# The Great Equalizer: Medicare and the Geography of Consumer Financial Strain

TThe Great Equalizer: Medicare and the Geography ofConsumer Financial Strain *Paul Goldsmith-Pinkham † Maxim Pinkovskiy ‡ Jacob Wallace § February 4, 2021

Abstract

We use a ﬁve percent sample of Americans’ credit bureau data, combined with a regressiondiscontinuity approach, to estimate the effect of universal health insurance at age 65—whenmost Americans become eligible for Medicare—at the national, state, and local level. We ﬁnda 30 percent reduction in debt collections—and a two-thirds reduction in the geographic varia-tion in collections—with limited effects on other ﬁnancial outcomes. The areas that experiencedlarger reductions in collections debt at age 65 were concentrated in the Southern United States,and had higher shares of black residents, people with disabilities, and for-proﬁt hospitals. * First version: April 22, 2017. This version: February 4, 2021. The views expressed are those of the authors and donot necessarily reﬂect those of the Federal Reserve Bank of New York or the Federal Reserve System. The authors wouldlike to thank Neale Mahoney, Matt Notowidigdo, Chima Ndumele, Mark Schlesinger, Julia Smith, Isaac Sorkin, EricZwick, and Trevor Gallen for helpful comments along with participants at the BU, Harvard, MIT Health EconomicsSeminar; Yale SOM Finance Lunch; Association for Public Policy Analysis and Management; Gerzensee EuropeanSummer Symposium in Financial Markets; NTA Annual Conference on Taxation; and Salomon Center/BPI Conferenceon Household Finance. Joseph Doran, Danno Lemu, Davy Perlman, and Lauren Thomas provided excellent researchassistance. † Yale School of Management. Email: [email protected] ‡ Federal Reserve Bank of New York. Email: [email protected] § Yale School of Public Health. Email: [email protected] a r X i v : . [ ec on . GN ] F e b Introduction

Why does consumer ﬁnancial strain vary so much across the United States (Keys, Mahoney andYang, 2020)? In this paper, we examine the role that health insurance plays in shaping the geogra-phy of ﬁnancial health. To do this, we use a ﬁve percent sample of Americans’ credit report data,combined with a regression discontinuity (RD) approach, to estimate the effect of universal healthinsurance at age 65—when most Americans become eligible for Medicare—at the national, state,and commuting zone level.We use our location-speciﬁc estimates of Medicare’s effects for three purposes. First, we showthat Medicare reduces geographic variation in debt collections by two-thirds at age 65. Second,we show that the gains in ﬁnancial health due to Medicare are greatest in the South, where ahigher share of the near-elderly (i.e., 55-64 year olds) are uninsured and the ﬁnancial health im-provements per newly-insured individual are largest. We show that the commuting zones (CZs)experiencing the largest gains in ﬁnancial health at age 65 had larger shares of black residents,people with disabilities, and for-proﬁt hospitals. Third, using shrinkage estimators, we constructforecasts of the causal effect of expanding coverage to the near-elderly in each of the 741 CZs inthe United States, which we use to evaluate existing policies and potential future expansions.To quantify how Medicare reduces geographic disparities in consumer ﬁnancial strain, weconstruct counterfactual estimates of consumer ﬁnancial outcomes at age 65, with and withoutMedicare, for each locality. With these location-speciﬁc counterfactuals, we construct an estimatorof the nationwide, cross-locality reduction in the variance of consumer ﬁnancial outcomes. Usingthis approach, we show that Medicare reduces the geographic variation in collections by two-thirds at age 65, highlighting an understudied aspect of the Medicare program—that it largelyeliminates geographic disparities in access to insurance (which fall by 93.5% at age 65), and sub-stantially reduces geographic disparities in collections-related ﬁnancial strain. However, we donot ﬁnd evidence that Medicare reduces geographic variation in the other ﬁnancial health out-comes we study (e.g., credit score, bankruptcy, delinquent debt), though our conﬁdence intervalsare quite wide.Second, we explore why the effect of Medicare on collections debt, where we ﬁnd a largereduction in geographic variation, differs so much across localities. We ﬁnd that reductions in1ollections debt are higher in areas that experienced larger increases in the insurance rate at age65, suggesting that the gains in ﬁnancial health are primarily driven by reducing the number ofuninsured, rather than changes in the composition of coverage. Motivated by this ﬁnding, wethen examine how CZ-level reductions in collections differ between areas on a per capita and pernewly-insured basis, with the latter done by scaling the change in ﬁnancial health estimates bythe estimated effect of Medicare on insurance rates. While our analysis suggests that a primary mechanism through which Medicare affects ﬁnancial strain is by reducing the uninsurance rate, itis unlikely that the exclusion restriction holds. As a result, we view this as an informative scalingexercise rather than an estimate of the causal effect of health insurance coverage on debt collec-tions. To understand why the effects of Medicare differ across areas, we examine the demographicand healthcare market characteristics associated with the 741 causal estimates of CZ-level reduc-tions in per capita collections debt at age 65. We ﬁnd that the effect of Medicare on collectionsdebt is larger in areas with larger shares of black residents, people with disabilities, and for-proﬁthospitals.Third, we construct forecasts of the causal effect of Medicare on ﬁnancial health in each CZ.This gives us a local approximation to the effect of lowering the Medicare eligibility age, a popularpolicy proposal. To reduce noise, we follow Chetty and Hendren (2018) and construct forecastsusing a shrinkage estimator that combines our unbiased RD estimates and a predicted effect foreach CZ based on its demographic and healthcare market characteristics. Maps of the forecastreductions in per capita and per newly-insured debt collections are strikingly similar, with thelargest values in both concentrated in the South. For example, a coverage expansion to the near-elderly is forecast to reduce collection balances by 53 dollars per capita in Raleigh, NC (one of thelargest forecast reductions). In contrast, the same treatment in San Francisco, CA, is only forecastto reduce collection balances by 8 dollars per capita. This is not simply because there are a greatershare of uninsured in Raleigh; in fact, the near-elderly uninsurance rates in the two places are Our state-level results imply a reduction in collection balances of $584 per newly-insured individual at 65, whichfalls within the range of estimates from prior work on the effects of Medicaid coverage (Finkelstein et al., 2012; Huet al., 2018). For individuals with no insurance prior to Medicare, turning 65 provides a big increase in risk protection. However,for individuals with insurance at age 65, the transition to Medicare changes premiums, beneﬁts, provider networks, andthe set of incentives their providers face (Clemens and Gottlieb (2017)). Hence, the treatment we study is a combinationof the effect of Medicare for those who were previously uninsured and those with other forms of coverage at age 64. The difference in the forecasts arises primarily due to large differences in the forecastreductions in collections balances per newly-insured individual in the two locations. In Raleigh,NC, the forecast reduction in collection balances was 956 dollars per newly-insured individual,785% higher than the analogous forecast in San Francisco, CA.Lastly, we examine how CZ-level forecasts changed due to the passage of the Affordable CareAct (ACA) in 2010, federal health reform legislation that substantially expanded coverage (e.g.,Frean, Gruber and Sommers, 2017). Forecasts of the causal effect of expanding coverage on collec-tions are smaller after the ACA’s implementation in 2014 and have become more geographicallyconcentrated in the South. This is because forecasted effects decreased by only 30% in the South(states like Mississippi, Texas, and Georgia) after the ACA’s implementation, while they decreasedby 50% in the rest of the country. Using a Kitagawa-Oaxaca-Blinder style decomposition, we showthat this was due to the uniformity of ACA coverage gains across regions for the near-elderly—despite higher rates of uninsurance in the South—and, within the South, the fact that near-elderlyuninsurance rates remain highest in areas where the ﬁnancial health gains of expanding cover-age per newly-insured are largest. These ﬁndings highlight a potential limitation of policies, suchas the ACA, that delegate states considerable latitude in policy implementation, and a relativeadvantage of programs, such as Medicare, that are federally-administered—speciﬁcally, that theformer may exacerbate geographic disparities while the latter tend to reduce them.This paper makes three primary contributions. First, we contribute to a small existing lit-erature that examines the ﬁnancial risk protection provided by Medicare to elderly Americans(Finkelstein and McKnight, 2008; Engelhardt and Gruber, 2011; Barcellos and Jacobson, 2015;Dobkin et al., 2018; Caswell and Goddeeris, 2019; Batty, Gibbs and Ippolito, 2020) and a broaderliterature on the risk protection provided by health insurance (e.g., Gross and Notowidigdo, 2011;Finkelstein et al., 2012; Mazumder and Miller, 2016; Brevoort, Grodzicki and Hackmann, 2020; Huet al., 2018) and the economic consequences of health shocks (e.g., Cochrane, 1991; Charles, 2003;Poterba, Venti and Wise, 2017; Dobkin et al., 2018; Meyer and Mok, 2019). We contribute to thisliterature in two ways. First, we examine the effect of Medicare at age 65 on a broad set of ﬁnan-cial health outcomes from administrative credit report data. These results expand the outcomes of The near-elderly uninsurance rates in Raleigh, NC and San Francisco, CA during the period 2014-2017 were 6.5 and5.9 percent, respectively.

In this section we brieﬂy describe the data we use to construct area-level ﬁnancial, demographic,health insurance, and healthcare market characteristics. Appendix A provides additional detail.4 .1 Financial outcomes data

The main dataset used in our analysis is the Federal Reserve Bank of New York’s Equifax Con-sumer Credit Panel (CCP). The CCP is a ﬁve percent random sample of all individuals in the U.S.with credit reports. The data include a comprehensive set of consumer credit outcomes quarterlyfrom 1999 to 2017, including credit scores (originating from Equifax Risk Score 3.0), unsecuredcredit lines, auto loans, and mortgages. The data also include year of birth and zip code. No otherdemographic information is available. A major virtue of the CCP is its large sample size, whichallows us to estimate the effect of Medicare separately for all 50 states and 741 commuting zonesin the country. Our datasets include ages 55-64 (“the near-elderly”) and 65-75 (“the elderly”).

For demographic and health insurance information, we draw on the American Community Sur-vey (Ruggles et al., 2019). All analyses use samples constructed from the Public Use MicrodataArea (PUMA) and state datasets, linked to the CZ- and state-level. Demographic and healthinsurance variables from the ACS allow us to test for covariate smoothness to validate our RDdesign and examine the correlates of geographic heterogeneity in our treatment effects.

We constructed additional characteristics at the PUMA-level using data from the Healthcare CostReport Information System (HCRIS) and the Dartmouth Atlas. From the HCRIS data, we con-struct PUMA-level measures of the share of hospital patient days at for-proﬁt hospitals, teachinghospitals, and public hospitals in addition to other healthcare market characteristics. From theDartmouth Atlas data, we measure the PUMA-level risk-adjusted Medicare spending per enrollee(Dartmouth Institute, 2019).

