results to bandwidth choice. We show estimates and 95 percent confidence intervals at bandwidths between 10 and 125 per centage points (incremented by 5 percentage points) around the 138 percent FPL cutoff for Medicaid (Figure 7A), non-group coverage (Figure 7B), and total personal medical spending (Figure 8). As these figures demonstrate, our estimates are robust across a wide range of bandwidths.

FIGURE 7.

Sensitivity of coverage estimates to bandwidth choice. Solid circles represent RDD estimates of Medicaid coverage (panel A) or private non-group coverage (panel B) from models using varying bandwidths of household income relative to the federal poverty line (FPL). Whiskers represent the 95 percent confidence interval for each estimate. A color version of this figure is available online.

FIGURE 8.

Sensitivity of total personal medical spending estimates to bandwidth choice. Solid circles represent RDD estimates of total medical spending from models using varying bandwidths of household income relative to the federal poverty line (FPL). Whiskers represent the 95 percent confidence interval for each estimate. A color version of this figure is available online.

E.4. DONUT RDD.

We also estimate donut RDD models that drop respondents close to the income eligibility threshold to test whether heaping in the measure of income relative to the FPL, caused, for example, by rounding of reported income, biases our results (Barreca, Lindo, and Waddell 2016). Online Appendix Table A5 reports discontinuity estimates for Medicaid and non-group coverage, dropping 0 to +/−5 percentage points around the cutoff. This range includes our main estimate, which excludes data at +/−2 percentage points of the threshold. Panel A shows the full expansion sample and panel B the sample without access to ESI. The discontinuities in coverage are not sensitive to small data exclusions around the 138 percent of the FPL threshold. Online Appendix Table A6 shows estimates from analogous models but for total personal medical and premium spending. As with coverage, results are robust to using all data, but estimates are more precise and larger in magnitude if we exclude a small range of data around the 138 percent FPL cutoff. This is likely because of noise in measuring adjusted gross income.

E.5. INCLUDING COVARIATES.

Online Appendix Table A7 shows RDD estimates for Medicaid, private non-group coverage, uninsured, total individual medical spending, and premium spending but controlling for covariates that changed discontinuously at 138 percent of the FPL. Specifically, we control for age and its square as well as the share of respondents that were white. Estimates are qualitatively similar.

E.6. GLOBAL POLYNOMIALS.

Online Appendix Table A8 shows estimates for coverage and for spending using a global polynomial RDD approach. We show estimates from both linear and quadratic polynomial models, where polynomials are allowed to vary on either side of the cutoff, and using a bandwidth of 75 percentage points.²³ Here again, we show that our estimates are not sensitive to modeling choice.

E.7. ALTERNATIVE HEALTH INSURANCE CODING.

Online Appendix Table A9 reports results for health insurance outcomes defined so that they are not necessarily mutually exclusive, and Online Appendix Table A10 reports results for health insurance outcomes that have the reverse, mutually exclusive hierarchical order as the main coding. The results from both alternative codings are generally similar to those in Table 1.

VI. Discussion

To recap our main results: we find a sharp shift from Medicaid to private non-group coverage but no change in uninsurance at 138 percent of the FPL in states that expanded Medicaid as well as a sizable, discontinuous increase in average personal medical spending of $341, driven by higher premium payments. The average increase in personal spending is driven by large increases in the right-hand tail of the distribution; increases at the 90th and 95th percentiles exceed $1,000. We find no evidence of increases in the take-up of other public benefits or changes in self-reported health or disability at this threshold; nor do we find any evidence of bunching in our data that would indicate individuals close to the eligibility threshold might be strategically manipulating their incomes.

