Three-Year Impacts Of The Affordable Care Act: Improved Medical Care And Health Among Low-Income Adults

: Significant policy uncertainty continues to surround the Affordable Care Act (ACA) at both the state and federal level. We assessed changes in health care use and self-reported health after three years of the ACA’s coverage expansion, using survey data collected from low-income adults through the end of 2016 in three states: Kentucky, which expanded Medicaid; Arkansas, which expanded private insurance to low-income adults using the federal marketplace; and Texas, which did not expand coverage. We used a difference-in-differences model with a control group, and an instrumental variables model to provide individual-level estimates of the effects of gaining insurance. By the end of 2016, the uninsured rate in the two expansion states had dropped by more than 20 percentage points relative to the non-expansion state. For uninsured individuals gaining coverage, this change was associated with a 41 percentage-point increase in having a usual source of care, a $337 reduction in annual out-of-pocket spending, significant increases in preventive health visits and glucose testing, and a 23 percentage-point increase in “excellent” self-reported health. Among adults with chronic conditions, we found improvements in affordability of care, regular care for those conditions, medication adherence, and self-reported health.


Introduction
The Affordable Care Act (ACA) has produced the largest gains in insurance coverage in nearly 50 years, 1 but its future is uncertain after the 2016 election. While initial attempts to repeal the law in 2017 failed in Congress, leading policymakers remain committed to substantial changes. Estimates are that as many as 20 million Americans have obtained insurance under the ACA, 2 with more than half via Medicaid and the remainder largely from health insurance marketplaces. 3 Meanwhile, several states that have yet to expand Medicaid are in the midst of a renewed debate over this possibility. 4 The impact of the ACA's coverage expansion on medical care and health is critical to evaluating future policy efforts related to the law.
National studies of the full non-elderly population have detected improvements in trends for coverage, satisfaction with insurance, and access to care. 5,6 Research specifically comparing Medicaid expansion versus non-expansion states have shown important clinical changes including increased primary care visits, 7 improved blood pressure control and Pap testing rates, 8 and suggestive evidence of improved self-reported health. 9,10 However, published analyses have been limited to 2014 or 2015 data, 11 and since insurance expansions extend their reach gradually, 12 updated analyses with more recent data could provide valuable insights.
Meanwhile, patients with chronic medical conditions may have the most to gain from coverage expansion. These conditions affect nearly half of all Americans, disproportionately those who have gained coverage under the ACA. 13 Given the high costs of care for this group 14 and the public health implications of these conditions, 15 there is substantial policy interest in whether expanding coverage improves quality of care and health in this population.
Our objective was to assess ongoing changes in health care use and self-reported health among low-income adults, including those with chronic conditions, after three full years of the ACA's coverage expansion. Using survey data in three states collected through the end of 2016, we provide timely evidence of the law's ongoing impact on patient care.