To estimate the causal impact of Medicare, we use an RD design that takes advantage of the sharpchange in eligibility at age 65. We compare individuals just above and below the age 65 eligi- Lee and Van der Klaauw (2010) show that the CCP is reasonably representative of the U.S. population. Our cross-walk from PUMA to CZ uses David Dorn’s crosswalks: . using equations of the following form: y i , l , t ( age ) = γ × ( age > ) + f ( age ) × ( age ≤ ) + g ( age ) × ( age > ) + (cid:101) i , l , t ( age ) . (1)where y i , l , t ( age ) is an outcome for individuals i in location l in time period t of a given age. Thefunctions f ( age ) and g ( age ) are the age proﬁle of y i , l , t for those below the age of 65 and thoseabove the age of 65, respectively, and control for the direct effect of age on outcomes. We denotethe national-level effect of Medicare at age 65 as γ . We denote a set of γ l , one for each location(either state or commuting zone) in our sample, as the location-level effect of Medicare at age 65,where f ( age ) and g ( age ) are allowed to vary by location.The estimation of Equation 1 is complicated by two features of our data; First, our runningvariable, age, is measured discretely by year in our sample; Second, because we only observe birthyear in the CCP data, and the data is quarterly, we measure age with noise. As a result, both theestimation and inference of f ( · ) , g ( · ) , and γ are more challenging, as error in measuring age 65forces us to omit age 65 in our estimation procedure and use a “donut” RD, and the discreteness ofthe age variable requires further extrapolation. To account for both issues, we follow the "honest"conﬁdence intervals approach outlined in Kolesár and Rothe (2018), and Armstrong and Kolesár(2018 b , a ). Brieﬂy, this approach bounds the second derivative of the true f ( · ) and g ( · ) functionsnear age 65, and uses this bound to estimate the maximum potential bias due to extrapolation.In our estimation, we report our point estimate and bias-adjusted 95% conﬁdence intervals. SeeAppendix B for additional details.Like all RD approaches, our design allows us to easily visualize the treatment effect γ usinggraphical methods. However, more uniquely in our setting, we also use our estimates to con- Commuting zones are groups of counties representing local labor markets (David and Dorn, 2013; Dorn, Hansonet al., 2019) The donut RD is a common solution to this problem in the literature (e.g., Barreca et al., 2011). l as y l ( − ) and y l ( +) , and deﬁne thecausal effect of Medicare on the variance of outcomes across locations as φ = Var ( y l ( − )) − Var ( y l ( +)) Var ( y l ( − )) ,where the variance is taken across locations. This measure captures the change in geographic vari-ance of our outcomes due to Medicare, rather than just the average level. We estimate standarderrors and bias-adjusted conﬁdence intervals for φ using the delta method following Armstrongand Kolesár (2018 a ), and report an estimated drop in variance for each of our outcomes along withbias-adjusted 95% conﬁdence intervals. See Appendix B for additional details. This section describes our approach to measuring the area-level factors associated with reductionsin consumer ﬁnancial strain at age 65 and constructing forecasts of the causal effects of Medicareby location.We begin by estimating bivariate regressions between our locality-level causal effects of Medi-care and location characteristics: ˆ γ l = α + X l ω + v l . (2)where X l is a scalar containing a single area-level characteristic (e.g. the share of black residents)and ˆ γ l is our RD estimate for location l . We separately estimate ω for each characteristic, weightingthe regression by each location’s near-elderly population.Given that many of the area-level characteristics we study are highly correlated, we re-estimateEquation 2 with the full set of area-level covariates:ˆ γ l = α + (cid:126) X l (cid:126) ω + v l . (3)where (cid:126) X l is the full set of area-level covariates. We estimate Equation 3 in two ways. First, weestimate the model using ordinary least squares (OLS) to recover the marginal association of eacharea-level characteristic with our locality-level causal estimates. Second, since the dimension of X l is large (and many of the covariates are highly correlated), we use Lasso to perform covariateselection on (cid:126) X l , and then re-estimate the model using OLS (Belloni and Chernozhukov, 2013).This Lasso procedure lets us trade off between including multiple characteristics and constructing7redictions of ˆ γ l with lower mean squared error.We next construct forecasts of the causal effect of expanding Medicare (i.e., a reduction in theMedicare eligibility age) in each commuting zone. While our RD estimates of γ l are unbiasedforecasts, many are estimated with substantial estimation error. To reduce noise, we build onChetty and Hendren (2018) and construct forecasts using a shrinkage estimator that combinesour unbiased RD estimates and a predicted effect for each commuting zone using the covariatesselected from our Lasso procedure (see Appendix B for additional details).We repeat our approach for a scaled estimate β l = γ l (cid:14) γ hl , where γ l is a location-speciﬁc esti-mate of the effects of Medicare on ﬁnancial health outcomes and γ hl is a location-speciﬁc estimateof the effects of Medicare on the insurance rate. This provides a measure of the effect of Medi-care on ﬁnancial health outcomes per newly-insured individual at each location. This helps accountfor the mechanical effect that areas with high near-elderly uninsurance rates will likely see largechanges in ﬁnancial health alongside increases in coverage. However, while we estimate β l usingfuzzy RD, it does not estimate the causal effect of insurance on ﬁnancial health, as the character-istics of health insurance coverage also change for the previously-insured as they enter Medicare(Card, Dobkin and Maestas, 2008). Figure 1 presents the effect of Medicare at age 65 at the national level and for each state. In solidred circles, we plot the average national outcome for each age. At age 65, we plot two points: inthe solid red triangle, we plot y ( − ) , the national average at age 65 without Medicare and inthe hollow red triangle we plot y ( +) , the national average at age 65 with Medicare. In gray,we repeat the same exercise for each of the states in our sample. For each outcome, we reportthe estimated national effect (with the bias-adjusted 95% conﬁdence interval) and the estimatedpercent reduction in variance across states (with the bias-adjusted 95% conﬁdence interval).We plot the share of the population with health insurance coverage in Panel A of Figure 1. Ashas been documented previously (e.g., Card, Dobkin and Maestas, 2008), the effect of Medicareeligibility on the share of individuals with any form of coverage at age 65 is large—rising by 7.98 igure 1: Changes in health insurance, ﬁnancial health, and covariates at age 65

Panel A:

Share with any coverage

Panel B:

Debt in collections

National change at 65:0.08 [0.06,0.09]Variance reduction at 65:93.2% [85.3,101.1]

Age

National change at 65:$−28.5 [−43.4,−13.7]Variance reduction at 65:67.3% [44.1,90.4]

Age

Panel C:

Credit Score

Panel D:

Bankruptcy (p.p.)

National change at 65:0.8 [−3.0,4.6]Variance reduction at 65:3.2% [−41.0,47.5]

Age

National change at 65:%−0.004 [−0.02,0.008]Variance reduction at 65:4.4% [−88.5,97.4]

Age

Panel E:

Share employed

Panel F:

Income

National change at 65:−0.03 [−0.1,0.04]Variance reduction at 65:18.1% [−240.2,276.4]

Age

National change at 65:880.7 [−2002.8,3764.2]Variance reduction at 65:−22.1% [−116.6,72.4]

Age

Note:

This ﬁgure plots the average outcomes by age at both the national level (in red) and across the ﬁfty U.S. states (in grey).The horizontal axis denotes age in years. The points plotted as circles reﬂect empirical averages from the data, while thetriangles reﬂect imputed counterfactual values at age 65 from a local linear regression from the left and right (following Kolesárand Rothe (2018) and Armstrong and Kolesár (2018 a )) with a bandwidth of 5. The imputed value without Medicare at age65 is plotted at age 64.9, while the imputed value with Medicare at age 65 is plotted at age 65.1. Panel A plots the share ofindividuals with any form of health insurance. Panel B plots the average debts in collections over the past 12 months. PanelC plots the average Equifax Risk Score 3.0. Panel D plots the share of individuals with new bankruptcies (p.p.). Panel E plotsthe share of individuals employed. Panel F plots the average total income. The sample includes individuals who were age55-75 between 2008 and 2017. See Section 2 for additional details on the outcomes and sample. Source: American CommunitySurvey, 2008-2017 and New York Fed Consumer Credit Panel/Equifax, 2008-2017. This suggests that Medicare, as expected,eliminates almost all variation across states in health insurance rates.In Panel B of Figure 1, we also estimate a large national reduction in collections debt at age65, with a sharp drop of 28.5 dollars (95% CI: -48.3 to -13.7). We also estimate a correspondingreduction of 67.3% (95% CI: 44.1%-90.4%) in the overall cross-state variance of collections debt atage 65, consistent with the drop in variance for health insurance, implying that Medicare reducedthe differences in collections debt across states by two-thirds.In contrast, for our other ﬁnancial health measures, we ﬁnd statistically insigniﬁcant effects oncredit score (Panel C) and bankruptcy (Panel D). The estimated variance reduction and nationaleffects are small with large conﬁdence bounds, suggesting that Medicare had limited effects oncross-state variance, despite large baseline differences across states. We also examine a variety ofother ﬁnancial health outcomes, including delinquent debt and foreclosure, and ﬁnd small, thoughnoisy, effects, with no corresponding reduction in national variance (Figure A2). In Appendix B wedemonstrate that these results (and lack thereof) are robust to alternative speciﬁcations (FiguresA3, A4, A5, A6, and Table A1).In Panels E and F of Figure 1, we test our key identifying assumption that non-Medicarecharacteristics that affect outcomes do not jump discontinuously at age 65. For example, giventhat many individuals tend to retire in their early to mid-60s, we test whether this coincides withthe age of eligibility for Medicare. We do not ﬁnd evidence of discontinuities in non-Medicarecharacteristics at the national or local level at age 65. For both ﬁgures, we cannot reject the nullthat the effect size is zero at the national level, and that there is no change in the variance acrossstates.In Appendix B we examine potential discontinuities in additional covariates and present ourestimates from state- and CZ-level covariate smoothness tests (Figures A7, A8, and A9). We ﬁndlittle evidence of discontinuities in the average values of covariates at age 65. Intuitively, while the Due to the asymptotic nature of these conﬁdence intervals, the 95% CI includes values larger than the maximumpossible value, 100%. We ﬁnd a decline in especially large collections balances (Figure A1), consistent with evidence that Medicare curbsthe upper tail of medical spending (Finkelstein and McKnight, 2008; Caswell and Goddeeris, 2019).

In this section, we seek to understand why the effect of Medicare on collections debt, where weﬁnd a large reduction in geographic variation, varies so much across localities.Prior to individuals gaining Medicare eligibility at age 65, we observe large differences incollections debt across areas, with particularly high levels in the South. This feature of the USlandscape of ﬁnancial health is apparent in Figure 2, where we map the commuting zone esti-mates of counterfactual collections debt ﬂows at age 65, with and without Medicare. It is clearfrom Panel A that, absent Medicare, collections debt for the near-elderly varies widely across thecountry, with low levels in the Midwest and Northeast, and with high levels of debt collectionconcentrated in the South. At age 65, we observe a large reduction in collections, concentrated inthe South. Panel B shows that much of the geographic variation in collections debt disappears atage 65, with lower (though still elevated) levels of collections debt in states like Mississippi, Texas,and Nevada. Why are collections so concentrated before Medicare and much less so after? One clear candi-date is geographic differences in the uninsurance rate for the near-elderly. Prior work documentsa link between health insurance coverage and collections debt (see e.g., Finkelstein et al., 2012),and we ﬁnd similar associations between area-level health insurance rates and ﬁnancial healthoutcomes among the near-elderly (Figure A10). We compare the estimated increase in health in-surance due to Medicare to the drop in debt collections at the state-level and CZ-level in FigureA11. This “extensive margin” effect of Medicare on coverage explains a surprisingly large shareof the variation ( R = 0.38), with small estimated reductions in collections for states with smallestimated changes in the insurance rate at age 65.These facts suggest that the effect of Medicare eligibility on debt collections may be driven byindividuals who gain coverage, rather than those whose primary source of coverage is changing. For this map, due to smaller sample sizes, we use an empirical Bayes approach to shrink each locality-level estimatetowards the overall average of the effects (see Appendix B). At the CZ-level, this reduction equates to a drop in the across-CZ variance in collections debt of 70% (very similarto our state-level estimate of 67.3%). igure 2: Counterfactual levels of collections debt by commuting zone at age 65 with andwithout Medicare

Panel A:

Without Medicare

Panel B:

With Medicare

Note:

This ﬁgure plots our counterfactual estimates of the ﬂow of newly reported collections debt (withinthe past year) per capita at age 65, with and without Medicare. The CZ-level variance reduction at age65—the difference in CZ-level variance between the two panels—was 63.8% (95% CI: 30.7%-97.0%). Thecounterfactuals are based on local linear regressions, done separately by commuting zone, using the meth-ods from Kolesár and Rothe (2018). These estimates are then shrunk using empirical Bayes, described inAppendix B. Darker regions correspond to higher counterfactual collections debt per capita. Source: Con-sumer credit outcomes are based on 137,340,577 person-year observations from the New York Fed Con-sumer Credit Panel/Equifax, 2008-2017.

Motivated by this ﬁnding, we construct a scaled version of our CZ-level estimates, β l , that mea-sures the reduction in collections debt per newly-insured , and examine it alongside our per capita estimates going forward. This allows us to compare how the effects of Medicare differ acrosslocations with different baseline levels of uninsurance among the near-elderly.We now examine what other area-level factors are associated with the reductions in collectionsdebt at age 65. We present evidence that commuting zones with larger shares of black residents,people with disabilities, and for-proﬁt hospitals experience the largest gains in ﬁnancial health atage 65, across a variety of estimation approaches.Figure 3 presents correlations between our area-level characteristics and estimated effects ofMedicare. The leftmost panel presents the coefﬁcients from separate bivariate OLS regressionsof our regression discontinuity estimates of CZ-level reductions in collections debt per capita (inred circles) and per newly-insured (in blue triangles) on CZ-level demographic and healthcare mar-ket characteristics, with bars representing the 95% conﬁdence intervals. Since many of the area-level characteristics are highly correlated, the center and right panels plot multivariate and post-Lasso analyses describing the partial correlations between the characteristics and our locality-12 igure 3: Commuting zone characteristics correlated with the reduction in collectionsdebt at age 65

Panel A:

Demographic characteristics

Bivariate Multivariate Post−Lasso−0.4 0.0 0.4 −0.4 0.0 0.4 −0.4 0.0 0.4Has Medicaid (%)Has any coverage (%)Physical disability (%)Median house valueIncome per capitaPoverty rate (%)Greater than high school education (%)Black (%)

Impact of 1 SD Change in Covariate

Per Capita Per Newly Insured

Panel B:

Healthcare market characteristics

Bivariate Multivariate Post−Lasso−0.5 −0.2 0.0 0.2 0.5−0.5 −0.2 0.0 0.2 0.5−0.5 −0.2 0.0 0.2 0.5Medicare spending per enrollee ($)Payment by charity care patients ($)Cost of charity care per patient day ($)Hospital beds per capitaHospital occupancy rate (%)Teaching hospitals (%)For−profit hospitals (%)

Impact of 1 SD Change in Covariate

Per Capita Per Newly Insured

Note:

