Does Medicare Reduce Medical Debt?

I.  Introduction

In 2016, Medicare insured 53 million Americans and accounted for roughly 20 percent of health-care spending (Barnett and Berchick 2017; CMS 2017). While some of those covered by Medicare are younger, a large majority of US residents become eligible for the program when they reach age 65. The share reporting Medicare coverage jumps from less than 20 percent at age 64 to more than 90 percent at age 66. Studies show that Medicare increases the utilization of medical care and at least in some circumstances improves health outcomes (Card, Dobkin, and Maestas 2008, 2009). Importantly, Medicare also decreases the risk of out-of-pocket medical expenditures (Finkelstein and McKnight 2008; Engelhardt and Gruber 2011; Barcellos and Jacobson 2015).

This paper contributes to the growing literature studying the effects of health insurance on individuals’ financial well-being using administrative data from consumer credit records. The use of credit records is complementary to studies of out-of-pocket spending and those that use self-reports of difficulty paying medical bills. Credit bureau records provide administrative data on actual unpaid bills from medical providers sent to third-party collectors, covering large samples that are representative of all consumers with credit records. Medical collections are unequivocally derogatory events that at a minimum can negatively impact consumers’ access to credit.

Most studies using credit bureau data to study the effects of health insurance on financial strain deal with expansions of Medicaid. No published study to date has used credit bureau data to study the effects of Medicare, as we do. We employ a regression discontinuity (RD) design to study effects of Medicare by exploiting the program’s eligibility age at 65 and the large discontinuous change in coverage at that age. We compare outcomes across individuals whose age is below and above the age eligibility threshold to identify the effect of Medicare on medical collections. Given that individuals may or may not have health insurance prior to age 65, and the large variation in generosity of coverage among the insured, the effect of Medicare on financial strain will vary by individuals’ pre-age-65 insurance status. The reduced-form effect we measure is an average over several possible underlying mechanisms specific to individuals near the Medicare eligibility age threshold.

Those covered only by Medicare Parts A (hospital insurance) and B (medical insurance) have less complete coverage than what is provided by Medicaid or most employer-sponsored plans. However, the large majority of Medicare beneficiaries have some form of supplemental coverage—only 14 percent had none in 2010 (Cubanski et al. 2015). This suggests that most people experience either little change at age 65, or an improvement in coverage, which is consistent with previous research findings that on average Medicare decreases out-of-pocket spending risk. The effects of Medicare on financial strain are likely to be largest for those who were uninsured prior to turning 65. We test this hypothesis by investigating variation in effects by geographies with different uninsured rates for those near age 65.

To preview our results, we find strong evidence, visual evidence in figures confirmed in statistical models, of discontinuities at age 65 in outcomes related to increases in medical collection balances during the 2011–13 period (pre-Affordable Care Act, ACA).¹ For example, we estimate over the entire sample that the probability of an increase in medical collection balance larger than $1,000 declines by 0.31 percentage points at age 65, a reduction of about 19 percent relative to the average value for those aged 60–64. Similarly, the change in medical collections falls by about $380 at the 99th percentile of the distribution, about a 23 percent reduction relative to the 99th percentile value for those aged 60–64. The mean change in medical collection balance falls by $25. If this effect is confined to the 10 percent who transition from uninsured to insured at age 65, it is an average of $250 per enrollee for that group, a sizeable “treatment on the treated” effect, though somewhat smaller than other researchers have estimated for the transition from uninsured to Medicaid in a nonelderly adult population spanning a broad range of ages. While our primary sample ends in 2013, we present some results that suggest a shrinking of the beneficial effect of Medicare on medical collections as insurance coverage expanded for nonelderly adults after implementation of the Affordable Care Act. This suggests that the ACA decreased financial strain related to health care for those near age 65 and supports the hypothesis that the main mechanism behind our results is the transition from uninsured to Medicare.

In what follows, Section II provides a brief literature review, Section III describes our data, and Section IV outlines our empirical methods. In presenting results in Section V, we first review data on changes in insurance coverage, employment, and incomes at age 65. We then show descriptive evidence on changes in medical collection balances around age 65 before turning to results from statistical models. We conclude in Section VI by summarizing what we find and relating our findings to other results in the literature.

II.  Literature Review

Beginning with Finkelstein et al. (2012), a rapidly growing literature has examined the effects of health insurance on financial strain using credit bureau records. Most of these studies have focused on the effects for low-income adults of gaining Medicaid coverage.² Finkelstein et al., in their study of the Oregon Health Insurance Experiment, used credit bureau records to estimate the effects on various measures of financial strain, including amounts in medical and nonmedical collections, of being randomly assigned access to Medicaid. Several recent studies using various identification strategies have exploited the natural experiment provided by the Medicaid expansion under the ACA to study effects of gaining Medicaid on financial strain (Caswell and Waidmann 2017; Hu et al. 2018; Brevoort et al. 2018). All of these studies find, to varying degrees and examining various outcome measures, that gaining Medicaid reduces the likelihood of having debt sent to collections. To date, no published studies have examined the effects of Medicare on financial outcomes using credit bureau data as we do in this paper.³

Several studies have used the age 65 threshold for Medicare eligibility in RD designs to study various effects of Medicare.⁴ For example, Card, Dobkin, and Maestas (2008) find that health-care use increases discontinuously at the age 65 threshold. The same authors found discontinuous declines in mortality at age 65 in groups of severely ill patients (Card, Dobkin, and Maestas 2009).⁵ The study most similar to ours is by Barcellos and Jacobson (2015). These authors use an RD design to study changes in out-of-pocket health-care spending at age 65, relying on data from two surveys covering the years 2007–10. They find a large drop in average out-of-pocket spending of $326 per year in the baseline estimate, a decline of about one-third from pre-age 65 levels. One of the surveys allows them to also study the effect of Medicare on several self-reported measures of financial strain, including contact from collection agencies and the amount of medical debt owed. The authors find that Medicare significantly decreases both contact by debt collectors and the amount owed in medical bills. Our access to credit bureau data gives us much larger samples than were available to Barcellos and Jacobson, and administrative data rather than self-reports.

