Mark Duggan Stanford University and NBER
Atul Gupta University of Pennsylvania, The Wharton School
Emilie Jackson Stanford University
May, 2018
Working Paper No. 18-026
THE IMPACT OF THE AFFORDABLE CARE ACT: EVIDENCE FROM
CALIFORNIA’S HOSPITAL SECTOR
1
The Impact of the Affordable Care Act:
Evidence from California’s Hospital Sector
Mark Duggan Stanford University and NBER
Atul Gupta University of Pennsylvania, The Wharton School
Emilie Jackson Stanford University
May 2018
Abstract
The Affordable Care Act (ACA) authorized the largest expansion of public health insurance coverage in the U.S. since the mid-1960s. Evidence on the effects of the ACA-induced increase in insurance coverage on patient health and health care providers is still emerging. We deploy administrative data from the universe of general acute care hospitals and emergency rooms in California over 2008-15 and present new empirical evidence on the effects on insurance coverage, health care utilization, and hospital finances. Our empirical approach utilizes regression discontinuity and differences-in-differences research designs, exploiting sharp changes in Medicaid coverage due to age-based eligibility restrictions and pre-ACA variation in un-insurance shares across hospitals and markets. We have three principal findings. First, we find that approximately half of the Medicaid expansion replaced existing county safety-net programs – implying a large transfer from federal taxpayers to those in California. Second, although we find substantial increases in utilization of hospital stays and ER visits as well as sorting toward better quality hospitals, we find no detectable effects on patient health. Third, we find heterogeneous effects on revenue – government owned ‘safety-net’ hospitals experienced large gains in revenue, while gains are modest for private hospitals. Additional revenue does not manifest in improved quality metrics or capital investment.
Acknowledgements: We would like to thank Mark Shepard and several seminar participants at Stanford, UC-Berkeley, Arizona, USC, Johns Hopkins, Columbia and Wharton along with the 2017 ASSA, fall NBER PE 2017 and MHEC 2018 meetings for helpful comments. We are also grateful to Betty Henderson-Sparks, Amy Peterson and Jon Teague of the California Office of Statewide Planning & Health Development for their assistance in providing the Hospital/ER discharge data and to Jack Coolbaugh for excellent research assistance. All remaining errors are our own.
2
I. INTRODUCTION
The Patient Protection and Affordable Care Act (ACA) was signed into law in March 2010 and
authorized the largest expansion of publicly funded health insurance coverage since the introduction of
Medicare and Medicaid in the mid-1960s. The main provisions of this legislation took effect on January 1,
2014, with Medicaid enrollment in the U.S. increasing by 18 million since then, representing a thirty percent
increase in enrollment since 2013. Enrollment in federally subsidized private health insurance exchanges
now exceeds 12 million.1 The Congressional Budget Office estimates that the ACA-induced increase in
coverage cost the federal government $120 billion in 2017 (CBO, 2017). This is a unique modern setting
to quantify the causal effects of public insurance coverage on health care utilization and patient health as
well as examine general equilibrium effects of such a large intervention in the hospital sector.
We use data on the universe of general acute care hospital stays and ER visits in California from
2008 through 2015. Our primary empirical approach exploits sharp discontinuities in public insurance
coverage at ages 21 and 65 (see Figure 1 Panel A) which occur due to Medicaid eligibility rules. Pre-ACA,
eligibility restrictions caused a large share of beneficiaries to lose Medicaid coverage when they reached
either of these age thresholds.2 For the young, the drop in Medicaid coverage led to an increase in un-
insurance, while the near-elderly gained insurance coverage due to the onset of nearly universal Medicare.
The pre-ACA period itself would offers stand-alone quasi-experimental setting to examine the effects of
gaining Medicare (at 65) and losing Medicaid (at 21) using a regression discontinuity (RD) approach,
similar to previous studies (Card et al., 2008; 2009; Anderson et al., 2012; 2014).
However, our focus is on the change in this relationship at the age 21 and age 65 thresholds. The
ACA substantially relaxed Medicaid eligibility restrictions for all non-elderly individuals in California,
leading to large increases in Medicaid coverage for all adults between the ages of 21 and 64. This offers a
second quasi-experiment and the one that is the focus of our study – where individuals aged 21-64
experienced a greater change in Medicaid coverage due to the ACA, relative to individuals under the age
of 21 or aged 65 and up. We use an RD differences-in-differences (RD-DD) research design focusing on
patients close to the two age thresholds - 21 and 65. Differencing out pre-ACA discontinuities at the
thresholds serves two purposes. First, it allows us to estimate effects of the ACA expansion specifically as
opposed to a generic increase in insurance coverage. Second, it eliminates the role of potentially
confounding time-invariant unobserved factors that may affect health care at these thresholds (e.g. increase
1 Medicaid enrollment obtained from Centers for Medicare and Medicaid Services (CMS), available at https://www.medicaid.gov/medicaid/program-information/medicaid-and-chip-enrollment-data/report-highlights/total-enrollment/index.html. ACA marketplace enrollment obtained from Kaiser Family Foundation (KFF). 2 A small share of individuals retain Medicaid coverage post-65 because they are eligible for both Medicaid and Medicare. Medicare is the primary insurer in these cases.
3
in alcohol consumption at age 21). Since changes in coverage at these two thresholds are affected by other
factors beyond just the ACA, we employ a fuzzy RD-DD design and obtain IV estimates for the effects of
health insurance coverage. Further, by examining young and near-elderly patients independently, we shed
light on heterogeneity in the effects of the insurance coverage expansion.
We begin by focusing on changes in insurance coverage following implementation of the ACA
Medicaid expansion and insurance exchanges. We find no evidence of a net increase in private coverage,
suggesting that the exchanges did not play an important role for these specific groups – and that the increase
in coverage was driven primarily by Medicaid. There is a large (and well-documented) decrease in the share
of self-pay patients. However, we find this decrease is lower than the increase in Medicaid coverage at both
age thresholds, implying there is some crowd-out of other forms of coverage. Indeed, in our setting
Medicaid replaces a substantial amount of last-resort coverage provided by counties to low income patients.
About 40-50% of the increase in Medicaid coverage for the young and the near-elderly replaces existing
county programs, implying that much of the incremental spending by Medicaid on hospital care was a
transfer from federal taxpayers to local taxpayers in California. These results corroborate recent evidence
that Medicaid tends to largely replace uncompensated or safety-net care by providers and is therefore under-
valued by beneficiaries (Finkelstein et al., 2015).
Next, we test if utilization of hospital care changes for patients in these age groups. The net effects
of coverage on quantity are not obvious ex-ante. A decrease in patient cost sharing may spur greater use of
health care while improved access to preventative and outpatient care may decrease the need for hospital
care. This has been referred to as the access vs. efficiency tradeoff (Dafny and Gruber, 2005). We find that
the access effect dominates, particularly for elderly patients. Our IV estimates imply that insurance
coverage causes a 10-15% increase in the rate of hospital care for young adults, while the magnitude is four
times as large for elderly patients. Importantly, we find a statistically significant and meaningful increase
in ER utilization for both patient groups, in contrast to evidence from the earlier Massachusetts reform
(Kolstad and Kowalski, 2012; Miller, 2012). Our RD-DD estimates for young adults are lower in magnitude
than those by Anderson et al. (2012) who looked at the loss of private coverage at age 19 and found a 40-
60% decrease in hospital care, and closely match those of Card et al. (2008) who examined the effects of
universal Medicare coverage at age 65. In order to highlight the potential for differences between a partial
and general equilibrium evaluation of coverage effects, we note that our estimates for the elderly are about
twice as large as those obtained in the case of the Oregon experiment (Finkelstein et al., 2012).3
In addition to a change in the quantity of hospital care, we find robust evidence that insurance
coverage causes a switch in the type and quality of hospital at which patients receive care. IV estimates
3 Their appendix table A.26 reports that hospital stays for elderly aged 50-64 increased by 30% due to Medicaid coverage. In comparison, we find a 60% increase in hospital stays.
4
imply that across both patient groups, individuals are 15% more likely to use care at a private hospital when
they gain coverage. We interpret this to be mostly demand driven since we find a similar magnitude of
switching in ER use, which is less likely to be driven by physician counsel or insurer constraints.
Furthermore, we find that after conditioning on hospital ownership, patients are more likely to receive care
at better quality hospitals (risk adjusted mortality and readmission rates) once they receive insurance
coverage. The IV estimates imply that elderly patients shift to hospitals with 0.15 s.d. lower mortality rates.
Based on estimates from previous studies for a comparable population, this is equivalent to the utility gain
associated with moving 4 miles (25%) closer to the serving hospital (Tay, 2003). Traditionally, Medicaid
has been valued based on its effects on mortality (Currie and Gruber 1996b; Goodman-Bacon, 2016) or its
effectiveness in decreasing financial risk for beneficiaries (Brevoort et al., 2017). However, Medicaid
coverage also appears to enable switching to better quality providers, which may not be captured in
mortality effects, and is an additional source of patient welfare gain.
Notwithstanding the above changes in utilization of health care, we fail to reject the null of no
significant effects on health, measured through indicators for two outcomes – in-hospital mortality and
whether the episode was potentially avoidable with appropriate primary care. The latter helps indirectly
infer improvements in primary care, which we do not observe – and which would improve efficiency of
health care delivery. In case of both metrics, the IV point estimates suggest large improvements but are
noisy enough that we cannot rule out such effects in both directions.
The large increase in insurance coverage implies that most hospitals received greater average
reimbursement per patient relative to what they did pre-ACA.4 In light of the large effects of previous public
insurance expansions on hospitals (Finkelstein, 2007), we examine effects of the ACA on hospitals and
deploy a differences-in-differences research design to do so, exploiting variation in pre-ACA un-insured
patient shares across hospitals. Hospitals that served a higher proportion of uninsured patients prior to the
ACA are likely to see a greater relative increase in mean reimbursement per patient, provided their patient
profile does not change dramatically. The Medicaid expansion replaced both self-pay and county safety-
net programs but these were distributed unevenly across hospitals prior to the ACA. Government owned
‘safety-net’ hospitals disproportionately served county indigent patients, while self-pay patients were more
uniformly distributed across all hospital types. Hence the insurance shock may affect hospital revenue
differentially, based on what was replaced. Accordingly, we also deploy a flexible specification allowing
effects on revenue to vary based on a hospital’s pre-ACA self-pay and county shares.
4 The ACA did influence hospital reimbursement on other dimensions as well. For example, the ACA reduced the growth rate of Medicare reimbursement rates and also funding through the DSH program (which differentially aided hospitals serving many low-income patients). So it is possible that average reimbursement did decline for some hospitals, however mean annual revenue per bed increased by ~20% post-ACA.
5
We find three robust patterns across our specifications, although the point estimates are noisily
estimated. First, hospitals with higher un-insurance in 2008 do receive ~2% greater total revenue each year
in the period 2014-17. The Medicaid expansion entirely drives this increase, and is more than twice as large
as the combined decrease in safety-net spending from counties and disproportionate share (DSH) payments.
Second, there is substantial heterogeneity in these effects based on hospital ownership and the nature of un-
insurance prior to ACA. City, county or district owned hospitals previously had the highest shares of un-
insured patients and experience large increases in annual total revenue, while private hospitals experience
a relative decline. A related narrative emerges when we examine differences based on whether the hospital
previously had a high share of self-pay or county indigent patients. Government owned ‘safety-net’
hospitals experience large gains due to the expansion, while gains associated with decrease in self-pay
patients are more evenly spread. Third, the additional revenue does not manifest itself in improved quality
of care or additional capital spending. For example, we find no evidence of positive spillovers of quality of
care to infra-marginal patients at these hospitals (e.g. infant mortality, elderly mortality). We acknowledge
that effects on investments may yet surface with additional years of data, given the long planning horizon.
However, based on current evidence, it is unclear how the additional revenue has been deployed.
We perform a series of robustness checks of the key assumptions we make in sample construction
and model specifications, as well as a falsification check based on a placebo insurance expansion. Our core
results remain valid under these tests. We also conduct a supplementary exercise using the sample of all
non-elderly adults, exploiting geographic variation in pre-ACA un-insurance rates – and obtain qualitatively
similar results. Nevertheless, there are four key limitations of our analyses. First, our results reflect the
experience of a specific state and thus may not generalize to other states (especially the 19 states that chose
not to expand Medicaid). Second, we cannot observe health care delivered outside of the hospital. Because
of this, we cannot directly test for improvements in access to preventative and outpatient care. Third, the
RD estimates are most relevant to individuals in narrow age groups around the thresholds (21 and 65).
Finally, we do not have information regarding the effects of the ACA on measures of economic well-being
such as out-of-pocket spending, credit scores, or bankruptcy.
This paper makes three contributions. To our knowledge, we are the first to deploy administrative
data on utilization of hospital care to examine the effects of the ACA on utilization and health. Previous
studies have mostly used survey data and find that individuals have better access to care and lower out of
pocket spending (Golberstein et al., 2015; Benitez, 2016; Sommers et al., 2016a; Sommers et al., 2016b;
Wherry and Miller, 2016; Courtemanche et al., 2017b).5 In addition, we find evidence of crowd-out due to
5 Studies that have used administrative data have so far focused on one aspect of medical care, such as emergency departments (Garthwaite et al., 2017) or drug prescriptions (Ghosh et. al., 2017). These studies suffer from some limitations, such as having data from specific providers (former) or pertaining to a much smaller sector of health care (hospitals account for a third of all spending on medical care, while prescription drugs contribute 10-15%). Literature showing effects of the ACA on insurance
6
the Medicaid expansion. This aspect has either not received much attention or has not been detected using
survey data and differences-in-differences based designs (Courtemanche et al., 2017a; Frean et al., 2017).
Second, we provide new quasi-experimental evidence on the effects of insurance coverage on
demand for medical care in the presence of general equilibrium effects. Current baseline estimates are
derived from randomized controlled field experiments (Manning et al., 1987; Finkelstein et al., 2012), but
were obtained in a partial equilibrium setting. Some recent studies have exploited the health care reform in
Massachusetts (Kolstad and Kowalski, 2012; Miller, 2012). This reform was on a much smaller scale than
the ACA (4% increase in coverage vs. 10% increase in Medicaid alone in California). We find robust
evidence that beneficiaries switch hospitals once they gain insurance, an aspect that is valuable and has not
been discussed in previous studies on the value of Medicaid. Our research design is similar to Card et al.
(2008; 2009) and Anderson et al. (2012; 2014) who also exploit sharp discontinuities in coverage at specific
age thresholds. We build on their approach but deploy a RD-DD design to focus on effects of the ACA
insurance expansion and eliminate unobserved heterogeneity.
Third, we focus on the supply side effects of the insurance expansion and provide new empirical
evidence consistent with recent studies that providers may be the primary beneficiaries of a Medicaid
expansion (Finkelstein et al., 2015) as well as link to older evidence on the supply side effects of insurance
expansions (Finkelstein, 2007).
The rest of the paper is structured as follows. Section II provides background on insurance coverage
in California and the insurance provisions of the ACA. Section III describes the data and presents
descriptive statistics. Section IV describes the empirical approach and presents results on effects on
insurance changes, utilization and health. Section V presents the empirical approach and results on hospital
finances. Section VI presents robustness and falsification tests. Section VI discusses limitations in
interpreting the results and section VIII concludes.
II. BACKGROUND
A. Insurance coverage pre-ACA
The health insurance landscape prior to 2014 was characterized by relatively high un-insurance
rates among specific sub-groups. According to data gathered by the American Community Survey (ACS),
about 18% of the California population was uninsured in 2012-13. While this indicates a high aggregate
level of un-insurance, it masks wide variation in insurance coverage across different age groups. Panel B
presents the share of different insurer categories in California, as reported to the ACS, pre-ACA (2012-13)
and post-ACA (2014-15) periods for three age groups – children (under 21), non-elderly adults (21-64) and
coverage is already too voluminous to summarize here (Sommers et al., 2014; Golberstein et al., 2015; Benitez and Creel, 2016; Sommers et al., 2016; Courtemanche et al., 2017a; Frean et al., 2017).
7
the elderly (65 and above). Pre-ACA un-insurance rates among non-elderly adults (25%) were more than
three times that of the remaining population (8%). The elderly benefited from universal coverage provided
by Medicare, while children were generously covered by Medicaid (nearly 40%).
Surveys like the ACS may overstate true un-insurance rates since they do not recognize last-resort
county insurance programs for the medically indigent. These programs fund medical care for a subset of
low-income individuals who are not eligible for Medicaid but cannot afford to buy insurance. They are not
considered equivalent to insurance since they often require individuals to provide proof of medical need or
have a chronic condition. Hadley et al. (2008) estimates that about 20% of total spending on the uninsured,
or about $11 billion dollars, was covered by such local programs.
In California, each county designs its indigent services program and thus there is substantial
variation in eligibility requirements (e.g. income, assets, residence, age, medical need and immigration
status) and services covered (California Health Care Foundation, 2009). California spent more than 2.1
billion dollars in one pre-ACA year to care for the uninsured through programs such as the Medically
Indigent Services Program (MISP), which provided care in 24 mostly urban counties, and the County
Medical Services Program (CMSP), which operated in 32 predominantly rural counties (Council of
Economic Advisers 2009). With the exception of some MISP counties, these services were available only
to non-elderly adults. Hence, a substantial fraction of non-elderly adults counted among the uninsured pre-
ACA were covered by county programs.
B. The Affordable Care Act
The ACA was signed into law in March 2010 with several key objectives: increasing access to
health care, introducing new consumer protections, and lowering the cost and improving the quality of
health care. This paper investigates the effect on health insurance coverage and the corresponding effect on
health care utilization and health outcomes.
There were two primary channels through which the ACA expanded access to health insurance,
both of which became effective on January 1, 2014. First, in all states, individuals in families with incomes
between 100 and 400 percent of the federal poverty level (FPL) who were not already eligible for affordable
health insurance, either from an employer or from Medicaid, were now eligible for premium subsidies
provided in the form of advanced tax-credits to purchase private health insurance. Second, the ACA
originally intended to expand Medicaid eligibility to all individuals below 133% of the FPL. However, legal
challenges allowed states the choice to opt out of expanding Medicaid. California is one of the original
twenty-six states (including DC) that chose to expand Medicaid in 2014. Six other states have since elected
to expand Medicaid. Duggan et al (2017) provide a more detailed summary of ACA-mandated expansions
in health insurance.
8
Several surveys estimate the number of uninsured in the United States at the quarterly or annual
level. Gallup and Sharecare interview 500 Americans over the age of 18 daily and ask if the individual has
health insurance coverage. Their well-being index shows that the percent of adults without health insurance
was trending steadily upward prior to 2014, peaked around 18% in late 2013 and then sharply dropped to
11% by the beginning of 2016. The increase in health insurance coverage is largely attributable to both key
health insurance initiatives, the Medicaid Expansion and subsidized insurance through exchanges.
Although nationally there is a substantial decrease in uninsured rates, there is substantial heterogeneity in
the magnitude of the reduction at the state level. States that elected to expand Medicaid tended to have
substantially larger increases in insurance coverage since 2013.
Even among states that chose to expand Medicaid, there is substantial variance in the impact on
Medicaid enrollment. This is driven by variation across states in baseline enrollment, due to states’ initial
generosity in eligibility criteria, as well as differences in the socio-economic composition of states. Figure
A. 1 shows the percent of the state population enrolled in Medicaid in late 2013 and the net change in
enrollment between late 2013 and October 2016. Compare California and New York, where almost one-
third of residents in both states are now covered through Medicaid. However, this was a much larger
increase in California, which saw an increase of 10 percentage points compared to an increase of 4
percentage points in New York. New York had more generous eligibility criteria that included childless
adults prior to 2014 whereas childless adults were generally not covered in California. Consequently, the
expansion of Medicaid had a larger impact in California. Figure 1 Panel C displays monthly Medicaid
enrollment in California over 2010-16 and directly illustrates the scale of the expansion. It shows that –
after trending up very slightly through 2012 – enrollment increased from about 8.5 million in mid-2013 to
13.5 million by mid-2016.6 The figure also plots enrollment on the newly established ACA individual
insurance exchange and shows that it plateaued at 1.3 million, or about one-fourth of the increase in
Medicaid. Hence, the ACA primarily expanded insurance coverage in California through Medicaid.
Returning to the survey data summarized in Figure 1 Panel B, notice that the elderly experience
virtually no changes in insurer shares between 2012-13 and 2014-15. Un-insurance rates decline by 4
percentage points among children, driven entirely by a corresponding increase in Medicaid coverage. There
is a ten percentage point decrease in un-insurance among non-elderly adults, driven mainly by the Medicaid
expansion and increase in private coverage. Excluding the elderly since they were unaffected by design,
this survey evidence suggests that the decrease in un-insurance (~8 pp) is entirely explained by Medicaid
expansion (5.5 pp) and increase in private coverage (~2.5 pp), with little or no crowd out of other insurers
by Medicaid.
6 The small jump in enrollment in 2013 is due to the transition of children from the Healthy Families Program to Medicaid.
9
Survey evidence has three important limitations. First, individuals typically under-report welfare
coverage in surveys by 20 percent or more (Klerman et al., 2005; Meyer et al., 2009) and hence this
evidence likely under-states both the initial share of Medicaid as well as its increase. Second, county
indigent coverage is unlikely to be recorded at all and hence potential crowd out under the ACA cannot be
determined. Third, survey data on coverage cannot guide us on utilization and spending effects if Medicaid
beneficiaries are more likely to use care than the average individual.
C. Age based discontinuities in public insurance
Public insurance programs commonly use age-based thresholds to determine eligibility. For
example, individuals can enroll in Medicare when they turn 65, but not earlier, unless they are enrolled in
the Social Security Disability Insurance program or have end stage renal disease. Similarly, children enjoy
relatively generous eligibility rules under Medicaid until age 18 (or 19 under some circumstances) but then
often lose coverage because the eligibility criteria become more restrictive. Prior to the ACA, two such
rules created discontinuities in insurance coverage at 21 and 65 in California. Appendix
Figure A. 2 presents an extract of California Medicaid eligibility requirements as of September 2007.
Welfare recipients and disabled individuals were relatively generously covered. However, to enroll based
on low income status (“medically indigent person or family”), individuals had to be under 21. Adults aged
21-64 were generally ineligible except under very specific circumstances such as pregnancies, nursing
home residence, or enrollment in the federal Supplemental Security Income program.
To examine the magnitude of this discontinuity, we turn to administrative hospital discharge data.
Note that this will reflect insurance coverage conditional on using hospital care rather than share of coverage
in the population. Figure 1 Panel A presents Medicaid’s share of hospital stays for patients aged 10-75
discharged from hospitals during 2012-15. In 2012, Medicaid coverage is high for children aged 10 (~45%)
and gradually declines until age 21 when it falls precipitously by 15 percentage points exactly at that age.
It then varies smoothly again until age 65 when there is another discontinuous drop of about 12 percentage
points. The discontinuities at both age thresholds are large and account for 35-40% of the mean coverage
for individuals just below the threshold. The pattern is similar in 2013, except for a mechanical increase in
coverage for children due to the movement of CHIP beneficiaries into Medicaid.
The ACA makes three key changes -- as seen in the 2014 and 2015 trend lines. First, it eliminates
the discontinuity at age 21. Second, it accentuates the discontinuity at age 65 since non-elderly adults are
now more likely to receive Medicaid coverage. Third, there is a small increase in the share of Medicaid for
children as well. Elderly adults are unaffected, by design. The large discontinuities in Medicaid coverage
at the two age thresholds and their interaction with the ACA motivates our use of a regression discontinuity
research design to examine the effects of gaining insurance coverage on a variety of outcomes.
10
III. DATA
Our main source of data contains the universe of hospital stays and emergency room (ER) visits at
non-federal hospitals in the state of California for the period 2008 through 2015, obtained from California's
Office of Statewide Health, Planning, and Development (OSHPD). These confidential data include
approximately 3.8 million discharges and 11 million ER visits each year. Each observation pertains to a
hospital stay or ER visit and provides information on the hospital, dates of service, patients’ primary insurer
type and basic demographics, a vector of up to 25 diagnoses and procedure codes, and patient zip code. As
is standard in such files, if an ER visit subsequently leads to hospitalization, then it only appears as a hospital
discharge, though the record indicates whether the stay originated as an ER visit. Crucially, we observe
both a patient’s birth date and admission date and hence we can precisely calculate a patient’s age at
admission.
We impose three restrictions to arrive at the master sample used in analyses involving the discharge
data. First, we focus our attention on short-term general acute care hospitals to decrease the likelihood of
small and specific hospitals (for example, rehabilitation or long-term care) driving the results. This
restriction decreases the number of hospitals from 450 to 370, but retains 95% of hospital stays and nearly
all ER visits. Second, since California Medicaid eligibility rules were already generous regarding pregnancy
and delivery cases before the implementation of the ACA, we exclude pregnancy-related hospital stays or
pregnancy-related ER visits from the analysis. Third, we exclude patients residing outside California or
with missing zip codes of residence.7
We organize the insurance coverage into five categories – Medicare, Medicaid, Private, County
and Self. The first two include both traditional and managed care sub-segments. Private includes
miscellaneous smaller insurers such as government employees and workers’ compensation. Self is
essentially un-insurance and includes so-called charity cases and those who pay for their care out-of-pocket.
Throughout the paper, we measure un-insurance share as the sum of county and self.
A. Specific age thresholds
In order to construct the RD sample for our preferred specifications we impose two further sample
restrictions. First, we exclude the years 2008-11 in order to focus on the two years just prior to and following
ACA implementation. In a falsification check, we use data from 2008 through 2011, all of which precede
the Medicaid expansion. Second, we limit the sample to patients admitted within 12 months of their 21st
(or 65th) birthday. We exclude individuals who arrived at the hospital less than thirty days of turning 21 or
7 1-2% of the discharge records in 2008 were of patients having either an out of state or missing zip code.
11
less than thirty days before turning 65. These two groups of patients exhibit a partial change in insurance
status and hence their inclusion amplifies measurement error. In robustness checks we explore the
sensitivity of our results to using other age bandwidths. Focusing on specific age groups dramatically
curtails the sample size, leaving approximately 104,000 (390,000) hospital stays and 1.4 mn (0.9 mn) ER
arrivals over the period 2012-15 for the young and near-elderly respectively. ER arrivals include both ER
visits and hospital stays that originated in the ER. Throughout the paper we prefer to analyze the sample of
ER arrivals since it enables analysis without conditioning on hospital admission decisions that could change
in response to the ACA, while recognizing that the composition of ER arrivals may change after ACA
implementation.
Table 1 Panel A summarizes descriptive statistics on the main RD analysis sample of hospital stays
and ER arrivals separately for the young and elderly. The table highlights the sharp increase in the share
insured by Medicaid and the corresponding decrease in un-insurance for patients in these age groups. We
compute utilization rates as hospital stays and ER arrivals per 1,000 person-years using California
population estimates by single year of age for these years obtained from the ACS.8 The normalization is
particularly helpful in the case of elderly adults where this period coincides with the transition of baby
boomers into Medicare and hence large increases in the underlying population.
B. All non-elderly adults
An important limitation of the RD results is that the estimates are local to the specific age groups
represented. Therefore we supplement these results using a larger sample of all non-elderly adults (ages 21-
64) and exploit geographic variation in un-insurance rates across Hospital Service Areas (HSAs)9; this is
similar to the approach used in other studies (Finkelstein, 2007; Courtemanche et al., 2017; Duggan et al.,
2017; Frean et al., 2017). HSAs are defined as “collections of contiguous zip codes whose residents receive
most of their hospitalizations from hospitals in that area”. There are roughly 220 HSAs in California, and
on average an HSA is smaller than a county but much larger than a zip code. One can think of these as
similar in concept to Commuting Zones, which are often used to analyze shocks that affect labor markets.
We use them as the unit of analysis to examine an insurance shock to hospital markets. Table 1 Panel B
presents summary statistics on this sample. We exclude 2008 as the baseline year and the resulting data has
8.8 million hospital stays. Reassuringly, the mean values for several important variables are similar to a 8 We obtained California population estimates for 2012-15 using 1-year ACS survey data available through www.socialexplorer.com. Although population figures are available for specific ages like 20 and 21 separately, we pool the population for the young (18-24) and elderly (62-69) separately and assume they are uniformly distributed in these groups. Hence, the population is allowed to vary over time (i.e. more 65 years olds in 2015 relative to 2012) but not across ages in any given year. This introduces measurement error but ensures there are no false positive jumps in utilization because of a large change in the denominator. 9 HSAs were defined by the Dartmouth Atlas Project. There are roughly 220 HSAs in California, of which 79 and 34 are in the LA and San Francisco metropolitan regions respectively.
12
plausible weighted average of the mean values in the RD sample. For example, Medicaid contributes 24%
share of stays in the non-elderly sample in 2009-13, while it is 34% and 13% for the young and elderly
respectively in the RD sample in 2012-13. We discuss the research design in more detail and present
corresponding results in Appendix B.
C. Hospital finances
OSHPD collects financial data on all hospitals in California and makes it publicly available on its
website. These reports are mandated by California law and provide details on hospital finances, utilization
and capital investments. We use files covering 2009-16 in order to examine the effects of the insurance
expansions on hospital finances. The financial data is available for a smaller number of hospitals (335-340
instead of 370) since Kaiser Permanente10 hospitals (approximately 30) do not report their finances
individually. We make two transformations to the revenue data in preparation to use it in our analysis. First
we convert all nominal values into real 2016 dollar values using the consumer price index for urban
consumers. Second, we normalize aggregate revenues for each hospital by its licensed number of beds.
Table 1 Panel C presents descriptive statistics on hospital revenue and the contributions of different
types of insurers. Hospitals received about 1 million dollars per bed in revenue over the 2009-13 period. Of
this, 21% was contributed by Medicaid. This share increased dramatically post-ACA to 28%. Nearly 90%
of the increase in Medicaid revenue was due to the rapid growth in managed care, where all newly insured
beneficiaries were placed. In contrast, revenue from county governments under indigent programs fell
precipitously post-ACA. In the new regime, Medicaid is comparable to Medicare in magnitude of revenue.
IV. ACA EFFECTS ON INSURANCE, UTILIZATION AND HEALTH
A. Empirical strategy
Consider a conceptual reduced form model of the effect of health insurance coverage on outcome
𝑌𝑌 as below:
𝑌𝑌𝑖𝑖 = 𝛼𝛼 + 𝛽𝛽 ⋅ 𝐼𝐼𝐼𝐼𝑠𝑠𝑖𝑖 + 𝜖𝜖𝑖𝑖 (1)
𝑌𝑌𝑖𝑖 denotes an outcome of interest (including utilization of care) for individual 𝑖𝑖 and 𝐼𝐼𝐼𝐼𝑠𝑠𝑖𝑖 is an
indicator set to 1 if the individual has health insurance coverage and 0 otherwise. 𝜖𝜖𝑖𝑖 represents all
unobserved factors that affect outcome 𝑌𝑌. The key challenge in obtaining an unbiased estimate of the causal
effect 𝛽𝛽 is that individuals choose to purchase or enroll in health insurance coverage based on private
10 Kaiser Permanente is the largest health maintenance organization (HMO) in the US and owns all its medical care facilities – primary care, hospitals and post-acute care. Kaiser plan members are supposed to receive all medical care within this network. Individual medical centers do not report financial results publicly. More details available at: https://share.kaiserpermanente.org/article/fast-facts-about-kaiser-permanente/.
13
information about their health risk as well as their appetite for risk.11 Table 2 illustrates this self-selection
problem by presenting key attributes for insured and un-insured individuals at age 21 (Panel A) and 65
(Panel B) using 2004-09 data from the National Health Interview Survey (NHIS). For example, insured
young adults are much more likely to be in school and much less likely to be married, employed or smokers.
Insured elderly are more likely to be married or employed, but less likely to be smokers. The differences
(presented in column 3) are both statistically significant and economically meaningful. These individuals
are likely to differ on important unobservable characteristics as well, implying that the required condition
𝔼𝔼(𝜖𝜖𝑖𝑖 |𝐼𝐼𝐼𝐼𝑠𝑠𝑖𝑖) = 0 will not be not satisfied.
Several recent studies (Card et al., 2008; 2009; Anderson et al., 2012; 2014) have overcome this
endogeneity concern by exploiting the presence of age based insurance eligibility restrictions and resulting
discontinuities in coverage in a fuzzy regression discontinuity framework. For example, in our setting we
can exploit the discontinuous drop in insurance coverage that existed pre-ACA at age 21 (discussed in
section II.C) to determine the causal effect of insurance coverage by estimating equations of the type shown
below.
Insi = α10 + 𝜃𝜃1𝑑𝑑𝑖𝑖 + 𝜆𝜆11(𝑎𝑎𝑖𝑖 − 21) + 𝜆𝜆12𝑑𝑑𝑖𝑖(𝑎𝑎𝑖𝑖 − 21) + ϵ1i (2𝑎𝑎)
Yi = α20 + 𝜃𝜃2𝑑𝑑𝑖𝑖 + 𝜆𝜆12(𝑎𝑎𝑖𝑖 − 21) + 𝜆𝜆22𝑑𝑑𝑖𝑖(𝑎𝑎𝑖𝑖 − 21) + ϵ2i (2𝑏𝑏)
Equation 2a is the first stage that models insurance status for patient 𝑖𝑖 as a function of her age 𝑎𝑎𝑖𝑖
and whether she is 21 and older (𝑑𝑑𝑖𝑖 = 1). Insurance status is assumed to vary linearly with age
(through 𝜆𝜆11), allowing for a different slope for individuals over the threshold (through 𝜆𝜆12). Equation 2b
presents the corresponding reduced form relationship between outcomes of interest (𝑌𝑌𝑖𝑖) and age status 𝑑𝑑𝑖𝑖.
These equations would be estimated using data from the pre-ACA period on patients aged close to 21, say
within 12 months. The fuzzy regression discontinuity estimator of the causal effect of insurance coverage
on outcome 𝑌𝑌 is then given by 𝛾𝛾𝑅𝑅𝑅𝑅 = 𝜃𝜃2/𝜃𝜃1, and is equivalent to a local average treatment effect (LATE)
estimator (Hahn et al., 2001).
However the primary goal of this paper is to quantify insurance coverage changes caused by the
ACA and the resulting effects on utilization of care and patient health. To do so, we build on the above
framework by exploiting the fact that Medicaid expansions under the ACA led to dramatic changes in the
previously noted discontinuities in insurance coverage at ages 21 and 65. This setting therefore lends itself
to an RD-differences-in-differences (RD-DD) research design. Accordingly we adapt the above estimating
equations as below:
11 Other factors would surely influence this as well, including the price and quality of health insurance.
14
Insit = α10 + δ1t + 𝜃𝜃11𝑑𝑑𝑖𝑖 + 𝜃𝜃12𝑑𝑑𝑖𝑖 ⋅ 𝑇𝑇𝑡𝑡 + Di′λ1G(𝑎𝑎𝑖𝑖) + ϵ1it (3𝑎𝑎)
Yit = α20 + δ2t + 𝜃𝜃21𝑑𝑑𝑖𝑖 + 𝜃𝜃22𝑑𝑑𝑖𝑖 ⋅ 𝑇𝑇𝑡𝑡 + Di′λ2G(𝑎𝑎𝑖𝑖) + ϵ2it (3𝑏𝑏)
Equations 3a and 3b represent the modified first stage and reduced form equations respectively. We now
define 𝑑𝑑𝑖𝑖 more generally in order to accommodate both age thresholds of interest. In the case of the young
it denotes those aged 21 or older, while in the case of the elderly it denotes those aged 64 or younger.
𝑑𝑑𝑖𝑖 = � 1(𝑎𝑎𝑖𝑖 ≥ 21) if young 1(𝑎𝑎𝑖𝑖 < 65) if elderly
The indicator 𝑇𝑇𝑡𝑡 = 1(𝑡𝑡 ≥ 2014) denotes whether the ACA has been implemented. Both equations allow
outcomes to vary flexibly with age using the vectors 𝐷𝐷𝑖𝑖 and 𝜆𝜆. 𝐷𝐷𝑖𝑖′ = [1 𝑑𝑑𝑖𝑖 ] is a 1x2 vector of patient-
specific dummies. 𝜆𝜆1 and 𝜆𝜆2 are the corresponding 2x𝑘𝑘 matrices of age coefficients to be estimated, where
𝑘𝑘 is the order of the age polynomial, 𝐺𝐺. In our main results we use a linear polynomial in age, i.e. 𝑘𝑘 = 1.
This specification can be made more flexible by modifying 𝐷𝐷𝑖𝑖 and 𝜆𝜆 to allow different slopes in the post-
ACA period as well. We present robustness checks with such permutations in the appendix.
The coefficients of interest in this model are 𝜃𝜃12 and 𝜃𝜃22 and they estimate the change in the
discontinuity at the threshold due to the ACA (i.e. post vs. pre). The RD-DD estimator is then given by
𝛾𝛾𝑅𝑅𝑅𝑅,𝑅𝑅𝑅𝑅 = 𝜃𝜃22/𝜃𝜃12 (Persson, 2017). Note that we can also recover 𝛾𝛾𝑅𝑅𝑅𝑅 = 𝜃𝜃21𝜃𝜃11
= 𝜃𝜃2/𝜃𝜃1 using the modified
estimating equations. In practice we also include a full set of age-month cell and year fixed effects, 𝛼𝛼𝑗𝑗 and
𝛿𝛿𝑡𝑡 respectively. For some outcomes we also include a vector of patient controls 𝑋𝑋𝒊𝒊 to account for observable
differences in patient sickness, such as reason for arrival and gender. We cluster standard errors by age-
month cell to account for possible correlated error terms among patients of the same age.
Identification
We obtain two different economic objects of interest in the strategy outlined above. First, the
change in insurance coverage caused by the ACA – quantified by 𝜃𝜃12 in equation 3a. Since this is essentially
a differences-in-differences estimator, the identification assumption is that in absence of the ACA there
would be no change to the pre-ACA discontinuity, or the lack of a differential pre-trend. We present
supporting evidence through a falsification exercise with a placebo insurance expansion in 2010 that
indicates little or no change in insurance coverage between 2008-09 and 2010-11. To the extent that
insurance coverage increases for the control groups (ages 20 and 65) as a result of the ACA, this would be
differenced out as a common shock. Hence this approach will tend to underestimate the increase in
15
insurance coverage caused by the ACA. This is a pertinent concern in case of young adults since Medicaid
coverage increased substantially for 20 year olds. In our empirical work we primarily use the differential
insurance coverage changes to examine potential crowd-out by Medicaid, which does not rely on changes
in levels.
The second object of interest is the change in utilization of health care and health outcomes caused
by the ACA – quantified by the RD-DD (IV) estimator 𝛾𝛾𝑅𝑅𝑅𝑅,𝑅𝑅𝑅𝑅. Since this is a derivative of the RD
estimator, 𝛾𝛾𝑅𝑅𝑅𝑅 we first consider identification assumptions under which this would be valid. As discussed
in Lee and Lemieux (2010), three assumptions enable a causal interpretation. First, relevant observable and
unobservable factors that could affect the outcomes of interest should vary smoothly at the age threshold.
For example, if individuals are disproportionately likely to graduate college or enter employment exactly
at age 21 or exit the labor force exactly at age 65, this would be a violation. Table 2 column 5 presents
population weighted estimates from the NHIS on discontinuities in school enrollment, marital status,
employment and a number of other factors at ages 21 (Panel A) and 65 (Panel B). Column 4 presents mean
values at the thresholds to serve as comparison. The evidence reassuringly indicates there is no statistically
significant jump in these factors – with the exception of alcohol consumption which jumps at age 21 -- at
either of these thresholds. We return to concerns over 21 being the age of alcohol maturity and its
confounding effect on identifying utilization or health effects.
The remaining two assumptions are common to all LATE estimators (Angrist and Imbens, 1994).
The instruments (age thresholds and timing) must satisfy the exclusion restriction i.e. they do not affect the
outcomes of interest other than through their effect on insurance coverage. This assumption is more credible
due to the evidence that age does not affect employment, marriage or school enrollment which could be
alternative pathways affecting utilization of care and health. Finally, we need to assume the absence of
individuals who ‘defy’ their predicted insurance transition. In our setting, that implies there are no
individuals who were eligible for Medicaid or some other insurance at age 21(64) pre-ACA who would
lose it post-ACA. This is plausible by construction due to relaxation of Medicaid eligibility and introduction
of subsidized private coverage under the ACA.
The RD-DD estimator has two benefits over the RD estimator discussed above. First, it allows us
to focus specifically on the effect of insurance expansions under the ACA rather than estimate the causal
effect of insurance coverage more generally. Second, it differences out time invariant unobservable
differences in discontinuities at the threshold due to confounding factors (such as alcohol consumption).
Since alcohol legislation and consumption patterns have not changed during this period, it is plausible that
their effects on health care use and outcomes are stable over this period and indeed differenced out.
The nature of our data poses an additional econometric challenge in obtaining an unbiased estimate
of the change in insurance coverage. We observe only those individuals who arrived at a hospital for an ER
16
visit or inpatient stay. If lack of insurance coverage decreases the probability of utilizing hospital care, this
data will under-state the true extent of un-insurance, since those patients are “missing”. In our setting this
implies that the observed decline in insurance coverage at age 21 and 64 in the pre-ACA period may actually
understate the true decline. Consequently, the increase in coverage seen post-ACA may understate the true
increase. Since modeling insurance coverage is the first stage in IV estimation, this bias contaminates the
IV estimates as well. This issue is discussed at length in Anderson et al. (2012), who propose a modified
estimator to correct for this potential bias. The intuition behind the correction is straightforward – we
assume that discontinuous decline in utilization at the threshold in the pre-ACA period was driven by
individuals who lost coverage due to the eligibility restrictions. This is consistent with the exclusion
restriction that crossing the age threshold affects utilization only through loss of insurance. The corrective
procedure then adjusts the insurance coverage estimate by adding these ‘missing’ individuals back in as un-
insured patients. We then extend this approach to apply in our RD-DD setting and present both raw and
bias-corrected estimates in our results. Appendix C discusses the bias correction approach in detail.12
B. Insurance coverage
We begin by analyzing the change in insurance coverage for patients discharged from hospitals and
emergency rooms in California’s hospitals over the 2012-15 period. We describe the changes in insurance
coverage post-ACA as well as any crowd-out of other insurers due to the Medicaid expansion.
i. Changes in insurance post-ACA
Figure 3 plots observed and predicted changes in insurance coverage in 2014-15 relative to 2012-
13 (solid lines) for the young (Panel A) and elderly (Panel B) respectively. The predicted values were
obtained by estimating equation 3a. In both groups, insurance coverage increases differentially for the
treated patient sample (i.e. 21 and 64 year olds) post-ACA. The differential increase is much larger among
the young (~14 percentage point) as compared to the elderly (6 pp). One approach to interpret the magnitude
of this change in coverage is to compare it to the pre-ACA jump in coverage between the affected and
unaffected patient groups, which was 15 pp and 7 pp respectively for the young and elderly. Hence the
ACA nearly eliminated the disparity in insurance coverage at these two age thresholds.
Appendix Figure A. 3 presents a detailed version of this figure by plotting relative changes for
different insurer types – Medicaid, Private, Self-pay and County indigent care. We do not present the change
12 An additional limitation with this strategy is that it ignores possible spillover effects to the control group. If, for example, the large increase in coverage among those aged 21 to 64 leads to an improvement in care for those 20 and younger and/or for those 65 and up, this will lead us to understate the effects of coverage. If instead there are capacity constraints at certain hospitals and the insurance-induced increase in demand leads to less care for those under age 21 or over age 64, this approach will lead us to overstate the effects of additional insurance coverage.
17
in Medicare and miscellaneous coverage types since there is essentially none. The appendix figure indicates
that Medicaid expansion drives this observed increase in insurance. Medicaid appears to be the only source
of increase in coverage for both 21 and 64 year old patients and the Medicaid expansion appears of similar
magnitude (14pp and 8pp respectively) as the increase in overall coverage reported above.
Figure 3 also presents – as a falsification exercise -- the corresponding observed and predicted
changes in insurance coverage over 2010-11 relative to 2008-09 (dashed lines). The estimated magnitudes
are very small, not economically meaningful, and of the opposite sign — -2 pp (young) and -0.5 pp (elderly)
respectively. These estimates imply a small differential pre-trend of decreasing insurance coverage among
those just over age 21 and those just under age 65, which would work against our finding an increase in
insurance coverage for these two groups post-ACA. Appendix Figure A. 4 presents the corresponding plot
of insurance coverage change for patients arriving at EDs in California in this period. The patterns are
qualitatively similar to those discussed in case of hospital stays, with similar sized differential increase in
coverage for both young and elderly patients (~9 pp). There is no evidence of a differential pre-trend for
emergency department visits in the 2008-11 period.
ii. Crowd-out
An important policy concern associated with the expansion of public insurance – especially means-
tested programs like Medicaid – is the potential crowd-out of other sources of insurance. This research
design is well suited to examine crowd-out since we are comparing changes in insurer shares for all patients
in the state in these age groups.
Table 3 presents formal estimates of the change in insurance coverage at the two age thresholds for
patients discharged from hospital stays, obtained by estimating equation 3a. Panels A and B present results
for the young and elderly respectively. Each value listed in the table is a coefficient from a different
regression. Within each panel, the top row presents coefficients obtained from a naïve or as-is estimation
of the data. The bottom row presents bias-corrected estimates accounting for potential undercounting of un-
insured patients. We mainly discuss the naïve estimated effects, selectively turning to the bias-corrected
values when they diverge meaningfully.
Table 3 Panel A indicates that overall insurance coverage for the young increases one-for-one with
Medicaid (~ 14 pp). However, self-insurance and charity care decrease only by about 6 pp, less than half
of the Medicaid expansion. It appears that the remainder of the Medicaid expansion (55%) replaces county
indigent programs, which decline by 8 pp. We refer to this as the crowd-out of county coverage by
Medicaid. Table 3 Panel B indicates crowd-out among the elderly as well. While Medicaid coverage
increases by 8 pp, self-pay and charity care decrease by only 2.5 pp or 30% of the Medicaid expansion.
18
The bias-corrected estimates diverge in two important ways for elderly patients. First, they indicate
no net crowd-out of private coverage implying that insurance coverage increased one-for-one with
Medicaid at ~8 pp. This also implies that in the case of the elderly, the bias-corrected IV estimates will be
smaller than the naïve IV estimates. Second, they indicate a greater decline in self-pay at 5.3 pp or two-
third of the Medicaid expansion, implying that the decline in county programs accounts for about 35% of
the Medicaid expansion.
On average, for these two specific patient age groups, we find that about 40-45% equivalent of the
Medicaid expansion replaced existing county programs. Appendix Table A. 1 presents corresponding
estimated changes in insurance coverage for all patient arrivals at ERs. The patterns are qualitatively similar
to those discussed in case of hospital stays, although the implied magnitude of crowd-out is slightly lower
at 35%. In the interest of brevity we do not discuss these estimates.
The insurance coverage results have two implications. First, Medicaid expansion drives the
increase in coverage and there is no evidence of a (net) change in private coverage. Hence, the ACA
exchange enrollments do not lead to a net increase among this specific patient group. When we discuss the
effects on utilization and health, we will primarily interpret them as due to Medicaid. Second, the crowd-
out of county programs implies that the decline in self-pay and charity care – a key goal of the ACA -- is
much smaller in magnitude than the increase in Medicaid. Assuming for the moment that coverage changes
in hospital care are equivalent to changes in spending on hospital care, federal taxpayers incurred about
$200 more in Medicaid spending to decrease the burden of hospital care for the uninsured by $100.13 The
remainder is a transfer to local governments and the state of California that previously financed county
indigent programs. There are distributional implications as well. If we ignore differences in the costs of
raising taxes at different levels of government, this transfer was essentially borne by federal taxpayers
including those residing in states that did not expand Medicaid under the ACA.14
C. Utilization of care
i. Volume
Since our data is conditional on the person using hospital care, we cannot study rate of use at the
individual level. For the purpose of estimating utilization effects we use log of hospital stays or ER arrivals
as the outcome of interest, while for figures it is more intuitive to use rate of hospitalization or ER arrival
13 We find 55% and 40% crowd-out for ages 21 and 64 respectively. Appendix Table A. 3 shows that hospital stays are nearly equal in number in 2012-13 for patients aged 21-49 and 50-64 respectively. If we extrapolate the above estimates to apply to these larger age groups, then a weighted average estimate of crowd out is just less than 50%. 14 Reimbursement for Medicaid may diverge from that provided by the county indigent programs. If, for example, Medicaid reimburses hospitals twice as generously as the county indigent program, then the effective crowd-out of local spending would be correspondingly lower. Both Medicaid and the county indigent programs require small or no co-payment so the effective price on the demand side is not different for the two programs.
19
per 1,000 person years, which we compute as described in Section III.A. We collapse the data to the age-
month (denoted by 𝑠𝑠) - year level and estimate the following models.
𝐼𝐼𝐼𝐼𝑠𝑠����st = γ3t + Ds′ λ3G(𝑎𝑎�𝑠𝑠𝑡𝑡) + 𝜃𝜃31𝑑𝑑𝑠𝑠 + 𝜃𝜃32𝑑𝑑𝑠𝑠 ⋅ 𝑇𝑇𝑡𝑡 + ϵ3st (4𝑎𝑎)
𝑦𝑦𝑠𝑠𝑡𝑡 = γ4t + Ds′ λ4G(𝑎𝑎�𝑠𝑠𝑡𝑡) + 𝜃𝜃41𝑑𝑑𝑠𝑠 + 𝜃𝜃42𝑑𝑑𝑠𝑠 ⋅ 𝑇𝑇𝑡𝑡 + ϵ4st (4𝑏𝑏)
The above equations are exact analogs of equations 3a and 3b, which are estimated on case level data.
𝐼𝐼𝐼𝐼𝑠𝑠����𝑠𝑠𝑡𝑡 ,𝑦𝑦𝑠𝑠𝑡𝑡 , and 𝑎𝑎�𝑠𝑠𝑡𝑡 denote the mean coverage rate, log of hospital stays or ER arrivals and age of patients
in the age-month cell 𝑠𝑠, 𝑡𝑡. 𝑑𝑑𝑠𝑠 is the corresponding indicator obtained by collapsing 𝑑𝑑𝑖𝑖 within each age-
month cell. The coefficients of interest are 𝜃𝜃32 and 𝜃𝜃42 – the estimated change in the discontinuity in
insurance coverage and in utilization, respectively. We obtain the corresponding RD-DD (IV) estimator as
𝛾𝛾𝑅𝑅𝑅𝑅,𝑅𝑅𝑅𝑅𝑢𝑢𝑡𝑡𝑖𝑖𝑢𝑢 = 𝜃𝜃42/𝜃𝜃32.
Figure 5 presents a scatter plot of the change in the rate of utilization of hospital stays post-ACA
for the young (Panel A) and elderly (Panel B) by age-month cell. In addition, we also plot fitted values
obtained by estimating equation 4b. Figure 5 presents the corresponding figures for changes in emergency
room visits. Table 4 presents formal estimated effects on utilization of care for both the young (Panel A)
and elderly (Panel B). Columns 1 and 4 present results for all hospital stays and ER arrivals respectively.
Columns 2 and 3 examine effects on hospital stays that originated through the ER or not separately since
they may respond differently to insurance coverage. Column 5 presents results only on ER arrivals that did
not result in an inpatient stay. Each panel presents both naïve and bias-corrected estimates. The key
difference in the bias-corrected estimates is the use of a bias-corrected first stage estimate for the change in
insurance coverage. Again, we will discuss bias corrected estimates only if they meaningfully diverge from
the naïve estimates. In each column and panel, we present both the reduced form and IV estimates. We
discuss young and elderly patients separately since their responses may be quite different.
Young patients - Figure 5 Panel A indicates that hospital stays differentially increase for 21 year olds by
less than 1 stay per 1,000 after ACA implementation. It also indicates that utilization does not meaningfully
change for young patients in general. Table 4 Panel A confirms that the ACA had very minor differential
effects on hospital stays for 21 year olds. The increase in hospital use is driven entirely by stays that did
not originate in the ER, consistent with the explanation that demand for such care is more price elastic.
Figure 5 Panel A indicates that ER use does increase differentially and noticeably for 21-year olds
by about 4 arrivals per 1,000 individuals, which is more than half the baseline gap in ER use between 20
and 21 year olds over 2012-13. The IV estimate implies an 11% increase (marginally significant) in ER
arrivals per unit increase in insurance. The plot also highlights how the RD-DD design may understate true
20
effects of the ACA – ER arrivals for 21 year olds actually increased by 35 per 1,000, but this is masked by
a slightly smaller increase for 20 year olds who were also affected by the ACA but are the “control” group
in this design. We return to this issue in section VI where we propose an alternative approach to overcome
this and other limitations.
Elderly patients - Figure 5 Panel B presents the corresponding plot on change in hospital stays for the
elderly. The rate of hospital stays has declined post-ACA for both 64 and 65 year olds, but the decline is
greater for patients aged 65. We find a differential increase among 64-year olds of 8 stays per 1,000
individuals, which eliminates 50% of the pre-ACA gap in hospital stays. Two-thirds of this increase is
driven by stays that did not originate in the ER. Figure 5 Panel B reports a differential increase in ER
arrivals for 64 year olds of about 14 per 1,000, which is 70% as large as the pre-ACA gap in ER use. To be
conservative, we will primarily consider the bias-corrected IV estimates. Hospital stays and ER arrivals
increased by 60% and 40% respectively per unit increase in insurance under the ACA. These increases are
highly statistically significant and economically meaningful. Further, the large increase in ER arrivals
indicates that there was substantial un-met demand for medical care among the elderly un-insured before
ACA implementation.
Our estimated effects on utilization for the young are lower than those reported by Anderson et al.
(2012) who examine the effects of losing private coverage at age 20. While there are several possible
explanations, we believe the primary reason is the difference in generosity (prices) and access to care
between private insurance and Medicaid. Similarly, we believe our estimated effects for the elderly are
slightly lower than those reported by Card et al. (2008) for the same reason. They examine the effects of
the onset of Medicare15 which is more generous and widely accepted than Medicaid. However, our
estimates are twice as large as those obtained in a partial equilibrium setting, such as the Oregon Medicaid
experiment (Finkelstein et al., 2012). Their LATE estimate for near-elderly individuals (50-63) indicates
that Medicaid coverage causes a 30% increase in likelihood of hospitalization (Table A.26). While some of
the disparity in estimates is due to differences in the setting, sample and methodology – we believe a key
reason could be responses by hospitals and physicians to the large-scale Medicaid expansion and increase
in reimbursement rates, resulting in greater utilization.
ii. Choice of hospital
15 Card et al. (2008) examined the effects of the onset of Medicare coverage using data from California, Florida and New York. They find an 8 percent increase in the rate of hospitalization at age 65, while we find a 5% increase post-ACA. The estimated effects on hospital stays by admission route are 5% vs. 3% for stays originating in ER and 14% vs. 8% for stays not through ER. They do not provide IV estimates.
21
We now examine whether patients receive care at different types of hospitals once they receive
Medicaid coverage. We explore hospital choice on two dimensions – ownership type and quality as
measured by risk adjusted mortality and readmission scores. A key benefit of expanding insurance could
be enabling patients to choose higher quality (real or perceived) care providers. If so, this is a source of
welfare improvement for patients irrespective of whether there are measurable improvements in patient
health, and has received little attention in previous studies quantifying the benefits of Medicaid.
A. Hospital owner type
Figure 6 presents a scatter plot of post-ACA change in observed share of stays at government
hospitals for young (Panel A) and elderly (Panel B) patients respectively. It also presents the corresponding
fitted values obtained by estimating equation 3b. Table 5 presents formal estimated effects on hospital share
by owner type (columns 1-3) and mean hospital quality (columns 4-5). It follows the same format as that
used in Table 4. Figure 6 Panel A indicates that patient volume has shifted away from government hospitals
toward private hospitals for groups with differential increases in insurance. The share of government
hospitals decreases by about 2 pp post-ACA for patients aged 21, while it remains stable for patients aged
20. This is a large swing away from government hospitals and eliminates the pre-ACA gap in government
hospital usage between the two patient groups. Table 5 Panel A columns 1-3 presents the corresponding
regression coefficients for young patients and confirms that the decrease in the share of government
hospitals and increase in share of for-profit hospitals are statistically significant, while the increase for
private non-profits is not. The IV estimate indicates that gaining (Medicaid) coverage decreases the
likelihood of using a government hospital by 14%. This is economically significant since it represents
nearly a 50% decrease in the probability of receiving care at a government hospital for an uninsured patient
(30 pp in 2012-13).
Figure 6 Panel B presents the corresponding plot for the elderly and suggests a similar pattern,
though smaller in magnitude (~1 pp, about half the pre-ACA gap). This is unsurprising since the change in
insurance coverage is much smaller for elderly patients. Table 5 Panel B presents corresponding regression
coefficients for elderly patients and suggests that for-profit hospitals have gained share at the expense of
government hospitals. The bias-corrected IV estimates indicate that elderly patients are about 10% less
likely to use a government hospital when they gain (Medicaid) coverage — slightly less elastic than young
patients.
Our research design cannot help us disentangle mechanisms behind this shift in hospital care
toward private hospitals. The most intuitive explanation is that it is driven by patient preferences. However,
it is an equilibrium outcome and one cannot rule out the role of hospital responses (such as greater targeting
of recently insured individuals, or greater propensity to admit these patients at the ER) in producing this
22
shift. We replicate this analysis on the sample of ER arrivals to learn more (See appendix Table A. 2). ER
arrival patterns are more likely to reflect patient preferences since they are presumably emergent. We find
strikingly similar patterns in the ER data — 12% and 21% decrease in likelihood of receiving care at a
government hospital on gaining (Medicaid) coverage among young and elderly respectively. We believe
this indicates the sorting is motivated by patient preference for private over public hospitals.
B. Hospital quality (scores)
Hospital ownership type is correlated with quality or perceived quality of care (for example,
academic medical centers are generally high quality and mostly non-profit), but probably also correlates
with non-quality factors. To examine if the observed sorting across hospitals is motivated by quality we use
two commonly accepted metrics – risk-adjusted 30-day mortality and readmission rates – as indicators of
hospital quality. We test if patient volume has shifted toward hospitals that were certified by CMS as having
better quality outcomes in 2009, comfortably before enactment of the ACA.
CMS calculates these measures for Medicare patients discharged from hospitals for a number of
serious conditions. The raw mortality and readmission rates are adjusted for patient risk history and
observed sickness at the time of admission.16 We start with the risk-adjusted rates for hospitals, as reported
by CMS in 2009, on three conditions: heart attack, heart failure and pneumonia. We then compute the mean
rate for each hospital and normalize it such that the distribution across hospitals is standard normal.
Figure 7 presents a scatter plot of mean normalized mortality scores and corresponding fitted values
obtained by estimating equation 3b on the Y-axis, against patient age-month cell on the X-axis. This figure
uses data on hospital stays over 2012-15 and presents results for both the young (Panel A) and elderly
(Panel B). Figure 7 indicates that the patterns for young and elderly patients are qualitatively similar — in
both cases, the treated group has seen a shift toward better quality hospitals relative to the control group.
The emphasis on relative is important since in the case of the elderly the mean hospital mortality score has
held constant for 64 year old patients post-ACA, but has worsened for 65 year olds. The relative shift
appears larger in the case of the young (0.02 s.d., almost twice as large as the corresponding pre-ACA gap).
Table 5 columns 4 and 5 present the formal estimated effects on mean mortality and readmission
scores respectively. Panels A and B present the results for young and elderly patients respectively. In order
to examine sorting on quality distinctly from the sorting on ownership we control for owner type in these
regressions. Both the reduced form and IV point estimates are not statistically significant. The tests do not
lack statistical power, for example in case of young patients we can reject a change in readmission score
16 More details on the methodology are available at https://www.cms.gov/Medicare/Quality-Initiatives-Patient-Assessment-Instruments/HospitalQualityInits/OutcomeMeasures.html. The mortality measures are available at https://data.medicare.gov/data/hospital-compare.
23
greater than 0.03 s.d. Other estimates can be tested even more precisely. The IV point estimates are large,
for example implying that gaining (Medicaid) coverage leads to a ~0.1 s.d. decrease in mortality score of
serving hospital for young patients. In case of elderly patients the IV estimates imply that gaining
(Medicaid) coverage leads to a shift to hospitals with 0.16 s.d. and 0.22 s.d. better mortality and readmission
scores respectively. This is equivalent to moving to a hospital with 0.26 (0.16*1.63) pp lower mortality or
a 2% decrease relative to the mean 30-day mortality rate for AMI, heart failure and Pneumonia. In case of
elderly patients the coefficients consistently imply that patients have shifted toward better quality hospitals.
To quantify the value of this effect of insurance coverage, we use revealed preference estimates of
the additional distance patients are willing to travel to receive care at better hospitals. There is a large
literature on hospital choice which has developed approaches to estimate these objects and a full review is
out of scope, but the closest reference is Tay (2003) which examines Medicare data from California, Oregon
and Washington. It estimates that younger, white male heart attack patients are willing to travel up to 8
miles further to receive care at a hospital with a 3% lower mortality rate. In our setting, patients now receive
care at a hospital with 0.2-0.3pp (~2%) lower mortality rate on average. Therefore, insurance coverage is
equivalent to moving patients ~6 miles closer to their hospital, 40% of the mean distance in our sample.
D. Health Outcomes
Policymakers typically consider measures of patient health and provider quality interchangeably.
For example, CMS rates hospitals and Medicare sets payments for hospitals based on risk adjusted
mortality, readmissions, hospital acquired infections and a host of other measures of patient health. Previous
studies of provider quality and patient health have also focused largely on mortality. We follow the literature
and use mortality as a key measure of patient health, specifically in-hospital mortality - a closely related17
metric to standard measures such as 30-day mortality. Since mortality is relatively rare for the young18, we
limit our discussion on patient health to the near-elderly, although we present results for the young as well.
A key argument used in favor of expanding insurance coverage is that greater immediate access to
preventative care will circumvent later wasteful use of expensive ER/hospital care. Hence, a natural second
outcome of interest to measure patient health is whether the ACA led to a decrease in the wasteful use of
17 Due to data limitations, we do not observe mortality outside the hospital. We obtained death-linked hospital discharge files over 2008-11 from California OSHPD to examine the link between in-hospital mortality and outcomes later. OSHPD creates these files by linking hospital discharge records with the state death register. Hence, we can observe standard short-term mortality outcomes like 7-day and 30-day mortality through November 2011. We find that in-hospital deaths accounted for 79% and 64% of 7-day and 30-day mortality respectively for patients in these age groups. In-hospital death is also highly predictive of 30-day mortality across hospitals, with an R-squared of over 0.9. Therefore, we believe in-hospital mortality is a reasonable proxy for 30-day mortality. 18 The mean mortality rate for the young in our sample is 6 per 1,000 hospital stays and also tends to be noisy across different age-month cells (coefficient of variation is 0.33). Therefore, it is difficult to visually infer a change at any specific age.
24
hospital care. Potentially avoidable episodes are identified for a subset of visits19 based on ICD-9 diagnosis
codes recorded in a patient’s discharge data and have previously been used for this purpose (Kolstad and
Kowalski, 2012).
i. In-hospital mortality
Figure 8 presents a scatter plot of changes in mean in-hospital mortality rates on the vertical axis
against mean age-month of patients on the horizontal axis. It also presents fitted values obtained by
estimating equation 3b, where the regression co-variates include patient characteristics. Panels A and B
present the figures for young and elderly patients respectively. The figure provides suggestive evidence that
for both young and elderly patients in-hospital mortality fell by more for the treated patient group. However,
as with the shifts in hospital quality, the mortality plots are also very diffuse.
Table 6 presents corresponding regression estimates on health outcomes for young (Panel A) and
elderly (Panel B) patients respectively. Columns 1 and 3 present estimated effects on in-hospital mortality,
while columns 2 and 4 present corresponding effects on the share of potentially avoidable stays/ER visits.
The pattern of results is similar across hospital stays and ER arrivals and so for brevity we discuss only
hospital stays. The IV estimates imply large reductions in hospital mortality rates among the two treated
groups (21 and 64-year olds). In both cases however, the estimates are imprecisely estimated such that we
cannot rule out large effects in either direction.
ii. Potentially avoidable care
Table 6 columns 2 and 4 presents corresponding estimated effects on share of stays and visits that
were potentially avoidable. The coefficients imply smaller improvements relative to the gains seen in
hospital mortality, nonetheless these are economically meaningful. For example, in the case of the elderly,
the IV estimate indicates that gaining insurance leads to a 3 pp decrease in the share of PAH stays. But as
with the mortality effects, none of the estimated effects on PAH are statistically significant.
Overall, we interpret the suggestive evidence on improvements in both mortality and PAH with
caution. The estimated effects are not statistically significant, hence we shy away from translating the
implied lives saved or avoided stays into dollars in a more concrete cost-benefit computation.
V. HOSPITAL FINANCES
19 Potentially avoidable care hospitalization is defined only for hospital care where the primary diagnosis code pertains to a condition of the endocrine, nervous, circulatory, respiratory, digestive or ill-defined systems. These categories account for about 40% and 55% of the total sample of young and elderly patients in 2012-15 respectively.
25
Given the unprecedented magnitude of the insurance expansion in California, hospitals could be
affected directly (through an increase in average prices and quantity) as well as indirectly (through
expansion of supply in response). Finkelstein (2007) shows that hospital spending in response to the
introduction of Medicare was far in excess of the increase in utilization of care predicted on the basis of a
consumer demand response alone. She argues that hospitals expanded supply in response to the insurance
expansion by investing in capacity additions and new technologies. Motivated by these results, we examine
effects on hospital finances below.
In order to do so we implement a differences-in-differences research design, which uses cross-
sectional variation in pre-ACA un-insurance rates across hospitals. The thought experiment is that hospitals
with a high pre-ACA share of uninsured patients would be affected by a greater insurance shock in 2014-
15 relative to hospitals that mostly saw insured patients. Figure 2 illustrates the magnitude of this variation
across hospitals. Panel A presents a histogram of hospital un-insurance shares as of 2008, calculated using
hospital discharge data. Un-insurance ranged then from approximately zero to about 50%. Panel B presents
the distribution in 2014 after the expansion of Medicaid. The range has noticeably shrunk to about 30%,
with most hospitals below 15%.
There are at least two reasons why the financial effect of the ACA may differ for hospitals based
on whether the insurance shock was to self-pay or county coverage. First, there were differences in the
distribution of these coverages across hospitals prior to 2014. These coverage types are positively correlated
i.e. a hospital with higher share of self-pay tends to also have higher share of county indigents. However,
county coverage is much more concentrated amongst government-owned hospitals while self-pay is
distributed somewhat more uniformly across all types of hospitals. This suggests that the financial effects
due to replacement of county coverage will be more concentrated among government hospitals. Second,
one would expect greater change in revenue per patient when Medicaid replaces self-pay as opposed to
county coverage given that the latter was already providing some reimbursement to hospitals.20 Hence, the
effect on finances may be larger per unit decrease in self-pay.
In order to highlight these differences, our empirical approach examines effects of the shock to
county and self-pay on finances separately. In our main specification, both coverage types enter in an
additively separable fashion. We deploy data on hospital finances collected by OSHPD over the period
2009-16, as described in section III.C and correspondingly this analysis is performed at the hospital-year
level, rather than at the patient level.
𝑌𝑌ℎ𝑡𝑡 = 𝛼𝛼ℎ + 𝛾𝛾𝑡𝑡 + 𝜒𝜒 ⋅ Uninsuredℎ08 ⋅ 𝑇𝑇𝑡𝑡 + 𝜖𝜖ℎ𝑡𝑡 (5𝑎𝑎)
𝑌𝑌ℎ𝑡𝑡 = 𝛼𝛼ℎ + 𝛾𝛾𝑡𝑡 + 𝜒𝜒1 ⋅ Selfℎ08 ⋅ 𝑇𝑇𝑡𝑡 + 𝜒𝜒2 ⋅ Countyℎ08 ⋅ 𝑇𝑇𝑡𝑡 + 𝜖𝜖ℎ𝑡𝑡 (5𝑏𝑏) 20 Of course, self-pay patients may also have been paying some non-zero amount initially.
26
Equation 5a presents the estimating equation that implements this approach. The key assumption for
identification is absence of differential pre-trends in finances across hospitals and the additive separable
nature of the two coverage types. We present estimates obtained from equation 5b in order to examine the
presence of pre-trends. Note that this analysis quantifies net effects of the ACA on hospital finances,
including patient sorting across hospitals discussed in Section IV.C above.
Yℎ𝑡𝑡 = αh + γt + � 𝜒𝜒1𝑠𝑠 ⋅ Selfh08 ⋅ I(t = s)𝑠𝑠=2016
𝑠𝑠=2009
+ � 𝜒𝜒2𝑠𝑠 ⋅ Countyh08 ⋅ I(t = s)𝑠𝑠=2016
𝑠𝑠=2009
+ ϵℎ𝑡𝑡 (5𝑐𝑐)
We present evidence for four types of patient revenue outcomes (Yℎ𝑡𝑡) – the sum of revenue from
Medicaid and County indigent programs, revenue from private insurers, disproportionate share (DSH)
payments to compensate hospitals for care provided to uninsured and Medicaid patients, and total patient
revenue. The first allows us to directly examine the benefit of Medicaid expansion after netting out
decreases in payments from counties. The second examines if hospitals received more money from private
insurers, perhaps due to the ACA exchanges. This will complement evidence in section IV.B using the RD
design that indicated no net increase in private coverage. The third directly tests if California withdrew
DSH payments in response to the ACA (some phased decreases were mandated by the ACA), and whether
any such decrease was comparable to the increased revenue from Medicaid. Finally the total revenue
clarifies if these three forces collectively resulted in more revenue. All revenue figures are deflated to be in
2016 dollars (in thousands) and are normalized by the hospital’s licensed number of beds.
Appendix Figure A. 5 presents the coefficients for each year on self-pay (Panel A) and county
(Panel B) shares. Panel A indicates that hospitals with a relatively high share of self-pay patients exhibit a
differential pre-trend of decreasing revenue. This makes it difficult to interpret the figure and is also at odds
with the identification assumption discussed above. Hence we prefer to discuss results from an enhanced
specification which includes hospital-specific linear time trends, i.e. this allows each hospital’s revenue to
develop on a different pre-ACA time trend and we focus on deviations from the trend post-ACA. Figure 9
presents coefficients estimated using this modified specification. Reassuringly, there is no evidence of
differential changes in trends in any revenue type prior to the ACA.
Table 7 Panel B presents coefficients obtained by estimating equations 5a (linear) and 5b (flexible)
on hospital-year level data. Columns 1-6 present results on hospital revenue (expressed in thousands of
2016$) contributed by different payers. The resulting patterns are qualitatively similar, although estimates
using the linear specification are dramatically noisier.
27
Based on the linear specification, hospitals that had a 10% greater share of uninsured patients in
2008 experienced a relative increase of $18,000 (177,000*0.1) per bed in total patient revenue over 2014-
16. This increase is driven entirely by the Medicaid expansion, since revenue from all other payers declined.
This represents a 2% increase relative to the mean revenue of 1 million per bed. These estimates suffer from
lack of precision. For example, to reject the null of no change in Medicaid revenue, the point estimate would
need to be larger than the revenue received from Medicaid by the mean hospital. Use of a linear
specification is not supported by trends in hospital revenue (not presented here). Hospitals with low levels
of pre-ACA un-insurance exhibit little or no change in trends while hospitals with the highest levels of pre-
ACA un-insurance experience a dramatic increase in revenue which further motivates the use of a more
flexible specification.
The non-parametric specification offers much more precise estimates as well as larger magnitudes.
Total revenue for hospitals having un-insurance rates among the top third of all hospitals in 2008 has
increased by (marginally significant) $95,000 per bed -- 10% of revenue for the mean hospital -- relative to
the remaining hospitals. To compare with the estimate from the linear specification, note that hospitals in
the top third have an un-insurance rate about 13 pp greater than the remaining hospitals. Medicaid accounts
for more than 70% of this increase, with the remainder due to an increase in Medicare and Private insurance.
These results strongly suggest that the expansion of insurance coverage led to a disproportionate
increase in revenue for hospitals that previously had high rates of un-insurance among patients. In other
results not presented here, we find no evidence of a statistically significant decrease in subsidies to publicly
owned hospitals (that typically had high un-insurance rates pre-ACA) post-ACA. This is in contrast to
previous studies (Duggan, 2000; Baicker and Staiger, 2005) that show government hospitals being operated
on a soft budget by local governments. Therefore, we believe this truly was an increase in revenue for
hospitals, including government hospitals. In addition, we do not find evidence of a significant change in
the total number of non-elderly patients served at hospitals with the greatest pre-ACA un-insurance share.
Hence, the increase in revenue seems to be driven mostly by an increase in revenue per patient.
To put the magnitude of this increase in perspective, note that hospitals operate on low margins. In
2008, hospitals in California reported a mean net income of $4 million out of a total patient revenue of $140
million, or a margin of 3%. This margin would be lower still if we were to include non-patient revenue in
the topline. Thus, for hospitals with the highest pre-ACA un-insurance levels, the average influx of money
has been substantially above their annual baseline net income. These results are consistent with the strong
position taken by hospital industry associations to prevent repeal of the ACA Medicaid expansion.21
21 See for example a letter by the President of the American Hospital Association (AHA) to US Congress opposing the American Health Care Act that repealed the ACA (available at http://www.aha.org/presscenter/pressrel/2017/030817-pr-acha.shtml). More details of its lobbying against ACA repeal discussed at http://www.modernhealthcare.com/article/20170317/NEWS/170319906.
28
Hospitals could potentially deploy the additional revenue to improve general quality of care, initiate
expansions, hire more or better staff or retain as surplus. We explore this line of reasoning by examining
effects on some unrelated outcomes such as mortality rates for infra-marginal patient groups (infants and
seniors) and capital investments. These two patient groups are more sensitive to changes in quality of care
than non-elderly adults or older children. Panel B presents results on these outcomes (columns 7-8). The
estimates are small and statistically insignificant. In the case of elderly mortality, we can rule out an effect
larger than 6% of the mean value. This suggests that the additional revenue has not generated substantial
spillover quality improvements for non-targeted patients. Similarly, we do not find any effects on spending
on capital expenditure, acknowledging that two years may be too short a time horizon to detect investment
responses. However, mortality is of course a blunt measure of quality and thus it is plausible that other
dimensions of quality did improve.
Finkelstein, Hendren and Luttmer (2015) suggest that hospitals and physicians receive substantial
value from Medicaid expansions since it mainly replaces uncompensated care that was already being
provided. Their conclusion was based on patient survey data and they lacked direct evidence from hospital
financial reports. Our results on hospital finances are consistent with their findings.
VI. ROBUSTNESS AND FALSIFICATION CHECKS (to be completed)
A. Placebo insurance expansion
An important identification concern in attributing changes post 2014 to the ACA is that the effects
may not be driven by the insurance expansions of 2014, but by other economic trends that preceded
implementation of the ACA. This is particularly relevant in the case of the estimated decrease in private
coverage, which is a larger trend observed in health care data since the great recession. In order to examine
whether pre-existing economic trends could be driving some of the changes in insurance coverage (our first
stage), we replicate our regression discontinuity analysis over the period 2008-11 i.e. before the ACA
insurance expansions were implemented.
Ideally, if our pre-ACA coefficient (2012-13) estimates a stable discontinuity in coverage under
the previous equilibrium, then we should find similar estimates in the 2008-09 period as well, i.e. 𝜃𝜃1108−09 =
𝜃𝜃1112−13. If the post-ACA coefficient captures changes only due to the ACA, then we would find a zero
effect in a placebo test, i.e. 𝜃𝜃1208−09 = 0.
Table 9 presents corresponding results on insurance coverage, utilization, hospital choice and
patient health for the sample of hospital stays corresponding to the period 2008-11. It has the same format
as Table 8 discussed above. The first row in both panels estimates the discontinuity in outcome over 2008-
09 at ages 21 and 65 respectively. All estimated effects pre-2010 for both the young and elderly are
29
qualitatively similar to the corresponding main results presented in previous tables for the period
immediately prior to the ACA. The estimated effects on hospital shares and patient health are not
statistically significant in the case of the young, but are of similar magnitude and sign as the main results.
The second row in both panels presents the estimated change in discontinuity in 2010-11 relative
to 2008-09. Several coefficients are not significantly different from zero or are very close to zero. Overall
the pattern of results does not mimic the post-ACA results. For example, we find an increase in self-
insurance both among the young and the elderly and no decrease in the share of government hospitals.
Taken together, these estimates further reinforce the interpretation that losing insurance coverage
causes sorting of patients towards government hospitals, and vice versa. Specifically, we do not find a
statistically significant decrease in private coverage for the near-elderly post 2010, allaying a key concern
with the crowd-out interpretation. Overall this exercise does not present evidence to challenge conclusions
drawn from our main results for the years following ACA implementation.
B. Alternative (age) bandwidth
We replicate all RD analyses using a sample defined by a wider (2-year) age bandwidth above and
below the age bandwidths of 21 and 65, i.e. 19-22.9 and 63-66.9 respectively….
VII. DISCUSSION
We note important caveats related to our data and research design that impose the following
limitations when interpreting the results:
A. Local average treatment effects
The results from the RD analysis recover parameters that pertain to individuals close to age 21 or
65. They are likely relevant to individuals in their late twenties and fifties as well, since patterns of health
and health care use are relatively stable over short ranges of age. Both groups of beneficiaries are very
policy relevant. The former group has seen the greatest increase in insurance coverage under the Medicaid
expansion, as evident from appendix Figure A. 6. The figure presents the share of 21-64-year-olds covered
by Medicaid and uninsured among different age groups. Panel A presents data from the ACS, while Panel
B presents corresponding data from hospital stays. Young adults, in their twenties, experienced the greatest
increase in Medicaid and decrease in un-insurance both in the population, as well as among individuals
30
discharged from hospitals. The latter group could benefit further from an increase in public insurance if
Medicare is expanded to include the near-elderly.22
The IV estimates are applicable only to compliers, i.e. individuals that aged out of Medicaid pre-
ACA and gained coverage under the ACA – hence, they are applicable to a subset even within this age
group. Compliers share some common attributes such as being low-income, not having children and not
having a disability or serious sickness. All of these factors precluded them from Medicaid coverage pre-
ACA. These features are not particularly restrictive, especially for the young.
Finally, the IV estimates pertain to obtaining public insurance. We may expect very different
estimates if the same individuals were awarded private coverage, for example. Anderson, Gross and Dobkin
(2012) find the likelihood of ER use decreases by 40% when individuals lose private insurance coverage at
age 19. This is much greater than our estimated effect on utilization for the young and could be driven by
the differences between private insurance and Medicaid, differences in the sample population, and the age
threshold being studied.
B. Adverse selection
Our research design and data do not allow us to quantify whether the newly insured Medicaid
beneficiaries are adversely or advantageously selected. We cannot distinguish between adverse selection
and moral hazard, both of which will result in an increase in utilization of care upon gaining insurance. In
the case of the young, the nature of Medicaid eligibility restrictions pre-ACA and the large increase in
coverage make it likely that the newly insured beneficiaries are healthier than existing Medicaid recipients.
The lack of increase in stays and ER arrivals among the young strengthens this argument.
In the case of the near-elderly, two factors suggest that the high elasticity could be driven by
infusion of high risk beneficiaries into Medicaid. First, a very small proportion of the near-elderly lacked
insurance pre-ACA. These could have been individuals priced out of individual insurance plans but not
eligible for Medicaid. Second, marginal near-elderly patients gaining insurance experience a large increase
not only in ER arrivals, but also in stays admitted through the ER, suggesting that this was a particularly
sick patient group.
C. Patient sorting across hospitals
The evidence on sorting of patients toward privately owned hospitals and hospitals with better
mortality scores is robust with economically significant estimates. This pattern is particularly compelling
22 Since the 1990s several unsuccessful legislative proposals have been floated to expand Medicare to cover near-elderly individuals aged 55-64. The latest one (still on-going) was introduced in August 2017 in the US Senate. See https://www.stabenow.senate.gov/news/senator-stabenow-announces-medicare-at-55-act for more details.
31
because it is also found in the ER arrivals data. We cannot quantify the role of different mechanisms causing
this sorting. The evidence indicates it is primarily driven by patient preferences for better (real or perceived)
hospital care. Another factor that could contribute to this sorting pattern is narrow provider networks
imposed on newly insured Medicaid and exchange patients that force them to obtain care at specific
hospitals and exclude certain hospitals. Haeder, Weimer and Mukamel (2015) examine the breadth, access
and quality of insurer networks offered on California’s ACA exchanges relative to commercial health plans.
They find that exchange plan networks are narrower and cover care at about 80% of the hospitals relative
to commercial plans. However, this reduction in network breadth does not correlate with hospital ownership
or quality and does not result in lower geographic access (in terms of distance). Thus, it seems unlikely that
this factor can drive the sorting pattern.
D. Short run effects
Our sample covers only two years of data after the insurance expansion (2014-15) on hospital care
and three years in the case of hospital finances. We find robust evidence of change in some aspects of
utilization of care and hospital choice, but not on patient health or spillover effects on quality. These results
should be seen as estimates of short-run effects of insurance expansion.
Depending on the importance of dynamics in generating changes in hospital care (e.g. newly
insured beneficiaries learn how to use preventative care or choose hospitals, or hospitals learn how to better
serve such patients), the long-run effects could be very different. Robust evidence of larger effects on
utilization of care and sorting across hospitals in 2015 relative to 2014 suggest that long-run effects may be
larger in magnitude.
VIII. CONCLUSION
The Affordable Care Act (ACA) authorized the largest expansion of publicly funded insurance
since the introduction of Medicare and Medicaid in the 1960s. This intervention in the health care system
offers an unprecedented opportunity to quantify the effects of public insurance coverage in a general
equilibrium setting on beneficiaries and hospitals alike. We exploit the presence of sharp discontinuities in
public insurance coverage at ages 21 and 65 in a regression discontinuity based research design to examine
several possible effects of insurance coverage in the context of the ACA. We conduct our analysis for the
state of California using the universe of all hospital stays and ER visits over 2008-15 and some
supplementary data files on hospital financials. To our knowledge, this is the first examination of the effects
of the ACA using administrative data on hospital care and financial reports.
We have three principal findings. First, we find that Medicaid expansion crowded out existing
county indigent safety-net programs in California for young and elderly patients. The crowd-out is large –
32
our estimates imply that only about half of incremental Medicaid spending on hospital care provided relief
to uninsured individuals. The remainder was transfer from federal taxpayers to entities (mostly counties)
that were previously bearing these costs. Second, we find heterogeneous effects on utilization of care across
young and elderly patients. Our IV estimates imply that gaining insurance coverage causes a 10-15%
increase in likelihood of hospital care for the young, while the increase is four times as large for the elderly.
The estimated effect for the elderly is twice as large as corresponding partial equilibrium estimates reported
from the Oregon experiment and illustrates the quantitative importance of general equilibrium effects.
Insurance coverage also enables patients to shift care toward private hospitals and better quality hospitals.
The sorting effects are economically significant and suggest an important source of welfare gain for
beneficiaries that has received less attention in previous studies on the value of Medicaid. Third, we find
robust evidence that the ACA insurance expansion has resulted in modest differential increase in revenue
for hospitals that previously had a high uninsured share of patients. The revenue gains are concentrated
among government owned ‘safety-net’ hospitals. We do not find evidence that increased revenue led to
spillover quality effects or investment responses. In addition to the main results, we provide supporting
evidence using a variety of robustness and falsification checks.
Our results have important limitations. The estimates are specific to California and local to
individuals in the specific age groups that we study. Since our main data source is hospital discharge data,
we do not observe health care effects outside of hospitals and any non-health care effects. For example, we
cannot comment on changes in access to care, quality of primary care or alleviation in consumer finances
post-ACA. We do not observe hospital responses on important dimensions such as technology adoption
and staffing. Exploring these aspects and linking them together in a comprehensive evaluation of this
landmark reform of the US health care system represent important directions for future research.
33
REFERENCES Anderson, Michael, Carlos Dobkin, and Tal Gross (2012). "The Effect of Health Insurance Coverage on the Use of
Medical Services." American Economic Journal: Economic Policy 4, no. 1: 1-27. Anderson, Michael L., Carlos Dobkin, & Tal Gross (2014). “The Effect of Health Insurance on Emergency
Department Visits: Evidence from an Age-Based Eligibility Threshold.” Review of Economics and Statistics, 96(1), 189-195.
Baicker, Katherine, and Douglas Staiger (2005). "Fiscal shenanigans, Targeted Federal Health Care Funds, and
Patient Mortality." The Quarterly Journal of Economics 120, no. 1: 345-386. Benitez, Joseph A., and Liza Creel (2016). "Kentucky’s Medicaid Expansion Showing Early Promise on Coverage
and Access to Care." Health Affairs 35.3: 528-534. Brevoort, Kenneth, Daniel Grodzicki and Martin Hackmann (2017). “Medicaid and Financial Health.” Working
Paper. Bureau of Labor Statistics (2009). “America's Youth at 21: School Enrollment, Training, and Employment
Transitions Between Ages 20 and 21.” Technical Report. Bureau of Labor Statistics (2010). “America’s Youth at 22: School Enrollment, Training, and Employment
Transitions Between Ages 21 and 22.” Technical Report. Cabral, Marika and Mark R Cullen (2016). “Estimating the Value of Public Insurance Using Complementary Private
Insurance,” NBER working paper 22583. California Health Care Foundation (2009). “County Programs for the Medically Indigent in California, 2009.”
Technical Report. Card, David, Carlos Dobkin, and Nicole Maestas (2008). "The Impact of Nearly Universal Insurance Coverage on
Health Care Utilization and Health: Evidence from Medicare." American Economic Review 98, no. 5: 2242-2258.
Card, David, Carlos Dobkin, & Nicole Maestas (2009). “Does Medicare Save Lives?” Quarterly Journal of
Economics, 124(2), 597-636. Congressional Budget Office (2017). “Federal Subsidies Under the ACA for Health Insurance Related to the
Expansion of Medicaid and Nongroup Health Insurance: Tables from CBO’s January 2017 Baseline.” Technical Report.
Council of Economic Advisors (2009). “The Impact of Health Insurance on State and Local Governments.”
Technical Report. Courtemanche, Charles, James Marton, Benjamin Ukert, Aaron Yelowitz, and Daniela Zapata. (2017a). “Impacts of
the Affordable Care Act on Health Insurance Coverage in Medicaid Expansion and Non-Expansion States.” Journal of Policy Analysis and Management, 36(1): 178-210.
Courtemanche, Charles, James Marton, Benjamin Ukert, Aaron Yelowitz, and Daniela Zapata. (2017b). “Early
Effects of the Affordable Care Act on Health Care Access, Risky Health Behaviors, and Self-Assessed Health.” NBER working paper 23269.
Currie, Janet and Jonathan Gruber (1996). “Health Insurance Eligibility, Utilization of Medical Care, and Child Health,” Quarterly Journal of Economics, 111(2): 431-466.
Currie, Janet and Jonathan Gruber (1996b). “Saving Babies: the Efficacy and Cost of Recent Changes in the
34
Medicaid Eligibility of Pregnant Women.” Journal of Political Economy, 104(6), 1263-1296. Dafny, Leemore (2005). “How do Hospitals Respond to Price Changes?” The American Economic Review, 95 (5),
1525–1547. Dafny, Leemore and Jonathan Gruber, (2005). “Public insurance and Child Hospitalizations: Access and Efficiency
Effects,” Journal of Public Economics, 89 (1), 109–129. Duggan, Mark G. (2000). “Hospital Ownership and Public Medical Spending,” The Quarterly Journal of
Economics, 115 (4), 1343–1373. Duggan, Mark, Gopi Shah Goda and Emilie Jackson. (2017) "The Effects of the Affordable Care Act on Health
Insurance Coverage and Labor Market Outcomes." NBER working paper 23607. Finkelstein, Amy (2007). "The Aggregate Effects of Health Insurance: Evidence from the Introduction of
Medicare." The Quarterly Journal of Economics 122, no. 1: 1-37. Finkelstein, Amy et al. (2012). “The Oregon Health Insurance Experiment: Evidence from the First Year,” The
Quarterly Journal of Economics, 127 (3), 1057–1106. Finkelstein, Amy, Nathaniel Hendren, and Erzo FP Luttmer (2015). “The Value of Medicaid: Interpreting Results
from the Oregon Health Insurance Experiment.” No. w21308. National Bureau of Economic Research. Frean, Molly, Jonathan Gruber, and Benjamin D. Sommers (2017). “Premium Subsidies, the Mandate, and Medicaid
Expansion: Coverage Effects of the Affordable Care Act.” Forthcoming, Journal of Health Economics. Gallagher, Emily, Radhakrishnan Gopalan, and Michal Grinstein-Weiss (2017). "The Effect of Health Insurance on
Home Payment Delinquency: Evidence from ACA Marketplace Subsidies." Garthwaite, Craig, Tal Gross, & Matthew Notowidigdo (2016). “Hospitals as Insurers of Last Resort.” American
Economic Journal: Applied Economics. Garthwaite, Craig, Tal Gross, Matthew Notowidigdo, & John Graves (2017). “Insurance Expansion and Hospital
Emergency Department Access: Evidence from the Affordable Care Act Insurance Expansion and Hospital Emergency Department Access.” Annals of internal medicine, 166(3), 172-179.
Ghosh, A., Simon, K. and Sommers, B.D., (2017). “The Effect of State Medicaid Expansions on Prescription Drug
Use: Evidence from the Affordable Care Act” NBER working paper 23044. Golberstein, Ezra, Gilbert Gonzales, and Benjamin D. Sommers (2015). "California’s Early ACA Expansion
Increased Coverage and Reduced Out-of-Pocket Spending for the State’s Low-Income Population." Health Affairs 34, no. 10: 1688-1694.
Goodman-Bacon, Andrew (2016). “Public insurance and mortality: Evidence from Medicaid Implementation.”
Journal of Political Economy, forthcoming. Hadley, Jack, John Holahan, Teresa Coughlin, and Dawn Miller (2008). “Covering the Uninsured in 2008: Current
Costs, Sources of Payment, and Incremental Costs.” Health Affairs, 27(5), w399-w415. Haeder, Simon F., David L. Weimer, and Dana B. Mukamel (2015). "California Hospital Networks are Narrower in
Marketplace than in Commercial Plans, but Access and Quality are Similar." Health Affairs 34, no. 5: 741-748.
Hahn, Jinyong, Petra Todd, and Wilbert Van der Klaauw (2001). “Identification and Estimation of Treatment
Effects with a Regression-Discontinuity Design.” Econometrica, 69(1), 201-209.
35
Imbens, Guido W., and Joshua D. Angrist (1994). "Identification and Estimation of Local Average Treatment Effects." Econometrica 62, no. 2: 467-475.
Klerman, Jacob A., Jeanne S. Ringel and Beth Roth (2005). Under-reporting of Medicaid and Welfare in the Current
Population Survey. RAND working paper WR-169-3. Kolstad, Jonathan T., and Amanda E. Kowalski (2012). "The Impact of Health Care Reform on Hospital and
Preventive Care: Evidence from Massachusetts." Journal of Public Economics 96, no. 11: 909-929. Lee, David S., and Thomas Lemieux (2010). “Regression Discontinuity Designs in Economics.” Journal of
Economic Literature, 48, 281-355. Manning, W. G., J. P. Newhouse, N. Duan, E. B. Keeler, and A. Leibowitz (1987). “Health Insurance and the
Demand for Medical Care: Evidence from a Randomized Experiment.” The American Economic Review, 251-277.
Meyer, Bruce D., Wallace K. Mok, and James X. Sullivan (2009). “The Under-Reporting of Transfers in Household
Surveys: its Nature and Consequences.” National Bureau of Economic Research No. w15181. Miller, Sarah (2012). “The Effect of Insurance on Emergency Room Visits: An Analysis of the 2006 Massachusetts
Health Reform,” Journal of Public Economics, 96 (11), pp. 893-908. Norton, Edward C., Jun Li, Anup Das, and Lena M. Chen (2017). “Moneyball in Medicare.” Journal of Health
Economics. Sommers, B.D., G.M. Kenney, and A.M. Epstein (2014). “New Evidence On The Affordable Care Act: Coverage
Impacts of Early Medicaid Expansions.” Health Affairs. 33(1):78-87. Sommers, B.D., K.P. Chua, G.M. Kenney, S.K. Long, and S. McMorrow (2015). “California’s Early Coverage
Expansion under the Affordable Care Act: A County-Level Analysis.” Health Services Research. 51(3):825-845.
Sommers, B. D., R. J. Blendon, E. J. Orav, and A. M. Epstein (2016). “Changes in Utilization and Health Among
Low-Income Adults After Medicaid Expansion or Expanded Private Insurance.” JAMA internal medicine, 176(10), 1501-1509.
Tay, Abigail (2003). “Assessing Competition in Hospital Care Markets: the Importance of Accounting for Quality Differentiation.” RAND Journal of Economics, 786-814.
Wherry, Laura, & Sarah Miller (2016). “Early Coverage, Access, Utilization, and Health Effects Associated with the
Affordable Care Act Medicaid Expansions: A Quasi-experimental Study.” Annals of Internal Medicine, 164(12), 795-803.
36
FIGURES AND TABLES
1a: Medicaid share of hospital stays
1b: Insurance coverage in California, ACS
1c: Medicaid and Exchange enrollment in California
Figure 1: Insurance coverage in California
Note: This figure presents primary insurance coverage in California over 2012-15 as reported in Medicaid share of hospital stays for ages 10-75 as recorded in hospital discharge data (Panel A), the American Community Survey (Panel B), and monthly enrollment in Medicaid and ACA exchange (Panel C). The sample used in Panel A excludes cases related to pregnancy and deliveries, is limited to General Acute Care hospitals and excludes individuals residing in zip codes outside California. In Panel B, if an individual reports Medicaid and Medicare, then we code Medicare as primary. Similarly, if an individual reports Private and Medicaid, then we code Medicaid as primary. This ensures coverage aggregates to 100%. Respondents are divided as approximately 21%, 59% and 12% in the three age groups respectively. Enrollment data obtained from CA Department of Health Care Services (Medicaid) and Covered California (Exchange) respectively.
37
2a: Hospital un-insurance distribution (2008)
2b: Hospital un-insurance distribution (2014)
Figure 2: Hospital un-insurance distribution
Note: This figure presents histograms of hospital share of patients that did not have insurance coverage, in 2008 (Panel A, pre-ACA) and 2014 (Panel B, post-ACA) respectively. These histograms were computed using the discharge data on hospital stays. It is based on the sample of non-elderly adults (aged 21-64) and excludes cases related to pregnancy or child birth. Only general acute care hospitals are included. Un-insurance share is top coded at 50% (one hospital in 2008).
38
3a: Young patients
3b: Elderly patients
Figure 3: Change in insurance coverage
Note: This figure presents change in insurance coverage among hospital stays and corresponding fitted values by age-month cell. These were obtained by estimating equation 3a on case level data as described in Section IV.A for the sample of young (Panel A) and elderly (Panel B) patients respectively. The treated groups are those aged 21 (young) and 64 (elderly). Both panels present the data for 2012-15 (circles, solid line) which includes the insurance expansions post-2014, and data from 2008-11 (squares, dashed line), which serves as a falsification exercise. The dependent variable is set to 1 if the individual is not self-insured, on charity or county indigent care. All models control linearly for age and include a full set of year and age-month cell fixed effects. The figures also present estimated change in discontinuity, which is the coefficient on 𝑑𝑑𝑖𝑖 .𝑇𝑇𝑡𝑡 in equation 3a. These coefficients are not bias-corrected. Standard errors are clustered by age-month cell.
39
4a: Young patients
4b: Elderly patients
Figure 4: Change in hospital stays Note: This figure presents the observed change in number of hospital stays post-2014, per 1,000 CA residents in each age-month cell. Panels A and B present data for the sample of young and elderly patients respectively. The treated groups are those aged 21 (young) and 64 (elderly). It also plots corresponding fitted values (dashed lines) obtained by estimating equation 4b as described in Section IV.C. California population estimates obtained for 2012-15 from ACS 1-year survey data. The discontinuity change estimates mentioned in the figure were obtained from regressions with log of stays as the dependent variables and hence can be interpreted as approximate percentage changes. All models control linearly for age and include a full set of year fixed effects. Robust standard errors reported.
40
5a: Young patients
5b: Elderly patients
Figure 5: Change in ER utilization Note: This figure presents the observed change in number of annual ER arrivals post-2014, per 1,000 CA residents in each age-month cell. Panels A and B present data for the sample of young and elderly patients respectively. The treated groups are those aged 21 (young) and 64 (elderly). It also plots corresponding fitted values (dashed lines) obtained by estimating equation 4b as described in Section IV.C. California population estimates obtained for 2012-15 from ACS 1-year survey data. The discontinuity change estimates mentioned in the figure were obtained from regressions with log of ER arrivals as the dependent variables and hence can be interpreted as approximate percentage changes. All models control linearly for age and include a full set of year fixed effects. Robust standard errors reported.
41
6a: Young patients
6b: Elderly patients
Figure 6: Change in government hospital share
Note: This figure presents observed post-ACA change in share of government hospital stays collapsed to age-month cells. It also plots fitted values obtained by estimating equation 3b on case level data as described in Section IV.A for the sample of young (Panel A) and elderly (Panel B) patients respectively. The treated groups are those aged 21 (young) and 64 (elderly). The regressions control linearly for age, and include age-month cell and year fixed effects. The figures also present estimated change in discontinuity, which is the coefficient on 𝑑𝑑𝑖𝑖 .𝑇𝑇𝑡𝑡 in equation 3b. Standard errors are clustered by age-month cell.
42
7a: Young patients
7b: Elderly patients
Figure 7: Change in quality of serving hospital
Note: This figure presents change in mean standardized mortality score (CMS Hospital Compare, 2009) for hospital stays collapsed to age-month cells. It also plots corresponding fitted values obtained by estimating equation 3b on case level data as described in Section IV.A for the sample of young (Panel A) and elderly patients (Panel B) respectively. The treated groups are those aged 21 (young) and 64 (elderly). The regressions linearly control for age, and include age-month cells and year fixed effects. The figures also present estimated change in discontinuity, which is the coefficient on 𝑑𝑑𝑖𝑖 .𝑇𝑇𝑡𝑡 in equation 3b. Standard errors are clustered by age-month cell.
43
8a: Young patients
8b: Elderly patients
Figure 8: Change in health outcomes
Note: This figure presents observed post-ACA change in in-hospital mortality by age-month cell. It also plots fitted values obtained by estimating equation 3b on case level data as described in Section IV.A for the sample of young (Panel A) and elderly (Panel B) patients respectively. The treated groups are those aged 21 (young) and 64 (elderly). The regressions control linearly for age, and include age-month cell and year fixed effects. The figures also present estimated change in discontinuity, which is the coefficient on 𝑑𝑑𝑖𝑖 .𝑇𝑇𝑡𝑡 in equation 3b. Standard errors are clustered by age-month cell.
44
9a: Shock to self-pay coverage
9b: Shock to county coverage
Figure 9: Effects on hospital finances
Note: This figure presents coefficients on the interaction of Selfℎ08 (Panel A) and Countyℎ08 (Panel B) with an indicator for each year 𝑠𝑠 from 2009-16, obtained by estimating equation 5b with various hospital revenue variables as outcomes. Selfℎ08 and Countyℎ08 are the shares of hospital h patients coded self-pay and county indigent respectively in 2008. All revenue values have been deflated to be in 2016 dollars. This model is estimated using hospital-year level finances data made available by OSHPD over 2009-16 and includes hospital and year fixed effects as well as hospital specific linear time trends. Hence we need to normalize relative to two pre-ACA years – 2009 and 2013. Figure A. 5 presents corresponding figure without including time trends.
45
Table 1: Summary Statistics
Note: This table presents descriptive statistics from the samples used in the main analyses of the paper. Panels A and B present statistics for the samples in the regression discontinuity analysis and all non-elderly (21-64) patients in the spatial variation analysis respectively. Both samples begin with the universe of all discharges and use three sample restrictions – 1) only general acute care hospitals 2) exclude pregnancy and delivery related cases and 3) exclude patients with missing or out-of-CA zip codes. Fraction uninsured includes patients coded as self-pay, charity or county indigent coverage. Panel A focuses on cases pertaining to ages 20-21 (both inclusive) or 64-65. ER arrivals include ER visits and hospital stays that originated in the ER. To calculate utilization, we normalize number of annual stays/ER arrivals by the population in relevant age group obtained from ACS 1-year estimates, hence these are measures of utilization per 1,000 person years. We obtain population estimates for each year in 2012-15 for the young (18-24) and elderly (62-69) and assume that it is uniformly distributed. Panel C presents mean values of key variables in the hospital finances data, primarily hospital revenue per licensed bed pre (2009-13) and post (2014-16) implementation of the Medicaid expansion. Revenue values are expressed in thousands of 2016 dollars. Finances data is only gathered for so-called “comparable” hospitals that excludes state-owned, Kaiser and Shriner group of hospitals. Hence the number of hospitals covered in financial analysis is about 40 lower than in the discharge data.
2012-13 2014-15 2012-13 2014-15 2012-13 2014-15 2012-13 2014-15All observations 51,744 48,626 189,028 185,193 624,153 684,535 426,894 472,062Admitted through ER 35,637 33,843 116,222 117,793 35,637 33,843 116,222 117,793Medicaid 0.34 0.50 0.12 0.17 0.28 0.44 0.12 0.19 Uninsured 0.18 0.05 0.04 0.02 0.30 0.16 0.09 0.05 Self-pay 0.12 0.05 0.03 0.01 0.27 0.16 0.07 0.04 Utilization per 1,000 popn. 22 22 135 126 260 301 305 322 Government hospital 0.18 0.18 0.11 0.12 0.17 0.16 0.15 0.15 In-hospital mortality 0.0060 0.0060 0.0265 0.0272 0.0008 0.0006 0.0117 0.0107 Potentially avoidable 0.22 0.20 0.21 0.19 0.18 0.15 0.21 0.18
2009-13 2014-15 2009-13 2014-15Observations 6,403,985 2,494,122 30,136,352 14,189,956Non-deferrable 1,137,862 503,128 3,540,435 1,703,846Medicaid 0.24 0.40 0.24 0.41Uninsured 0.14 0.04 0.26 0.12Self-pay 0.08 0.03 0.21 0.11Government hospital 0.16 0.15 0.19 0.17In-hospital mortality 0.0159 0.0164 0.0011 0.0009Potentially avoidable 0.18 0.20 0.17 0.15
2009-13 2014-16Total revenue 1,054 1,245 Medicaid 225 345 Traditional 160 173 Managed Care 64 170 Medicare 319 357 Private 464 517 County Indigent 15 5 Others 22 14 Number of hospitals 339 335
Panel C: Hospital revenue per bed (2016$ '000s)
Panel B: Non-elderly sample (21-64)
Panel A: Regression discontinuity sample
Hospital Stays
All ER arrivalsHospital staysAges 20.0 - 21.9 Ages 64.0 - 65.9 Ages 20.0 - 21.9 Ages 64.0 - 65.9
ER arrivals
46
Table 2: Population attributes at age thresholds (National Health Interview Survey)
Note: This table presents population weighted descriptive statistics and regression discontinuity estimates at ages 21 and 65 using data from the National Health Interview Survey (NHIS) person and sample adult files from 2004-2009. Data is limited to individuals within 12 months of their 21st and 65th birth month, excluding individuals interviewed in their month of birth. There are 11,321 and 6,883 such individuals in the person files. The outcomes percent days alcohol in past 12 months, smoking status and flu shot in past 12 months are taken from the sample adult files which have 4,375 and 3,587 individuals respectively. Standard errors (in brackets) are adjusted to account for sampling stratification as recommended by NHIS documentation. Mean value at threshold pertains to the mean value for individuals aged 20 and 65 respectively. RD estimate indicates difference in mean for individuals aged 21 and 64 (the treatment group) respectively. RD estimate obtained using OLS including linear polynomial in age and year fixed effects.
(1) (2) (3) (4) (5)Insured Uninsured Difference Mean value RD estimatemean mean at threshold at threshold
Panel A: Ages 20-21Married 0.08 0.13 0.044 0.07 -0.003
(0.008) (0.013)Employed 0.61 0.66 0.047 0.60 0.004
(0.012) (0.023)In school 0.23 0.07 -0.160 0.21 -0.028
(0.010) (0.019)Percent days alcohol 0.12 0.11 -0.010 0.09 0.033
(0.008) (0.015)Smoker 0.21 0.36 0.148 0.23 0.059
(0.021) (0.041)Flu shot past 12 months 0.14 0.09 -0.056 0.13 0.015
(0.014) (0.026)No insurance coverage - - - 0.29 0.056
(0.022)Panel B: Ages 64-65Married 0.69 0.50 -0.1908 0.67 0.010
(0.025) (0.027)Employed 0.37 0.35 -0.0205 0.34 -0.007
(0.026) (0.029)In school 0.00 0.00 -0.0005 0.00 0.002
(0.000) (0.002)Percent days alcohol 0.16 0.09 -0.0662 0.15 -0.015
(0.020) (0.025)Smoker 0.17 0.30 0.1343 0.17 0.012
(0.036) (0.031)Flu shot past 12 months 0.51 0.25 -0.2672 0.51 -0.066
(0.032) (0.042)No insurance coverage - - - 0.03 0.062
(0.016)
47
Table 3: Insurance coverage (hospital stays)
Note: This table presents regression coefficients obtained using case level data as described in Section IV.A for the sample of young (Panel A) and elderly (Panel B) patients respectively. The dependent variable is coverage by different insurers or self-pay/county indigent coverage. Each cell presents a coefficient from a different regression. Miscellaneous includes Medicare, Government employees and workers’ compensation. This table pertains to hospital stays only. Appendix Table A. 1 pertains to ER visits including those resulting in hospital stays. The naïve estimate of change in discontinuity at threshold is the coefficient on 𝑑𝑑𝑖𝑖 ⋅ 𝑇𝑇𝑡𝑡 in equation 3a. The bias corrected estimate follows the procedure described in Appendix C. All models include a full set of age-month cell and year fixed effects. Standard errors are clustered by age-month cell.
Panel A: Ages 20 - 21 (1) (2) (3) (4) (5) (6)Medicaid Private Miscellaneous Insured County Self-Pay
Naïve estimate 0.143 0.007 -0.009 0.141 -0.084 -0.057(0.006) (0.006) (0.003) (0.004) (0.003) (0.003)
Bias-corrected estimate 0.137 0.007 -0.008 0.136 -0.081 -0.055(0.011) (0.015) (0.004) (0.026) (0.002) (0.026)
2012-13 mean for age 21 0.27 0.40 0.08 0.74 0.10 0.16Observations 100,272
Panel B: Ages 64 - 65Medicaid Private Miscellaneous Insured County Self-Pay
Naïve estimate 0.079 -0.017 -0.002 0.06 -0.035 -0.025(0.003) (0.004) (0.004) (0.002) (0.001) (0.001)
Bias-corrected estimate 0.079 -0.003 0.008 0.084 -0.032 -0.053(0.004) (0.008) (0.007) (0.016) (0.001) (0.016)
2012-13 mean for age 64 0.19 0.42 0.30 0.92 0.04 0.05Observations 373,974
48
Table 4: Patient volume
Note: This table presents regression coefficients obtained using age month-year level data as described in Section IV.C, using the sample of young(Panel A) and elderly (Panel B) patients. The dependent variable is log of hospital stays or ER arrivals in each age-month-year. Each cell presents a coefficient from a different regression. Estimated change in the discontinuity post-ACA is the coefficient on 𝑑𝑑𝑠𝑠 ⋅ 𝑇𝑇𝑡𝑡 in equation 4b. The RD-DD (IV) estimate is computed as the ratio of the change in discontinuity in utilization to that in insurance coverage, which is the first stage. The bias corrected estimate follows the procedure described in Appendix C and employs a modified first stage estimator. All models include a full set of year fixed effects. Robust standard errors reported.
(1) (2) (3) (4) (5)
All Through ER Not through ER All arrivals ER visitsPanel A: Ages 20 - 21Naïve estimate Change post-ACA 0.020 0.004 0.053 0.011 0.012
(0.016) (0.022) (0.024) (0.005) (0.006) First stage (Ins) 0.096 0.087
(0.002) (0.005) RD-DD (IV) estimate 0.139 0.025 0.374 0.116 0.137
(0.111) (0.146) (0.162) (0.053) (0.061)Bias-corrected estimate First stage (Ins) 0.104 0.114
(0.003) (0.007) RD-DD (IV) estimate 0.145 0.032 0.381 0.100 0.110
(0.080) (0.123) (0.129) (0.035) (0.038)
2012-13 mean (@21) per 1,000 23.8 16.8 7.1 284.5 267.7
Panel B: Ages 64 - 65Naïve estimate Change post-ACA 0.049 0.025 0.082 0.042 0.047
(0.010) (0.011) (0.015) (0.006) (0.007) First stage (Ins) 0.087 0.086
(0.002) (0.004) RD-DD (IV) estimate 0.813 0.412 1.373 0.484 0.546
(0.152) (0.177) (0.222) (0.067) (0.0799)Bias-corrected estimate First stage (Ins) 0.109 0.116
(0.012) (0.012) RD-DD (IV) estimate 0.592 0.301 0.999 0.387 0.409
(0.102) (0.110) (0.217) (0.048) (0.051)
2012-13 mean (@64) per 1,000 140.7 89.1 51.6 328.6 239.5
0.060(0.002)
0.083(0.016)
Hospital stays ER data
0.141(0.004)
0.136(0.026)
49
Table 5: Hospital choice
Note: This table presents estimated effects on choice of hospital on two dimensions – owner type and quality of care. Panels A and B present results for the young and elderly respectively, using data on hospital stays. The dependent variables are indicators for non-profit, for-profit or government ownership (Columns 1-3) and standardized 2009 quality scores reported by CMS (Columns 4-5). Quality scores are available for a subset of hospitals and hence fewer observations in sample (~84,000 and 310,000 for young and elderly respectively). Estimated change in discontinuity post-ACA is the coefficient on 𝑑𝑑𝑖𝑖 ⋅ 𝑇𝑇𝑡𝑡 in equation 3b. The RD-DD (IV) estimate is computed as the ratio of the change in discontinuity in share of hospital owner-type or mean quality score to that in insurance coverage, which is the first stage. All models control linearly for age, and include age-month cell and year fixed effects. Regressions on hospital quality score control for hospital owner type to determine sorting conditioning on hospital owner type. Standard errors are clustered by age-month cell. Table A. 2 presents corresponding estimates using ER arrivals data.
(1) (2) (3) (4) (5)
Non-profit For-profit Govt. Mortality ReadmissionPanel A: Ages 20 - 21
Change post-ACA 0.008 0.012 -0.020 -0.017 0.014(0.007) (0.004) -0.007 (0.010) (0.015)
First stage (Ins)
RD-DD (IV) estimate 0.058 0.081 -0.139 -0.120 0.100(0.051) (0.027) (0.050) (0.067) (0.103)
2012-13 mean at age 21 0.66 0.15 0.20 0.10 0.08
Panel B: Ages 64 - 65
Change post-ACA 0.001 0.008 -0.009 -0.010 -0.013(0.003) (0.002) (0.002) (0.007) (0.008)
First stage (Ins)
RD-DD (IV) estimate 0.011 0.135 -0.145 -0.159 -0.215(0.055) (0.038) (0.040) (0.118) (0.129)
2012-13 mean at age 64 0.72 0.15 0.13 0.06 -0.03
(0.002)
Owner type Quality score
(0.004)
0.060
0.141
50
Table 6: Health outcomes
Note: This table presents estimated effects on two health outcomes – in-hospital mortality and share of stays/visits that were potentially avoidable. Panels A and B present results for the young and elderly respectively. Columns 1-2 and 3-4 present results for hospital stays and ER arrivals respectively. The dependent variables are indicators for in-hospital death (Columns 1 and 3) and potentially avoidable episode (Columns 2 and 4). Estimated change in discontinuity post-ACA is the coefficient on 𝑑𝑑𝑖𝑖 ⋅ 𝑇𝑇𝑡𝑡 in equation 3b. The RD-DD (IV) estimate is computed as the ratio of the change in discontinuity in health outcome to that in insurance coverage, which is the first stage and presented in Table 3. All models control linearly for patient age, include indicators for gender and reason for arrival, hospital owner type, and age-month cell and year fixed effects. Standard errors are clustered by age-month cell.
(1) (2) (3) (4)
Mortality PAH Mortality PAHPanel A: Ages 20 - 21
Change post-ACA -0.001 -0.001 -0.0001 -0.0017(0.001) (0.006) (0.000) (0.002)
First stage (Ins)
RD-DD (IV) estimate -0.010 -0.007 -0.001 -0.017(0.006) (0.042) (0.0006) (0.020)
2012-13 mean at age 21 0.007 0.23 0.001 0.18
Panel B: Ages 64 - 65
Change post-ACA -0.002 -0.002 -0.0011 0.0023(0.001) (0.004) (0.000) (0.002)
First stage (Ins)
RD-DD (IV) estimate -0.025 -0.031 -0.013 0.027(0.014) (0.062) (0.004) (0.025)
2012-13 mean at age 64 0.027 0.21 0.012 0.20
Hospital stays ER arrivals
0.141 0.096(0.004) (0.002)
0.060(0.002)
0.087(0.002)
51
Table 7: Hospital finances and spillovers
Note: This table presents regression coefficients on hospital finances and spillover effects. All revenue variables are expressed in thousands of 2016 $. Panel A presents average effects across all hospital types without including linear time trends, while Panel B presents corresponding results including linear trends. The coefficient on un-insurance is obtained by estimating equation 5a, while the coefficients on self-pay and county indigent are obtained simultaneously from equation 5b. DSH refers to Disproportionate Share payments made by state and federal governments to hospitals if they disproportionately serve Medicaid and un-insured patients. Under the ACA, DSH payments were mandated to decrease. All models include a full set of hospital and year fixed effects. Standard errors are clustered by hospital. Dependent variable mean values computed pre-ACA i.e. over 2009-13. To interpret coefficients, note that mean share of self-pay and county indigent patients across hospitals in 2008 was 0.07 and 0.04 respectively.
Total rev Medicaid County (Mdcd+Cty) Private DSH Infant Elderly Capital exp.per bed per bed per bed per bed per bed per bed mortality mortality per bed('000 $) ('000 $) ('000 $) ('000 $) ('000 $) ('000 $) per 1k deliv. per 1k stays ('000 $)
Panel A: Without trends Un-insurance 189.81 508.47 -120.50 387.96 -173.26 -89.46 2.07 -14.95 81.50
(300.38) (196.18) (56.78) (198.69) (183.60) (62.75) (3.05) (9.48) (145.50)
Self-pay -1062.38 -219.35 -164.89 -384.24 -480.51 -46.60 6.83 -29.27 413.73(640.39) (274.03) (71.61) (249.82) (448.95) (69.55) (3.46) (13.38) (488.04)
County indigent 783.33 853.44 -99.46 753.98 -27.63 -109.77 -0.04 -8.16 -75.97(249.21) (197.82) (81.97) (194.58) (162.62) (76.09) (4.10) (11.16) (111.30)
Panel B: Including trends Un-insurance 250.01 387.23 -284.20 103.03 62.01 -200.79 2.67 -24.83 -67.48
(177.89) (218.15) (129.53) (186.26) (94.76) (62.46) (5.91) (16.63) (124.50)
Self-pay 58.02 -154.69 -200.28 -354.97 311.81 -196.02 9.44 -18.95 -254.37(319.84) (311.80) (150.28) (343.92) (223.90) (101.72) (6.45) (13.05) (339.03)
County indigent 341.03 644.18 -323.99 320.19 -56.43 -203.06 -0.30 -27.61 21.13(230.61) (229.12) (148.54) (177.14) (93.32) (71.59) (7.90) (20.25) (106.31)
Observations 2,878 1,577 2,202 2,878Dep. Var. mean 978 189 13 202 416 18 3.5 52.9 83
Other outcomesHospital revenue
52
Table 8: Robustness of insurance coverage change (hospital stays)
Note: This table presents estimated change in discontinuity in insurance coverage at ages 21 (Panel A) and 64 (Panel B) under alternative bandwidth and specification assumptions. Due to space constraints only three different insurance types are presented – Medicaid, Private and County. These are the principal coverage types of interest, at least for the discussion on crowd-out (Section IV.B.ii). The standard specification models insurance as a linear function of age but holds the slope constant pre and post-ACA. The flexible specification relaxes this constraint. Columns 1-3 present estimates obtained using a sample with 1 year bandwidth around the threshold, columns 4-6 use a 2 year bandwidth (i.e. ages 19-22.9 and 63-66.9), and columns 7-9 correspond to a 3 year bandwidth. Naive estimates are coefficients on 𝑑𝑑𝑖𝑖 ⋅ 𝑇𝑇𝑡𝑡 in equation 3a. Bias-corrected estimates follow the approach described in Appendix B. All regressions include age-month cell and year fixed effects. Standard errors are clustered by age-month cell.
Panel A: Young patientsNaïve estimate Medicaid County Private Medicaid County Private Medicaid County PrivateStandard spec 0.143 -0.084 0.007 0.151 -0.084 0.006 0.153 -0.084 0.009
(0.006) (0.003) (0.006) (0.005) (0.002) (0.004) (0.004) (0.001) (0.003)Flexible spec 0.154 -0.083 0.006 0.143 -0.084 0.003 0.145 -0.085 0.001
(0.009) (0.008) (0.011) (0.008) (0.004) (0.008) (0.007) (0.003) (0.007)Bias-corrected estimateStandard spec 0.138 -0.081 0.008 0.154 -0.081 0.016 0.15 -0.081 0.012
(0.012) (0.002) (0.015) (0.009) (0.001) (0.010) (0.008) (0.001) (0.009)Flexible spec 0.148 -0.081 0.007 0.146 -0.08 0.014 0.142 -0.083 0.004
(0.015) (0.006) (0.02) (0.011) (0.003) (0.014) (0.010) (0.003) (0.011)
Observations 100,272 203,913 305,4952012-13 Mean for patients age 21 0.268 0.097 0.401
Panel A: Elderly patientsAs-is estimates Medicaid County Private Medicaid County Private Medicaid County PrivateStandard spec 0.079 -0.035 -0.017 0.081 -0.036 -0.015 0.085 -0.037 -0.017
(0.003) (0.001) (0.004) (0.002) (0.001) (0.003) (0.002) (0.001) (0.002)Flexible spec 0.077 -0.033 -0.03 0.076 -0.033 -0.02 0.074 -0.034 -0.014
(0.005) (0.001) (0.011) (0.004) (0.001) (0.007) (0.003) (0.001) (0.005)Bias-corrected estimatesStandard spec 0.079 -0.032 -0.003 0.091 -0.032 0.017 0.099 -0.032 0.024
(0.004) (0.001) (0.008) (0.004) (0.001) (0.009) (0.004) (0.001) (0.010)Flexible spec 0.077 -0.03 -0.015 0.086 -0.029 0.013 0.088 -0.029 0.026
(0.006) (0.001) (0.012) (0.005) (0.001) (0.011) (0.005) (0.001) (0.016)
Observations 373,974 612,439 1,137,3302012-13 Mean for patients age 64 0.188 0.038 0.424
BW=1 year BW=2 years BW=3 years
BW=1 year BW=2 years BW=3 years
53
Table 9: Falsification exercise
Note: This table presents results of a falsification exercise using data from 2008-11 (pre-ACA) supposing a placebo Medicaid expansion in 2010. Regression coefficients are obtained using data on hospital stays as described in Section IV.A for the sample of young (Panel A) and elderly (Panel B) patients respectively. This table provides equivalent estimates to the main estimates on insurance coverage, utilization, hospital choice and health outcomes. Estimated change in discontinuity post 2010 is the coefficient on 𝑑𝑑𝑖𝑖 ⋅ 𝑇𝑇𝑡𝑡 in equation 3a (insurance) and 3b (all others). Case level models include a full set of age-month cell and year fixed effects. When examining sorting on hospital quality, models control for hospital owner type. When examining effect on patient health, models control for hospital owner type. Standard errors are clustered by age-month cell.
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)Utilization Health
Panel A: 20.0 - 21.9 Medicaid Private Insured County Self-Pay Stays ER arrivals Govt. RA Mort. RA Readm. Mortality
Standard spec -0.02 0.009 -0.017 0.004 0.014 -0.017 -0.0002 0.011 -0.008 -0.006 0.001(0.005) (0.006) (0.005) (0.002) (0.004) (0.013) (0.005) (0.006) (0.010) (0.013) (0.001)
Flexible spec -0.032 0.024 -0.029 0.012 0.017 -0.024 -0.0167 0.016 -0.028 0.024 0.000(0.008) (0.013) (0.012) (0.003) (0.011) (0.027) (0.010) (0.012) (0.022) (0.027) (0.002)
Mean at 21 (2008-09) 0.27 0.39 0.75 0.09 0.17 24 269 0.21 0.07 0.06 0.007Observations 104,546 88,644 86,850
Panel B: 64.0 - 65.9
Standard spec 0.007 -0.006 -0.005 0.002 0.003 0.11 0.1243 0.002 0.009 -0.015 0.002(0.002) (0.004) (0.002) (0.001) (0.001) (0.014) (0.013) (0.002) (0.008) (0.007) (0.001)
Flexible spec 0.004 -0.014 -0.014 0.005 0.008 0.014 0.0291 0.005 0.002 -0.012 -0.000(0.004) (0.008) (0.004) (0.001) (0.003) (0.013) (0.014) (0.003) (0.010) (0.017) (0.002)
Mean at 64 (2008-09) 0.18 0.48 0.93 0.03 0.04 154 295 0.12 0.04 -0.03 0.028Observations 340,011 281,381 285,415
Insurance coverage Hospital choice
54
A. ADDITIONAL FIGURES AND TABLES
A.1a: Medicaid share in expansion states
A.1b: Medicaid share in non-expansion states
Figure A. 1: Medicaid share in expansion and non-expansion states
Note: This figure presents Medicaid share of state population for states that expanded Medicaid under the ACA (Panel A) and those that did not (Panel B). Medicaid share as of July-Sept 2013 (i.e. pre-ACA) is depicted in blue and the change through October 2016 is plotted in red. In both figures, states are sorted in ascending order of share of population in 2013. Comparable baseline data was not available for Connecticut (expanded) and Maine (did not expand).
55
Figure A. 2: California Medicaid eligibility requirements Note: This figure presents an extract from an official notice on California Medicaid (Medi-Cal) eligibility requirements. This is available at http://www.dhcs.ca.gov/formsandpubs/forms/Forms/MCED/Info_Notice/MC002_ENG_0907.pdf and pertains to 2007. The top right portion discusses age thresholds for a person to be eligible for Medicaid under the “indigent” category, i.e. not disability or welfare recipient. Childless adults were usually ruled out unless they had special circumstances such as pregnancy (in the case of women) or were in a nursing home.
56
A.3a: Insurance change for the young
A.3b: Insurance change for the elderly
Figure A. 3: Insurance coverage changes (details)
Note: This figure presents observed coverage rates for different insurers, collapsed to age-month bin and corresponding fitted values (dashed line) obtained by estimating equation 3a on case level data as described in Section IV.A. It is a more detailed version of Figure 3. Self-pay includes charity care. The figure pertains to hospital stays in the RD sample for young (Panel A) and elderly (Panel B) patients respectively. All models control linearly for age and include age-month cell and year fixed effects. Standard errors are clustered by age-month cell.
57
A.4a: Insurance change for ER patients (Young)
A.4b: Insurance change for ER patients (Elderly)
Figure A. 4: Insurance change for ER patients
Note: This figure presents change in insurance coverage among patient arrivals at ERs and corresponding fitted values by age-month cell. These were obtained by estimating equation 3a on case level data as described in Section IV.A for the sample of young (Panel A) and elderly (Panel B) patients respectively. The treated groups are those aged 21 (young) and 64 (elderly). Both panels present the data for 2012-15 (circles, solid line) which includes the insurance expansions post-2014, and data from 2008-11 (squares, dashed line), which serves as a falsification exercise. The dependent variable is set to 1 if the individual is not self-insured, on charity or county indigent care. All models control linearly for age and include a full set of year and age-month cell fixed effects. The figures also present estimated change in discontinuity, which is the coefficient on 𝑑𝑑𝑖𝑖 .𝑇𝑇𝑡𝑡 in equation 3a. These coefficients are not bias-corrected. Standard errors are clustered by age-month cell.
-0.1
0
0.1
0.2
0.3
20 20.5 21 21.5 22Cha
nge
in in
sura
nce
cove
rage
Age at admission (months)
2012-15 2008-11
Period RD-DD (S. E.)2012-15 0.096 (0.001)2008-11 -0.033 (0.002)
58
A.5a: Shock to self-pay
A.5b: Shock to county coverage
Figure A. 5: Effects on hospital finances
Note: This figure presents coefficients on the interaction of Selfℎ08 (Panel A) and Countyℎ08 (Panel B) with an indicator for each year 𝑠𝑠 from 2009-16, obtained by estimating equation 5b with various hospital revenue variables as outcomes. Selfℎ08 and Countyℎ08 are the shares of hospital h patients coded self-pay and county indigent respectively in 2008. All revenue values have been deflated to be in 2016 dollars. This model is estimated using hospital-year level finances data made available by OSHPD over 2009-16 and includes hospital and year fixed effects. This specification is identical to that used in Figure 9 except that it does not include hospital specific time trends.
59
A.6a: ACS survey data, California
A.6b: Hospital discharge data
Figure A. 6: Medicaid and un-insurance by age group, California
Note: This figure uses two different data sources from California to examine the increase in share of Medicaid and decrease in un-insurance among individuals aged 21-64. Panel A presents the share of Medicaid and un-insurance as reported by the American Community Survey (ACS). Note that these are weighted averages and represent population share. Panel B presents an identical plot, but using data on hospital discharges. Panel B is therefore conditional on use.
60
Table A. 1: Insurance coverage (ER arrivals)
Note: This table presents regression coefficients obtained using case level data as described in Section IV.A for the sample of young (Panel A) and elderly (Panel B) patients respectively. The dependent variable is coverage by different insurers or self-pay/county indigent coverage. Miscellaneous includes Medicare, Government employees and workers’ compensation. This table pertains to ER arrivals i.e. ER visits including those that resulted in hospital stays. presents corresponding results on hospital stays. The naïve estimate of change in discontinuity at threshold is the coefficient on 𝑑𝑑𝑖𝑖 ⋅ 𝑇𝑇𝑡𝑡 in equation 3a. The bias corrected estimate follows the procedure described in Appendix x. All models include a full set of age-month cell and year fixed effects. Standard errors are clustered by age-month cell.
Panel A: Ages 20 - 21 (1) (2) (3) (4) (5) (6)Medicaid Private Miscellaneous Insured County Self-Pay
Naïve estimate 0.097 0.005 -0.006 0.096 -0.036 -0.06(0.002) (0.002) (0.001) (0.002) (0.001) (0.003)
Bias-corrected estimate 0.104 0.016 -0.005 0.116 -0.035 -0.081(0.003) (0.004) (0.001) (0.007) (0.001) (0.008)
2012-13 mean for age 21 0.21 0.36 0.06 0.64 0.05 0.31Observations 1,305,681
Panel B: Ages 64 - 65Medicaid Private Miscellaneous Insured County Self-Pay
Naïve estimate 0.100 -0.010 -0.004 0.087 -0.043 -0.043(0.002) (0.002) (0.002) (0.002) (0.001) (0.001)
Bias-corrected estimate 0.102 0.002 0.005 0.109 -0.041 -0.068(0.004) (0.006) (0.005) (0.012) (0.001) (0.012)
2012-13 mean for age 64 0.18 0.40 0.27 0.85 0.05 0.10Observations 897,504
61
Table A. 2: Hospital choice (ER arrivals)
Note: This table presents estimated effects on choice of hospital on two dimensions – owner type and quality of care. Panels A and B present results for the young and elderly respectively, using data on all ER arrivals. The dependent variables are indicators for non-profit, for-profit or government ownership (Columns 1-3) and standardized 2009 quality scores reported by CMS (Columns 4-5). Quality scores are available for a subset of hospitals and hence fewer observations in sample (~1.1 mn and 725,000 for young and elderly respectively). Estimated change in discontinuity post-ACA is the coefficient on 𝑑𝑑𝑖𝑖 ⋅ 𝑇𝑇𝑡𝑡 in equation 3b. The RD-DD (IV) estimate is computed as the ratio of the change in discontinuity in share of hospital owner-type or mean quality score to that in insurance coverage, which is the first stage. All models control linearly for age, and include age-month cell and year fixed effects. Regressions on hospital quality score control for hospital owner type. Standard errors are clustered by age-month cell.
(1) (2) (3) (4) (5)
Non-profit For-profit Govt. Mortality ReadmissionPanel A: Ages 20 - 21Naïve estimate Change post-ACA 0.009 0.002 -0.011 0.001 0.000
(0.002) (0.001) (0.001) (0.004) (0.004) First stage (Ins)
RD-DD (IV) estimate 0.0936 0.026 -0.119 0.011 0.005(0.017) (0.012) (0.016) (0.044) (0.040)
2012-13 mean at age 21 0.68 0.14 0.17 0.23 0.06
Panel B: Ages 64 - 65Naïve estimate Change post-ACA 0.012 0.007 -0.019 -0.001 -0.007
(0.002) (0.001) (0.002) (0.004) (0.005) First stage (Ins)
RD-DD (IV) estimate 0.139 0.075 -0.214 -0.016 -0.077(0.023) (0.017) (0.020) (0.047) (0.053)
2012-13 mean at age 64 0.70 0.13 0.17 0.16 0.00
0.087(0.002)
Owner type Quality Score
0.096(0.002)
62
Table A. 3: Discharge records by age and year
Note: This table presents the number of hospital stays and ER arrivals (i.e. ER visits and hospital stays originating in the ER) by age group over different two-year periods from 2008-15. The RD analysis focuses on patients aged 20-21 and 64-65 (highlighted). The non-elderly sample used in the geographical analysis focuses on patients aged 21-64. The RD sample and non-elderly sample cover about 5% and 50% of the total hospital stays respectively. In the case of ER arrivals, the analysis samples cover 5% and 55% of all available records.
Age 2008-09 2010-11 2012-13 2014-15 2008-09 2010-11 2012-13 2014-150-4 204,725 193,509 172,122 159,230 2,549,353 2,536,691 2,525,177 2,598,756 5-9 70,262 70,128 68,932 68,594 1,009,162 1,054,114 1,174,867 1,328,809 10-14 73,406 71,061 68,763 68,706 899,326 891,320 944,282 1,057,023 15-19 122,318 117,073 109,005 102,694 1,344,089 1,341,702 1,353,781 1,461,857 20 26,727 28,366 26,655 24,794 303,292 324,043 324,409 353,802 21 26,376 27,693 27,274 25,879 293,690 313,836 326,816 360,470 22-29 227,743 233,570 236,569 235,468 2,263,397 2,388,402 2,585,435 2,948,660 30-34 160,508 167,276 168,499 168,976 1,226,926 1,342,541 1,502,243 1,687,523 35-39 207,304 194,674 184,238 185,655 1,239,140 1,236,095 1,330,798 1,519,107 40-44 280,236 268,467 248,943 230,913 1,321,394 1,352,207 1,408,115 1,486,697 45-49 375,300 361,555 324,560 300,495 1,437,028 1,468,252 1,497,807 1,586,650 50-54 424,877 431,589 415,826 403,937 1,337,020 1,448,396 1,582,252 1,731,409 55-63 790,000 831,580 830,159 858,887 1,854,649 2,086,173 2,365,820 2,723,207 64 82,252 91,416 94,254 94,555 158,111 190,741 220,298 248,733 65 90,385 90,159 102,692 98,366 162,834 173,575 224,817 243,634 66-74 764,768 792,285 791,630 823,566 1,267,951 1,400,027 1,586,741 1,831,380 75-79 449,845 436,236 416,979 410,007 699,740 726,558 783,214 850,939 80-85 455,789 439,070 406,646 381,712 712,183 733,496 768,144 796,834 86-90 407,003 416,811 393,337 371,704 644,859 706,884 748,409 774,619 Total 5,239,824 5,262,518 5,087,083 5,014,138 20,724,144 21,715,053 23,253,425 25,590,109
Hospital Stays All ER arrivals
63
B. SPATIAL VARIATION The RD design yields our main set of results on utilization and crowd-out, but it trades off external validity
for internal validity. RD designs rely on weaker identification assumptions which are partially testable,
relative to differences-in-differences which relies on untestable assumptions. However, the RD estimates
are mainly representative of young adults and elderly individuals. While these age groups are important
targets of health policy, they are not representative of all non-elderly adults. Further, to the extent that 20
and 65 year olds are also affected by the ACA, the RD estimates will be biased downward. In order to
address these limitations, we supplement the RD analysis using an alternative approach that uses the entire
21-64 age sample and relies on variation across hospital markets rather than age thresholds.
We deploy a differences-in-differences research design using two sources of policy driven
variation. First, we use cross-sectional variation in un-insurance levels across Hospital Service Areas in
2008 (much before the ACA was passed). Regardless of other potential effects, the ACA was designed to
increase insurance coverage among lower income families and individuals. In 2013, the un-insurance rate
among California adults aged 19-64 with income below the federal poverty level was 36%, while only about
18% of adults above the poverty level were uninsured.23 Lower income neighborhoods had higher rates of
un-insurance prior to the ACA and would experience greater decreases in un-insurance due to the ACA.
Our thought experiment treats such HSAs as being hit by a more intense “insurance shock” than others.
Second, we leverage within-HSA time-series variation created due to the introduction of the ACA in 2014.
The DD estimator will quantify the impact of the ACA as the change in outcome of interest for HSAs with
high pre-ACA un-insurance relative to HSAs with low pre-ACA un-insurance.
We calculate un-insurance rates for each hospital service area in 2008 (𝑈𝑈𝐼𝐼𝑖𝑖𝐼𝐼𝑠𝑠𝑗𝑗08) as the observed
proportion of hospital stays among patients residing in an HSA that were self-insured or covered by county
indigent programs. Equation 6a represents the basic model to be estimated.
𝑌𝑌𝑖𝑖𝑗𝑗𝑡𝑡 = 𝛼𝛼𝑗𝑗 + 𝛾𝛾𝑡𝑡 + 𝜉𝜉 ⋅ 𝑈𝑈𝐼𝐼𝑖𝑖𝐼𝐼𝑠𝑠𝑗𝑗08 ⋅ 𝑇𝑇𝑡𝑡 + 𝜖𝜖1𝑖𝑖𝑗𝑗𝑡𝑡 (6𝑎𝑎)
𝑌𝑌𝑖𝑖𝑗𝑗𝑡𝑡 = 𝛼𝛼𝑗𝑗 + 𝛾𝛾𝑡𝑡 + 𝜉𝜉1 ⋅ 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑗𝑗08 ⋅ 𝑇𝑇𝑡𝑡 + 𝜉𝜉2 ⋅ 𝐶𝐶𝐶𝐶𝐶𝐶𝐼𝐼𝑡𝑡𝑦𝑦𝑗𝑗08 ⋅ 𝑇𝑇𝑡𝑡 + 𝜖𝜖2𝑖𝑖𝑗𝑗𝑡𝑡 (6𝑏𝑏)
𝑌𝑌𝑖𝑖𝑗𝑗𝑡𝑡 is an outcome of interest for patient 𝑖𝑖 in HSA 𝑗𝑗 in year 𝑡𝑡. 𝑇𝑇𝑡𝑡 is an indicator for years 2014 and later. The
coefficient of interest is 𝜉𝜉 which estimates the change in outcome 𝑌𝑌 post-ACA (2014-15) versus pre-ACA
(2009-13) for a market with baseline un-insurance of 100% compared to a market with zero baseline un-
insurance. We include a full set of HSA and year fixed effects, 𝛼𝛼𝑗𝑗 and 𝛾𝛾𝑡𝑡 respectively. Some specifications
account for observable differences in patient characteristics by including a vector 𝑿𝑿𝒊𝒊.
23 Reported by the Kaiser Family Foundation based on Current Population Survey (CPS) 2014 data.
64
To interpret the coefficient 𝜉𝜉 as a causal effect of insurance expansion, we need to make two
identification assumptions. First, outcomes in HSAs would evolve in a similar fashion in the absence of the
insurance expansions. To test the presence of possible pre-trends, we estimate and present results from
models allowing effects 𝜉𝜉𝑠𝑠 to vary flexibly by year from 2009 through 2015, as depicted in equation 6b
below. This approach does not treat 2014 and 2015 differently than the other years.
Y𝑖𝑖𝑗𝑗𝑡𝑡 = αj + γt + � 𝜉𝜉𝑠𝑠 ⋅ 𝑈𝑈𝐼𝐼𝑖𝑖𝐼𝐼𝑠𝑠j08 ⋅ I(t = s)𝑠𝑠=2015
𝑠𝑠=2009
+ ϵ3𝑖𝑖𝑗𝑗𝑡𝑡 (6𝑐𝑐)
Y𝑖𝑖𝑗𝑗𝑡𝑡 = αj + γt + � 𝜉𝜉1𝑠𝑠 ⋅ 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆j08 ⋅ I(t = s)𝑠𝑠=2015
𝑠𝑠=2009
+ � 𝜉𝜉2𝑠𝑠 ⋅ 𝐶𝐶𝐶𝐶𝐶𝐶𝐼𝐼𝑡𝑡𝑦𝑦j08 ⋅ I(t = s)𝑠𝑠=2015
𝑠𝑠=2009
+ ϵ4𝑖𝑖𝑗𝑗𝑡𝑡 (6𝑑𝑑)
Note that this analysis uses a different source of identifying variation relative to the regression
discontinuity analysis, in addition to using a larger patient sample. The estimates obtained in this analysis
inform us about changes in coverage, utilization or health outcomes for patients in high un-insurance areas
relative to patients in other markets. Though all patients included in the sample are potentially exposed to
the insurance expansions, this approach will not inform us on changes in the mean across all hospital
markets and it may understate true effects of the insurance expansions. Below we present results on two
key outcomes – changes in insurance coverage and patient sorting across hospitals.
i. Insurance coverage
TBD
ii. Hospital choice
TBD
65
Panel A: Change in insurance coverage
Panel B: Change in nature of patients admitted
Note: This table presents regression coefficients obtained by estimating equation 6a on case level data for the entire sample of non-elderly adults. Each coefficient is obtained from a different regression. We present results obtained using linear specifications, where the linear specification uses un-insurance share of patients in Hospital Service Area (HSA) in 2008 as the key regressor of interest. Models include a full set of HSA and year fixed effects. Standard errors are clustered by HSA. Dependent variable mean values are computed over 2012-13. Mean un-insurance share across HSAs is 0.12. Mean self-pay and county share is 0.075 and 0.045 respectively.
Insurance changesMedicaid Private Miscellaneous Insured Self County Uninsured
Basic specificationUninsured * Post 0.5635*** 0.3103*** -0.1160 0.7578*** -0.3889*** -0.3690*** -0.7578***
(0.1009) (0.0926) (0.1046) (0.0684) (0.0468) (0.0784) (0.0684)
Including county-specific trendsUninsured * Post 0.6658*** 0.2425*** -0.1693** 0.7390*** -0.2864*** -0.4527*** -0.7390***
(0.0990) (0.0585) (0.0698) (0.0745) (0.0447) (0.0774) (0.0745)
Simultaneously using self and county sharesSelf * Post 0.5612*** 0.2887*** -0.1998* 0.6502*** -0.7994*** 0.1491* -0.6502***
(0.1230) (0.1097) (0.1161) (0.0896) (0.0539) (0.0770) (0.0896)
County * Post 0.5658*** 0.3319*** -0.0321 0.8657*** 0.0224 -0.8881*** -0.8657***(0.1313) (0.1030) (0.1099) (0.0906) (0.0664) (0.1001) (0.0906)
Mean value in 2012-13 0.25 0.38 0.22 0.85 0.08 0.06 0.15Observations 8,898,107 8,898,107 8,898,107 8,898,107 8,898,107 8,898,107 8,898,107
Shift in reason for admission?Non-deferrable Through ER Neoplasms Injury
Basic specificationUninsured * Post 0.0451 -0.0558 0.0444*** 0.0343
(0.0299) (0.0776) (0.0170) (0.0277)
Including county-specific trendsUninsured * Post 0.0776*** -0.0501 0.0436*** 0.0237
(0.0222) (0.0547) (0.0123) (0.0349)
Simultaneously using self and county sharesSelf * Post -0.0255 -0.0798 0.0818*** 0.0504
(0.0457) (0.0921) (0.0230) (0.0412)
County * Post 0.1158* -0.0317 0.0193 0.0207(0.0637) (0.0908) (0.0222) (0.0403)
Mean value in 2012-13 0.19 0.66 0.07 0.22Observations 8,898,107 8,898,107 8,898,107 8,898,107
66
Panel C: Volume of care
Panel D: Choice of hospital
Note: This table presents regression coefficients obtained by estimating equation 6a on case level data for the entire sample of non-elderly adults. Each coefficient is obtained from a different regression. We present results obtained using linear specifications, where the linear specification uses un-insurance share of patients in Hospital Service Area (HSA) in 2008 as the key regressor of interest. Models include a full set of HSA and year fixed effects. Standard errors are clustered by HSA. Dependent variable mean values are computed over 2012-13. Mean un-insurance share across HSAs is 0.12. Mean self-pay and county share is 0.075 and 0.045 respectively.
VolumeLog(total) Log(through ER) Log (not ER) Log (NonDef)
Basic specificationUninsured * Post 0.9829*** 0.7734*** 1.0268*** 1.0184***
(0.1873) (0.2698) (0.2743) (0.2688)
Including county-specific trendsUninsured * Post 0.9344*** 0.6799*** 0.9620*** 1.0915***
(0.1813) (0.2487) (0.2138) (0.2462)
Simultaneously using self and county sharesSelf * Post 1.1818*** 1.1040*** 1.1171*** 1.0510***
(0.2222) (0.2978) (0.3398) (0.2837)
County * Post 0.7141*** 0.3267 0.9047*** 0.9742**(0.2742) (0.3731) (0.2941) (0.4385)
Mean value in 2012-13Observations 1,463 1,463 1,463 1,463
Choice of hospitalGovt. hosp Non-profit For-profit R A Mortality
Basic specificationUninsured * Post -0.1572* 0.2215*** -0.0643 0.1163
(0.0859) (0.0712) (0.0716) (0.2217)
Including county-specific trendsUninsured * Post 0.0401 0.0879 -0.1280 0.0146
(0.0568) (0.0726) (0.0784) (0.1880)
Simultaneously using self and county sharesSelf * Post -0.3001*** 0.2422** 0.0579 0.0402
(0.0913) (0.0974) (0.0876) (0.2284)
County * Post -0.0139 0.2008** -0.1868* 0.3676(0.1211) (0.0946) (0.0982) (0.2690)
Mean value in 2012-13 0.16 0.68 0.17 0.06Observations 8,898,107 8,898,107 8,898,107 7,619,890
67
Panel E: Health outcomes
Note: This table presents regression coefficients obtained by estimating equation 6a on case level data for the entire sample of non-elderly adults. Each coefficient is obtained from a different regression. We present results obtained using linear specifications, where the linear specification uses un-insurance share of patients in Hospital Service Area (HSA) in 2008 as the key regressor of interest. Models include a full set of HSA and year fixed effects. Standard errors are clustered by HSA. Dependent variable mean values are computed over 2012-13. Mean un-insurance share across HSAs is 0.12. Mean self-pay and county share is 0.075 and 0.045 respectively. Interpreting mortality coefficient: Mean value of %uninsured = 0.12 in 2008 Effect for average HSA = 0.12*-0.048 = -0.006 Relative to mean in-hospital mortality in CA = -0.006/0.045 ~13%!! Alternatively, can compare to differences in mortality rates across HSAs in 2012-13: - S.D. in mortality across HSAs in 2012-13 was 0.009 - Difference between top and bottom quartile was 0.056-0.035 = 0.021 So decrease in mortality for average HSA was very large relative to difference between ‘good’ and ‘bad’ HSAs. Can zoom in on 50+, potentially effects are coming from this group.
Health outcomes
All patients Non-deferrable All patients Non-deferrableBasic specificationUninsured * Post -0.0052 -0.0312* -0.1762*** -0.0403**
(0.0056) (0.0179) (0.0557) (0.0193)
Including county-specific trendsUninsured * Post -0.0110*** -0.0484*** -0.1683*** -0.0280*
(0.0040) (0.0139) (0.0325) (0.0149)
Simultaneously using self and county sharesSelf * Post -0.0140 -0.0584** -0.2311*** -0.0249
(0.0095) (0.0290) (0.0589) (0.0313)
County * Post -0.0026 -0.0297 -0.1336** -0.0301(0.0093) (0.0307) (0.0582) (0.0240)
Mean value in 2012-13 0.016 0.0454 0.18 0.07Observations 8,898,107 1,640,990 4,226,585 845,761
Mortality Potentially avoidable
68
B.1a: Insurance coverage (basic specification)
B.1b: Insurance coverage (including county specific trends)
Figure B. 1: Insurance coverage increases but county cover decreases
Note: This figure presents coefficients 𝜉𝜉𝑠𝑠 on the interaction of 𝑈𝑈𝐼𝐼𝑖𝑖𝐼𝐼𝑠𝑠𝑗𝑗08 and indicator for each year 𝑠𝑠 from 2009-15, obtained by estimating equation 6b with insurance coverage (Panel A) and hospital choice (Panel B) variables as outcomes. 𝑈𝑈𝐼𝐼𝑖𝑖𝐼𝐼𝑠𝑠𝑗𝑗08 is the share of uninsured hospital stays for people residing in HSA j in 2008. This model is estimated using case level data from the sample of all non-elderly adults over 2009-15, about 9 million observations and includes HSA and year fixed effects.
69
B.2a: Mortality (basic specification)
B.2b: Mortality (including county specific trends)
Figure B. 2: In-hospital mortality results
Note: This figure presents coefficients 𝜉𝜉𝑠𝑠 on the interaction of 𝑈𝑈𝐼𝐼𝑖𝑖𝐼𝐼𝑠𝑠𝑗𝑗08 and indicator for each year 𝑠𝑠 from 2009-15, obtained by estimating equation 6b with insurance coverage (Panel A) and hospital choice (Panel B) variables as outcomes. 𝑈𝑈𝐼𝐼𝑖𝑖𝐼𝐼𝑠𝑠𝑗𝑗08 is the share of uninsured hospital stays for people residing in HSA j in 2008. This model is estimated using case level data from the sample of all non-elderly adults over 2009-15, about 9 million observations and includes HSA and year fixed effects.
70
B.3a: Health outcomes (basic specification)
B.3b: Health outcomes (including county specific trends)
Figure B. 3: Health outcomes
Note: This figure presents coefficients 𝜉𝜉𝑠𝑠 on the interaction of 𝑈𝑈𝐼𝐼𝑖𝑖𝐼𝐼𝑠𝑠𝑗𝑗08 and indicator for each year 𝑠𝑠 from 2009-15, obtained by estimating equation 6b with insurance coverage (Panel A) and hospital choice (Panel B) variables as outcomes. 𝑈𝑈𝐼𝐼𝑖𝑖𝐼𝐼𝑠𝑠𝑗𝑗08 is the share of uninsured hospital stays for people residing in HSA j in 2008. This model is estimated using case level data from the sample of all non-elderly adults over 2009-15, about 9 million observations and includes HSA and year fixed effects.
71
C. BIAS CORRECTION (to be completed)