This ﬁgure plots bivariate OLS estimates (left panel), multivariate OLS estimates (center panel), and post-Lassomultivariate estimates (right panel) of CZ-level estimated reductions in collections debt per capita on a set of CZ-levelcharacteristics. We standardize all the variables so the coefﬁcients reﬂect the strength of the association between a onestandard deviation change in the covariate and the estimated reduction in collections debt at age 65. The horizontal barsare 95% conﬁdence intervals. The multivariate OLS regression results and post-Lasso multivariate regression resultsare both run on the full set of characteristics in Panels A and B. For post-Lasso, we ﬁrst estimate a Lasso regressionon the full set of characteristics and then report the results of multivariate OLS run on the characteristics chosen bythe Lasso regression. Tabular versions of these results are in Table A4. Source: Consumer credit outcomes are basedon 137,340,577 person-year observations from the New York Fed Consumer Credit Panel/Equifax, 2008-2017. CZ-leveluninsurance rates are from the American Community Survey, 2008-2017. Healthcare market characteristics are fromthe Healthcare Cost Report Information System (HCRIS) and the Dartmouth Atlas. For additional details on the datasee Section 2. per capita and per newly-insured measures, we standardize all the area-level correlates, and thendivide the coefﬁcients by the respective national per capita or per newly-insured estimates. Hence,plotted coefﬁcients in Figure 3 correspond to the effect of a one standard deviation change in thecovariate on a percentage change in the reduction in debt collections at age 65. Hence, a coefﬁ-cient of 0.1 implies a 10% increase in the effect of Medicare in reducing debt collections from a onestandard deviation change in the covariate.Panel A of Figure 3 shows that the share of high school graduates, income per capita, andmedian house values in a CZ were all associated with smaller reductions in per capita collectionsdebt at age 65. Unsurprisingly, the share of residents with health insurance (or Medicaid), wasalso associated with a smaller reduction in per capita collections debt at age 65. The share ofblack residents, the poverty rate, and the share of people with disabilities, on the other hand,were associated with larger reductions in collections at age 65. In multivariate and post-Lassoanalyses, only the near-elderly health insurance rate and the high school graduation rate wereassociated with smaller reductions in per capita collections, while the share of black residents andpeople with disabilities in a CZ were consistently associated with larger reductions in per capitacollections. Including Census region or division ﬁxed effects—and using only the within-area,across-CZ variation that remains—does not qualitatively change our results (Figure A12). Oncewe restrict to only using variation across CZs, but within states, however, the association betweenthe share of black residents and per capita reductions in collections is severely attenuated.Panel B of Figure 3 presents the healthcare market characteristics correlated with our esti-mates. In the bivariate OLS model, a higher share of for-proﬁt hospitals and higher risk-adjustedspending per Medicare beneﬁciary were both associated with larger reductions in collections debtat age 65, while a higher share of teaching hospitals and higher hospital occupancy rates wereassociated with smaller reductions in collections at age 65. In multivariate OLS and post-Lassoanalyses, only the CZ-level share of for-proﬁt hospitals was associated with our causal CZ-level While the Lasso procedure did not select the near-elderly health insurance rate, we note a very high correlationbetween that measure and the high school graduation rate ( ρ = per capita collections at 65. Unlike not-for-proﬁt hospitals, for-proﬁtsare not required to provide charity (discounted or free) care and evidence suggests that they offerless charity care than not-for-proﬁt hospitals (e.g., Horwitz, 2005; Schlesinger and Gray, 2006; Val-dovinos, Le and Hsia, 2015). Figure A12 demonstrates that the relationship between for-proﬁthospital share and CZ-level reductions in per capita collections debt at age 65 is robust to theinclusion of ﬁxed effects for census regions or divisions, but not states.Figure 3 also examines the demographic and healthcare market characteristics associated withreductions in collections debt per newly-insured . This exercise accounts for the change in the unin-surance rate at age 65, to identify CZs that experienced larger or smaller reductions in debt col-lections not mechanically driven by Medicare’s “extensive margin” effect on coverage. The shareof people with disabilities and for-proﬁt hospitals in a CZ is consistently associated with largerreductions in collections debt per newly-insured. In addition, CZs with a larger share of blackresidents—where we see large reductions in per capita collections debt—also appear to experiencelarger reductions in collections per newly-insured . The other area-level characteristics were notconsistently associated with the estimated reductions in collections debt per newly-insured.

Given that the effect of Medicare varies substantially across localities, where would the effectsof a broad expansion of coverage to the near-elderly (i.e., by lowering the Medicare eligibilityage) be the largest? In what follows, we discuss our forecasts of CZ-level causal effects, how thoseforecasts have changed post-ACA, and their implications for future potential coverage expansions.In Panel A of Figure 4, we map the per capita mean-square error minimizing forecast causaleffects, ˆ γ fl , across CZs for the near-elderly, with darker colors depicting areas predicted to ex-perience larger reductions in collections-related strain associated with an expansion of (nearly)universal health insurance to the near-elderly. The largest forecast reductions are concentrated inthe South, ranging from $20-$50 in most CZs. The opposite is true in the Midwest, where fore-cast reductions in consumer ﬁnancial strain are small across all CZs. In Panel B, we map the In addition, hospitals in markets with a higher share of for-proﬁts respond to competition by reducing charity careand trying to avoid the uninsured (Frank, Salkever and Mitchell, 1990). Table A2 lists the forecasts for the 50 commuting zones with the largest near-elderly populations (accounting for53.2% of the near-elderly US population during our sample period). igure 4: Forecasts of causal reductions in collections debt from expanding health insur-ance to the near-elderly by commuting zone

Panel A:

Per capita, 2008-2017

Panel B:

Per newly insured, 2008-2017

Note:

This ﬁgure plots mean square error (MSE)-minimizing forecasts of the reductions in collections debtper capita (Panel A) and the reduction in collections debt per newly-insured (Panel B). We construct theMSE-minimizing forecasting by ﬁrst running a Lasso regression to predict the CZ-level reductions in col-lections debt per capita (or per newly-insured) as discussed in Section 3. This generates a prediction foreach CZ, which we call ˆ γ l . Following Chetty and Hendren (2018) we then combine the ˆ γ l estimates withour estimates of γ l to construct the MSE-minimizing forecast for each commuting zone, γ fl . Source: Con-sumer credit outcomes are based on 137,340,577 person-year observations from the New York Fed Con-sumer Credit Panel/Equifax, 2008-2017. CZ-level uninsurance rates are from the American CommunitySurvey, 2008-2017. Healthcare market characteristics are from the Healthcare Cost Report Information Sys-tem (HCRIS) and the Dartmouth Atlas. For additional details on the data see Section 2. forecasts per newly-insured at age 65, ˆ β fl . Despite large geographic differences in the near-elderlyuninsurance rate, the maps are strikingly similar, with the largest forecast reductions in collec-tions debt per newly-insured also concentrated in the South. This is consistent with the Lassoprocedure selecting similar area-level characteristics when predicting changes in per capita andper newly-insured debt collections at age 65. We next examine how these forecasts have changed due to the ACA, federal health reformlegislation that substantially expanded coverage (Frean, Gruber and Sommers, 2017). Panel A ofFigure 5 presents the forecasts using the sample before and after the implementation of the ACA.On the x-axis, we plot the pre-ACA per capita forecast reductions in collections, and on the y-axis, the post-ACA per capita forecast reductions. The forecast reductions post-ACA are generallysmaller than pre-ACA, which results in the majority of commuting zones below the 45-degree line.Rather than having a uniform effect across CZs on collections forecasts, which would appear as avertical shift in the cloud of points, the ACA led to a “rotation” in the forecasts; CZs with larger We plot the relationship between the two forecasts across CZs in Figure A13 and ﬁnd an R of 0.8213. igure 5: Forecasts of causal reductions in collections debt at the commuting zone levelbefore and after the Affordable Care Act (ACA)

Panel A: Pre and Post-ACA Per Capita Forecasts Panel B: Decomposing Changes in Forecasts

Slope = 0.59Slope = 0.250306090 0 30 60 90

Pre−ACA Effects

SouthAll Others

Post−ACA Effects

Note:

Panel A of this ﬁgure plots mean square error (MSE)-minimizing forecasts of the CZ-level reductions in collectionsdebt per capita in the pre-ACA period on the x-axis against the analogous post-ACA period forecast on the y-axis. Weseparately plot commuting zones in the South in orange circles and all other commuting zones with blue triangles.Fitted lines are constructed using bivariate OLS. We construct the MSE-minimizing forecasting by ﬁrst running a Lassoregression to predict the CZ-level reductions in collections debt per capita separate for each period as discussed in Section3. This generates a prediction for each CZ in each period, which we call ˆ γ l . Following Chetty and Hendren (2018) wethen combine the ˆ γ l estimates with our estimates of γ l to construct the mean square error-minimizing forecast for eachcommuting zone in each period, γ fl . Panel B of this ﬁgure plots the average percentage decline in per capita forecastedeffects for the South and non-South regions, and the decomposition described in the text and Appendix B. The top barspresent the average percentage decline in per capita forecasted effects for the South and non-South regions. The secondset of bars present the decline due to the change in insurance coverage rates, holding ﬁxed the per newly-insured effect.The third set of bars present the decline due to the change in the per newly-insured effect, holding ﬁxed the impact oninsurance coverage rates. The last set of bars is due to the change in covariance between the per newly-insured effect andthe effect on insurance. In curly braces for each set of bars we report the difference between the non-South and Southaverages. Source: Consumer credit outcomes are based on 137,340,577 person-year observations from the New YorkFed Consumer Credit Panel/Equifax, 2008-2017. CZ-level uninsurance rates are from the American Community Survey,2008-2017. For additional details on the data see Section 2. pre-ACA forecasts experienced larger changes in forecast pre- to post-ACA. This is consistent withthe increase in health insurance coverage due to the ACA (Frean, Gruber and Sommers, 2017).However, the degree of rotation varied signiﬁcantly by geography. The forecasts fell less in theSouth than elsewhere. As a result, the effect of Medicare on collections-related ﬁnancial strainhave become much more geographically-concentrated in the South, and particularly the “DeepSouth” region comprised of Louisiana, Alabama, Mississippi, Georgia, South Carolina, and partsof Texas and Florida (Figure A14).Why did the forecast reductions in collections fall less in the South than elsewhere? Panel Bof Figure 5 documents that average CZ-level forecasts decreased by only 30% in the South after17he ACA’s implementation, but 50% elsewhere. We decompose the differential change in forecasts(i.e., the 30% vs. 50%) using a Kitagawa-Oaxaca-Blinder style decomposition (Kitagawa, 1955;Oaxaca, 1973; Blinder, 1973). Intuitively, this decomposes the change in the forecast reductionsin collections debt per capita into changes in the extensive margin “any health insurance” effect,changes in the intensive margin “reduction in collections per newly-insured” effect, and a residualterm that reﬂects the changing covariance between the two (see Appendix B for additional details).We ﬁnd that the differential change in forecasts between regions was driven by all three pieces.First, uniform gains in ACA coverage across regions for the near-elderly (Figure A16)—despitehigher pre-existing rates of uninsurance in the South—led to smaller reductions, percentage-wise,in uninsurance rates in the South as compared to elsewhere. This accounted for one-quarter ofthe differential change in forecasts between regions (0.051/0.2). Second, the forecast reductions incollections debt per newly-insured increased in the South (where they were larger to begin with)after the ACA, but remained unchanged elsewhere (Figure A17). This accounted for two-ﬁfthsof the differential change in forecasts (0.083/0.2). Lastly, in the non-south CZs, the covariancebetween the per newly-insured effect and the effect on the share of individuals covered decreased post-ACA, whereas it was unchanged in the South. This suggests that the ACA expansions in theSouth were not as well targeted (on this dimension) as those elsewhere in the country. The poorertargeting explained the remaining third (0.065/0.2) of the differential change in forecasts betweenregions.

This paper examines the relationship between health insurance and ﬁnancial health by studyingﬁnancial outcomes for individuals as they age onto Medicare at 65. We ﬁnd a 30 percent reductionin debt collections—and a two-thirds reduction in the geographic variation in collections—at age65, with limited effects on other ﬁnancial outcomes. Areas that experienced larger gains in ﬁnan-cial health at age 65 had higher shares of black residents, people with disabilities, and for-proﬁthospitals.Our data suggest that the ﬁnancial health beneﬁts of potential future coverage expansions to This left the uninsured near-elderly population more concentrated in the South after the implementation of theACA (Figure A15). eferences

Armstrong, Timothy, and Michal Kolesár. a . “Simple and honest conﬁdence intervals in non-parametric regression.” Cowles Foundation Discussion Paper. Armstrong, Timothy B, and Michal Kolesár. b . “Optimal inference in a class of regressionmodels.” Econometrica , 86(2): 655–683.

Barcellos, Silvia Helena, and Mireille Jacobson.

American Economic Journal: Economic Policy , 7(4): 41–70.

Barreca, Alan I, Melanie Guldi, Jason M Lindo, and Glen R Waddell.

The Quarterly Journal of Economics ,126(4): 2117–2123.

Batty, Michael, Christa Gibbs, and Benedic Ippolito.

AEI Economics Working Paper , , (June).

Belloni, Alexandre, and Victor Chernozhukov.

Bernoulli , 19(2): 521–547.

Bhattacharya, Jay, and Darius Lakdawalla.