For several reasons, the transition to Medicare at age 65 might have different effects on financial strain than gaining Medicaid coverage in a population of more diverse ages. For example, there are no income limits for Medicare eligibility, while Medicaid is usually available only to those with very low incomes (an upper limit of 138 percent of the poverty level for the ACA Medicaid expansion). Medicaid covers a broader set of services with in most cases zero or very small copayments. Most of those who transition to Medicare coverage at age 65 were already insured, whereas those who gain Medicaid coverage were more likely to have been uninsured. On the other hand, older individuals are more likely to experience adverse health events that may lead to large expenditures.

III.  Data

Data are derived from a nationally representative 2 percent sample of US consumers with a credit history, obtained from one of the three major credit bureaus.⁶ The universe for the analysis sample is the credit bureau’s records on approximately 250 million total consumers in each year. Our primary (baseline) sample includes data from annual archives, covering 2011 through 2013, which were culled at the end of August for each year. If the credit bureau knows an individual’s exact date of birth, age is coded as age in completed years at the time of the archive. Observations for which only year of birth is known are treated as though the birth date comes after the archive. Age is in those cases coded as the archive year − year of birth − 1.⁷ Our sample is an unbalanced panel, where consumers appear in each year of the data for which they have a record with the credit bureau. We study consumers who are near the Medicare eligibility age (65 years old), and our main subsample (discussed below) includes more than 1.1 million unique consumers aged 55–75, per year. Some analyses also use archives from the years 2014 through 2016 to examine whether the implementation of the ACA changed the effect of Medicare on medical debt.

The primary outcomes we study relate to debt in collections, meaning debt that was sent to a third-party collector or assigned to a creditor’s internal collections department (Consumer Financial Protection Bureau 2014). Debt in collections is normally at least 180 days past due, where the creditor acted to collect the unpaid debt but was unsuccessful. Our data include for each consumer the total amount of unpaid debt in collections at the time of each archive, divided into medical and nonmedical collections. We refer to these as measures of collection balances. Collections are defined as medical if the original creditor is identified as a medical provider. We define an individual’s nonmedical collection balance as the difference between the total collection balance and the medical collection balance. Thus, some medical collections may be misclassified as nonmedical if information about the original creditor is missing.⁸ Medical bills that are paid using a credit card but ultimately go to collections would also appear as nonmedical collections. For consumers who had a medical bill sent to collections at some point, we also observe the number of months since the most recent medical collection, which allows us to determine whether any new collections were incurred since the previous archive.⁹

Two important limitations of our data are that (1) we do not observe flows of new collections, only the stock (collection balance), and (2) a debt appears at the time it is reported to the credit bureau as being in collections, not at the time the debt is initially incurred. Medicare does not influence debts incurred prior to obtaining Medicare, but it may influence the likelihood or magnitude of new debt sent to collections. We construct a proxy measure for the flow of new collections from changes in the stock:

We study the distribution of changes in collection balances, Δy. Having medical debt in collections is not uncommon. For example, about 14 percent of our baseline sample, aged 55 to 75, has a positive medical collection balance in 2013. Nonetheless, most often $\mathrm{\Delta }{y}_{\mathit{it}}=0$, and zero values become more likely at older ages, even prior to Medicare eligibility. We study the mean of Δy_it, the likelihood of any positive value ($\mathrm{\Delta }{y}_{\mathit{it}}>0$) and of large positive values (e.g., $\mathrm{\Delta }{y}_{\mathit{it}}>\$\mathrm{1,000}$), and quantile values in the right tail of the distribution. For mean collections, we top/bottom code at the 0.5th and 99.5th percentiles, by year, throughout the analysis to ensure our results are not influenced by extreme values.

In the credit bureau data we do not observe any individual characteristics other than age and place of residence (state, county, and zip code). We link individual credit records to information on the share of those aged 55–64 residing in the same zip code who are uninsured. We use five-year estimates (2009–13) from the American Community Survey (ACS) of uninsured rates by zip code, accessed through the American FactFinder, Table B27001 (US Census Bureau 2018). We also use national data from the ACS accessed through IPUMS USA to examine changes in insurance coverage around age 65 and smoothness in age profiles of employment and incomes over the time period covered by this study (Ruggles et al. 2017).

IV.  Methods

To estimate the effect of Medicare on medical collections we estimate RD models that take the following general form:

To estimate the effect on the average change in medical collections, on the probability of a positive change, or on the likelihood of changes larger than specified sizes, we estimate ordinary least squares (OLS) models with standard errors clustered by age. To estimate effects at specific quantiles of the distribution (e.g., 98th and 99th percentiles), we estimate quantile regressions using Stata (version 15) with the command qreg, and compute standard errors by block bootstrapping, blocking by age (250 replications).¹²

Our statistical models exclude consumers age 65, making them “donut” RD models. Our motivation for excluding age 65 is twofold. First, because we observe only annual archives, for almost all individuals coded as age 65 at time t the time between the t − 1 and t archives includes some time when the individual was age 64, prior to Medicare eligibility age. Second, collections that are reported at age 65 partially reflect collections arising from services received prior to age 65. As noted above, unpaid medical bills are not instantly sent to third-party collectors. Rather, creditors first generally try to collect unpaid bills, and should they eventually refer the debt to third-party collectors there can be lags in time between when a debt is purchased by a collector and when it is reported to a credit bureau (Brevoort et al. 2018). Thus, we believe that the age 65 year belongs with neither the pre-Medicare nor the post-Medicare regime.