These results speak to several different policy debates. Our results on personal medical spending suggest that private non-group coverage imposes significantly higher costs on beneficiaries than does Medicaid, which as discussed above should inform discussions about whether to extend premium tax credits to adults living in poverty or to hold out for Medicaid expansion in the remaining 12 states that have not yet done so. To put our findings in perspective, the $341 increase in personal medical spending (or $386 for the sample without ESI access) is the reduced-form effect of the switch from Medicaid to subsidized private insurance. The group affected by this switch is fairly small, between 5.5 and 7 percentage points. Scaled by this “first stage,” the change in total personal medical spending at the mean is between $5,600 and $6,200 or 25–28 percent of income at this threshold for a household of two in 2016.²⁴ Most of the higher personal medical spending associated with private coverage is premiums, not out-of-pocket medical spending. This is likely due in part to the ACA’s cost-sharing reductions, which subsidize care at the point of service for individuals with family income up to 250 percent of the FPL. The cost-sharing reductions do not lower the (quite high) cap on total out-of-pocket spending that beneficiaries may face, however, a fact that is reflected in the right tail of the out-of-pocket spending distribution when scaled by the magnitude of the coverage change estimated in the first stage. For example, scaling implies that out-of-pocket spending increases by between $1,400 and $2,000 at the 75th percentile for those with private non-group coverage relative to those with Medicaid.²⁵ That said, we urge some caution in overinterpreting the magnitude of these scaled estimates. The minimum $F$-statistic for proper inference is now recommended as 104.7 (Andrews, Moreira, and Stock 2006; Lee et al. 2020). The corresponding $F$-statistic for Medicaid coverage in our data is 103.4 for the full sample and 27 for the sample without access to ESI.

Irrespective of scaling, our results imply that low-income individuals just above the Medicaid threshold do face high premium costs, even after subsidies. From a risk protection perspective, higher premiums are not concerning since they are predictable (Barcellos and Jacobson 2015). Moreover, we find no evidence that these premiums lead individuals in this income range to opt out of insurance altogether, which if it were occurring would potentially undermine the functioning of the private non-group market (Hackmann, Kolstad, and Kowalski 2012). Nonetheless, concern about the cost burden facing low-income households with marketplace coverage has spurred proposals to increase the generosity of premium tax credits. The American Rescue Plan Act changed the design of the premium subsidy in 2021 and 2022 so that individuals below 200 percent of the FPL may pay nothing in premiums for private non-group coverage (subject to judicious choice of plan). Moreover, the redesigned subsidy schedule does not simply shift the notch—that is, moving the discontinuity higher up the income distribution—it actually gets rid of the notch by smoothly phasing in subsidies above 200 percent of the FPL. The Biden administration has proposed making this change permanent, which would require further Congressional action.

An unsolved puzzle remains. Existing evidence demonstrates that individuals are sensitive to premiums in Medicaid (Dague 2014; Cliff et al. 2022) and on the exchanges (Abraham et al. 2017; Drake and Anderson 2020). Yet, faced with clear evidence that subsidized private insurance imposes higher costs, largely through premiums, than Medicaid, why don’t individuals in lower-income households appear to strategically sort into Medicaid? One possibility is that enrollment frictions, combined with uncertainty about income for near-poor households, may make such sorting difficult. Another possibility, albeit one we are unable to quantify, is that private coverage provides some benefit that might make its higher cost worthwhile. The most likely benefit is better access to and/or quality of care. Multiple audit studies have shown that patients are more likely to be able to make appointments with private coverage than with Medicaid (for a review, see Hsiang et al. [2019]). Allen et al. (2021) document fewer emergency visits and more outpatient visits, as well as higher quality of care, associated with private coverage rather than Medicaid.²⁶

Data limitations prevented us from using the RDD framework to measure differences in access to care. This is because there is no data set with a sufficiently large sample that has both detailed income data and measures of access to care. The CPS does not include any measures of access to care; adding such measures would allow us to examine this outcome, as well as opening up many other opportunities for researchers to understand changes in access to care in the wake of the new sources of coverage introduced by the ACA. The National Health Interview Survey (NHIS) and the Behavioral Risk Factor Surveillance System (BRFSS) include measures of access to care, but income is not measured finely enough to capture an income-based eligibility threshold for Medicaid or premium tax credits; refining income measurement in the NHIS and BRFSS would therefore open up new avenues for research on income-based programs and access to care.

Our results for other outcomes—health outcomes and use of other public programs—make a targeted, unique contribution to the literature on the impacts of health insurance coverage. Numerous studies have attempted to document whether having health insurance—as opposed to being uninsured—leads to improvements in health, with the preponderance of evidence suggesting small, positive effects (see Guth, Garfield, and Rudowitz [2020] for a review). A smaller literature has addressed whether expansions of Medicaid lead to increases in the use of other public programs (Baicker et al. 2014; Burns and Dague 2017; Schmidt, Shore-Sheppard, and Watson 2019, 2020; Soni et al. 2017). Our analysis, in contrast, is designed to address the question of whether type of insurance—public versus private—conditional on having some coverage, affects these outcomes. We find no evidence that it does.