Journal ofPublic Economics , 90(1-2): 277–292.

Blinder, Alan S.

Journal ofHuman resources , 436–455.

Brevoort, Kenneth, Daniel Grodzicki, and Martin B Hackmann.

Journal of Public Economics , 187: 104203.

Calonico, Sebastian, Matias D Cattaneo, and Rocio Titiunik.

R Journal , 7(1): 38–51.

Card, David, Carlos Dobkin, and Nicole Maestas.

American Economic Review ,98(5): 2242–58.

Caswell, Kyle J., and John H. Goddeeris.

AmericanJournal of Health Economics , 0(ja): null.

Cattaneo, Matias D, Richard K Crump, Max H Farrell, and Yingjie Feng. arXiv preprint arXiv:1902.09608 . Charles, Kerwin Koﬁ.

Journal of human Resources , 38(3): 618–646.

Chetty, Raj, and Nathaniel Hendren.

The Quarterly Journal of Economics , 133(3): 1163–1228.

Chetty, Raj, Nathaniel Hendren, Patrick Kline, and Emmanuel Saez.

The QuarterlyJournal of Economics , 129(4): 1553–1623. 20 lemens, Jeffrey, and Joshua D Gottlieb.

Journal of Political Economy , 125(1): 000–000.

Cochrane, John H.

Journal of political economy ,99(5): 957–976.

Cooper, Zack, Stuart V Craig, Martin Gaynor, and John Van Reenen.

The Quarterly Journal of Eco-nomics , 134(1): 51–107.

Dartmouth Institute.

David, H, and David Dorn.

American Economic Review , 103(5): 1553–97.

Dobkin, Carlos, Amy Finkelstein, Raymond Kluender, and Matthew J Notowidigdo.

American Economic Review , 108(2): 308–52.

Dorn, David, Gordon Hanson, et al.

American Economic Review: Insights , 1(2): 161–78.

Duggan, Mark, Atul Gupta, and Emilie Jackson.

Engelhardt, Gary V, and Jonathan Gruber.

American Economic Journal: Economic Policy , 3(4): 77–102.

Finkelstein, Amy, and Robin McKnight.

Journal of public economics ,92(7): 1644–1668.

Finkelstein, Amy, Matthew Gentzkow, and Heidi L Williams.

Finkelstein, Amy, Matthew Gentzkow, and Heidi Williams.

The quarterly journal of economics ,131(4): 1681–1726.

Finkelstein, Amy, Sarah Taubman, Bill Wright, Mira Bernstein, Jonathan Gruber, Joseph PNewhouse, Heidi Allen, Katherine Baicker, and Oregon Health Study Group.

The Quarterly journal of eco-nomics , 127(3): 1057–1106.

Frandsen, Brigham R, Markus Frölich, and Blaise Melly.

Journal of Econometrics , 168(2): 382–395.

Frank, RG, DS Salkever, and J Mitchell.

Advances in health economics and health servicesresearch , 11: 159.

Frean, Molly, Jonathan Gruber, and Benjamin D Sommers.

Journal of HealthEconomics , 53: 72–86. 21 ross, Tal, and Matthew J Notowidigdo.

Journal of Public Economics , 95(7): 767–778.

Horwitz, Jill R.

Health affairs , 24(3): 790–801.

Hu, Luojia, Robert Kaestner, Bhashkar Mazumder, Sarah Miller, and Ashley Wong.

Journal of publiceconomics , 163: 99–112.

Imbens, Guido, and Stefan Wager.

Review ofEconomics and Statistics , 101(2): 264–278.

Keys, Benjamin J, Neale Mahoney, and Hanbin Yang.

Kitagawa, Evelyn M.

Journal of the americanstatistical association , 50(272): 1168–1194.

Kolesár, Michal, and Christoph Rothe.

American Economic Review , 108(8): 2277–2304.

Lee, David S, and David Card.

Journal of Econometrics , 142(2): 655–674.

Lee, Donghoon, and Wilbert Van der Klaauw.

Mazumder, Bhashkar, and Sarah Miller.

American Economic Journal: Economic Policy , 8(3): 284–313.

McClellan, Mark, and Jonathan Skinner.

Journal of PublicEconomics , 90(1-2): 257–276.

Meyer, Bruce D, and Wallace KC Mok.

Journal of Public Economics , 171: 51–69.

Miller, Sarah, and Cindy K Soo.

Morris, Carl N.

Journal ofthe American statistical Association , 78(381): 47–55.

Oaxaca, Ronald.

International eco-nomic review , 693–709.

Poterba, James M, Steven F Venti, and David A Wise.

TheJournal of the Economics of Ageing , 9: 172–184.

Ruggles, Steven, Sarah Flood, Ronald Goeken, Josiah Grover, Erin Meyer, Jose Pacas, andMatthew Sobek. chlesinger, Mark, and Bradford H Gray.

Health affairs , 25(Suppl1): W287–W303.

Valdovinos, Erica, Sidney Le, and Renee Y Hsia.

Health Affairs ,34(8): 1296–1303. 23 or Online PublicationAppendix for:

The Great Equalizer: Medicare and the Geography of ConsumerFinancial StrainA Study data

A.1 Financial outcomes data

The main dataset used in our analysis is the Federal Reserve Bank of New York’s Equifax Con-sumer Credit Panel (CCP). The CCP is a ﬁve percent random sample of all individuals in the U.S.with credit reports. The CCP data is a representative sample of all individuals with a credit reportbut it does not include the roughly 11 percent of the U.S. population without credit reports. Asa result, the CCP data is more representative for high-income individuals than for low-incomeindividuals, and it is more representative for older than younger people. Lee and Van der Klaauw(2010) show that the CCP is reasonably representative of the U.S. population with the possible ex-ception of very young adults, suggesting that sample representativeness should not be a concernin our application.The data include a comprehensive set of consumer credit outcomes from quarterly from theﬁrst quarter of 1999 to the fourth quarter of 2017, including information on credit scores (origi-nating from Equifax Risk Score 3.0), unsecured credit lines, auto loans, and mortgages. The dataalso include year of birth and precise geographic location at the census block level. No otherdemographic information is available at the individual level.A major virtue of the CCP is its large sample size, which allows us to measure ﬁnancial out-comes at granular geographic levels with precision. This is key to our RD estimation strategyacross geographies, where we estimate the effect of Medicare separately for 50 states and 741commuting zones in the country. For our analyses, we aggregate the data to locality-by-age-by-year cells and weight by the underlying population in each cell. Since we only observe birth year,and the data is quarterly, age is measured with noise. For example, all individuals with birth year1940 are measured as age 65 in the ﬁrst quarter of 2005, while in reality some of these individualswill turn 65 later in the year. We address this using a “donut” RD procedure which we discuss inmore detail in Section 3.The ﬁnancial health variables that we focus on from the CCP are the size of accounts sent tocollection agencies (usually, these are accounts that have been delinquent for over 90 days), thesize of accounts that are delinquent, the Equifax Risk Score, as well as additional ﬁnancial healthoutcomes (e.g., bankruptcy). Refer to Appendix Table A3 for the deﬁnitions of each of the ﬁnancialhealth outcomes we use from the CCP.We examine the impact of Medicare on the distributions of three of our outcome measures:amount of debt in collections, total amount of debt in delinquency, and amount of credit card debtin delinquency. For all three of these cases, we would expect that large out-of-pocket expenseswould cause increases in the right tail of the distribution. To examine this, we calculate the shareof the population in a county-year-age bin that has amounts in the following bins: 1-500, 501-1,000,1,001-2,500, 2,501-5,000, 5,001-10,000, and greater than 10,000 dollars. The residual category is anyperson with 0 dollars. We use the share within a given bin as the outcome variable in our mainspeciﬁcation, so that our estimate is the change in the relative share of individuals within each bindue to Medicare eligibility. An alternative approach would be to directly estimate quantile treatment effects using regression discontinuity, .2 Demographic and health insurance data For demographic and health insurance information, we draw on the American Community Sur-vey (Ruggles et al., 2019). All analyses use samples constructed from the PUMA and state datasets,linked to the Commuting Zone (CZ)- and state-level. Our cross-walk from PUMA to CommutingZones uses David Dorn’s crosswalks (available here: ). Demographic data . We construct demographic variables from the ACS at the PUMA-by-age-by-year level and then crosswalk to the CZ- and state-level to test for covariate smoothness invalidating our RD design and to examine the correlates of geographic heterogeneity in our treat-ment effects. From the ACS, we measure the homeownership rate, marital status, race, educationalattainment, employment status, usual hours worked per week, total personal income, social secu-rity income, poverty status, and disability rate.

Health insurance data . The ACS also allows us to construct health insurance variables from theACS at the PUMA-by-age-by-year level to test for changes in health insurance at age 65 and toexamine the correlates of geographic heterogeneity in our treatment effects. From the ACS, wemeasure the share of individuals in each cell with any health insurance coverage.

A.3 Additional area-level characteristics data

In addition to the CCP and ACS data, we constructed additional characteristics at the PUMA-level.These characteristics are drawn from several places, including the Healthcare Cost Report Infor-mation System (HCRIS) and the Dartmouth Atlas. From the HCRIS data, we construct PUMA-level measures of the share of hospital patient days at for-proﬁt hospitals, teaching hospitals, andpublic hospitals. We also measure the average hospital occupancy rate at the PUMA-level and thehospital beds per capita. In addition, for the period 2010-2017 we measure PUMA-level reportedcharity care costs per patient day and payments recovered by hospitals from charity care patientsby patient day. From the Dartmouth Atlas data, we measure the PUMA-level risk-adjusted Medi-care spending per enrollee (Dartmouth Institute, 2019).

B Additional Methods and Robustness

B.1 Honest RD and Shrinkage Estimators

Robustness using Honest RD

As discussed in Section 3, we account for discreteness and mea-surement error in our running variable, age, by following the "honest" conﬁdence intervals ap-proach outlined in Kolesár and Rothe (2018), and Armstrong and Kolesár (2018 b , a ). This methodrequires an additional tuning parameter, K , which imposes an upper bound on the absolute valueof the second derivative of the conditional expectation function. Intuitively, this method places abound on how quickly the functions f ( · ) and g ( · ) can change. To choose our value of K for ourmain estimates, we follow an approach similar to the approach advocated in Imbens and Wager(2019). We take a large window to the left of the RD cutoff and ﬁt a quadratic function of age tothe data. We take the coefﬁcient on the quadratic term (the second derivative), take the absolutevalue, and multiply it by four. We take this as our estimate of K . Similar to robustness exerciseswith bandwidths in previous RD methods, we present additional robustness tests which vary thevalue of K by changing the number that we scale this second derivative by in Appendix FiguresA3 and A5. such as those proposed in Frandsen, Frölich and Melly (2012). However, we are not able to easily account for thediscrete running variable in our estimation process using quantile treatment effects. As a result, we focus on our share-based approach. nference for variance reduction Our estimate of cross-state and cross-CZ variance reductiondue to the Medicare is a non-linear functional of the different estimates of a local non-parametricestimate. Speciﬁcally, we are interested in T ( f , g ) = φ ( f , g ) , where φ ( f , g ) = − (cid:0) g ( ) (cid:48) g ( ) − ( L − g ( ) (cid:48) ι ) ) (cid:1) ( f ( ) (cid:48) f ( ) − ( L − f ( ) (cid:48) ι ) )) = − L − (cid:16) ∑ l g l ( ) − (cid:0) L − ∑ m g m ( ) (cid:1) (cid:17) L − (cid:16) ∑ l f l ( ) − ( L − ∑ m f m ( )) (cid:17) ,where f and g are the vector of functions f l and g l that are estimated using local linear regres-sion and ι is an L × To construct conﬁdence intervals for this estimate that correctly account for the discretenessof our outcome variable, we apply the delta method following Appendix B.1.1 of Armstrong andKolesár (2018 a ). Let the numerator and denominator of φ be A and B , respectively. Then, dA / dg = g ( ) (cid:48) − ( L − g ( ) (cid:48) ι ) L − ι (cid:48) and dB / d f = f ( ) (cid:48) − ( L − f ( ) (cid:48) ι ) L − ι (cid:48) . The cross-derivatives arezero. Hence, d φ ( f , g ) / d f = − d ( AB − ) / d f = − ( dA / d f ) B − and d φ ( f , g ) / dg = − d ( AB − ) / dg = − A ( dB − ) / dg = A ( B − ) dB / dg .Thus, our bias term will be B = ∑ l | φ (cid:48) l B l | , where B l is the bias determined from the underlyingestimation.We next consider the covariance matrix Σ of our stacked f and g . Since we estimate each f l and g l separately, Σ is simply a diagonal matrix of the S l estimates for each f l ( ) and g l ( ) estimate.Hence, our variance estimate is S = φ (cid:48) ( f , g ) (cid:48) Σ φ (cid:48) ( f , g ) .Finally, to calculate the conﬁdence intervals around our estimate ˆ T ( f , g ) , we follow Arm-strong and Kolesár (2018 a ) and calculate the 95% conﬁdence intervals around our estimate of theratio as ˆ T ( f , g ) ± cv ( t ) · ˆ se ,where t = B / S is our bias-sd ratio and ˆ se = √ S . We note that cv ( t ) is the quantile of thefolded normal distribution with mean equal to t (see the note in Table 1 of Armstrong and Kolesár(2018 a )). Shrinkage Estimators

Due to smaller sample sizes, the locality-level estimates are noisier thanestimates of the overall national effects (or counterfactuals). Hence, our estimates of γ l , y l ( − ) ,and y l ( +) have more inherent noise and variation than the true underlying estimates due toestimation error (in part due to smaller sample sizes). Here we provide additional details on theshrinkage estimator we use to address this.Formally, using our estimates of state-level discontinuities as an example, we calculate theshrinkage estimator by assuming that the γ s ∼ N ( γ , σ ) . We estimate these two parametersdirectly. Then, using the standard errors estimated for each γ s , ˆ σ s , and following the standardJames-Stein estimator approach (Morris, 1983), we construct B = ˆ σ ˆ σ s + ˆ σ , and our shrinkage esti-mator is ˜ γ s = B ˆ γ s + ( − B ) γ . The CZ-level counterfactuals in Figure 2, for example, are shrunkusing this method. B.2 Robustness checks

In this section, we discuss our approach to assessing the robustness of our results to alternativespeciﬁcations and bandwidths. As discussed in Section B, our empirical methodology requirestwo tuning parameters: the bandwidth (standard regression discontinuity applications) and our Note that we additionally population-weight our estimates. We omit this notation here for simplicity’s sake.