Study Design
We conducted a difference-in-differences analysis, which compares pre-versus postexpansion outcomes in two expansion states, with one non-expansion state as the control group.
Our study states were Kentucky, which expanded coverage to low-income adults via Medicaid beginning in 2014; Arkansas, which used federal Medicaid funding to provide private insurance from the health insurance marketplace to low-income adults beginning in 2014; and Texas, which did not expand coverage to low-income adults. Thus, our study captures elements of both private marketplace insurance expansion and Medicaid expansion under the ACA; for brevity, we refer below to "ACA expansions" to describe Kentucky's Medicaid expansion and Arkansas's "private option" approach. In sensitivity analyses, we test for differences between these two approaches.
We selected Texas as a comparison state for Kentucky and Arkansas because it is a nonexpansion state in the same Census region that had a similarly restrictive set of Medicaid eligibility criteria prior to the ACA (Texas covered parents only up to 26% of poverty as of 2012, compared to 17% for Arkansas and 59% in Kentucky, versus the national median of 81%). 16 Our study design relies on the assumption that if not for Medicaid expansion, trends in coverage and other outcomes would have been similar across these three states. Data from the U.S.
Census Bureau show that coverage trends for our study population were similar in all three states from 2010-2013 (see Appendix Figure 1), offering support for this assumption. 27 For our primary difference-in-differences model, we separately identified the expansion effects for each year of expansion -2014, 2015, and 2016 -all compared to the pre-expansion 2013 data. This approach allows us to trace out differential changes over time. We also tested a model that pooled 2015-2016 together to increase statistical power.
We then conducted an instrumental variables (IV) analysis, in order to estimate individual-level changes in health care outcomes for uninsured individuals who acquired coverage under the ACA. We repeated this analysis among the subset reporting at least one chronic condition. The IV approach uses a quasi-experimental source of variation in a key predictor to identify the treatment effect of that predictor; 17 in our case, the variation was the state ACA expansion policy, and the key predictor of interest was having health insurance. This approach builds on the underlying quasi-experimental design of our difference-in-differences model, but provides more directly interpretable estimates of patient-level outcomes, similar to the local average treatment effect estimated using an IV model in the Oregon Health Insurance Experiment. 18 This approach does not change the causal inference for our study, which still relies on the assumption that in the absence of the ACA expansion, trends in our outcomes would have been similar in Texas versus the expansion states.
One key assumption for an IV analysis is that the instrument has a significant relationship with the predictor of interest; here, the state expansion decisions clearly had large effects on insurance coverage. The other key assumption is that the instrument (state expansion) only affects outcomes via the predictor of interest (insurance coverage). While this seems plausible in our case, other potential aspects of coverage expansions might affect access to care and health, even among those who did not gain insurance. For instance, there could be positive spillovers of expansion via better funding to safety net institutions, 19 which would bias our IV estimates upwards; or negative spillovers via reduced health system capacity to care for previouslycovered populations, 20 which would have the opposite effect on our estimates. In addition, some people may not have gone from uninsured to insured, but simply switched types of coverage due to expansion. While these influences are likely swamped by the individual-level effects of gaining insurance, they are nonetheless potential sources of bias.

Data
We contracted with a research firm to conduct a random-digit dialing telephone survey from November to December each year, from 2013-2016. The survey sample contained U.S. citizens ages 19-64, with family incomes below 138% of the federal poverty level (FPL), the ACA's Medicaid expansion eligibility threshold. The survey was available in Spanish and English, and the sample included cellphone and landline users. Each year, we recruited a new sample split equally across our three study states. Annual sample sizes ranged from 2,209 to 3,011, for an overall total of 10,885. The study was exempted from review by the Harvard Chan Institutional Review Board since the investigators only had access to deidentified data.
The overall response rate was 22%, which compares favorably to several other surveys that have been used to evaluate the ACA. 21-23 Previous research demonstrates that the use of population-weighting in random-digit telephone surveys can mitigate non-response bias and produce estimates similar to those from government surveys. 24-26 Accordingly, our analyses were weighted to demographic targets for low-income adults in our study states based on age, gender, education, marital status, race/ethnicity, geographic region, population density, and cellphone use. Our survey has been previously validated against two large government-sponsored sources, the American Community Survey and the Behavioral Risk Factor Surveillance System.
In that validation, we compared estimates for low-income adults in our three study states for coverage and several measures of access to care in our survey and the government datasets. We found moderate-to-strong correlations and a range of absolute differences in estimates consistent with analogous differences between various federal surveys. 9 Our study outcomes spanned seven domains: 1) health insurance (uninsured, Medicaid, and private insurance, in which each individual was assigned a primary type of insurance [see Appendix Methods], 27 as well as any coverage changes within 12 months); 2) access to care (having a personal doctor, usual location of care, difficulty obtaining primary care and specialty appointments, and reasons for ED use); 3) affordability (skipping needed care or medications due to cost, trouble with medical bills, and medical out-of-pocket spending); 4) utilization (outpatient, emergency department, and inpatient care in the prior 12 months); 5) preventive care (receipt of a check-up, cholesterol test, or glucose test in the prior 12 months); 6) quality of care (cholesterol and glucose testing for high-risk patients, regular care for chronic conditions, and self-rated quality of care); and 7) health status (self-reported health on a five-point scale 28 and a two-item depression score 29 ).
We then examined the same outcomes (other than coverage) in an instrumental variables (IV) analysis. We used a two-stage least-squares regression, 30 in which the first stage predicted the likelihood of a person having any health insurance (i.e. 1 -uninsured) as a function of state expansion decisions and the year, using the difference-in-differences model described above (see Appendix Table 1). 27 The second stage then provided an estimate of the impact of gaining insurance from the ACA expansion on individual-level health care outcomes. This method also has the advantage of using all four years of data simultaneously to produce a single estimated policy effect from expansion.
We used Stata 14.0 for all analyses. All regression models used county-level robust clustered standard errors to account for the non-independence of observations within the same state and county.

