The three‐year impact of the Affordable Care Act on disparities in insurance coverage

Objective To estimate the impact of the major components of the ACA (Medicaid expansion, subsidized Marketplace plans, and insurance market reforms) on disparities in insurance coverage after three years. Data Source The 2011‐2016 waves of the American Community Survey (ACS), with the sample restricted to nonelderly adults. Design We estimate a difference‐in‐difference‐in‐differences model to separately identify the effects of the nationwide and Medicaid expansion portions of the ACA using the methodology developed in the recent ACA literature. The differences come from time, state Medicaid expansion status, and local area pre‐ACA uninsured rates. In order to focus on access disparities, we stratify our sample separately by income, race/ethnicity, marital status, age, gender, and geography. Principal Findings After three years, the fully implemented ACA eliminated 43% of the coverage gap across income groups, with the Medicaid expansion accounting for this entire reduction. The ACA also reduced coverage disparities across racial groups by 23%, across marital status by 46%, and across age‐groups by 36%, with these changes being partly attributable to both the Medicaid expansion and nationwide components of the law. Conclusions The fully implemented ACA has been successful in reducing coverage disparities across multiple groups.

While gains in insurance coverage after the ACA have been well documented, few papers in this literature examine how the ACA affected coverage disparities. One recent paper estimates the first-year impact of the ACA on coverage using difference-indifference-in-differences (DDD) models where the differences come from time, state Medicaid expansion decisions, and pre-ACA local area uninsured rates. 4 This strategy leverages the propensity for universal coverage initiatives to provide the most intense "treatment" in local areas with the highest prereform uninsured rates. 5,6 Using data from the American Community Survey (ACS), the authors find that the ACA increased coverage by an average of 5.9 percentage points in Medicaid expansion states compared to 2.8 percentage points in nonexpansion states in 2014. 4 In subsample analyses, they show that the fully implemented ACA (including the Medicaid expansion) reduced the coverage disparity between college graduates and those with a high school diploma or less by 11.4%, and that between whites and nonwhites by 14%. 4 The paper also finds greater gains in coverage for young adults and unmarried individuals, which had lower pre-ACA coverage rates than older adults and married individuals, respectively. 4 Another recent paper uses the same research strategy and data from the Behavioral Risk Factor Surveillance System (BRFSS), finding that the ACA reduced the coverage disparity between those with incomes above versus below the median by 38%. 7 A third paper uses ACS data through 2015 and leverages variation in state Medicaid expansion decisions, pre-ACA eligibility requirements, and subsidy rates across the income distribution. 8 They find that coverage gains from the Medicaid expansion and premium subsidies are larger among childless adult couples than among single adults or adults with children, but the increase from the individual mandate is largest among singles. 8 Other studies focus only on the ACA's Medicaid expansion, using simpler difference-in-differences (DD) models to compare changes in insurance coverage over time between Medicaid expansion and nonexpansion states. One paper includes 2015 data from the ACS and shows that the Medicaid expansion reduced the coverage disparity between 19-to 26-and 56-to 64-year-olds by 15%, while the disparity between Hispanics and non-Hispanic whites only fell by 4%. 9 Another paper, also using data through 2015, finds that the Medicaid expansion led to smaller gains among low-income Hispanics than other low-income individuals, implying a widened disparity. 10 Other papers provide evidence that the Medicaid expansion increased insurance coverage among those with low incomes or levels of education, implying reduced disparities relative to individuals with higher socioeconomic status. 11,12 One study's focus is on the impact of the ACA in a single state, Kentucky, finding that much of the reduction in the state's uninsured rate is due to large coverage gains from areas with higher concentrations of poverty. 13 We contribute to this literature by using the DDD method described above and elsewhere 4,7,14 to uncover the causal impact of the 2014 ACA provisions, both with and without the Medicaid expansion, on coverage disparities after three years. Changes in coverage disparities are evaluated by stratifying our sample by income, race/ethnicity, marital status, age, gender, and geography. Data come from the American Community Survey (ACS) between 2011 and 2016. The ACS includes multiple categories of insurance coverage, allowing us to evaluate how the ACA affected coverage disparities via changes to both private and public coverage. In addition, the ACS is a large enough survey to precisely estimate the effects for states and many local areas, given that it includes approximately 3 000 000 observations per year and relatively narrow geographic identifiers. Finally, the mandatory nature of the ACS reduces concerns about sample selection among respondents.
Our primary hypothesis is that, in its first three years, the ACA significantly reduced insurance coverage disparities across demographic groups. We contribute to the literature on the ACA's coverage effects in multiple ways. First, we are, to our knowledge, the first to quantify the impacts of the ACA on disparities using three years of post-ACA implementation data (2014-2016). One recent study examines the effect on the uninsured rate after three years using the BRFSS, but does not specifically examine disparities. 14 Second, in contrast to the BRFSS, the ACS allows us to examine how changes in sources of coverage, such as employer-sponsored and individually purchased private coverage and Medicaid, drove any changes in disparities.
Third, in contrast to other recent work, 9,10 our approach allows us to estimate the impact of the fully implemented ACA, rather than just focusing on the Medicaid expansion. Fourth, we examine disparities along a new dimension: residence in rural vs. urban locations.