RDRobust package from Calonico, Cattaneo and Titiunik (2015). Wereport this estimate in Column 8.Our estimates are similar across various estimation methodologies. However, many outcomesdo appear statistically signiﬁcant when we cluster on the running variable, unlike in our mainspeciﬁcation. This is likely due to incorrect coverage, as highlighted in Kolesár and Rothe (2018).When using heteroskedasticity-robust standard errors, there are also additional signiﬁcant esti-mates, but fewer, and they are not consistent across various parametric forms. Our estimates arequite similar, qualitatively, to using the Calonico, Cattaneo and Titiunik (2015) method, but ourpreferred estimate’s point estimate is larger in magnitude and the conﬁdence interval is smallerfor our collection estimates.

B.3 Forecasting the causal effects of Medicare by location

This section provides additional details about how we forecast the causal effects of Medicare bylocation. We are interested in the forecastable components of both γ l and β l , where β l is our fuzzy-RD estimates of CZ-level reductions in collections debt per newly-insured (see Section 3 for details).We are interested in the best predictions of γ l and β l .Ideally, each forecast would be the unbiased causal estimate for the location from our RD de-sign. However, in many locations, the near-elderly population is small and the estimates are noisy.To reduce noise, we follow Chetty and Hendren (2018) and construct forecasts using a shrinkageestimator that combines our unbiased RD estimates and the predicted effect for each commutingzone based on its demographic and healthcare market characteristics. Since the dimension of ourset of predictors, (cid:126) X l , is large (and many of the covariates are highly correlated), we use our Lassopredictions from Section 3.2 in order to minimize over-ﬁtting.We denote our predictions of γ l and β l estimated using our Lasso model as ˆ γ l and ˆ β l , re-spectively. Brieﬂy, the Lasso estimation procedure penalizes covariates and shrinks terms in theestimated ω l towards zero, in order to minimize mean squared error. As a result, the estima-tion procedure will select a subset of the covariates in (cid:126) X l , to have non-zero parameters, and setthe remaining parameters to zero. We implement this using a ten-fold cross-validation over thepenalization parameter, implemented using R glmnet package.To forecast the causal effects of Medicare by location, we then combine the Lasso estimatestogether with our RD estimates of γ l order to construct the mean square error-minimizing forecastfor each location, deﬁned as ˆ γ fl . This MSE-minimizing forecast is constructed using the following4ormula (Chetty and Hendren, 2018): ˆ γ fl = (cid:18) χ χ + s l ( γ l − γ l ) + s l χ + s l τ ( ˆ γ l − ˆ γ l ) (cid:19) + γ l , (4)where γ l is the average RD prediction across locations, ˆ γ l is the average Lasso prediction acrosslocations, τ = Cov ( ˆ γ l , γ l ) (cid:14) Var ( ˆ γ l ) is the coefﬁcient of a regression of γ l on ˆ γ l , χ is the residualplace effect variation after subtracting off the variance due to estimation of γ l , and s l the squaredstandard error of the γ l . For the purposes of the shrinkage, we demean our estimates and then addthe overall mean back, such that the shrinkage is around the variation around the overall mean.We estimate τ using linear regression of the demeaned values, and calculate χ as χ = Var ( γ l − τ ( ˆ γ l − ˆ γ l )) − E ( s l ) ,where E ( s l ) is the average sampling variance across locations. In all calculations, we weight bythe precision of the ﬁxed effect estimates (1/ s l ) to maximize efﬁciency.Note that this approach will shrink our estimates towards the predicted ˆ γ l when the originalestimate is noisy and the shrinkage will only occur if the lasso prediction has predictive power for γ l . If this prediction has limited value, then τ will be zero, and the shrinkage will shrink towardsthe overall mean. By a similar argument, as s goes to zero, the forecasted estimate will be exactly γ l . We follow the same procedure to construct forecasts for β l , deﬁned as ˆ β fl .To calculate the prediction errors of the forecasts for Table A2, we follow Chetty and Hendren(2018), where the root mean-squared error of the prediction is: (cid:113) e l = (cid:118)(cid:117)(cid:117)(cid:116) s l + χ .Note that as the variance for our unbiased estimate ( s l ) grows, χ places an upper bound onthe size of the root MSE. In contrast, if the sampling error gets very small, the forecast will placeall the weight on the unbiased estimate, and send the root MSE to zero. B.4 Estimating the effects of Medicare eligibility before and after implementation ofthe ACA

We brieﬂy describe the methods used to document differences in the effect of Medicare eligibilitybefore and after the ACA. We also document our approach to quantifying the changes in healthinsurance and ﬁnancial health for the near-elderly from the ACA.To examine the relationship between Medicare eligibility, health insurance, and ﬁnancial healthbefore and the after the implementation of the ACA we re-estimate our primary speciﬁcation sep-arately, pre- and post-ACA: y i , l , t ( age ) = γ T × ( age > ) + f T ( age ) × ( age ≤ ) + g T ( age ) × ( age > ) + (cid:101) i , l , t ( age ) . (5)where T indexes the pre-ACA (2008-2013) and post-ACA (2014-2017) periods. The coefﬁcients ofinterest are γ T which measure the change in health insurance and ﬁnancial strain at age 65 before See Appendix D of Chetty and Hendren (2018) for the explicit derivation of this approach. Our approach deviatesfrom Chetty and Hendren (2018) in that we use the Lasso predicted estimate, rather than the estimated mean value ofresidents (as Chetty and Hendren (2018) do). This extension is discussed in their Appendix D. Additionally, since ourestimates are not mean zero by construction, we demean our estimates for the purposes of the shrinkage, and then addthe overall mean back in. Our approach is otherwise identical. T =

0) and after ( T =

1) the implementation of the ACA. As before, the above speciﬁcationallows for ﬂexible age trends on both sides of the discontinuity, with standard errors constructedusing methods outlined in Kolesár and Rothe (2018) and discussed previously.To quantify the changes in health insurance and ﬁnancial health for the near-elderly fromthe ACA, we follow Duggan, Gupta and Jackson (2019) and estimate a regression discontinuitydifference-in-differences (“difference-in-discontinuity”) research design. Intuitively, this designexploits the fact that for 65-year-olds (and older), the expansion of the ACA was limited (com-pared to the near-elderly). To implement this approach, we construct ∆ y i , l , t ( age )) = y i , l , t ( age )) − ¯ y i , l , t ,2010 − ( age ) , where ¯ y i , l , t ,2010 − ( age ) is the average outcome in a given location-age from2010-2013. We then re-estimate our regression discontinuity approach using this modiﬁed out-come variable: ∆ y i , l , t ( age ) = ˜ γ × ( age > ) + ˜ f ( age ) × ( age ≤ ) + ˜ g ( age ) × ( age > ) + ˜ (cid:101) i , l , t . (6)with the standard errors constructed using methods outlined in Kolesár and Rothe (2018) anddiscussed previously. B.5 Decomposing the change in forecast reductions in per capita collections beforeand after implementation of the ACA

In this section, we describe how we use our estimates to provide insight into why forecast reduc-tions in collections have become increasingly concentrated in the Deep South. We are interested inunderstanding why the changes in the average CZ-level forecast from pre- to post-ACA differedbetween the South and other regions of the country.To formally decompose this, we deﬁne the relative percentage change before and after theACA , η = (cid:0) E ( γ Postl ) − E ( γ Prel ) (cid:1) (cid:14) E ( γ Prel ) , for both regions, South and All Others. This change canbe written as three parts: the change in the effect of Medicare on health insurance rates, the changein the effect of Medicare on collections debt per newly-insured, and the change in the covariancebetween the two. Formally, η = η + η + η E ( γ Prel ) η = (cid:16) E ( β Postl ) − E ( β Prel ) (cid:17) E ( γ h , Postl ) η = E ( β Prel ) (cid:16) E ( γ h , Postl ) − E ( γ h , Prel ) (cid:17) η = Cov ( β Postl , γ h , Postl ) − Cov ( β Prel , γ h , Prel ) This derivation follows from the fact that E ( γ l ) = E ( γ hl β l ) = E ( γ hl ) E ( β l ) + Cov ( γ hl , β l ) , where γ l is the reduction in collections debt per capita at age 65, γ hl is the change in the insurance rate at age65, and β l is the reductions in debt collections per newly-insured at age 65. We then rearrange termsto derive the above expression. An important note is that since we are focusing on the forecasts,rather than the underlying parameters, there are small differences because we use the shrinkageestimates. E.g. γ l = γ hl β l , but we use ˆ γ l . As a result, it is useful to rewrite ˆ γ l = γ l + (cid:101) γ , l , and notethat ˆ γ l = γ hl β l + (cid:101) γ , l . We can use these approximations to redeﬁne our approximation in termsof the forecasted estimates, which will leave us with additional error terms. In our results, theseterms are captured in our last decomposition piece, η .For transparency, we present the underlying quantities in Appendix Table A5 and plot thethree components of the decomposition (the η s) in Panel B of Figure 5.6 ppendix Figure A1: Changes in the distribution of collections debt at age 65

Panel A:

Share $1-500

Panel B:

Share $500-1,000

National change at 65:7e−05 [−0.004,0.004]Variance reduction at 65:−4.1% [−44.9,36.7]

Age

National change at 65:−0.001 [−0.003,3e−04]Variance reduction at 65:21.6% [−22.5,65.7]

Age

Panel C:

Share $1,000-2,500

Panel D:

Share $2,501-5,000

National change at 65:−0.002 [−0.004,3e−04]Variance reduction at 65:35.9% [−4.1,75.8]

Age

National change at 65:−0.001 [−0.002,−3e−04]Variance reduction at 65:59.2% [30.1,88.4]

Age

Panel E: $5,001-10,000

Panel F:

Share over $10,000

National change at 65:−6e−04 [−9e−04,−3e−04]Variance reduction at 65:58.1% [23.2,92.9]

Age

National change at 65:−6e−04 [−8e−04,−3e−04]Variance reduction at 65:79.9% [59.9, 99.8]

Age

Note:

This ﬁgure plots the effect of Medicare eligibility at age 65 on the distribution of new collections debtswithin the past year. A local linear regression is ﬁt on each side of the Medicare eligibility threshold usingmethods from Kolesár and Rothe (2018). We include hollow points that are the predicted counterfactualoutcomes with and without Medicare at 65. The blue hollow dot is the predicted outcome without Medicareat age 65 and the red hollow dot is the predicted consumer credit outcome with Medicare at age 65. PanelA plots the share of individuals with collections debt between $1-500 by age. Panel B plots the share ofindividuals with collections debt between $500-1,000 by age. Panel C plots the share of individuals withcollections debt between $1,000-2,500 by age. Panel D plots the share of individuals with collections debtbetween $2,501-5,000 by age. Panel E plots the share of individuals with collections debt between $5,001-10,000 by age. Panel F plots the share of individuals with collections debt greater than $10,000 by age. Thesample includes individuals who were age 55-75 between 2008 and 2017. See Section 2 for additional detailson the outcomes and sample. Source: The ﬁnancial health outcomes are based on 137,340,577 person-yearobservations from the New York Fed Consumer Credit Panel / Equifax, 2008-2017 ppendix Figure A2: Additional outcomes for changes in ﬁnancial health at age 65

Panel A:

Total debt past due

Panel B:

Mortgage debt past due

National change at 65:$−221.0 [−935.8,493.8]Variance reduction at 65:23.1% [−23.8,70.0]

Age

National change at 65:$−191.3 [−810.1,427.6]Variance reduction at 65:24.6% [−19.8,69.0]

Age

Panel C:

Credit card debt past due

Panel D:

Foreclosure

National change at 65:$−13.9 [−60.9,33.0]Variance reduction at 65:13.4% [−59.2,85.9]