The choice of bandwidth—in this case the age range on either side of age 65 to include in the analysis—is an important issue in an RD study. After exploring a range of bandwidths, we chose to limit the sample to those aged 55 to 75 in a given year for our preferred specification, resulting in a subsample of more than 1.1 million unique consumers per year for our main analysis.¹³

Because we hypothesize that protective effects of Medicare will be most pronounced for those who transition from uninsured to insured, we expect to find larger discontinuities in places where the uninsured rate for those younger than 65 is higher. We explore this by separately estimating models that include only individuals residing in zip codes in specific quartiles of the distribution of uninsured rates among those aged 55–64.

A number of important features of the ACA were implemented in 2014, including the Medicaid expansion in most states that expanded Medicaid, the individual mandate, and the availability of insurance exchanges in the individual market, subsidized for those with incomes between 100 and 400 percent of the poverty level. To explore whether the ACA modified the effect of Medicare on medical collections, we also estimate a version of the baseline model that includes the 2014–16 archives (covering changes in collections between 2014 and 2015 and between 2015 and 2016, omitting 2013–14 as a transitional year). This model includes a post-ACA dummy variable, which is fully interacted with G65_it and the other age-related variables, to test whether the discontinuity at age 65 changed over these periods. We also estimate this model separately for the highest and lowest uninsured rate quartiles.

As a form of placebo test, we estimate models where the dependent variable is a function of the change in nonmedical collections. This is not a pure placebo test, because other research has shown that changes in health insurance coverage can also have effects on nonmedical collections (Dobkin et al. 2018), perhaps in part because some collections of a medical origin appear in the data as nonmedical collections. Nonetheless, if discontinuities in collections near age 65 are truly due to Medicare we expect them to show up more strongly for medical than for nonmedical collections.

As in other regression discontinuity studies of the effects of Medicare (Barcellos and Jacobson 2015; Card, Dobkin, and Maestas 2008, 2009), central to our analysis are assumptions that insurance coverage changes discontinuously at age 65 while age profiles of other variables that might affect outcomes evolve more smoothly. We use data from the American Community Survey to examine the age profiles of insurance coverage, employment, and incomes over the time period covered by this study. In addition, a lack of smoothness in the distribution of the running variable—age in our case—at the cutoff can be a concern in RD analyses (McCrary 2008). We explore this issue within our sample and in comparison with US Census Bureau estimates of the population by age during the relevant time period.¹⁴

V.  Results

A.  Insurance Coverage, Employment, Income, and Age Distribution

To explore the plausibility of our RD strategy, we turn first to the ACS data. We consider two time periods: 2011–13, our baseline period, and 2014–16, the ACA post-implementation period. We use all observations for adults aged 55–75, weighted by the person weight provided in the survey.

Figure 1 examines the age profile in health insurance coverage over the two time periods. The left panel considers the share with Medicare coverage by age. While some people gain Medicare coverage before age 65 (mainly by qualifying for Social Security disability insurance or because they have end-stage renal disease), we see a large jump in that share at age 65 when almost everyone else becomes eligible. The age profiles for the two time periods are virtually indistinguishable. The middle panel examines the share with insurance from any source. Although more than 88 percent of 64-year-olds had some form of insurance in the earlier period, there is a jump upward of more than 9 percentage points at age 65, and about 10 percentage points between age 64 and age 66. This panel also provides some evidence about the likely effect of the ACA on insurance coverage, in the comparison of the two profiles. The share with insurance at age 64 was a full 4 percentage points higher after implementation of the ACA than before, while the share with insurance at age 65 increased only slightly.

Most of those with Medicare have some form of supplementary coverage, which we examine in the right panel of Figure 1. It shows age profiles of the share who report having insurance coverage from more than one source. This share makes a large jump at age 65. The share is slightly lower at age 64 in the post-ACA period as compared with the pre-ACA period, and it is also a bit lower at all ages over 65 in the post-ACA period. All of the findings in Figure 1 for the baseline period closely mirror findings in previous studies using other data sources and earlier years (Barcellos and Jacobson 2015; Card, Dobkin, and Maestas 2008, 2009).

Figure 2 explores the smoothness of other variables across the age 65 threshold. The left panel displays the age profile of employment. The share employed declines from nearly 60 percent at age 60 to about 20 percent at age 70, but visually the decline appears to happen rather smoothly. The right panel shows the age profile of the mean of personal income, adjusted to 2014 dollars using the consumer price index.¹⁵ The ACS questions ask about income received in the previous 12 months, so income does not match up perfectly with the age at which it was received. We plot incomes reported by those aged 61 at age 60, and so forth. There is some visual evidence of an upward shift in the profile of income at age 65, in both time periods.