| DATA
The ACS is a nationally representative survey administered by the Census Bureau sampling approximately 1% of the U.S. population annually. Participation is mandatory, and the survey can be completed online or through the mail. In terms of geography, the ACS identifies all 50 states and the District of Columbia, and additionally identifies localities known as Public Use Microdata Areas (PUMAs). PUMAs are approximately 2300 areas of at least 100 000 people nested entirely within a state. Our primary sample consists of 19-to 64-year-olds from calendar years 2011 to 2016, which results in over 3 000 000 individuals per year. By starting our sample period in 2011, we aim, as in other recent work, 4 to measure only the effects of the package of ACA provisions taking effect in 2014, as opposed to also capturing the effect of the 2010 dependent coverage mandate that required insurers to allow dependents to remain on their parents' insurance plans until the plan year following their 26th birthday. This mandate has already been studied extensively in prior research. [15][16][17][18][19] We create several binary outcome variables based on the ACS insurance coverage questions: any insurance, any private insurance (either employer sponsored or directly purchased), employersponsored insurance, directly purchased insurance, Medicaid, and any other coverage. We define other coverage as neither private nor Medicaid coverage. These categories are not mutually exclusive due to the possibility of multiple sources of coverage.
In order to exploit within-state variation in pre-ACA uninsured rates in 2013 to identify the impact of the national components of the ACA, we would ideally simply use the PUMA classification system included in the ACS. Unfortunately, the PUMA definitions changed during our sample period due to new boundaries introduced in the 2010 Census. To address this problem, we follow a recent paper 4 and use both the old and new PUMA classification systems to identify core-based statistical areas (CBSAs), which we then use as our local areas. If a CBSA spans multiple states, we define a different local area for the parts of the CBSA in each state. In addition, we create additional local areas for the non-CBSA portion of each state, in order to prevent respondents who do not live in a CBSA from being dropped from the sample. We classify non-CBSA local areas as "rural" and CBSA local areas as "urban." Our dataset Our "economic" controls consist of dummies for education (high school degree, some college, and college graduate, with less than a high school degree as the omitted category), household income   and whether her local area's pretreatment uninsured rate was above or below the median for individuals in the sample. Table 1 shows that 79% of the sample was covered by some type of insurance in the baseline year of 2013, including 11% with Medicaid and 60% with employer-provided coverage. For both the high-and low-uninsured rate subgroups, individuals in Medicaid expansion states were slightly more likely to be covered by some type of insurance in 2013 than those in nonexpansion states, with the differences being driven entirely by Medicaid coverage.
Our DDD model will account for such baseline differences.

| ME THODS
In order to uncover the causal impact of the ACA on coverage dis- The term POST t is not separately included in Equation (1) since it is absorbed by the time fixed effects, while the terms UNINSURED as × MEDICAID st are not separately included since they are absorbed by the local area fixed effects.
The effect of the ACA without the Medicaid expansion is given by γ 1 × UNINSURED as , which means it is assumed to be zero in a (hypothetical) area with a 0% uninsured rate at baseline and to increase linearly as the pre-ACA uninsured rate rises. 4 Similarly, the effect of the Medicaid expansion alone is given by  affect coverage in an area with a 0% baseline uninsured rate. The effect of the "fully implemented" ACA, that is, in Medicaid expansion states, combines the impacts of the Medicaid and non-Medicaid components: 1 × UNINSURED as + 3 × UNINSURED as . In our results, we report the predicted effect of the ACA at the sample mean pretreatment uninsured rate. Formally, this predicted effect is given by 1 × UNINSURED as in nonexpansion states and 1 × UNINSURED as + 3 × UNINSURED as in expansion states. For each subsample of interest, we re compute the pre treatment uninsured rate using only individuals within that particular subsample. 4 Tables 2-4 report the implied effects of the ACA at the average pre-ACA uninsured rate based on coefficient estimates from the DDD regression described by equation (1) for each coverage outcome.

| RE SULTS
TA B L E 2 Implied effects of the ACA at mean pre treatment uninsured rate for full sample and income and race subsamples The other two panels of Table 2  According to the first column of Table 2 We next examine the race stratification in the third panel of  Figure S1   Another limitation is that our disparity analyses assume that the subsamples are exogenously determined. Income is one source of stratification that might seem particularly likely to adjust endogenously in response to the 2014 ACA provisions. Two recent papers found little impact of the ACA Medicaid expansions on work effort, implying that the effect on income should be minimal. 11,22 Another found that while labor market outcomes in the aggregate were not significantly affected by the ACA, labor force participation reductions in areas with higher potential exchange enrollment were offset by increases in labor force participation in areas with higher potential Medicaid enrollment. 23 In order to examine whether our particular ACA treatment variables influence income, the first column of Table S8 presents the results of our baseline regression model and employer mandates, subsidies, and health insurance exchanges that served as the model for the ACA-had only small effects on marriage and divorce rates. 24 In Table S8, we estimate our baseline regression model separately with indicators of being married, of being newly married during the past 12 months, of being newly divorced during the past 12 months, and of being newly married or divorced in the past 12 months as dependent variables. The results

| D ISCUSS I ON
suggest that there was no statistically significant effect of our ACA treatment variables on these outcomes. In addition, we replicate our main analyses after dropping individuals from the sample that had any change in their marital status in the last 12 months. The results, reported in Table S9, are very similar to the results presented previously. While the available evidence therefore suggests that our assumption of exogenous stratification is plausible, it is of course not possible to establish this definitively.
With those limitations in mind, our results are broadly consistent with those reported in the Medicaid expansion literature 9 in that both the Medicaid expansion and the fully implemented ACA generally reduce but do not eliminate coverage disparities.
These results imply that full repeal of the ACA would exacerbate these disparities. Additionally, it is possible that changes to the ACA after 2016, including regulatory changes, such as Medicaid work requirements, and the elimination of the individual mandate, would lead to further changes in disparities. For example, our finding that the Medicaid expansion eliminated 43% of the coverage gap across income groups is likely to change if Medicaid work requirements, that would be expected to potentially reduce enrollment, 25 are widely implemented. Thus, more work is needed to examine the impact of the ACA as economic conditions change and the ACA itself changes.