Age

National change at 65:%−0.005 [−0.02,0.006]Variance reduction at 65:−41.9% [−261.5,177.6]

Age

Panel E:

Share of credit card debt past due

Panel F:

Share of mortgage debt past due

National change at 65:−0.003 [−0.01,0.005]Variance reduction at 65:14.4% [−41.2,70.1]

Age

National change at 65:−0.003 [−0.01,0.005]Variance reduction at 65:17.3% [−19.2,53.7]

Age

Note:

This ﬁgure plots consumer credit outcomes by age. The horizontal axis denotes age in years. A local linearregression is ﬁt on each side of the Medicare eligibility threshold using methods from Kolesár and Rothe (2018). Weinclude hollow points that are the predicted counterfactual outcomes with and without Medicare at 65. The blue hollowdot is the predicted outcome without Medicare at age 65 and the red hollow dot is the predicted consumer credit outcomewith Medicare at age 65. Panel A plots average amount of debt that is more than 30 days past due by age. Panel B plotsthe average amount of mortgage debt that is more than 30 days past due by age. Panel C plots the average amountof credit card debt that is more than 30 days past due by age. Panel D plots the share of individuals experiencing aforeclosure by age. Panel E reports the share of credit card that is more than 30 days past due. Panel F reports theshare of mortgage debt that is more than 30 days past due. The share debt past due outcomes are calculated as theaverage individual debt past due, divided by the average total debt of all individuals of the same age living in that state.We divide by this average, rather than individuals’ own debt levels, to avoid the divide-by-zero problem. The sampleincludes individuals who were age 55-75 between 2008 and 2017. See Section 2 for additional details on the outcomesand sample. Source: The ﬁnancial health outcomes are based on 137,340,577 person-year observations from the NewYork Fed Consumer Credit Panel / Equifax, 2008-2017. ppendix Figure A3: Robustness of Age RD Estimates to Bound Scaling Factor

Panel A:

Share with coverage

Panel B:

Total Collections ($)

Panel C:

Risk Score

Bound Scaling Factor −50−40−30−20−100 2.5 5.0 7.5 10.0

Bound Scaling Factor −4−20246 2.5 5.0 7.5 10.0

Bound Scaling Factor

Panel D:

Bankruptcy

Panel E:

Total debt past due

Panel F:

Mortgage debt past due −0.010.00 2.5 5.0 7.5 10.0

Bound Scaling Factor −1000−5000500 2.5 5.0 7.5 10.0

Bound Scaling Factor −5000500 2.5 5.0 7.5 10.0

Bound Scaling Factor

Panel G:

CC debt past due

Panel H:

Share debt past due

Panel I:

Foreclosure −75−50−2502550 2.5 5.0 7.5 10.0

Bound Scaling Factor −0.010−0.0050.000 2.5 5.0 7.5 10.0

Bound Scaling Factor −0.010.00 2.5 5.0 7.5 10.0

Bound Scaling Factor

Note:

This ﬁgure plots the robustness of our regression discontinuity estimates to the choice of the boundscaling factor used in the Kolesár and Rothe (2018) estimation procedure. Panel A plots the robustness ofthe share of the population with any coverage estimates. Panel B plots the robustness of the average debtin collections in dollars RD estimates. Panel C plots the robustness of the risk score RD estimates basedon the Equifax Riskscore 3.0. Panel D plots the robustness of the bankruptcy RD estimates. Panel E plotsthe robustness of the average debt past due RD estimates. Panel F plots the robustness of the averagemortgage debt past due RD estimates. Panel G plots the robustness of the average credit card debt past dueRD estimates. Panel H plots the robustness of the share of debt past due RD estimates. Panel I plots therobustness of the foreclosure RD estimates. The sample includes individuals who were age 55-75 between2008 and 2017. See Section 2 for additional details on the outcomes and sample. Source: The ﬁnancialhealth outcomes are based on 137,340,577 person-year observations from the New York Fed ConsumerCredit Panel / Equifax, 2008-2017. ppendix Figure A4: Robustness of Age RD Estimates to Bandwidth Selection

Panel A:

Share with coverage

Panel B:

Total Collections ($)

Panel C:

Risk Score

Bandwidth −50−40−30−20−100 4 6 8 10

Bandwidth −5.0−2.50.02.55.0 4 6 8 10

Bandwidth

Panel D:

Bankruptcy

Panel E:

Total debt past due

Panel F:

Mortgage debt past due −0.02−0.010.000.010.02 4 6 8 10

Bandwidth −1500−1000−50005001000 4 6 8 10

Bandwidth −1000−5000500 4 6 8 10

Bandwidth

Panel G:

CC debt past due

Panel H:

Share debt past due

Panel I:

Foreclosure −80−40040 4 6 8 10

Bandwidth −0.015−0.010−0.0050.0000.005 4 6 8 10

Bandwidth −0.02−0.010.00 4 6 8 10

Bandwidth

Note:

This ﬁgure plots the robustness of our regression discontinuity estimates to the bandwidth selectionused in the Kolesár and Rothe (2018) estimation procedure. Panel A plots the robustness of the share ofthe population with any coverage estimates. Panel B plots the robustness of the average debt in collectionsin dollars RD estimates. Panel C plots the robustness of the risk score RD estimates based on the EquifaxRiskscore 3.0. Panel D plots the robustness of the bankruptcy RD estimates. Panel E plots the robustnessof the average debt past due RD estimates. Panel F plots the robustness of the average mortgage debt pastdue RD estimates. Panel G plots the robustness of the average credit card debt past due RD estimates.Panel H plots the robustness of the share of debt past due RD estimates. Panel I plots the robustness ofthe foreclosure RD estimates. The sample includes individuals who were age 55-75 between 2008 and 2017.The regressions include 26,120,830 person-year-quarter observations for 2,977,952 unique individuals. SeeSection 2 for additional details on the outcomes and sample. Source: New York Fed Consumer Credit Panel/ Equifax. ppendix Figure A5: Robustness of Variance Reduction Estimates to Bound Scaling Fac-tor

Panel A:

Share with coverage

Panel B:

Total Collections

Panel C:

Risk Score

Bound Scaling Factor

Bound Scaling Factor −0.30.00.30.6 2.5 5.0 7.5 10.0

Bound Scaling Factor

Panel D:

Bankruptcy

Panel E:

Total debt past due

Panel F:

Mortgage debt past due −101 2.5 5.0 7.5 10.0

Bound Scaling Factor

Bound Scaling Factor

Bound Scaling Factor

Panel G:

CC debt past due

Panel H:

Share debt past due

Panel I:

Foreclosure −1.0−0.50.00.51.0 2.5 5.0 7.5 10.0

Bound Scaling Factor −0.50.00.5 2.5 5.0 7.5 10.0

Bound Scaling Factor −4−202 2.5 5.0 7.5 10.0

Bound Scaling Factor

Note:

This ﬁgure plots the robustness of our regression discontinuity estimates to the choice of the boundscaling factor used in the Kolesár and Rothe (2018) estimation procedure. Panel A plots the robustness ofthe share of the population with any coverage estimates. Panel B plots the robustness of the average debtin collections in dollars RD estimates. Panel C plots the robustness of the risk score RD estimates basedon the Equifax Riskscore 3.0. Panel D plots the robustness of the bankruptcy RD estimates. Panel E plotsthe robustness of the average debt past due RD estimates. Panel F plots the robustness of the averagemortgage debt past due RD estimates. Panel G plots the robustness of the average credit card debt past dueRD estimates. Panel H plots the robustness of the share of debt past due RD estimates. Panel I plots therobustness of the foreclosure RD estimates. The sample includes individuals who were age 55-75 between2008 and 2017. See Section 2 for additional details on the outcomes and sample. Source: The ﬁnancialhealth outcomes are based on 137,340,577 person-year observations from the New York Fed ConsumerCredit Panel / Equifax, 2008-2017. ppendix Figure A6: Robustness of Variance Reduction Estimate to Bandwidth Selection

Panel A:

Share with coverage

Panel B:

Total Collections

Panel C:

Risk Score

Bandwidth

Bandwidth −0.250.000.250.50 5 6 7 8 9 10

Bandwidth

Panel D:

Bankruptcy

Panel E:

Total debt past due

Panel F:

Mortgage debt past due −2−1012 5 6 7 8 9 10

Bandwidth −0.30.00.30.6 5 6 7 8 9 10

Bandwidth −0.30.00.30.6 5 6 7 8 9 10

Bandwidth

Panel G:

CC debt past due

Panel H:

Share debt past due

Panel I:

Foreclosure −0.40.00.40.81.2 5 6 7 8 9 10

Bandwidth −0.50.00.5 5 6 7 8 9 10

Bandwidth −3−2−10123 5 6 7 8 9 10

Bandwidth

Note:

This ﬁgure plots the robustness of our regression discontinuity estimates to the bandwidth selectionused in the Kolesár and Rothe (2018) estimation procedure. Panel A plots the robustness of the share ofthe population with any coverage estimates. Panel B plots the robustness of the average debt in collectionsin dollars RD estimates. Panel C plots the robustness of the risk score RD estimates based on the EquifaxRiskscore 3.0. Panel D plots the robustness of the bankruptcy RD estimates. Panel E plots the robustnessof the average debt past due RD estimates. Panel F plots the robustness of the average mortgage debt pastdue RD estimates. Panel G plots the robustness of the average credit card debt past due RD estimates.Panel H plots the robustness of the share of debt past due RD estimates. Panel I plots the robustness ofthe foreclosure RD estimates. The sample includes individuals who were age 55-75 between 2008 and 2017.The regressions include 26,120,830 person-year-quarter observations for 2,977,952 unique individuals. SeeSection 2 for additional details on the outcomes and sample. Source: New York Fed Consumer Credit Panel/ Equifax. ppendix Figure A7: Smoothness of covariates at age 65

Panel A:

Share homeowners by age in years

Panel B:

Share married by age in years

National change at 65:0.002 [−0.03,0.04]Variance reduction at 65:0.4% [−30.4,31.3]

Age

National change at 65:3e−04 [−0.01,0.01]Variance reduction at 65:22.3% [−23.8,68.3]

Age

Panel C:

Usual hours worked per week

Panel D:

Social Security income

National change at 65:−1.0 [−4.0,2.0]Variance reduction at 65:14.1% [−308.9,337.0]

Age

National change at 65:1870.1 [−761.4,4501.7]Variance reduction at 65:33.7% [−329.2,396.7]

Age

Note:

This ﬁgure plots a series of individual covariates by age. The horizontal axis denotes age in years. Alocal linear regression is ﬁt on each side of the Medicare eligibility threshold using methods from Kolesárand Rothe (2018). We include hollow points that are the predicted counterfactual outcomes with and with-out Medicare at 65. The blue hollow dot is the predicted covariate without Medicare and the red hollowdot is the predicted covariate with Medicare. Panel A plots homeownership rates by age. Panel B reportsthe share married by age. Panel C plots weekly hours worked by age. Panel D plots social security incomeby age. The sample includes individuals who were age 55-75 between 2008 and 2017. See Section 2 foradditional details on the outcomes and sample. Source: American Community Survey, 2008-2017. ppendix Figure A8: Covariate smoothness by state

Panel A:

Homeowner

Panel B:

Married

Panel C:

Employed

SDNDOKNMKYIDILUTNHCTVAGAINTXMOORAZFLCAPANJMTWVAROHMASCNYMETNMINCCOWANEMDMNKSMSVTWIIAWYALHINVLADERIAK −0.2 −0.1 0.0 0.1

Sig. Coef

FALSETRUE WYSDNEIDWVNDMTRIKYALUTHICANJNVORVTCOCTARGAOKNMTNMSOHMDMAKSAZLAMIILFLTXNYPAVASCMNIANCWAWIINNHDEMOMEAK −0.1 0.0 0.1

Sig. Coef

FALSETRUE AKDENMSCARGALAAZRIMIMDTXMOKYNVCONYMSWVCAHISDILNCNDFLUTALOKMTMAVAIDORPAOHTNCTNJVTIAINWIMEWYMNWAKSNENH −0.2 0.0 0.2 0.4

Sig. Coef

FALSE

Panel D:

Income

Panel E:

Hours worked

Panel F:

Social Security Income

SDDEARVAMTNJWVMDVTKSNYUTLAWIKYMIILMAGANHWAAZNMCAFLMEIDTNAKALNETXMOCOCTMSPAHIOKSCINNCNVOHMNORIAWYRIND−20000 −10000 0 10000

Sig. Coef

FALSETRUE AKSDDEWYARNMMTSCGAILAZWVMSMIUTLAKYTXMDMACAOKMOFLNYCTMEIAORALWIVTVANJPATNNCCORINVOHWANDINKSHIIDMNNHNE −10 0 10 20

Sig. Coef

FALSE MACTNHMERINENJHIKSVTMTNDWYILIATXWAUTMNVAOKCOMDSDNYWIORCAINLAPANMGAARFLMOALTNOHMSNCNVWVAZSCDEKYIDMIAK −10000 −5000 0 5000 10000

Sig. Coef

FALSE

Note:

This ﬁgure plots the estimated discontinuities at age 65 in individual covariates by state. A local lin-ear regression is ﬁt on each side of the Medicare eligibility threshold using methods from Kolesár and Rothe(2018). The red bars indicate statistically signiﬁcant results at the 5% level. Panel A plots the estimated dis-continuities in homeownership rates by state. Panel B plots the estimated discontinuities in marriage ratesby state. Panel C plots the estimated discontinuities in employment by state. Panel D plots the estimateddiscontinuities in income by state. Panel E plots the estimated discontinuities in hours worked by state.Panel F plots the estimate discontinuities in social security income by state. The sample includes individ-uals who were age 55-75 between 2008 and 2017. See Section 2 for additional details on the outcomes andsample. Source: American Community Survey, 2008-2017. ppendix Figure A9: Covariate smoothness by commuting zone

Panel A:

Homeowner

Panel B:

Married

Share Sig.: 0.047

Estimated EffectSig. Coef

FALSETRUE

Share Sig.: 0.03

Estimated EffectSig. Coef

FALSETRUE

Panel C:

Employed

Panel D:

Income

Share Sig.: 0.028

Estimated EffectSig. Coef

FALSETRUE

Share Sig.: 0.065

Estimated EffectSig. Coef

FALSETRUE

Panel E:

Hours worked

Panel F:

Social Security Income

Share Sig.: 0.016

Estimated EffectSig. Coef

FALSETRUE

Share Sig.: 0.008

Estimated EffectSig. Coef

FALSETRUE

Note:

This ﬁgure plots histograms of normalized estimated discontinuities at age 65 in individual covariates bycommuting zone (CZ). A local linear regression is ﬁt on each side of the Medicare eligibility threshold by CZ usingmethods from Kolesár and Rothe (2018). The red bars indicate statistically signiﬁcant results at the 5% level. Theblue bars indicate statistically insigniﬁcant results at the 5% level. Panel A plots the estimated discontinuitiesin homeownership rates by state. Panel B plots the estimated discontinuities in marriage rates by state. PanelC plots the estimated discontinuities in employment by state. Panel D plots the estimated discontinuities inincome by state. Panel E plots the estimated discontinuities in hours worked by state. Panel F plots the estimatediscontinuities in social security income by state. The sample includes individuals who were age 55-75 between2008 and 2017. See Section 2 for additional details on the outcomes and sample. Source: American CommunitySurvey, 2008-2017. ppendix Figure A10: Consumer credit outcomes and uninsurance rates across states

Panel A:

Debt collections

Panel B:

Bankruptcy

AKAL ARAZCACOCTDCDE FLGAHIIA IDILINKSKY LAMA MD MEMIMN MO MSMTNCNDNENH NJ NMNVNY OH OKORPARI SCSD TN TXUTVAVT WAWI WV WY

Slope: 1390SE:165R−squared:0.591

Uninsurance Rate

AKAL ARAZCACOCTDCDE FLGAHIIA IDILINKSKY LAMA MD MEMIMN MO MSMTNCNDNENH NJ NMNVNY OH OKORPARI SCSD TN TXUTVAVT WAWI WV WY

Slope: 0.243SE:0.159R−squared:0.0456

Uninsurance Rate

Panel C:

Share debt past due

Panel D:

Credit Score

AKAL ARAZCACOCTDCDE FLGAHIIA IDILINKSKY LAMA MD MEMIMN MO MSMTNCNDNENH NJ NMNVNY OH OKORPARI SCSD TN TXUTVAVT WAWI WV WY

Slope: 0.174SE:0.0585R−squared:0.152

Uninsurance Rate

Slope: −312SE:41.7R−squared:0.533

Uninsurance Rate

Note:

This ﬁgure plots consumer credit outcomes against state uninsurance rates for individuals aged 55-64years old. Panel A plots the dollar value of new collections debt reported on credit reports annually. PanelB plots the annual rate of new bankruptcies in percentage points. Panel C plots the share of debt that ismore than 30 days past due. The share debt past due is calculated as the average individual’s debt morethan 30 days past due, divided by the average total debt of all individuals of the same age living in thatstate. Panel D plots credit score data using the Equifax Risk Score 3.0. See Section 2 for additional detailson the outcomes and sample. Source: Consumer credit outcomes are based on 137,340,577 person-yearobservations from the New York Fed Consumer Credit Panel / Equifax, 2008-2017. State-level uninsurancerates are from the American Community Survey, 2008-2017. ppendix Figure A11: Effect of Medicare eligibility on the level of collections debt at age65 vs. effects on insurance

Panel A:

State-level estimates

Panel B:

CZ-level estimates

AKAL ARAZCACOCTDCDE FLGAHIIA IDIL INKS KY LAMA MD MEMIMN MO MS MTNCND NE NHNJ NM NVNY OH OKORPARI SCSD TN TXUTVAVT WAWI WVWYSlope: 584SE:110R−squared:0.37

Estimated Effect on Insurance Rate

Slope: 409SE:55.2R−squared:0.0692

Estimated Effect on Insurance Rate

Note:

This ﬁgure plots point estimates of the reduction in the ﬂow of newly-reported collections debt(within the past year) and the increase in the insurance rate at age 65 based on local linear regressions,done separately by state and commuting zone (CZ), using the methods from Kolesár and Rothe (2018).Panel A plots state-level estimates. Panel B plots CZ-level estimates, where the dark points are binnedaverages constructed using the binsreg command from Cattaneo et al. (2019). The horizontal axes arethe estimated effect on the insurance rate at age 65 by locality. The vertical axes are the reduction in theﬂow of collections debt at age 65 by locality. Source: Consumer credit outcomes are based on 137,340,577person-year observations from the New York Fed Consumer Credit Panel / Equifax, 2008-2017. State-leveluninsurance rates are from the American Community Survey, 2008-2017. ppendix Figure A12: Correlates with reduction in collections debt at age 65, with FixedEffects

Panel A:

Area-level demographic characteristics

Bivariate Multivariate Post−Lasso−0.6 −0.3 0.0 0.3 0.6−0.6 −0.3 0.0 0.3 0.6−0.6 −0.3 0.0 0.3 0.6Has Medicaid (%)Has any coverage (%)Physical disability (%)Median house valueIncome per capitaPoverty rate (%)Greater than high school education (%)Black (%)

Impact of 1 SD Change in CovariateFixed Effects

No FE Region Division State

Panel B:

Healthcare market characteristics

Bivariate Multivariate Post−Lasso−0.25 0.00 0.25 0.50 −0.25 0.00 0.25 0.50 −0.25 0.00 0.25 0.50Medicare spending per enrollee ($)Payment by charity care patients ($)Cost of charity care per patient day ($)Hospital beds per capitaHospital occupancy rate (%)Teaching hospitals (%)For−profit hospitals (%)

Impact of 1 SD Change in CovariateFixed Effects

No FE Region Division State

Note:

This ﬁgure plots bivariate OLS regression results (left panel), multivariate OLS regression results (center panel),and post-Lasso multivariate regression results (right panel) of CZ-level estimated reductions in collections debt percapita on a set of CZ-level characteristics. We standardize all the variables so the coefﬁcients reﬂect the strength of theassociation between a one standard deviation change in the covariate and the estimated reduction in collections debtat age 65. The horizontal bars are 95% conﬁdence intervals. The multivariate OLS regression results and post-Lassomultivariate regression results are both run on the full set of characteristics in Panels A and B. For post-Lasso, we ﬁrstestimate a Lasso regression on the full set of characteristics and then report the results of multivariate OLS run on thecharacteristics chosen by the Lasso regression. For each correlate we report our primary results where there are area-level ﬁxed effects (termed “No FE”) and then results where we include ﬁxed effects for Census Region, Census Division,and state, respectively. Tabular versions of these results are in Table A4. Source: Consumer credit outcomes are basedon 137,340,577 person-year observations from the New York Fed Consumer Credit Panel / Equifax, 2008-2017. CZ-leveluninsurance rates are from the American Community Survey, 2008-2017. Healthcare market characteristics are fromthe Healthcare Cost Report Information System (HCRIS) and the Dartmouth Atlas. For additional details on the datasee Section 2. ppendix Figure A13: Effect of Medicare eligibility on per-capita collections debt vs.per-newly insured collections debt at the CZ-level

Slope: 12SE:0.527R−squared:0.413 −1000010002000 −100 0 100 200

Estimated Effect on Per Capita Debt Collections

Note:

This ﬁgure plots point estimates of the reduction in the ﬂow of per-capita collections debt (withinthe past year) vs. point estimates of the reduction in the ﬂow of per-capita collections debt (within the pastyear) per newly-insured. The estimates are based on local linear regressions, done separately by commut-ing zone (CZ), using the methods from Kolesár and Rothe (2018). The dark points are binned averagesconstructed using the binsreg command from Cattaneo et al. (2019). The horizontal axis is the estimatedeffect on per capita debt collections at age 65 by CZ. The vertical axis is the estimated reduction in the ﬂowof collections debt per newly-insured at age 65 by CZ. Source: Consumer credit outcomes are based on137,340,577 person-year observations from the New York Fed Consumer Credit Panel / Equifax, 2008-2017.State-level uninsurance rates are from the American Community Survey, 2008-2017. ppendix Figure A14: Forecasts of causal reductions in collections debt from expandinghealth insurance to the near-elderly by commuting zone

Panel A:

Pre-ACA, 2008-2013

Panel B:

Post-ACA, 2014-2017

Note:

This ﬁgure plots mean square error (MSE)-minimizing forecasts of the reductions in collections debtper capita in the pre-ACA (Panel A) and post-ACA period (Panel B). We construct the MSE-minimizingforecasting by ﬁrst running a Lasso regression to predict the CZ-level reductions in collections debt percapita separate for each period. This generates a prediction for each CZ in each period, which we callˆ γ l . Following Chetty and Hendren (2018) we then combine the ˆ γ l estimates with our estimates of γ l toconstruct the mean square error-minimizing forecast for each commuting zone in each period, γ fl . Source:Consumer credit outcomes are based on 137,340,577 person-year observations from the New York Fed Con-sumer Credit Panel / Equifax, 2008-2017. CZ-level uninsurance rates are from the American CommunitySurvey, 2008-2017. Healthcare market characteristics are from the Healthcare Cost Report Information Sys-tem (HCRIS) and the Dartmouth Atlas. For additional details on the data see Section 2.. ppendix Figure A15: Counterfactual health insurance rates by commuting zone at age65 without Medicare, pre- and post-ACA

Panel A:

Pre-ACA, 2008-2013

Panel B:

Post-ACA, 2014-2017

Note:

This ﬁgure plots our counterfactual estimates of the share of the population with health insurancecoverage at age 65, without Medicare, before and after the full implementation of the Affordable CareAct in 2014. The counterfactuals are based on local linear regressions, done separately by commutingzone, using the methods from Kolesár and Rothe (2018). These estimates are then shrunk using empiricalBayes, described in Section B. Panel A. presents the counterfactuals from the pre-ACA period, 2008-2013.Panel B. presents the counterfactuals from the post-ACA period, 2014-2017. Darker shading corresponds tostates with higher counterfactual health insurance rates. Source: CZ-level uninsurance rates are from theAmerican Community Survey, 2008-2017. ppendix Figure A16: Difference-in-discontinuities estimates of increases in the near-elderly health insurance rate due to the ACA

Note:

This ﬁgure plots CZ-level estimates of the increase in health insurance coverage due to the Afford-able Care Act using a difference-in-discontinuities design similar to Duggan, Gupta and Jackson (2019). Wecompare the discontinuity in health insurance at age 65 in the post-ACA period (2014-2017) to the disconti-nuity in health insurance at age 65 prior to its full implementation (2008-2013). CZ-level uninsurance ratesare from the American Community Survey, 2008-2017. ppendix Figure A17: Forecasts of causal reductions in collections debt per newly-insured near-elderly person by commuting zone, pre- and post-ACA

Panel A:

Pre-ACA, 2008-2013

Panel B:

Post-ACA, 2014-2017

Note:

This ﬁgure plots mean square error (MSE)-minimizing forecasts of the reductions in collections debtper newly-insured for the pre-ACA period (Panel A) and post-ACA period (Panel B). We construct the MSE-minimizing forecasting by ﬁrst running a Lasso regression to predict the CZ-level reductions in collectionsdebt per newly-insured. This generates a prediction for each CZ, which we call ˆ β l . Following Chetty andHendren (2018) we then combine the ˆ β l estimates with our estimates of β l to construct the mean squareerror-minimizing forecast for each commuting zone, β fl . Source: Consumer credit outcomes are based on137,340,577 person-year observations from the New York Fed Consumer Credit Panel / Equifax, 2008-2017.State-level uninsurance rates are from the American Community Survey, 2008-2017. Healthcare marketcharacteristics are from the Healthcare Cost Report Information System (HCRIS) and the Dartmouth Atlas.For additional details on the data see Section 2. ppendix Table A1: Robustness of estimated changes in ﬁnancial outcomes at age 65