Sensitivity and Subgroup Analyses
We repeated our IV analysis for the subset of respondents who reported having been diagnosed with any of nine chronic conditions: hypertension, coronary artery disease, stroke, asthma / chronic obstructive pulmonary disease (COPD), kidney disease, diabetes, depression, cancer (other than skin cancer), or substance abuse. We also tested whether the prevalence of these conditions changed in association with expansion status.
We conducted several sensitivity analyses. We tested the impact of pooling 2015 and 2016 expansion state data together to increase statistical power. To account for multiple hypothesis testing within each domain of outcomes, we estimated "family-wise" p-values using a step-down bootstrapping approach similar to other recent analyses. 11,31 We also tested a spatial correlation model described below.
Finally, we repeated our primary model with the expansion states divided into Kentucky's Medicaid expansion versus Arkansas's private option. This produced separate estimates for expansion effects in Arkansas and Kentucky.

Limitations
Our analysis has several limitations. Our study includes only three states. This means our results may not generalize to the nation as a whole. It also affects our estimation of standard errors and the possibility of idiosyncratic changes in any given state exerting an outsized influence on our findings. In studies with a larger number of states, the use of state-clustered standard errors can limit this risk, but standard methods to estimate correlation within states are biased when there are only three states. 32 Instead, we used county-level clustering to estimate standard errors as in our previous work with this dataset, 9 and also present results using spatial correlation across counties similar to other health care analyses using small numbers of states. 33 Our main findings were similar under both alternatives.
As discussed earlier, the response rate for random digit telephone surveys like ours is lower than that for government interview surveys. However, we believe that the tradeoff of timeliness and ability to design our own comprehensive survey outweighed those concerns, particularly given the previous validation of our survey instrument. 9 Other limitations are inherent to our study's quasi-experimental design, which helps control for secular trends and takes advantage of a comparison group using a non-expansion state, but is still subject to unmeasured confounders that vary over time across states. Our IV analyses produced estimates with fairly wide confidence intervals, which means that the exact magnitudes of change should be interpreted cautiously. Finally, the possibility of recall or social desirability bias may affect our survey-based results. However, our findings in several domains are consistent with ACA studies using non-survey data such as pharmacy claims, 34,35 lab results, 36 and community health center reports. 8

RESULTS
Exhibit 1 presents descriptive statistics by state for our full sample and for those with chronic conditions. Respondents in Texas were disproportionately Latino and urban compared to Arkansas and Kentucky. Chronic conditions affected 69% in Arkansas, 72% in Kentucky, and 55% in Texas. Changes in disease prevalence between 2013 and 2016 by state were nonsignificant for all but kidney disease, which showed a small decline in expansion states (-2.2 percentage points, p=0.06) (Appendix Table 2). 27 Individuals with chronic conditions were older and less likely to be male or Latino. Among those with a condition, the mean number of conditions ranged from 2.0 to 2.3 by state, with depression (57-64%), hypertension (52-54%), asthma/COPD (32-43%), and diabetes (22-26%) the most common. in Kentucky, 11.7% in Arkansas, and 28.2% in Texas.
Exhibit 3 presents regression-based estimates for differential changes in our study outcomes, comparing expansion to non-expansion states (Appendix Table 3  Appendix Tables 4 and 5 present sensitivity analyses for our difference-in-differences model. 27 When we used bootstrapped "family-wise" p-values that accounted for multiple variables within each domain of outcomes, we continued to find significant changes in 2016 for outcomes related to coverage, access, affordability, and prevention (p<.05) and quality (p<.10), but not for utilization and self-reported health. 27 Pooling 2015-2016 data together strengthened the statistical significance of some 2016 findings such as private insurance gains, having a usual source of care, out-of-pocket spending, and excellent self-reported health, and outcomes in 5 of 7 domains were significant at p<.05 and the other two at p<.10 using family-wise p-values.
Difficulty obtaining an appointment to see a specialist was no longer significant in the pooled model. In models using spatially-correlated standard errors, several estimates were affected by the lack of weighting, but overall this approach yielded similar precision as the main model, offering support for our primary method using county-level clustering. Exhibit 4 presents individual-level estimates of changes in these outcomes for patients acquiring insurance using our IV model. For the full sample, we estimated that expansion led to significant changes including a 41 percentage-point increase in having a usual source of care among those gaining coverage, a $337 reduction in medical out-of-pocket spending, a 28 percentage-point reduction in the likelihood of any emergency department visits, and a 25 percentage-point increase in glucose testing. The proportion in excellent health increased by nearly 23 percentage points.
Exhibit 4 also shows IV results for adults with chronic conditions. While out-of-pocket spending and cholesterol and glucose testing among high-risk patients (those with diabetes, stroke, hypertension, or heart disease) did not change significantly, we otherwise found similar results for most outcomes as in the full sample, including a 51 percentage-point decrease in skipping medications due to cost and a 20 percentage-point increase in excellent health. In a question asked only of this subgroup, we estimated a 56 percentage-point increase in obtaining regular care for chronic conditions.

Comparisons of the 2016 effects of private versus public insurance approaches in
Arkansas and Kentucky (Appendix Table 6) showed no significant differences for most outcomes. 27 As expected, health insurance type differed, with more private coverage gains in Arkansas and more Medicaid in Kentucky. The only other significant difference was a greater decline in "fair/poor quality of care" in Arkansas compared to Kentucky. Both expansions were associated with significant improvements in numerous outcomes compared to Texas, including access to a personal doctor and medications, trouble with medical bills, checkups and cholesterol testing, and self-reported health.

DISCUSSION
In our analysis of survey data of low-income adults in three states, we note three key contributions to the growing body of research on the ACA. First, we provide the earliest The validity of these estimates is supported by their similarity to those from the IV analyses in the randomized Oregon Health Insurance Experiment, which showed an average reduction of $390 in medical debt, a 34 percentage-point increase in office-based usual source of care, and a 13 percentage-point change in the share in excellent, very good, or good health. 31 These latter results are particularly noteworthy given policy interest in the impact of the ACA on health status. For context, prior research indicates that a self-reported health rating of fair or poor confers a mortality risk 2-4 times higher than those in the healthiest category. 28 Our finding of improved self-reported health is consistent with results in the Oregon study and other pre-ACA Medicaid expansions, 37 though the evidence on similar changes under the ACA has been more mixed. 6,10,11,38 In part, this likely reflects differences in sample frame and timing.
Studies that have not found significant changes in self-reported health after the Medicaid expansion have typically used only 1-2 years of post-expansion data and have studied expansionrelated coverage gains on the order of 3 to 8 percentage points. 6,11,38 Here, we assessed 3 full years of post-expansion data and studied a population experiencing a much larger coverage change of over 20 percentage points.
Adults with chronic conditions experienced numerous improvements in both access and quality of care, including more checkups, improved adherence to medications, higher rates of regular care for chronic disease, and -perhaps as a consequence of these changes -improved self-reported health. These findings build on a previous study using national data through 2014 showing gains in two access measures for adults with chronic conditions (having a check-up and no cost-related delays in care); 39 however, our study included a much richer set of outcomes and two additional years of data.
We detected an increased rate of difficulty obtaining specialist appointments in 2016 in the expansion states, particularly in Kentucky. This is consistent with a recent national study that showed an increase in appointment wait times after expansion, 7 as well as some studies showing greater barriers to specialty care in Medicaid than primary care. 40 However, in part this may also reflect that patients without coverage are less likely to attempt to make appointments with specialists; thus, coverage expansion may increase the share who try but experience difficulties in doing so, even as their overall access to care has improved.
Our results also offer insights into alternative state approaches to coverage expansion.
With increased interest under the new administration in state flexibility and innovation, we find that either a private insurance expansion via marketplace coverage (as in Arkansas) or a Medicaid expansion (as in Kentucky) produce similar benefits across most study outcomes.
Consistent with prior comparisons, 41 the results imply that coverage expansion is quite important for patients, but the type of coverage obtained is less critical.

Conclusion
Over three years of coverage expansion in two states, the ACA was associated with statistically significant and clinically relevant improvements for low-income adults' access to care, preventive services utilization, and self-reported health. Among those with chronic conditions, coverage expansion was linked to improved medication adherence, more regular communication with physicians, and improved perceived health status. As policymakers debate the ACA's future and additional states consider whether to expand Medicaid, our findings demonstrate the benefits associated with coverage expansion for two particularly vulnerable populations -low-income adults and those with chronic conditions.

18.
Baicker K, Taubman S, Allen H, et al.  *** p<0.01, ** p<0.05, *p<0.10 ED = Emergency Department. Results show differences-in-differences estimates for expansion states (Arkansas and Kentucky) versus Texas, by year. All analyses adjusted for sex, age, race/ethnicity, marital status, family size, education, income, urban vs. rural residence, county annual unemployment rate, state, and year.
a -All estimates are reported as percentage-point changes for binary outcomes, other than number of office and ED visits and out-of-pocket spending.
b -Usual source of care was grouped into 3 categories -those reporting an office-based usual source of care, those without any usual source of care, and those using the ED as the usual source of care. c -Sample limited to patients reporting heart disease, stroke, diabetes, or hypertension (n=5,611). d -Sample limited to patients reporting a history of diabetes (n=2,213). e-Sample limited to patients reporting at least one of the following conditions: hypertension, heart attack/coronary artery disease, stroke, asthma/COPD, kidney disease, diabetes, depression, cancer, and substance abuse (n=7,734).

Exhibit 4 (Table)
Caption: Instrumental Variables (IV) Analysis -Individual-Level Change Per Person Gaining Insurance Source/Notes: SOURCE: From a telephone survey of U.S. citizens ages 19-64, with family income below 138% of the federal poverty level, in Arkansas, Kentucky, and Texas. The sample contained 10,885 adults (minus item non-response for each specific outcome), except where otherwise noted below.

NOTES:
*** p<0.01, ** p<0.05, *p<0.10 ED = Emergency Department. IV = Instrumental Variables. Results show local average treatment effect from gaining coverage via expansion in Arkansas and Kentucky, compared to Texas (which did not expand) using two-stage least squares IV regression. All analyses adjusted for sex, age, race/ethnicity, marital status, family size, education, income, urban vs. rural residence, state, and year.
a -All estimates are reported as percentage-point changes for binary outcomes, other than number of office and ED visits and out-of-pocket spending.
b -Sample limited to patients (n=7,734 adults) reporting at least one of the following conditions: hypertension, heart attack/coronary artery disease, stroke, asthma/COPD, kidney disease, diabetes, depression, cancer, and substance abuse.
c -Usual source of care was grouped into 3 categories -those reporting an office-based usual source of care, those without any usual source of care, and those using the ED as the usual source of care. d -Sample limited to patients reporting heart disease, stroke, diabetes, or hypertension (n=5,611). e -Sample limited to patients reporting a history of diabetes (n=2,213). 2.0 -6.9* -1.8

SOURCE:
From a telephone survey of U.S. citizens ages 19-64, with family income below 138% of the federal poverty level, in Arkansas, Kentucky, and Texas. The sample contained 10,885 adults (minus item non-response for each specific outcome), except where otherwise noted below. NOTES: *** p<0.01, ** p<0.05, *p<0.10 ED = Emergency Department. Results show differences-in-differences estimates for expansion states (Arkansas and Kentucky) versus Texas, by year. All analyses adjusted for sex, age, race/ethnicity, marital status, family size, education, income, urban vs. rural residence, county annual unemployment rate, state, and year.
a -All estimates are reported as percentage-point changes for binary outcomes, other than number of office and ED visits and out-of-pocket spending.
b -Usual source of care was grouped into 3 categories -those reporting an office-based usual source of care, those without any usual source of care, and those using the ED as the usual source of care. c -Sample limited to patients reporting heart disease, stroke, diabetes, or hypertension (n=5,611). d -Sample limited to patients reporting a history of diabetes (n=2,213). e-Sample limited to patients reporting at least one of the following conditions: hypertension, heart attack/coronary artery disease, stroke, asthma/COPD, kidney disease, diabetes, depression, cancer, and substance abuse (n=7,734). a -All estimates are reported as percentage-point changes for binary outcomes, other than number of office and ED visits and out-of-pocket spending. b -Sample limited to patients (n=7,734 adults) reporting at least one of the following conditions: hypertension, heart attack/coronary artery disease, stroke, asthma/COPD, kidney disease, diabetes, depression, cancer, and substance abuse.

Exhibit 4: Instrumental Variables (IV) Analysis -Individual-Level Change
c -Usual source of care was grouped into 3 categories -those reporting an office-based usual source of care, those without any usual source of care, and those using the ED as the usual source of care. d -Sample limited to patients reporting heart disease, stroke, diabetes, or hypertension (n=5,611). e -Sample limited to patients reporting a history of diabetes (n=2,213).  respectively. The error terms, ε icst , were assumed to be correlated between individuals within counties, which we addressed using robust clustered standard errors. For the 1.9% of our sample that did not report a county of residence, we created a residual county for each state (e.g.

Regression Equation -Year-by-Year Difference-in-Differences Model (Exhibit 3)
"missing county in Arkansas") that served as its own unit for clustering. We used linear regression models to allow for easy interpretation of the magnitude of the interaction terms β 7β 9 .

Regression Equations -Instrumental Variables (Exhibit 4)
1 st Stage:   Table 1. The IV model was implemented using the "ivregress 2sls" command in Stata. Table 5) to-year differences in effect sizes.

Regression Equations -Medicaid Expansion in Kentucky vs. Private Option in Arkansas
(Appendix Table 6) Our comparison between Arkansas and Kentucky used the following regression equation: where the terms were all defined as in Equation 1, but with the replacement of the Expansion State * Year variables (which pooled the two expansion states together) with two sets of interaction terms in β 7 -β 9 (which captures the year-by-year changes in Arkansas, compared to Texas) and β 10 -β 12, (which captures the year-by-year changes in Kentucky, compared to Texas).
We then used a post-estimation Wald test to determine whether β 9 and β 12 (the 2016 estimates) differed significantly from one another; these results are reported in the last column of Appendix Table 6. Comparisons between Arkansas and Kentucky in 2014 and 2015 have been published previously -see references 9 and 41.

Spatial Correlation Model
As an alternative to our primary model using county-based robust clustered standard errors, we also test the effects of using a spatial correlation model described by Conley (1999), 1 and following Mazumder and Miller (2016). 2 This approach measures the distance between each county's centroid and uses that distance to account for geographical correlation in outcomes; for observations with missing county identifiers (1.9%), we used the state's overall centroid for this calculation. We adapted Stata code from Hsiang (2010) to implement this model. 3 One limitation of this approach is that it does not allow for the use of survey weights, so while it provided a sensitivity test of our standard errors, it also affected our point estimates due to the lack of weighting. We used a spatial correlation cutoff of 100 kilometers and a time lag of 4 years to allow for correlation across all years of our study.

Family-Wise P-Values
To account for multiple hypothesis testing within each domain of our survey, we conducted the step-down bootstrapping approach developed by Westfall and Young (1993), 4 adapting Stata code written and made available by Finkelstein and colleagues (2012). 5 We used 1000 iterations within each family to generate the p-values listed in Appendix Tables 4 and 5.

Survey Outcome Measurement: Health Insurance and Out-of-Pocket Spending
Respondents were asked to answer "yes/no" to whether they currently had any of seven care, employer or union coverage, insurance from a health insurance marketplace (with state specific name), health plan purchased directly from an insurance company, and other insurance.
Respondents answering no to all seven were asked a confirmation question, "Does this mean you have no health insurance of any kind?" before being classified as "uninsured." Respondents who indicated having more than one type of coverage were assigned a primary insurance type based on the following hierarchy: Medicaid, Medicare, employer-sponsored insurance, Marketplace coverage, non-Marketplace direct purchase, and other. The exception to this hierarchy was for individuals in Arkansas and Texas who reported both marketplace and Medicaid coverage. In these states, which did not expand Medicaid but did offer marketplace coverage to some (Texas) or all (Arkansas) adults with incomes below 138% of the federal poverty level, such individuals were classified as having marketplace insurance.
Out-of-pocket spending was asked in 6 categories: <$50, $50-$99, $100-$199, $200-$499, $500-$999, and $1000 or more. We took the midpoint of each bin and top-coded the higher bin as $1250. In previous work, we have tested the sensitivity of our results to alternative approaches, including using the minimum amount in each bin, and results are largely similar.
For all regression analyses, we excluded observations with missing values for the dependent variables. Appendix

NOTES:
a -All estimates are reported as percentages for binary outcomes, other than number of office and ED visits and out-of-pocket spending.
b -Usual source of care was grouped into 3 categories -those reporting an office-based usual source of care, those without any usual source of care, and those using the ED as the usual source of care. c -Sample limited to patients reporting heart disease, stroke, diabetes, or hypertension (n=4,446). d -Sample limited to patients reporting a history of diabetes (n=1,768). e -Sample limited to patients reporting at least one of the following conditions: hypertension, heart attack/coronary artery disease, stroke, asthma/COPD, kidney disease, diabetes, depression, cancer, and substance abuse (n=6,103). c -p-values derived from county-level spatial correlation models, following Conley (1999) and using Stata code from Hsiang (2010), which does not allow for survey weights. See Appendix footnotes for additional detail.

Appendix
d -Usual source of care was grouped into 3 categories -those reporting an office-based usual source of care, those without any usual source of care, and those using the ED as the usual source of care. e -Sample limited to patients reporting heart disease, stroke, diabetes, or hypertension (n=5,611). f -Sample limited to patients reporting a history of diabetes (n=2,213). g-Sample limited to patients reporting at least one of the following conditions: hypertension, heart attack/coronary artery disease, stroke, asthma/COPD, kidney disease, diabetes, depression, cancer, and substance abuse (n=7,734). Notes: ED = Emergency Department. Results show estimated change per year of expansion in Arkansas and Kentucky, compared to the baseline time trend in Texas, which did not expand. All analyses adjusted for sex, age, race/ethnicity, marital status, family size, education, income, urban vs. rural residence, state, and year. The sample contained 10,885 adults (minus item non-response for each specific outcome).

Appendix
a -All estimates are reported as percentage-point changes for binary outcomes, other than number of office and ED visits and out-of-pocket spending.
b -Family-Wise p-values use step-down bootstrapping within each family of outcomes, following Finkelstein et al. (2012). See Appendix Methods for additional detail.
c -p-values derived from county-level spatial correlation models, following Conley (1999) and using Stata code from Hsiang (2010), which does not allow for survey weights. See Appendix footnotes for additional detail.
d -Usual source of care was grouped into 3 categories -those reporting an office-based usual source of care, those without any usual source of care, and those using the ED as the usual source of care. e -Sample limited to patients reporting heart disease, stroke, diabetes, or hypertension (n=5,611). f -Sample limited to patients reporting a history of diabetes (n=2,213). g-Sample limited to patients reporting at least one of the following conditions: hypertension, heart attack/coronary artery disease, stroke, asthma/COPD, kidney disease, diabetes, depression, cancer, and substance abuse (n=7,734).