We test for discontinuities in the age profiles of employment and income in Table 1. Results are based on regressions pooling all years 2011–16 and including all individuals aged 55–75. In columns 1 and 3, explanatory variables include a quadratic in age, fully interacted with a dummy variable for age ≥ 65, and year dummies. Standard errors are clustered by age. The table reports the coefficient on the intercept shift at age ≥ 65 (first row), as well as that coefficient as a percentage of the mean of the dependent variable in the age 60–64 population (fourth row). Results show no significant discontinuity in employment, but an increase in personal income at age 65.¹⁶ While relatively small in percentage terms, the upward discontinuity in income might contribute to better credit outcomes after age 65. While we know of no estimates of the effect of an income increase of this size on credit outcomes, in the Appendix we explore the correlation between income and medical collections in the Health Tracking Household Survey (a data source used in Barcellos and Jacobson 2015), finding suggestive evidence that the effect would be small relative to discontinuities we find at age 65.

	Employment (%)		Personal income ($)
	(1)	(2)	(3)	(4)
Coefficient age ≥ 65	0.06	0.02	$1,601a	$1,636a
Standard error	0.91	0.90	194	198
Mean, age 60–64	52.4%	52.4%	$43,796	$43,796
Relative change (%)	0.1	0.0	3.7	3.7
Birth year ≥ 1946	—	−0.45b	—	$227
Standard error	—	0.19	—	255
Relative change (%)	—	−0.9	—	0.5
N	4,704,287	4,704,287	4,538,126	4,538,126

Note. Sample includes all adults age 55–75 from the ACS, pooling years 2011–16. All models include quadratic in age that differs before and after age 65, and year dummies. Models presented in columns 2 and 4 additionally include a dummy variable indicating whether ACS respondents were born after 1945 (baby boomer generation). “—” indicates not applicable. Estimation by ordinary least squares, standard errors clustered by age. Relative change is coefficient as a percent of age 60–64 mean. Personal income measured in 2014 dollars. Significant at

^a 1 percent,

^b 5 percent,

^c 10 percent levels.

View Table Image

Turning to our credit bureau data, examining the age distribution in our sample revealed a drop in sample size at age 66 for $t=2012$ and at age 67 for $t=2013$. As we show in Appendix Figure A13, however, the distribution of the US population by age changed during this time, and the discontinuities in the age distribution are associated with the increase in births from the post–World War II baby boom beginning in 1946.¹⁷ Although our baseline sample appears to be representative of the population by age, the fact that all those younger than age 65 are baby boomers while nearly all of those older than age 65 are not could still be a concern for our RD design if there are differences in socioeconomic status between baby boomers and others at the same age born slightly earlier. We explore this by adding an indicator variable for baby boomer status to the ACS models in Table 1 (columns 2 and 4). The indicator takes the value 1 for those born after 1946, as inferred from their age when surveyed. We find a quite small (negative) effect on employment and a small and insignificant effect on income.¹⁸

B.  Age Profiles of Changes In Medical Collections, 2011–13

Figure 3 plots the age profiles of the one-year change in medical collection balances from our baseline sample for several features of the distribution. These are changes between 2011 and 2012 and between 2012 and 2013, studied simultaneously and treated as one distribution. We are interested in discontinuities in the distribution at age 65, which might be effects of Medicare (with the caveat that the age 65 year itself reflects a mixture of things that happened before and after age 65). It is clear in the figure that the likelihood of a positive change in balance and percentile values in the right tail of the distribution are declining prior to age 65. The decline in the likelihood of experiencing a new medical collection in the past 12 months (Figure 3, top left panel) appears to happen rather smoothly across ages, so there is no strong visual evidence of discontinuous change in that variable due to Medicare.¹⁹ That likelihood was falling steadily toward 5 percent at age 65, which is consistent with the 95th percentile value of changes in collection balance being zero at all ages 65 and above (Figure 3, top right panel). Further in the right tail, however, at the 99th percentile (bottom left panel), there does appear to be a discontinuity downward, most evident at age 66.

The age profile of the average change, shown in the bottom right panel of Figure 3, is a somewhat different and more complex story. While the average change over the entire sample is about $5, the figure provides compelling visual evidence of a drop from positive to negative values beginning at age 66, followed later by a rise toward zero, especially evident at age 72. Table 2 shows some additional features of the distribution of changes in medical collections in the baseline sample, providing some insight into the behavior of the mean. Only about 11 percent of consumers experienced a nonzero change, roughly split in half, positive (5.6 percent) and negative (5.2 percent). That is, the average is determined roughly by the bottom and top 6 percent of the distribution, which include much larger values compared with the mean. For example, the average change among those whose medical collection balance increased was approximately $1,400.²⁰

	< $0	= $0	$0	Mean	p1	p5	p95	p99
	(%)	(%)	(%)	($)	($)	($)	($)	($)
One-year change in medical collections balance	5.2	89.3	5.6	5	−1,327	−25	38	1,442

Note. p[X] denotes the [X]th percentile of the distribution. N = 2,387,833, min = $486,589, max = $672,830.

View Table Image

It is tempting to assert that changes in collection balances in the right tail of the distribution are dominated by new collections, but this issue warrants closer consideration. Recalling equation 1, we are interested in the effect of Medicare on the distribution of new collections, $y_{\mathit{it}}^{\mathrm{new}}$, but in general we observe only $\mathrm{\Delta }{y}_{\mathit{it}}=\mathrm{\Delta }{y}_{\mathit{it}}^{\mathrm{old}}+y_{\mathit{it}}^{\mathrm{new}}$. If there are discontinuities near age 65 in the distribution of $\mathrm{\Delta }{\mathrm{y}}_{\mathit{it}}^{\mathrm{old}}$, our inferences about the effect of Medicare on the distribution of $y_{\mathit{it}}^{\mathrm{new}}$could be biased.

To gain some insight into this issue, we focus on the group for whom we do observe $y_{\mathit{it}}^{\mathrm{new}}$—those who begin a period with a medical collection balance equal to zero, which implies $\mathrm{\Delta }{\mathrm{y}}_{\mathit{it}}^{\mathrm{old}}=0$. This group includes a large majority of the full baseline sample (82 percent overall, increasing with age), but a much smaller share of those who experience a new medical collection (30 percent).²¹ If we condition on experiencing a new collection during the period, however, the general age profile looks similar for those with and without positive initial balances. We illustrate this point in Figure 4. Here values at the 90th percentile are somewhat larger for those who began with a positive balance, but the downward shift, most evident at age 66, is quite similar for the two groups. This downward shift across both groups is also apparent at other percentiles above the median (not reported).

While this analysis is not definitive, the similarity in discontinuity at age 66 between a group for whom we observe $y_{\mathit{it}}^{\mathrm{new}}$ directly and a group for whom it is comingled with changes in old balances suggests that changes in age profiles in the right tail of the distribution of Δy primarily reflect changes in new collections. In what follows we do not condition on experiencing a new collection, because the probability of experiencing one could be influenced by Medicare. In the unconditional distribution the value at the 90th percentile is zero, so we pay special attention to values further in the right tail.

C.  Modeling Changes in Collection Balances, 2011–13

We next consider RD models of changes in collection balances in the 2011–13 period based on equation 2. To look at outcomes in the right tail of the distribution we take two approaches, both of which tell similar stories. One approach uses quantile regression; the other models the probability of changes in balances greater than a specific size. In this section we estimate models for medical and nonmedical collections using the full sample and also separately using the subsamples of individuals residing in the top and bottom quartiles of the distribution of uninsured rates by zip code. Table 3 shows features of the distribution of uninsured rates by zip code for those aged 55–64, based on ACS five-year estimates, 2009–13. The mean percentage uninsured is 12.6 percent, the median is 11.0 percent, and the 25th and 75th percentile values are 6.9 and 16.4 percent respectively.

Statistic	Percentage uninsured
Mean	12.6
10th percentile	4.1
25th percentile	6.9
Median	11.0
75th percentile	16.4
90th percentile	22.9

Note. Based on ACS 5-year estimates of the uninsured rate by zip code, 2009–13, age 55–64, weighted by observations in baseline sample.

View Table Image

Our main results for both medical and nonmedical collections balances are summarized in Figures 5 through 10 and Tables 4 and 5.²² We turn first to results for the full sample, shown graphically in Figures 5 and 6. The figures include values at each age taken from the data as well as curves from the fitted models. The sample value for age 65 is included in the figures, although observations at age 65 are not used in estimating our donut regression discontinuity models.

Panel a: Full sample
	Mean ($)	> $0 (%)	> $1,000 (%)	> $2,000 (%)
	Medical collections
Coefficient age ≥ 65	−25a	−0.24a	−0.31a	−0.26a
Standard error	5	0.08	0.04	0.04
Mean, age 60–64	12	5.9	1.6	0.8
Relative change (%)	−211.9	−4.1	−19.3	−30.3
	Nonmedical collections
Coefficient age ≥ 65	−6	−0.06	−0.12b	−0.10b
Standard error	9	0.12	0.06	0.04
Mean, age 60–64	−25	7.6	2.6	1.6
Relative change (%)	26.0	−0.8	−4.8	−6.4
N	2,261,791	2,261,791	2,261,791	2,261,791
Panel b: Highest quartile of zip codes by uninsured rate
	Mean ($)	> $0 (%)	> $1,000 (%)	> $2,000 (%)
	Medical collections
Coefficient age ≥ 65	−31b	−0.51a	−0.70a	−0.52a
Standard error	12	0.13	0.08	0.08
Mean, age 60–64	13	8.5	2.5	1.3
Relative change (%)	−235.3	−6.0	−28.2	−38.6
	Nonmedical collections
Coefficient age ≥ 65	7	−0.02	−0.09	−0.10
Standard error	12	0.27	0.18	0.13
Mean, age 60–64	−25	11.8	4.0	2.4
Relative change (%)	−27.4	−0.2	−2.3	−4.1
N	552,350	552,350	552,350	552,350
Panel c: Lowest quartile of zip codes by uninsured rate
	Mean	> $0	> $1,000	> $2,000
	Medical collections
Coefficient age ≥ 65	−20a	−0.21b	−0.13b	−0.20a
Standard error	3	0.07	0.04	0.03
Mean, age 60–64	10	0.03	0.01	0.00
Relative change (%)	−192.4	−6.4	−15.8	−48.2
	Nonmedical collections
Coefficient age ≥ 65	−15	0.00	−0.15c	−0.12a
Standard error	13	0.14	0.08	0.03
Mean, age 60–64	−19	4.2%	1.4%	0.9%
Relative change (%)	79.1	−0.1	−10.5	−12.9
N	555,420	555,420	555,420	555,420

Note. Dependent variable equals 1 if the change in collection balances is greater than the specified size ($0, $1,000, $2,000), and the corresponding coefficient is expressed in percentage points. For the column labeled Mean, the dependent variable equals the change in collections, and the coefficient is expressed in dollars. Model includes quadratic in age that differs before and after age 65, and year dummy for 2012–13. Estimation by ordinary least squares, standard errors clustered by age. Full sample includes all aged 55–75 present in contiguous years, age 65 omitted. Panel b includes those residing in highest quartile of zip codes by uninsured rate, based on 2009–13 ACS. Panel c includes those residing in lowest quartile.

^a Significant at 1 percent level.

^b Significant at 5 percent level.

^c Significant at 10 percent level.

View Table Image

Panel a: Full sample
	98th	99th	99.5th
	Medical collections
Coefficient age ≥ 65	−211a	−381a	−1,050a
Standard error	58	120	221
Percentile, age 60–64	735	1,685	3,357
Relative change (%)	−28.6	−22.6	−31.3
	Nonmedical collections
Coefficient age ≥ 65	−134	−257	29
Standard error	218	499	976
Percentile, age 60–64	1,456	3,561	7,413
Relative change (%)	−9.2	−7.2	0.4
N	2,261,791	2,261,791	2,261,791
Panel b: Highest quartile of zip codes by uninsured rate
	98th	99th	99.5th
	Medical collections
Coefficient age ≥ 65	−391a	−1,088b	−1,921a
Standard error	133	538	614
Percentile, age 60–64	1,290	2,717	5,259
Relative change (%)	−30.3	−40.1	−36.5
	Nonmedical collections
Coefficient age ≥ 65	−179	125	208
Standard error	451	1,331	2,136
Percentile, age 60–64	2,481	5,321	9,842
Relative change (%)	−7.2	2.3	2.1
N	552,350	552,350	552,350
Panel c: Lowest quartile of zip codes by uninsured rate
	98th	99th	99.5th
	Medical collections
Coefficient age ≥ 65	−54	−139	−448
Standard error	35	145	1,429
Percentile, age 60–64	195	740	1,655
Relative change (%)	−27.9	−18.8	−27.1
	Nonmedical collections
Coefficient age ≥ 65	−87	−200	−819
Standard error	147	372	1,166
Percentile, age 60–64	556	1,787	4,699
Relative change (%)	−15.6	−11.2	−17.4
N	555,420	555,420	555,420

Note. Coefficients expressed in dollars. Model includes quadratic in age that differs before and after age 65, and year dummy for 2012–13. Estimation by quantile regression, qreg in Stata 15, age-based block bootstrapped standard errors, 250 replicates. Full sample includes all aged 55–75 present in contiguous years, age 65 omitted. Panel b includes those residing in highest quartile of zip codes by uninsured rate, based on 2009–13 ACS. Panel c includes those residing in lowest quartile.

^a Significant at 1 percent level.

^b Significant at 5 percent level.

^c Significant at 10 percent level.

View Table Image

Figure 5 illustrates the estimated discontinuities at age 65 in mean changes in collection balances, in the probability of a positive change and in the probability of a change greater than $1,000. As we saw in Figure 3, the mean change in medical collection balance drops discontinuously after age 65 before rising again at age 72. A discontinuity is less evident for the nonmedical balance. The age profiles for positive changes and increases of more than $1,000 in collections look rather similar for medical and nonmedical (albeit at a higher level for nonmedical), but a discontinuity after age 65 is most evident for the large increase in medical collections.

Panel a of Table 4 shows estimates of the discontinuity at age 65 in the models shown in Figure 5 and in addition in models for changes in collection balances greater than $2,000. The coefficient on the age ≥ 65 variable corresponds in the respective figure to the difference between the height of the right-hand line segment and the left-hand line segment at age 65. The table also includes the value of the dependent variable in the subsample aged 60–64 to aid in interpreting the magnitude of the discontinuity.

For medical collections, we see statistically significant negative effects in each column, including a reduction in the mean change of $25, and a 0.31 percentage point decrease in the probability of a change above $1,000. Interestingly, the discontinuity in the probability of any increase in medical collection balance, in absolute terms, is more than fully accounted for by the discontinuities for increases of large amounts, more than $1,000 and more than $2,000. The changes relative to the baseline values for those aged 60–64 are 19 percent for increases above $1,000 and 30 percent for increases above $2,000. Point estimates of discontinuities in increases in nonmedical collections balances are also negative in each column, but the magnitudes are much smaller and the estimates are only significant for large changes.

Figure 6 illustrates the estimated discontinuities at age 65 at the 98th, 99th, and 99.5th percentiles of changes in medical and nonmedical collection balances. While there do appear to be discontinuities for medical collections balances in each case, most apparent at age 66, they are less apparent for nonmedical collections balances. Panel a of Table 5 presents discontinuity estimates at age 65 corresponding to Figure 6. For medical collection balances, Table 5 shows evidence of a downward shift at age 65 in the value at each of these percentiles, with the discontinuity increasing in size as we move further into the tail of the distribution, ranging from $211 at the 98th percentile to $1,050 at the 99.5th percentile. Relative to the value at each percentile in the group aged 60–64, the size of the discontinuity is rather stable, in the range of a one-quarter to one-third reduction. Estimated discontinuities for nonmedical collection balances are smaller in magnitude and not statistically significant.

Next, we turn to results among those who resided in zip codes with relatively high and low rates of uninsured adults near age 65. Figure 7, and panel b of Table 4, use the sample of residents of the top quartile of uninsured rates by zip code, whereas Figure 9, and panel c of Table 4, use residents of the bottom quartile of uninsured rates.²³ The discontinuity in the mean is largest in the high-uninsured sample and smallest in the low-uninsured sample, and the discontinuities in the probability of changes above various sizes follow the same pattern. Results for nonmedical collection balances are, in almost every case, smaller in magnitude than those for medical collections.

Figures 8 and 10, and panels b and c of Table 5, again use the samples from high and low quartiles of uninsured rates by zip code, this time presenting discontinuities at specific quantiles of the dependent variable. Estimated discontinuities in medical collection balances at each percentile are largest in the high-uninsured-rate sample, and smallest in the low-uninsured-rate sample and not statistically significant in the low-uninsured sample. Estimated relative changes are somewhat more comparable across samples. Roughly speaking, it appears that the whole distribution of changes shifts to the left as the uninsured rate in an area increases, and the discontinuity at age 65 moves along with it. For example, the estimated downward discontinuity at age 65 in the high-uninsured-rate sample at the 98th percentile ($391) is about the same as the discontinuity at the 99th percentile in the full sample ($381), and only a little smaller than the discontinuity at the 99.5th percentile in the low-uninsured sample ($448). Discontinuities in changes in nonmedical collection balances are not as visually apparent and are never statistically significant (in all cases they are smaller in magnitude than their standard errors).

VI.  Discussion

This work provides compelling evidence that Medicare reduces financial strain related to medical care use among Americans at age 65, the eligibility age at which most transition into the program. We exploit the program’s eligibility age and compare collection balances, reflecting unpaid bills from medical providers, among those just younger and older than age 65 using a regression discontinuity design. Our main results suggest that Medicare reduced the mean increase in medical collection balance by about $25 per enrollee per year during our baseline period (2011–13). While this estimate is expressed on a per enrollee basis, this effect was concentrated among the much smaller percentage who experience large collections. We found that Medicare reduced the probability of an increase in medical collection balance larger than $1,000 by 0.31 percentage points, a 19 percent reduction relative to the probability at ages 60–64. Taking a slightly different approach, the change in medical collection balance at the 99th percentile decreased by $381, a 23 percent decrease relative to the value among those aged 60–64. We also examined discontinuities in nonmedical collection balances at age 65 and generally found much smaller effects, mostly statistically insignificant. The stronger effects on medical than nonmedical collections support our interpretation that the effects we estimate are due to Medicare.

We further hypothesize that Medicare reduces financial strain at age 65 mainly among those who were previously uninsured and completely exposed to medical care spending risk. To investigate the role of health insurance prior to age 65, we first looked separately at individuals who resided in zip codes with the highest and lowest shares of uninsured aged 55–64 (Figures 7–10 and panels b and c of Tables 4 and 5). We found that decreases in collection balances at age 65 were larger in the high-uninsured sample than in the low-uninsured sample, generally supporting our hypothesis. These results may only be suggestive, however, of how the effects of Medicare on collections vary, causally, with the baseline uninsured rate.²⁵

The second approach we used to study the importance of health insurance status, prior to age 65, exploits implementation of the Affordable Care Act’s major health insurance coverage provisions in 2014. The ACA led to substantial declines in the rate of uninsured for those younger than age 65, which suggests that the role of Medicare at age 65 on financial strain should decrease. Consistent with our hypothesis, we find suggestive evidence that discontinuities at age 65 in medical collections contracted the second year after ACA implementation (Table 6 and Appendix Table A3).

The most direct comparison to our results from elsewhere in the literature is a finding in Barcellos and Jacobson (2015) about the effect of Medicare on the probability of being contacted by a collection agency about overdue medical bills in the last 12 months. Their results, which are based on self-reported survey responses, appear graphically in their Figure 5, Panel B, and they report an estimate of a 2.8 percentage point (standard $\mathrm{error}=0.9$) decline at age 65 using their preferred model. In our data we examine the age profiles of the share with a new medical collection in the last 12 months (Figure 3, upper left panel) and the share with a positive increase in medical collection balance (Figure 5, middle graph). Our two approaches yield very similar age profiles (both are shown on the same graph in Appendix Figure A10). Values are somewhat lower than those in Barcellos and Jacobson (2015) Figure 5 prior to about age 70, but in general our age profiles look quite similar to theirs. The points by age describe smoother profiles in our much larger sample than is the case for Barcellos and Jacobson, and we find very little visual evidence of a discontinuity at age 65 (Figure 5). We nonetheless estimate in Table 4 a discontinuity of −0.24 (standard $\mathrm{error}=0.08$) percentage points, an order of magnitude smaller than the result they report. We emphasize, however, that conceptual differences of “collection” across data sources do complicate a direct comparison. Collections are precisely defined by the credit bureaus, whereas survey responses require respondent interpretation and most likely reflect a broader concept of debt.²⁶ Another distinction may arise should bill collectors find it optimal to focus their resources on collecting moderate to large debts, relative to small ones, suggesting that collection agency contact may not reflect the extensive margin of collection balances.²⁷

As noted above, we find that the decline in the probability of an increase in medical collection balances at age 65 is entirely accounted for by the decline in increases of more than $1,000. This result is similar to a finding of Brevoort et al. (2018) regarding the effects of the ACA Medicaid expansions on financial outcomes. They find much larger effects of the expansions on the probability of medical collections larger than $1,000 than on smaller-sized collections, even though the smaller collections are much more common.²⁸ Relatedly, Brevoort and colleagues report (their Table 2) that 38 percent of all new medical collections are removed from an individual’s record within a year, with less than 8 percent of those clearly marked as repaid, which may indicate that many small collections reflect clerical errors or disputes that are often resolved in the consumer’s favor.

It is also of interest to compare our findings with estimates in the literature of the effects of gaining Medicaid coverage on medical collections. Collections are more common in the younger populations targeted by Medicaid expansions; the likelihood of a large increase in collections (medical or nonmedical) falls rather sharply with age even prior to age 65 (as seen, for example, in Figure 5), despite the fact that costly medical events are more common at older ages. Rates of uninsurance also decline with age prior to age 65, however, which might largely explain declines in collections (Batty, Gibbs, and Ippolito 2018).

To make a meaningful comparison with Medicaid expansions we need an estimate of the effect of “treatment on the treated.” To do so we must scale our results to the share of the population for whom transitioning to Medicare has an effect on financial strain. We make the simplifying assumption that the treated are only those who transition from uninsured to Medicare coverage (along with, in some cases, some form of supplemental coverage). As discussed in the Introduction, this assumption is probably not entirely true, as insurance coverage changes in significant ways at the time of Medicare eligibility, even for many of those who were already insured, but the effects are likely to be largest for those who were previously uninsured.²⁹

Based on ACS data cited above, uninsured rates dropped by 10 percentage points between age 64 and age 66 during the time of our baseline sample. We estimate the mean reduction in medical collection balances at age 65 to be $25. Dividing by the change in the share uninsured yields an estimate of $250 for the effect of treatment on the treated (i.e., per new and previously uninsured Medicare enrollee). Finkelstein et al. (2012), in analyzing the Oregon Health Insurance Experiment, estimated reductions in medical collection balances for those who gained Medicaid coverage of $390 per person over an 18-month period, or roughly $260 per person per year. More recent estimates of treatment on the treated based on the ACA Medicaid expansion are somewhat larger. Brevoort et al. (2018) report an estimate of $609. Hu et al. (2018) report $1,140 as an estimate of the effect on total debt in collections (medical and nonmedical). Thus, our back-of-the-envelope estimate of $250 is quite similar to the Oregon estimate (in nominal dollars), but smaller than recent estimates based on the ACA Medicaid expansion.

We also carried out treatment on the treated calculations separately for each of the four quartiles of uninsured rates by zip code. The numerators were estimated effects on the mean increase in medical collection balance as seen in Table 4 and Appendix Table A6. For the denominator in each quartile, we took the difference between the average uninsured rate among those aged 65–74 and those aged 55–64, based on ACS data, 2009–13. Uninsured rates at the zip code level are not available in finer age groupings, so we multiplied these differences by the following:

In addition to unpaid debt sent to collections, Medicare may also affect out-of-pocket medical expenditures, direct payments for care not covered by insurance or other third parties. Barcellos and Jacobson (2015) find a mean reduction of $326 in out-of-pocket spending at age 65, much larger than the mean effect on medical collections of $25 that we estimate in the pre-ACA period. While we do not have data on out-of-pocket spending, it is very plausible that the effect of transitioning to Medicare would be larger on out-of-pocket spending than on debt in collections, because out-of-pocket spending is much more common. Barcellos and Jacobson (2015) find annual mean and median out-of-pocket spending to be $1,003 and $464 respectively in their pre-age 65 sample, whereas we find that less than 6 percent of those just under age 65 experience a medical collection of any size during a year.³⁰

Understanding the consequences of changes in medical collections is an active area of research (Finkelstein, Mahoney, and Notowidigdo 2018). In contrast to out-of-pocket payments, which by definition subtract from a household’s resources available for other uses, medical debt in collections is for the most part never repaid, and so may have a less direct and immediate effect on a household’s financial position (Finkelstein, Mahoney, and Notowidigdo 2018). Having debt in collections, however, has other negative consequences with regard to access to credit, disutility from dealing with debt collectors, and increased probability of bankruptcy. For example, in their study of the ACA Medicaid expansion, Brevoort et al. (2018) find that Medicaid coverage leads to better credit terms on other debt, in addition to decreases in medical out-of-pocket spending, and that the former effect is 69 percent as large as the latter.

Medical debt incurred by the uninsured and never repaid is part of the “implicit insurance” provided within the US health-care system. To the extent that Medicare reduces the amount of such debt incurred, it therefore benefits whoever would have borne the costs of that implicit insurance. While the incidence of those costs is a complex and unsettled question, several studies of the ACA Medicaid expansion show that it reduced the volume of uncompensated care provided by hospitals, improving the financial position of hospitals in expansion states relative to nonexpansion states (Blavin 2016; Dranove, Garthwaite, and Ody 2016; Camilleri 2017).

Finally, in conjunction with previous literature our findings suggest multiple implications for public policy proposals that influence the rate of uninsured among those near the current Medicare eligibility age. For example, policies that increase the eligibility age would increase the rate of uninsured for those whose age is above 65, yet lower than the newly proposed eligibility age.³¹ Consequently, out-of-pocket spending risk would increase for the uninsured below the new eligibility age, as would unpaid medical bills ultimately sent to third-party collection agencies. This implies a commensurate increase in uncompensated care to providers. Using our calculations above, increasing the eligibility age to 67 equates to roughly a $175 million increase in medical collections per year.³² Conversely, policies that increase the rate of insured near the Medicare eligibility age decrease out-of-pocket spending risk, medical collections, and uncompensated care, which we observed to some degree in this work with the implementation of the major health insurance provisions of Affordable Care Act in 2014. While these calculations do not represent net benefits or a welfare analysis of changes in the Medicare eligibility age, they do underscore additional costs and benefits for consideration when choosing policy options that influence the rate of uninsured among Americans who are near retirement age.

Funding Information

The Urban Institute supported this research.