Main Linear with Quad. with Cubic. with Linear with Quad. with Cubic. with Local linearEstimate robust SEs robust SEs robust SEs clustered SEs clustered SEs clustered SEs with rdrobust(1) (2) (3) (4) (5) (6) (7) (8)Share with any coverage 0.076 ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( )[ [ [ [ [ [ [ [ ] ] ] ] ] ] ] ] Debt in collections − ∗ − ∗ − ∗ − ∗ − ∗ − ∗ − ∗ − ∗ ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( )[ − [ − [ − [ − [ − [ − [ − [ − − ] − ] − ] − ] − ] − ] − ] − ] Credit score 0.818 1.199 0.645 0.682 1.199 ∗ ∗ ∗ ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( )[ − [ − [ − [ − [ [ [ [ − ] ] ] ] ] ] ] ] Bankruptcy (pp) − − − − ∗ − ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( )[ − [ − [ − [ − [ − [ − [ − [ − ] ] ] ] ] ] ] ] Share employed − − ∗ − − − ∗ − − − ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( )[ − [ − [ − [ − [ − [ − [ − [ − ] − ] ] ] − ] ] ] ] Income 880.67 458.575 1572.808 343.741 458.575 1572.808 ∗ ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( )[ − [ − [ − [ − [ − [ [ − [ − ] ] ] ] ] ] ] ] Total debt past due − − − − − ∗ − ∗ − ∗ − ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( )[ − [ − [ − [ − [ − [ − [ − [ − ] ] ] ] − ] − ] − ] ] Mortgage debt past due − − − − − ∗ − ∗ − ∗ − ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( )[ − [ − [ − [ − [ − [ − [ − [ − ] ] ] ] − ] − ] − ] ] Credit card debt past due − − ∗ − − − ∗ − ∗ − − ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( )[ − [ − [ − [ − [ − [ − [ − [ − ] − ] ] ] − ] − ] ] ] Foreclosure − − − − − − − ∗ − ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( )[ − [ − [ − [ − [ − [ − [ − [ − ] ] ] ] ] ] − ] ] Share of mortgage debt past due − − ∗ − − − ∗ − ∗ − ∗ − ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( )[ − [ − [ − [ − [ − [ − [ − [ − ] − ] ] ] − ] − ] − ] ] Share of cc debt past due − − ∗ − − − ∗ − ∗ − − ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( )[ − [ − [ − [ − [ − [ − [ − [ − ] − ] ] ] − ] ] ] ] Note:

This table reports the sensitivity of our main regression discontinuity estimates to alternative speciﬁcations.Column 1 reports the point estimate, standard error, and bias-adjusted 95% conﬁdence interval from a local lin-ear regression using techniques from Kolesár and Rothe (2018). Columns 2-4 report the results of estimating thediscontinuity at 65 using three parametric models and robust standard errors with linear, quadratic, and cubicage trends, respectively. Columns 2-4 report the results of estimating the discontinuity at 65 using three para-metric models and clustering standard errors by age (the running variable) as in Lee and Card (2008) with linear,quadratic, and cubic age trends, respectively. Column 8 reports the results of estimating the discontinuity usingthe local linear regression model as in Calonico, Cattaneo and Titiunik (2015). The sample includes individualswho were age 55-75 between 2008 and 2017. Credit score data used is from Equifax Riskscore 3.0. See Section 2 foradditional details on the outcomes and sample. Source: The ﬁnancial health outcomes are based on 137,340,577person-year observations from the New York Fed Consumer Credit Panel / Equifax, 2008-2017. ppendix Table A2: Location-speciﬁc estimates and forecasts for 50 largest CZs

Per capita Per newly-insured γ fl RMSE β fl RMSEState CZ (1) (2) (3) (4)Arizona Phoenix -25 11 -273 180California Los Angeles -29 11 -206 171California Sacramento -17 8 -259 112California San Diego -27 7 -443 158California San Francisco -8 10 -108 176California San Jose -14 12 -243 247Colorado Denver -25 6 -409 87Connecticut Bridgeport -15 3 -272 61District of Columbia Washington DC -22 10 -468 208Florida Jacksonville -42 13 -614 248Florida Miami -62 12 -419 139Florida Orlando -44 13 -472 267Florida Port St. Lucie -58 13 -719 243Florida Sarasota -53 13 -485 256Florida Tampa -53 13 -484 222Georgia Atlanta -43 12 -548 206Illinois Chicago -17 9 -277 159Indiana Indianapolis -41 11 -614 194Maryland Baltimore -23 12 -599 265Massachusetts Boston -14 8 -403 263Michigan Detroit -21 11 -417 242Michigan Grand Rapids -8 12 -106 233Minnesota Minneapolis -3 9 -152 225Missouri Kansas City -40 13 -412 240Missouri St. Louis -30 10 -436 180Nevada Las Vegas -53 12 -591 196New Hampshire Manchester -11 11 -148 194New Jersey Newark -18 7 -221 98New Jersey Toms River -14 10 -280 191New York Buffalo -14 3 -449 62New York New York City -12 6 -197 114North Carolina Charlotte -57 13 -729 268North Carolina Raleigh -53 11 -956 167Ohio Cincinnati -26 9 -456 140Ohio Cleveland -16 10 -281 199Ohio Columbus -37 6 -759 61Ohio Dayton -24 11 -381 208Oregon Portland -10 5 -139 82Pennsylvania Philadelphia -18 6 -344 84Pennsylvania Pittsburgh -21 11 -289 243Rhode Island Providence -2 7 -46 171Tennessee Nashville -52 8 -611 78Texas Austin -43 13 -436 248Texas Dallas -54 13 -502 227Texas Fort Worth -50 12 -563 125Texas Houston -49 12 -435 153Texas San Antonio -43 13 -305 254Utah Salt Lake City -31 13 -276 245Washington Seattle -25 10 -467 213Wisconsin Milwaukee -28 10 -532 215

Note:

This table reports the mean square error (MSE)-minimizing forecasts of the reductions in collections debtper capita and the reduction in collections debt per newly-insured for the 50 most populous CZs based on theirnear-elderly population. We construct the MSE-minimizing forecasting by ﬁrst running a Lasso regression topredict the CZ-level reductions in collections debt per capita (or per newly-insured). This generates a predictionfor each CZ, which we call ˆ γ l . Following Chetty and Hendren (2018) we then combine the ˆ γ l estimates with ourestimates of γ l to construct the mean square error-minimizing forecast for each commuting zone, γ fl , which wepresent in Column 1. Column 2 presents the root-mean-square error (RMSE) which is calculated using methodsfrom Chetty and Hendren (2018). Column 3 reports the mean square error-minimizing forecast of the reduction incollections debt per newly-insured associated with a (nearly) universal health insurance expansion, β fl . Column 4presents the RMSE for β fl . Source: Consumer credit outcomes are based on 137,340,577 person-year observationsfrom the New York Fed Consumer Credit Panel / Equifax, 2008-2017. CZ-level uninsurance rates are from theAmerican Community Survey, 2008-2017. Healthcare market characteristics are from the Healthcare Cost ReportInformation System (HCRIS) and the Dartmouth Atlas. For details on the data see Section 2. ppendix Table A3: Description of the Federal Reserve Bank of New York’s EquifaxConsumer Credit Panel (CCP)

Variable DescriptionAmount in collections the total collection amount of these 3rd party collection accounts.Number of delinquent accounts the count of all non-current loans.Amount delinquent the sum of all non-current loan balances.Total credit card balance past due the difference of total bankcard balance and current bankcard balance.Total mortgage account balance past due the difference of total mortgage account balance (incl. home equity installment) and current mortgage balance.Foreclosure ﬂag for if an individual recorded a foreclosure in the past 24 months.New Foreclosure Number of people that recorded a foreclosure in the current quarter, but not the two previous quarters.Bankruptcy ﬂag for if an individual recorded a bankruptcy in the past 24 months.New Bankruptcy Number of people that recorded a bankruptcy in the current quarter, but not the two previous quarters.Equifax Risk Score always refers to Equifax Risk Score 3.0.

Note:

This table reports deﬁnitions for the ﬁnancial variables used from the New York Fed Consumer Credit Panel. The datasetconsists of 137,340,577 person-year observations from the New York Fed Consumer Credit Panel / Equifax, 2008-2017. ppendix Table A4: Correlates with reduction in collections debt at age 65

Bivariate Multivariate Post-LassoCovariate Estimate Type Estimate S.E. Estimate S.E. Estimate S.E.Black (%) Per Capita -7.17 (2.77) -5.74 (2.28) -6.23 (2.08)Greater than high school education (%) Per Capita 11.30 (1.74) -2.47 (3.52) 4.86 (2.46)Has any coverage (%) Per Capita 12.00 (1.86) 7.09 (2.94)Has Medicaid (%) Per Capita 6.75 (1.65) 3.24 (2.87)Hospital beds per capita Per Capita -1.09 (1.4) 1.86 (1.48)Income per capita Per Capita 11.90 (1.79) 6.86 (5.01)Median house value Per Capita 10.70 (1.88) -2.25 (2.49)Hospital occupancy rate (%) Per Capita 6.56 (1.68) -0.90 (3.12)Physical disability (%) Per Capita -11.90 (2) -5.60 (3.21) -7.41 (2.56)Poverty rate (%) Per Capita -7.01 (2.34) -0.01 (3.24) 1.02 (2.16)Payment by charity care patients ($) Per Capita -1.52 (1.65) -1.78 (1.46) -2.70 (1.53)Medicare spending per enrollee ($) Per Capita -6.48 (2.08) -0.63 (2.98)For-proﬁt hospitals (%) Per Capita -10.20 (1.96) -4.96 (2.17) -8.29 (1.97)Teaching hospitals (%) Per Capita 9.69 (1.51) 6.14 (3.32)Cost of charity care per patient day ($) Per Capita 0.07 (3.1) -0.96 (2) -1.26 (2.21)Black (%) Per Newly Insured -62.20 (37.3) -54.20 (33.7) -53.50 (31.8)Greater than high school education (%) Per Newly Insured 76.10 (25) -47.20 (49.8) -5.16 (39.9)Has any coverage (%) Per Newly Insured -3.87 (29.3) -127.00 (52.6)Has Medicaid (%) Per Newly Insured 91.80 (28.4) 101.00 (52.8)Hospital beds per capita Per Newly Insured 21.00 (38.2) 46.80 (38.7)Income per capita Per Newly Insured 95.50 (24.9) 125.00 (64.3)Median house value Per Newly Insured 97.20 (30.9) -58.50 (44.9)Hospital occupancy rate (%) Per Newly Insured 45.20 (34.5) -6.50 (47.8)Physical disability (%) Per Newly Insured -113.00 (24.6) -39.50 (50.8) -123.00 (38.6)Poverty rate (%) Per Newly Insured -19.30 (28.6) -66.60 (44.4) 48.80 (26.3)Payment by charity care patients ($) Per Newly Insured -39.30 (33.5) -38.00 (32.4) -49.20 (31.5)Medicare spending per enrollee ($) Per Newly Insured -29.50 (32) -40.20 (45.1)For-proﬁt hospitals (%) Per Newly Insured -70.00 (28.6) -65.30 (30.3) -69.40 (26.6)Teaching hospitals (%) Per Newly Insured 110.00 (18.6) 92.70 (55.1)Cost of charity care per patient day ($) Per Newly Insured 29.50 (37.2) -59.30 (39.1) 0.70 (31.4)

Note:

This table reports the CZ-level correlates with our RD estimated reductions in collections debt at age 65 plotted in Figure 3.The “Estimate Type” column indicates whether the row presents correlates with our “per capita” or “per newly-insured” estimates.For each row, we we present the estimates and standard errors for bivariate, multivariate, and post-Lasso models. We standardize allthe variables so the coefﬁcients reﬂect the strength of the association between a one standard deviation change in the covariate andthe estimated reduction in collections debt at age 65. The multivariate OLS regression results and post-Lasso multivariate regressionresults are both run on the full set of characteristics. For post-Lasso, we ﬁrst estimate a Lasso regression on the full set of characteristicsand then report the results of multivariate OLS run on the characteristics chosen by the Lasso regression. Source: Consumer creditoutcomes are based on 137,340,577 person-year observations from the New York Fed Consumer Credit Panel / Equifax, 2008-2017.CZ-level uninsurance rates are from the American Community Survey, 2008-2017. Healthcare market characteristics are from theHealthcare Cost Report Information System (HCRIS) and the Dartmouth Atlas. For additional details on the data see Section 2. ppendix Table A5: Changes in the forecast reductions in collections debt at age 65, preand post-ACA, used for decomposition in Figure 5

Per Capita Insurance Effect Per Newly Insured Covariance DecompositionLocation Pre Post Diff Pre Post Diff Pre Post Diff Pre Post Diff η η η η South 43.61 30.66 -12.95 0.13 0.08 -0.05 365.95 408.95 43.0 -0.95 -0.95 0.00 -0.3 -0.38 0.08 0.00All Others 14.36 7.23 -7.13 0.09 0.06 -0.04 168.76 167.66 -1.1 0.19 -0.30 -0.49 -0.5 -0.43 0.00 -0.06Difference -29.25 -23.43 5.82 -0.03 -0.02 0.01 -197.20 -241.30 -44.1 1.15 0.66 -0.49 -0.2 -0.05 -0.08 -0.06

Note: