+1(978)310-4246 credencewriters@gmail.com
  

https://www.rwjf.org/en/library/research/2019/02/medicaid-the-basics.html

Journal of Health Economics 53 (2017) 1–16
Contents lists available at ScienceDirect
Journal of Health Economics
journal homepage: www.elsevier.com/locate/econbase
Medicaid, family spending, and the financial implications of
crowd-out夽
Marcus Dillender
W.E. Upjohn Institute for Employment Research, 300 S. Westnedge Ave., Kalamazoo, MI 49007-4686, United States
a r t i c l e
i n f o
Article history:
Received 7 September 2016
Received in revised form
22 December 2016
Accepted 6 February 2017
Available online 16 February 2017
JEL classification:
D12
I13
Keywords:
Medicaid eligibility
Crowd-out
family spending
a b s t r a c t
A primary purpose of health insurance is to protect families from medical expenditure risk. Despite this
goal and despite the fact that research has found that Medicaid can crowd out private coverage, little is
known about the effect of Medicaid on families’ spending patterns. This paper implements a simulated
instrumental variables strategy with data from the Consumer Expenditure Survey to estimate the effect of
an additional family member becoming eligible for Medicaid on family-level health insurance coverage
and spending. The results indicate that an additional family member becoming eligible for Medicaid
increases the number of people in the family with Medicaid coverage by about 0.135–0.142 and decreases
the likelihood that a family has any medical spending in a quarter by 2.7 percentage points. As previous
research often finds with different data sets, I find evidence that Medicaid expansions crowd out some
private coverage. Unlike most other data sets, the Consumer Expenditure Survey allows for considering
the financial implications of crowd-out. The results indicate that families that transition from private
coverage to Medicaid are able to spend significantly less on health insurance expenses, meaning Medicaid
expansions can be welfare improving for families even when crowd-out occurs.
© 2017 Elsevier B.V. All rights reserved.
Medicaid provides free or cheap health insurance to individuals with low incomes and has the potential to transform families’
spending patterns. According to the 2014 Consumer Expenditure
Survey (CEX), 23 percent of households with at least one person with Medicaid coverage have any quarterly medical spending,
while 44 percent of households without Medicaid do. Meanwhile,
only 38 percent of households with at least one person with Medicaid coverage have any quarterly health insurance spending, while
73 percent without Medicaid do. These numbers are consistent
with Medicaid being a financial boon to households. But as Medicaid is a means-tested program, these numbers could also reflect
that poorer families that are able to spend less on medical care and
health insurance are more likely to be eligible for Medicaid.
Despite a key goal of health insurance being to protect families from medical expenditure risk, little is known about the effect
of Medicaid on financial outcomes. As Buchmueller et al. (2015a)
explain in their recent review of economics research on Medicaid,
“Given that a fundamental purpose of health insurance is to protect
individuals and families from the financial burden of large medical expenditures, there is surprisingly little research on the effect
夽 I thank Joelle Abramowitz, Erik Nesson, and seminar participants at the 2016
ASHEcon Conference for discussions and comments.
E-mail address: dillender@upjohn.org
http://dx.doi.org/10.1016/j.jhealeco.2017.02.002
0167-6296/© 2017 Elsevier B.V. All rights reserved.
of Medicaid on financial outcomes.” The research on the financial
effects of Medicaid that exists often focuses on extreme spending
events and finds that Medicaid reduces bankruptcies and the number of bills going to collections (Finkelstein et al., 2012; Gross and
Notowidigdo, 2011; Hu et al., 2016). In contrast to the literature on
the financial effects of Medicaid, the literature that studies Medicaid coverage crowding out private coverage is large,1 but a key
issue that has received little attention is what the financial implications of this crowd-out are for families.
The goals of this paper are to understand how Medicaid eligibility affects families’ spending and to consider the welfare
implications of spending effects. To estimate the effect of Medicaid
eligibility on families’ spending, I implement a simulated instrumental variables (IV) strategy using data from the CEX, which is a
data set collected by the Bureau of Labor Statistics (BLS) that tracks
the expenditures of U.S. households over the past quarter of the
year. I focus on families with incomes less than 200 percent of the
federal poverty level (FPL) and use variation in Medicaid eligibil-
1
A sample of the literature, much of which I discuss later, includes Bronchetti
(2014), Buchmueller et al. (2005), Busch and Duchovny (2005), Cutler and Gruber
(1996), Gruber and Simon (2008), Ham and Shore-Sheppard (2005a), Ham et al.
(2014), Hamersma and Kim (2013), Koch (2013, 2015), LoSasso and Buchmueller
(2004), and Shore-Sheppard (2008).
2
M. Dillender / Journal of Health Economics 53 (2017) 1–16
ity due to legislative changes during the 2000s.2 During this time
period, several states changed their eligibility rules for children,
but the majority of changes come from states expanding Medicaid
coverage for parents, meaning that the bulk of the identifying variation comes from parents becoming eligible for Medicaid. The CEX
provides two main advantages over many other data sets. First, it
allows for studying how Medicaid eligibility affects medical spending for a nationally representative sample of the U.S. population.
Second, the CEX allows for studying how crowd-out affects families’
spending on health insurance.
I first consider the effect of Medicaid eligibility on health insurance coverage using the CEX data. I estimate that an additional
person in a family being eligible for Medicaid increases the number of people in the family with Medicaid coverage by 0.135–0.142
and decreases the number of private health insurance plans paid
for by families by an average of 0.053–0.055. These estimates are
similar to estimates that I find using the March Current Population
Surveys (CPS) and the Survey of Income and Program Participation
(SIPP) and fall within the range of other estimates of the effect of
Medicaid eligibility on health insurance coverage.
I then consider how Medicaid eligibility affects spending. I find
that an additional person being eligible for Medicaid reduces the
likelihood that a family has any spending on medical care in a
quarter by 2.7 percentage points. The bulk of this decrease appears
to come from families who had relatively small quarterly medical
expenditures prior to Medicaid as decreases in the likelihood of
having positive spending in a quarter that is less than $100 drive
the results.
It is not immediately clear that the crowd-out of private health
insurance coverage will lead to families spending less on health
insurance. According to the 2014 CEX, more than 25 percent of
families with private health insurance do not have any quarterly
spending on health insurance, while only 30 percent spend more
than $1000 per quarter on health insurance. A cottage industry has developed around helping firms sign up Medicaid-eligible
employees for Medicaid to save the firms money (Kim, 2016), and
some research suggests that firms are successfully able to capitalize on Medicaid expansions to lower their health insurance costs
(Buchmueller et al., 2005). If the crowd-out comes from workers
who had low spending on health insurance, crowd-out may not
result in cost savings to families.
Taking advantage of the CEX’s questions about health insurance
spending, I find that an additional person being eligible for Medicaid
reduces the likelihood that a family has any health insurance expenditures by 4.3 percentage points. This decrease appears to come
from families that were spending more than $100 per quarter on
health insurance prior to Medicaid and reduces average spending
on health insurance in a quarter by $47. Under the assumption that
Medicaid only affects health insurance costs for those who experience crowd-out, the estimates imply that the reductions in private
insurance from Medicaid expansions save the switching families an
average of $4124–4284 per year, meaning the families that experience crowd-out were paying significant amounts of the premiums
for their private insurance. The results from this study suggest that
Medicaid eligibility makes families better off even when crowd-out
occurs.
The paper contributes to the small literature on the spending effects of Medicaid in a number of ways. One contribution is
that the current study focuses on Medicaid expansions that affect
families, whereas other studies have mainly focused on Medicaid
2
I use variation in both Medicaid and the State Children’s Health Insurance Program (CHIP) eligibility. As is common in this literature, I refer to both Medicaid and
CHIP as Medicaid for ease of discourse, even though the two are distinct programs
in many states.
coverage for childless adults, who comprise a much smaller share
of the Medicaid population. Another contribution of this study is
that it uses a nationally representative data set that is designed to
measure expenditures and that allows for considering the effect of
Medicaid eligibility on a wider variety of financial outcomes than
previous research has considered. Importantly, the paper estimates
the financial impacts of crowd-out on health insurance spending.
Whether families or employers capture monetary savings from
crowd-out is an open question that has important welfare implications.
The paper proceeds as follows. The next section provides a
brief overview of Medicaid and discusses previous research on
crowd-out from Medicaid expansions and on the spending effects
of Medicaid. Section 2 discusses a simple conceptual framework
for how Medicaid eligibility could affect spending and what the
implications of spending effects are for welfare. Section 3 describes
the variation in Medicaid eligibility, the CEX data, and the empirical strategy. Section 4 presents the empirical results. Section 5
discusses the results and concludes.
1. Background
1.1. Medicaid
Medicaid is a state-run program that is jointly financed by the
federal government and by states that provides health insurance to
people with low incomes, people with disabilities, and the elderly in
long-term care with low incomes.3 In 1997, State Children’s Health
Insurance Program (CHIP) legislation expanded eligibility of children for public health insurance beyond the existing limits of the
Medicaid program. In the years that followed, states have expanded
eligibility for both the CHIP program and Medicaid. States typically
require family incomes to be lower for parents to be eligible than
they do for children to be eligible.
Although federal Medicaid rules require states to cover major
services including physician and hospital care, the rules do not
require states to pay for other services such as prescription drugs
or dental care. Despite flexibility in coverage options, most states
cover most basic categories of health spending. For instance, all
states cover prescription drugs and optometrist services. While
almost all cover dental services for children, not all states cover
dental care for adults. Covered services are provided with little or
no copayment required (Ross et al., 2009).
1.2. Prior research on crowd-out from Medicaid expansions
The health insurance effects of Medicaid eligibility are relevant
for assessing the magnitude and channels of any potential effects of
Medicaid eligibility on spending. In contrast to the literature on the
financial impacts of Medicaid, a large literature has examined Medicaid take-up and the effect of Medicaid on private health insurance
coverage. Much of this research builds on Cutler and Gruber (1996),
who estimate the effect of Medicaid eligibility on Medicaid takeup and on private health insurance coverage using data from the
1988 to 1993 March CPS. To deal with the endogeneity of Medicaid
eligibility, they use a simulated IV strategy and find that Medicaid
eligibility increases Medicaid coverage by 23.5 percentage points
and decreases private insurance coverage by 7.4 percentage points.
3
Medicaid has traditionally provided coverage for low-income families with children rather than all low-income adults. As part of the Affordable Care Act, Medicaid
was expanded in many states to include low-income childless adults. Some states
had already expanded Medicaid to low-income childless adults prior to the passage
of the Affordable Care Act.
M. Dillender / Journal of Health Economics 53 (2017) 1–16
Since the seminal work by Cutler and Gruber (1996), other studies using simulated IV strategies with March CPS data have tended
to estimate effects of Medicaid eligibility that are smaller in absolute value both on Medicaid take-up and on the decrease in private
coverage. For example, Shore-Sheppard (2008) finds that adding
controls for age over time causes the estimated effect of Medicaid eligibility on Medicaid coverage to fall to between 15 and
19 percentage points and the estimated effect on private coverage to be close to zero and statistically insignificant. LoSasso and
Buchmueller (2004) estimate that 9 percent of the newly eligible
from the CHIP expansions received coverage and that half of this
increase came from those with private coverage.
Papers that use simulated IV strategies with the SIPP find similarly mixed estimates. For example, Ham and Shore-Sheppard
(2005a) estimate that Medicaid eligibility increases take-up by 11.8
percentage points but does not affect private coverage. Gruber and
Simon (2008), on the other hand, estimate that an additional person being eligible for Medicaid increases the number of people in
a household with Medicaid coverage by between 0.109 and 0.156
people and reduces the number of people with private coverage by
between 0.066 and 0.122. They find that most of the decrease in
private coverage comes from employer-sponsored coverage.
In addition to results varying by data set and specification, the
effect of Medicaid eligibility on the crowd-out of private coverage differs by group and time period. For instance, Hamersma and
Kim (2013) use SIPP data from 1996 to 2007 and find that a parent becoming eligible for Medicaid increases the likelihood that
the parent has Medicaid coverage by 14.8 percentage points and
has no effect on private coverage. Using CPS data from 1996 to
2002, Busch and Duchovny (2005) find similar take-up rates and
only weak, suggestive evidence that parental Medicaid expansions
crowd out private coverage. In contrast to these studies, McMorrow
et al. (2016) use National Health Interview Survey data from 1998
to 2010 and find evidence that about one-third of people who
take up parental Medicaid expansions previously had private coverage. Wagner (2015) finds that Medicaid eligibility expansions
for the disabled crowd out much private coverage, while early
results for childless expansions suggest that the degree of crowdout may vary by state (Sommers et al., 2014). As Gruber and Simon
(2008) summarize, early results appear to differ by data set and are
often sensitive to specification, while the literature on the more
recent expansions tends to find more consistent crowd-out effects.
Although there is no consensus about the degree of crowd-out, the
Congressional Budget Office considers the crowd-out to lie between
25 and 50 percent (Congressional Budget Office, 2007).4 Refer to
Bitler and Zavodny (2014), Buchmueller et al. (2015a), and Gruber
and Simon (2008) for excellent reviews of the literature.5
4
While most of the literature is agnostic about reasons that crowd-out might
occur, two papers explore endogenous health insurance offerings by firms. ShoreSheppard et al. (2000) use firm-level data and find that a firm having more workers
eligible for Medicaid is associated with the firm being less likely to offer dependent
coverage. However, they do not find that the share of Medicaid-eligible workers in
a firm has an effect on the premiums the workers have to pay. Buchmueller et al.
(2005) find some evidence that firms whose workers are likely to have been affected
by Medicaid expansions raise employees’ contributions for family coverage. Neither paper finds that firms are less likely to offer health insurance after Medicaid
expansions.
5
This literature review focuses largely on studies that use simulated IV strategies.
Other papers use different methods, including difference-in-differences and regression discontinuity designs. For examples, refer to Blumberg et al. (2000), Card and
Shore-Sheppard (2004), Dague et al. (2011), De La Mata (2012), Koch (2013), Koch
(2015), and Yazici and Kaestner (2000).
3
1.3. Prior research on financial impacts of Medicaid
The literature on the financial impacts of Medicaid is smaller
than the literature on crowd-out and has generally focused on an
inability to pay medical bills. Gross and Notowidigdo (2011) study
the effect of Medicaid expansions in the 1990s on bankruptcies.
They find that a 10-percent increase in Medicaid eligibility reduces
personal bankruptcies by 8 percent. Finkelstein et al. (2012) study
Oregon allocating slots to an oversubscribed Medicaid program for
adults using a lottery. Using administrative data from the Consumer
Credit Database and from the credit bureau TransUnion, they find
that Medicaid coverage is associated with a significant decline in
the likelihood that a medical bill is sent to collections but no significant decline in bankruptcies or liens. Hu et al. (2016) study the
effect of the Affordable Care Act (ACA) Medicaid expansion to childless adults using data from the Federal Reserve Bank of New York’s
Consumer Credit Panel. They find that low-income zip codes experienced a reduction in the number of unpaid bills and the amount of
debt sent to third-party collection agencies in states that expanded
Medicaid.
Although these extreme financial outcomes are important, they
do not represent a full picture of the spending effects of Medicaid.
Contemporaneous spending has received less attention, though the
Finkelstein et al. (2012) Oregon study is an exception. Finkelstein
et al. supplement their analysis of administrative data with survey
data and find intent-to-treat estimates that suggest that winning
the Medicaid lottery decreases the likelihood of having any medical spending in the last six months by 5.8 percentage points and
of having an outstanding medical bill by 5.2 percentage points. The
current study differs from Finkelstein et al. in three main ways.
First, the current study focuses on Medicaid eligibility for families,
while Finkelstein et al. study the effect of Medicaid on childless
adults, who may respond to Medicaid differently than families.
Medicaid eligibility for childless adults is still relatively new, and
the vast majority of people covered by Medicaid are in families with
children.6 Second, the CEX contains a much wider set of spending
outcomes than is available in the survey used in Finkelstein et al. A
particularly important spending outcome available in the CEX that
the Oregon health insurance experiment does not address is health
insurance spending. Third, the current study uses a simulated IV
strategy, while Finkelstein et al. study the effects of an experiment.
While experiments are often thought to be the gold standard of
research, they are expensive and rarely done with health insurance,
which precludes them as an option for most studies of Medicaid.
As is the case with the Oregon health insurance experiment, they
are also often restricted in their geography.
Another study that considers the financial implications of Medicaid is Sommers and Oellerich (2013), who estimate the impact of
Medicaid with the Census Bureau’s Supplemental Poverty Measure
by stochastically drawing counterfactual medical expenditures
from propensity-score-matched individuals without Medicaid.
They find that Medicaid reduces out-of-pocket medical spending
from $376 to $871 per beneficiary and decreases poverty rates by
1.0 percent among children, 2.2 percent among disabled adults,
and 0.7 percent among elderly individuals. My study differs from
Sommers and Oellerich in that I estimate the effect of Medicaid eligibility using variation from natural experiments and consider a
wider set of outcomes.7
6
According to 2015 March CPS data, less than a quarter of Medicaid recipients
had no children in their household.
7
Gruber and Yelowitz (1999) produce an early study of the effect of Medicaid
eligibility on savings and consumption using CEX data from the 1980s and early
1990s and find evidence that Medicaid eligibility increases non-health-related consumption and decreases savings. As Medicaid typically had asset tests prior to the
4
M. Dillender / Journal of Health Economics 53 (2017) 1–16
2. Conceptual framework
Medicaid eligibility has the potential to affect health insurance
coverage in two main ways. First, it can allow uninsured people to
become insured. Second, it can allow people who were receiving
private coverage to drop that coverage so that they can receive
Medicaid.
For the previously uninsured, Medicaid lowers the price of
health care. As health care is a normal good, this lower price means
people should theoretically consume more of it, a prediction that
has found broad empirical support.8 If all health care expenditures are covered by Medicaid, the effect of Medicaid on spending
on health care should be negative. If non-covered care complements covered care, Medicaid eligibility could potentially increase
spending on medical care. As lowering the cost of medical care can
increase the amount of medical services received, reduce medical expenditures, and lower the risk of catastrophic medical bills,
Medicaid likely increases welfare for the previously uninsured.
For those who previously had private coverage, Medicaid may
allow families to spend less on health insurance premiums, especially if they were paying a high share of the premiums for the
private health insurance. However, if the coverage they had was
poor or if employers can successfully capture the cost savings from
Medicaid expansions, households’ spending on health insurance
may not fall dramatically. If the coverage people lose was poor,
then Medicaid would affect medical spending in a way similar as it
does for the previously uninsured.
The welfare implications of crowd-out depend on both the quality of the private coverage that is crowded out and what share the
family was paying for the private coverage. Consider the following
three pre-Medicaid-eligibility cases:
1. The family was paying very little for good coverage.
2. The family was paying full price for good coverage.
3. The family was paying very little for poor coverage.
Medicaid eligibility crowding out private coverage for case (1)
does not necessarily help the family. If an employer was paying
large amounts for the worker’s health insurance, then the worker
leaving private employer-sponsored coverage for Medicaid lowers
employers’ costs. These lower costs for employers may or may not
later translate into higher wages for the worker.9
1990s, their analysis pertains to a different vintage of the Medicaid program than
exists now. In Appendix B, I consider the effect of Medicaid eligibility on non-healthrelated consumption and discuss Gruber and Yelowitz further.
8
However, it should be noted that people may ultimately consume less medical
care if Medicaid allows them to obtain more preventive care that prevents the need
for more health care later. Examples of research that find that Medicaid increases
the use of health care services include Aizer (2007), Baicker et al. (2013), Bronchetti
(2014), Buchmueller et al. (2015b), Burns et al. (2014), Currie et al. (2008), Dafny
and Gruber (2005), De La Mata (2012), DeLeire et al. (2013), Finkelstein et al. (2012),
Lipton and Decker (2015), and Taubman et al. (2014). This finding holds for private
insurance health insurance (Anderson et al., 2012, 2014), Medicare (Card et al., 2008,
2009), and health insurance expansions coming from broad health insurance reform
(Kolstad and Kowalski, 2012; Miller, 2012) as well.
9
It is unclear whether workers or firms bear the cost of health insurance benefits. If workers value money spent on health insurance as much as they would value
the money itself, if employers can perfectly identify which people are eligible for
Medicaid, and if employers can perfectly and instantly adjust compensation, theory
predicts that employers will pass the costs of health insurance to employees in the
form of lower wages (Gruber, 1994; Summers, 1989). Therefore, Medicaid expansions have the potential to increase wages if employers no longer have to provide
workers with health insurance coverage. However, these assumptions may not hold
for a variety of reasons. For example, identifying Medicaid-eligible people is likely
difficult since employers would have to know total family income. Furthermore,
Finkelstein et al. (2012) find that the Oregon health insurance experiment’s Medicaid recipients only value Medicaid at 20–40 percent of Medicaid’s cost, suggesting
that many low-income workers may not fully value health insurance.
Crowd-out of (2) or (3), on the other hand, will still make the
family better off. If the family was paying full price for good coverage (case 2), then Medicaid functions as an income transfer to
low-income families. As Bitler and Zavodny (2014) explain, these
transfers are still welfare enhancing on average if societal preferences put more weight on income at the bottom end of the income
distribution than on income at the top end of the income distribution. Under certain assumptions, risk aversion results in crowd-out
of (3) being more welfare improving than crowd-out of (2), but in
either case, Medicaid still makes these families better off. Refer to
Appendix A for a discussion of these assumptions and the model
that underlies these predictions.
To summarize the empirical predictions and implications, Medicaid likely lowers medical expenditures, but there are also ways
that Medicaid could increase medical expenditures. Medicaid
crowding out private coverage would mean that Medicaid eligibility should have a non-positive effect on health insurance spending.
Finding that Medicaid eligibility reduces private health insurance
coverage but not health insurance spending means that the family’s
private insurance likely had very low benefits (and was therefore
not costly) or that the cost-savings from Medicaid were passed
through to firms.10 Finding a negative effect of Medicaid eligibility on spending on health insurance suggests that the family was
paying the premiums for private insurance prior to switching to
Medicaid.
3. Medicaid variation, data, and estimation
3.1. Variation in Medicaid Eligibility
This paper uses variation in Medicaid eligibility that arises due
to changes to the income thresholds for parents and children.11
As most of the CHIP expansions for children occurred in the late
1990s, the majority of the variation in this paper comes from expansions for parents. In 2000, the mean income threshold for parental
Medicaid eligibility for states in the CEX was 76.7 percent of the
FPL. By 2014, the mean income threshold was 100.7 percent of
the FPL. From 2000 to 2014, fourteen states more than doubled
their income thresholds. Many states, like Arizona, Connecticut, and
Illinois, increased their threshold dramatically in the early to mid2000s. Other states, like Arkansas and Kentucky, did not experience
large increases until they adopted the ACA’s Medicaid expansion in
2014. A third set of states, including Missouri and Washington, lowered their eligibility requirements from 2000. Despite the increases
in parental Medicaid eligibility since 2000, all states limit coverage
to parents in families with incomes under 250 percent of the FPL,
and all states except three limit coverage to parents in families with
incomes under 200 percent of the FPL.
While most of the expansions for children occurred in the late
1990s, several states changed their income thresholds for children
during the time period studied. For instance, Arkansas and Tennessee both lowered their income thresholds in the early 2000s to
under 200 percent of the FPL before they increased the thresholds
a few years later. Both Hawaii and Texas expanded CHIP in 2000,
while other states altered their eligibility thresholds in the mid2000s. Table 1 shows the states with the five largest increases and
decreases to adult income thresholds during the time period stud-
10
Alternatively, finding an effect of Medicaid eligibility on private health insurance coverage but not on health insurance spending could also mean that firms
distribute the cost savings from Medicaid expansions to employees in other ways,
such as through higher wages or better working conditions. In the empirical section,
I find that families experience a reduction in the amount that they pay for health
insurance, but I cannot rule out adjustments along other margins.
11
Data on Medicaid’s income thresholds come from Hamersma and Kim (2013),
the Kaiser Family Foundation (2016a,b), and LoSasso and Buchmueller (2004).
M. Dillender / Journal of Health Economics 53 (2017) 1–16
Table 1
Largest changes in Medicaid eligibility.
State
Panel A: parents
Largest increases
Arkansas
Arizona
Nevada
Maryland
Illinois
Largest decreases
New Jersey
Washington
Maine
Wisconsin
Missouri
Panel B: Children ages 1–5
Largest increases
Oregon
Illinois
Indiana
Louisiana
Hawaii
Panel C: Children ages 6–18
Largest increases
Hawaii
Texas
Oregon
Louisianna
Wisconsin
2000
5
Table 2
Summary statistics.
2013
2014
19
32
32
34
35
16
106
84
122
139
138
138
138
138
138
200
200
157
185
107
200
71
200
200
35
138
138
105
100
24
170
185
150
150
185
300
200
250
250
300
305
318
255
255
313
100
100
170
150
185
300
200
300
250
300
313
206
305
255
306
Sources: Data on Medicaid’s income thresholds come from Hamersma and Kim
(2013), the Kaiser Family Foundation (2016a,b), and LoSasso and Buchmueller
(2004).
ied as well as the states with the five largest increases to the income
thresholds for younger and older children. For each of these states,
Table 1 displays the thresholds in 2000, 2013, and 2014.12
3.2. Data
The main data used in the analysis come from the 2000–2014
CEX Interview Surveys. The CEX is a rotating panel that interviews
roughly 5000 consumer units about their spending in the previous quarter every three months over five calendar quarters. Every
quarter 20 percent of the sample is replaced by new consumer
units. A consumer unit is not necessarily a household and instead
includes household members who share expenditures. I use the
terms household and family to mean consumer unit.
I first use the CEX to estimate the effect of Medicaid eligibility
on measures of health insurance coverage. The CEX asks about the
number of people in a family covered by Medicaid as well as each
plan that a family has. Unlike most other data sets that ask about
health insurance, the CEX also asks about spending on each private
plan. I focus on the number of family members covered by Medicaid
and on the number of various types of private health insurance
plans that the family spends money on in a quarter. I then study
four main categories of quarterly health-related spending: medical,
prescription drugs, dental, and health insurance. Next, I draw on the
CEX’s limited information on medical debt to estimate the effect of
Medicaid eligibility on medical debt. In Appendix B, I use the CEX’s
information on labor market outcomes and other sources of income
to consider possible effects on non-health spending and on program
participation.
As the focus of this study is on low-income families, I restrict
the sample to families who make less than 200 percent of the FPL. I
12
I display the thresholds in 2013 in Table 1 in addition to the thresholds at the
end of the study period because many states adjusted their adult eligibility in 2014
as they adopted the Medicaid expansions of the ACA.
Male head of household
Age of head of household
White head of household
Black head of household
Asian head of household
Two parents present
Number of high school graduates in family
Number in family with some college
Number of college graduates in family
Family income as a fraction of FPL
Anyone in family with Medicaid
Number of people in family receiving Medicaid
Any private health insurance plan
Number of private health insurance plans
Mean
St. Dev.
0.340
36.940
0.723
0.207
0.045
0.586
1.115
0.618
0.190
1.006
0.355
0.992
0.316
0.374
0.474
8.733
0.447
0.405
0.207
0.492
0.898
0.792
0.501
0.637
0.479
1.602
0.465
0.614
Notes: The data come from the 2000 to 2014 CEX. The sample contains 30,752
observations from 14,024 families.
choose 200 percent since most Medicaid expansions for parents are
to less than 200 percent of the FPL, but I evaluate the sensitivity of
the results to the cutoff in Section 4.3. For the primary analysis, I use
income data as reported by families but drop families with people
who work but have no income since they likely did not fully report
income.13 I calculate all expenditures in 2012 dollars and exclude
families whose state identifiers are suppressed in the CEX due to
confidentiality reasons.14 The final sample contains 30,752 observations from 14,024 families. Summary statistics for the sample are
shown in Table 2.15 As would be expected with a low income sample, these households have lower average education levels than the
rest of the nation in addition to lower incomes. Slightly more than
one-third of the sample has someone who receives Medicaid.
Table 3 summarizes spending patterns for families with incomes
less than 200 percent of the FPL as well as for families above 200
percent of the FPL. All of the means are statistically different from
each other at the one-percent level except for the mean percent of
household spending on medical care.16
3.3. Estimation
The primary goal of this paper is to estimate the effect of an
additional person in a family becoming eligible for Medicaid on
health insurance and spending outcomes. I estimate models of the
following form:
yifst = ˛ + fs + ıft + ˇXifst + unempst + ELIGifst + ifst ,
(1)
where i indexes the family, f indexes the family type, s indexes
the state, t indexes the year, y is a measure of health insurance or
spending, X is a set of individual covariates, unemp is the stateyear unemployment rate from the BLS’s Local Area Unemployment
Statistics, are state fixed effects, ı are year fixed effects, ELIG is
the number of people in the family eligible for Medicaid, and is
13
Prior to 2004, the CEX did not impute income, so families who declined to report
income or did not know their income show up in the microdata as having no income.
Beginning in 2004, the CEX began imputing the income of non-responders. I use
reported income data and drop families with people who work but have no income to
have as consistent a sample across time as possible. When the CEX begins imputing
income in 2004, almost all families with earners report positive income or the BLS
imputes income for them. I consider the robustness of the results to a variety of
ways of handling income in Section 4.3.
14
Families in the following states are not interviewed or their states are not identified for confidentiality reasons: Arkansas, Iowa, Mississippi, Montana, New Mexico,
North Dakota, South Dakota, West Virginia, and Wyoming.
15
I do not show summary statistics for the ethnicity of Medicaid recipients because
the CEX does not begin asking about Hispanic status until 2003.
16
Refer to Levy and DeLeire (2008) for a full consideration of spending patterns
by income with CEX data.
6
M. Dillender / Journal of Health Economics 53 (2017) 1–16
Table 3
Quarterly spending patterns in the CEX.
Spending categories
Health insurance
Medical care
Prescription drugs
Dental care
Food
Clothes
Housing
Transportation
Entertainment
Retirement
Other
Total
Incomes 200% of FPL or less
Incomes over 200% of FPL
Means
St. Devs.
Mean % of spending
Means
St. Devs.
Mean % of spending
223
118
39
40
1793
349
3603
1831
419
172
2480
11,069
536
493
157
238
1074
514
2659
4362
893
679
2912
8813
1.8%
0.8%
0.3%
0.3%
19.4%
3.4%
35.5%
12.6%
3.6%
1.3%
21.1%
544
297
81
118
2653
630
6801
3697
1138
578
7170
23,708
749
792
207
419
1524
1142
5011
6923
2163
2456
6358
16,127
2.7%
1.0%
0.4%
0.5%
12.9%
2.6%
30.5%
13.1%
4.6%
2.1%
29.6%
Notes: The data come from the 2000 to 2014 CEX. There are 30,752 observations with family incomes at 200 percent or less of the FPL and 62,029 observations with incomes
above 200 percent of the FPL. All expenditures are in 2012 dollars. The mean family size for families with incomes less than or equal to 200 percent of the FPL is 3.9, while
the mean family size for families with incomes greater than 200 percent of the FPL is 3.8. All of the means are statistically different from each other at the one-percent level
except for the mean percent of household spending on medical care.
the error term. The X vector includes family income as a percent
of the FPL, the square of family income as a percent of the FPL, an
indicator for the head of the household being male, indicators for
the race of the head of the household, the age of the head of the
household, the number of working adults in the household, and
the number of people in the household with high school degrees,
with some college, and with bachelors’ degrees.
I impute ELIG using family structure, family income, the FPL, and
the state’s income eligibility rules. The coefficient is the effect of
an additional family member becoming eligible for Medicaid on the
dependent variable. Because I focus on family measures and the
data are at the family level, this specification is slightly different
than much of the related literature that conducts analysis at the
individual level. This model is similar to Gruber and Simon (2008),
who consider family-level eligibility in addition to individual-level
eligibility and argue that family-level analysis is important since
health insurance decisions are made at the family level. I define
family type using the number of parents, the number of school-aged
children (ages 6 through 18), and the number of young children
(under age 6) in a family. With this set up, a family with two parents
and a young child, a family with two parents and a teenager, and
a family with one parent, one young child, and one teenager are
all different family types. This model accommodates the fact that
multiple people in a family can become eligible for Medicaid at
the same time. For instance, all two-parent families will have two
people become eligible for Medicaid at the same time if parental
Medicaid is expanded, and the family’s ELIG increases by two.17
Estimating Eq. (1) with OLS would yield biased estimates of
because Medicaid eligibility is endogenous as families with lower
incomes are more likely to be eligible for Medicaid. Since health
insurance and health care are normal goods, families with lower
incomes will likely consume less of them and therefore spend less
on them. Estimating Eq. (1) with OLS would falsely attribute this
lower ability to spend on health insurance and health care to cost
savings from Medicaid eligibility.
To overcome the endogeneity of Medicaid eligibility, I use a
simulated IV strategy to estimate the effect of an additional per-
17
Differential effects based on the number of people becoming eligible at the
same time is potentially an interesting source of heterogeneity. Unfortunately,
results from analysis that examines this heterogeneity are too imprecise to draw
meaningful conclusions about differences. Another potentially interesting source of
heterogeneity is the effect of an additional parent becoming eligible for Medicaid
versus the effect of an additional child becoming eligible for Medicaid. However,
there is generally not enough variation in child Medicaid eligibility for precise estimates of the effect of children separately.
son in a family becoming eligible for Medicaid. This approach was
pioneered by Cutler and Gruber (1996) and involves creating a
measure that represents a state’s Medicaid generosity that is correlated with Medicaid eligibility and that is correlated with the
outcome of interest only through its correlation with Medicaid eligibility. To compute the instrument, I apply the income eligibility
criteria in each state-year to the full sample for that year and calculate the average number of people in each family type that would
be eligible under that state’s Medicaid rules. Conditional on family type, states with more generous Medicaid eligibility rules will
have a higher value of the instrument than states with less generous Medicaid rules. Since I use the same sample to compute each
state’s instrument in a given year, variation in the instrument across
states comes solely from differences in Medicaid eligibility rules. I
estimate Eq. (1) with two-stage least squares (2SLS) and use the
simulated measure as an instrument for ELIG. The instrument is
highly correlated with the number of people in a family eligible
for Medicaid. The first stage coefficients on simulated eligibility
are close to one (0.95–0.99 dependent on specification), and the
F-statistics are over 1000.
While the simulated IV approach provides a useful way to
summarize complicated variation in Medicaid eligibility rules, the
exogeneity of the instrument relies on the exogeneity of the Medicaid eligibility rules. If states that expand Medicaid are changing
differently from states that do not in unobservable ways, then
the estimates from the simulated IV strategy may be biased.
For instance, if firms increase employees’ share of premiums for
spousal coverage over time, then states with more two-parent
households would see average health insurance spending increase.
If states with more two-parent households were also more likely
to expand Medicaid eligibility, then a failure to account for these
unobserved trends would incorrectly attribute the increase in
health insurance spending from rising employee premiums to Medicaid eligibility.
I take several steps to mitigate concerns about unobserved
trends. First, I allow each family type to have different state fixed
effects ( coefficients) and different year fixed effects (ı coefficients). Thus, identification of comes from how an outcome
changes after a change to eligibility relative to the initial differences for that family type in that state and relative to how outcomes
change for that family type over time. This approach addresses
the criticism of early use of the simulated IV strategy from ShoreSheppard (2008), who shows that the implicit assumption that
everyone has the same year effect and that all people in a state
have the same baseline state effects can lead to biased estimates.
M. Dillender / Journal of Health Economics 53 (2017) 1–16
Second, for each outcome, I estimate two specifications: one
set with minimal controls and another set with extensive controls.
The rationale for this approach is that many potential unobserved
trends would likely be correlated with observable characteristics
like unemployment and income levels. For example, suppose that
a state increases Medicaid eligibility because one of its main industries experienced increased competition from abroad, which led to
many people losing their jobs and health insurance over time. Since
the important, unobserved trend occurs at the state level, it would
not be captured by the year fixed effects. While this trend is unobservable to the researcher, it would likely be correlated with income
and unemployment, which means including income and unemployment rates as controls would alter the estimates because they
would be correlated with the error term. However, if the instrument really is exogenous, including the controls described above
should not alter the estimates.
I return to the issues of unobserved trends and of policy endogeneity in Section 4.4 to consider other ways to address the
possibility that unobserved trends confound the simulated IV strategy.
4. Results
4.1. The effect of Medicaid eligibility on health insurance coverage
4.1.1. Estimates from the CEX
I begin by showing estimates of the effect of an additional family
member becoming eligible for Medicaid on various health insurance measures from the CEX. Table 4 displays the results. Each cell
under the 2SLS estimates subheading is the effect of an additional
person becoming eligible for Medicaid from separate regressions
of Eq. (1) estimated with 2SLS. In the first column of estimates, the
regressions control for year fully interacted with family type and for
state fully interacted with family type. In the second column of estimates, I supplement the regressions with the various demographic
and labor force controls discussed earlier.
The first row of Table 4 considers the effect of an additional
person in a family becoming eligible for Medicaid on the number of people in the family with Medicaid. The estimated effect
is 0.135 when only state-by-family-type and year-by-family-type
fixed effects are included. When demographic and labor market
controls are included, the estimate is 0.142.
The remainder of Table 4 considers changes in private health
insurance coverage. In row 2, the dependent variable is an indicator variable equal to one if the family pays for any private health
insurance. The share of families with expenditures on private health
insurance falls by 4.4–4.6 percentage points when an additional
person becomes eligible for Medicaid. In row 3, the dependent variable is the number of private health insurance plans paid for by the
family. The results indicate that an additional person in the family
becoming eligible for Medicaid decreases the average number of
health insurance plans paid for by the family by 0.053–0.055. The
bulk of this decrease appears to come from a drop in the number
of employer-sponsored plans paid for by the family, which falls by
0.038–0.040. The estimated coefficients are statistically indistinguishable from zero when the dependent variable is the number of
privately purchased plans paid for by the family.
The last two rows of Table 4 consider the effect of an additional
person becoming eligible for Medicaid on the number of special
purpose health insurance plans paid for by the family. Special purpose health insurance plans include dental, vision, and prescription
drug coverage. The results provide no evidence that Medicaid eligibility affects the number of special purpose plans paid for by the
family. The last row of Table 4 focuses on dental coverage, which
is the most common special purpose plan, and shows no evidence
7
that an additional person becoming eligible for Medicaid affects the
number of dental plans paid for by the family.
4.1.2. Estimates from the March CPS and SIPP
The most commonly used data sets to estimate health insurance
changes from Medicaid expansions are the March CPS and SIPP. To
consider how estimates from the CEX compare to estimates from
the March CPS and the SIPP over this time period and with the states
available in the CEX, I next replicate the health insurance analysis
with both March CPS and SIPP data.
For the March CPS analysis, I use data from the 2001 to 2015
March CPS. As the March CPS asks about health insurance coverage in the previous year, these data provide information on health
insurance coverage from 2000 to 2014. I impute eligibility based
on the household’s income in the prior year and the state’s eligibility rules in the prior year. Although it is a common data set for
studying the health insurance effects of Medicaid, the March CPS
is not without problems. Partly because people may not be able
to remember their insurance from the previous year, as many ten
percent of Medicaid recipients in the previous year may not report
that they received Medicaid (DeNavas-Walt et al., 2013; Klerman
et al., 2009; Lewis et al., 1998). This Medicaid undercount can bias
estimates of the effect of Medicaid eligibility on take-up towards
zero. A related issue is that some people report having had multiple
sources of health insurance in the previous year. I code people based
on their responses, meaning that some people may be counted
as having had both private health insurance and Medicaid in the
previous year.
For the SIPP analysis, I use data from the 2001, 2004, and 2008
waves of the SIPP, which provide information from 2001 to 2013.
The SIPP interviews people every four months and asks about each
month since the previous interview. As Chetty (2008) finds that
people report the same responses during one interview for all four
months associated with the interview period, I restrict the sample
to observations from the interview month.
I aggregate information in both data sets to the family level so
that the three analysis data sets have similar structures. This aggregation involves calculating the number of people in each family
with various sources of health insurance as well as the number
of workers in the household and the number of people with high
school degrees, with some college, and with at least a bachelor’s
degree. I also control for the age, race, and gender of the head of
the household. As with the CEX analysis, I restrict the sample to
families making less than 200 percent of the FPL in the past year
and to the 42 states that are publicly identifiable in the CEX. I create the instruments as described earlier and then instrument for
the number of people in the household eligible for Medicaid using
2SLS.
Table 5 displays estimates of Eq. (1) using these two data sets.
Panel A displays the March CPS results, while Panel B displays the
SIPP results. The estimates suggest that an additional person in a
family becoming eligible for Medicaid increases the average number of people in the family with Medicaid by 0.126–0.151. The point
estimates for the fall in private coverage range from 0.028 to 0.096.
Only one is statistically significant.
Although the evidence for a fall in overall private coverage is
weak, the results show a clear reshuffling of insurance. The point
estimates for the change in employer-sponsored coverage range
from −0.095 to −0.053 and are statistically significant at at least the
ten-percent level in three specifications. This decrease appears to
come from both employer-sponsored health insurance as a dependent and as a policyholder. Neither data set provides evidence that
Medicaid eligibility affects the number of people in the family with
privately purchased health insurance.
8
M. Dillender / Journal of Health Economics 53 (2017) 1–16
Table 4
The effect of Medicaid eligibility on health insurance coverage: estimates from the CEX.
Dependent variables
Means
Number of people receiving Medicaid
0.992
Any private health insurance plans
0.316
Number of private health insurance plans
0.374
Number of employer-sponsored plans
0.300
Number of privately purchased plans
0.055
Number of non-employer-group plans
0.019
Number of special purpose plans
0.049
Number of private dental insurance plans
0.031
2SLS estimates
0.135**
(0.055)
−0.044***
(0.014)
−0.053**
(0.023)
−0.038**
(0.018)
−0.012
(0.011)
−0.003
(0.003)
−0.001
(0.010)
0.006
(0.005)
0.142***
(0.050)
−0.046***
(0.013)
−0.055***
(0.020)
−0.040**
(0.015)
−0.012
(0.011)
−0.003
(0.003)
−0.001
(0.009)
0.006
(0.005)
Additional controls
x
Notes: The data come from the 2000 to 2014 CEX. The sample contains 30,752 observations from 14,024 families. Each cell is the coefficient on the number of people in
a family eligible for Medicaid from Eq. (1). All regressions control for family-type-year and family-type-state fixed effects. The second column of estimates also includes
controls for family income as a percent of the FPL, the square of family income as a percent of the FPL, an indicator for the head of the household being male, indicators for
the race/ethnicity of the head of the household, the number of working adults in the family, the age of the head of the household, the annual state unemployment rate, and
the number of family members with high school degrees, with some college, and with bachelors’ degrees. Standard errors are clustered at the state level and are shown in
parentheses below the estimates.
**
Significance at 5%.
***
Significance at 1%.
Table 5
The effect of Medicaid eligibility on health insurance coverage: estimates from the March CPS and SIPP.
Dependent variables
Panel A: March CPS
Means
Number of people with Medicaid
1.62
Number of people with private coverage
1.37
Number of people with ESHI
1.03
Number of people with ESHI through their own employers
0.34
Number of people with ESHI as dependents
0.70
Number of people with privately purchased coverage
0.17
Additional controls
Panel B: SIPP
2SLS estimates
**
0.126
(0.051)
−0.028
(0.036)
−0.053
(0.035)
−0.015*
(0.008)
−0.041
(0.029)
0.004
(0.014)
***
0.133
(0.048)
−0.035
(0.032)
−0.060*
(0.030)
−0.018***
(0.006)
−0.046*
(0.026)
0.005
(0.015)
x
Means
2SLS estimates
1.45
0.151***
(0.045)
−0.096*
(0.055)
−0.095**
(0.041)
−0.034***
(0.012)
−0.068**
(0.033)
−0.001
(0.026)
1.47
1.30
0.46
0.86
0.17
0.142***
(0.042)
−0.071
(0.051)
−0.076*
(0.040)
−0.030**
(0.011)
−0.052
(0.033)
0.005
(0.024)
x
Notes: The data for the estimation in Panel A come from the 2001 to 2015 March CPS. The sample contains 128,385 observations. The data for the estimation in Panel B come
from the 2001, 2004, and 2008 SIPP. The sample contains 77,388 observations from 18,949 families. Each cell is the coefficient on the number of people in a family eligible
for Medicaid from Eq. (1). All regressions control for family-type-year and family-type-state fixed effects. The second column of estimates also includes controls for family
income as a percent of the FPL, the square of family income as a percent of the FPL, an indicator for the head of the household being male, indicators for the race/ethnicity
of the head of the household, the number of working adults in the household, the age of the head of the household, the annual state unemployment rate, and the number
of family members with high school degrees, with some college, and with bachelors’ degrees. Standard errors are clustered at the state level and are shown in parentheses
below the estimates.
*
Significance at 10%.
**
Significance at 5%.
***
Significance at 1%.
4.1.3. Comparison of estimates
All three sets of estimates are generally consistent with each
other and with the previous literature. The point estimates of the
increase in Medicaid coverage from an additional person becoming eligible for Medicaid in the CEX fall within the bounds of the
estimates from the March CPS and SIPP, and none are statistically
different from each other. As these estimates are at the family level,
they are comparable to Gruber and Simon (2008). Their estimates of
the effect of an additional person in the family becoming eligible for
Medicaid on the number of people with Medicaid coverage range
from 0.109 to 0.156, which encompass the estimates presented in
this paper.
The point estimates from my study for the fall in employersponsored insurance from the March CPS and SIPP range from 0.053
to 0.096, while the point estimates in Gruber and Simon (2008)
range from 0.066 to 0.121. Although these results for employersponsored health insurance are not directly comparable to the CEX
results since the CEX results are about plans that may cover multiple people, they suggest similar patterns among the three data
sets.
The results support the idea that there is more crowd out from
more recent expansions for parents than from earlier expansions as
suggested by the results of Busch and Duchovny (2005), Hamersma
M. Dillender / Journal of Health Economics 53 (2017) 1–16
and Kim (2013), and McMorrow et al. (2016).18 One possible reason
that crowd-out could be higher in more recent years is that firms
have pushed more of the cost of health insurance onto employees as health insurance costs have risen (Kaiser Family Foundation,
2016a,b). In short, the results about the effect of Medicaid expansions on health insurance coverage found in the CEX are largely
corroborated by the March CPS and the SIPP and are generally consistent with other studies.
4.2. The effect of Medicaid eligibility on spending on health
insurance and medical care
I next consider the implications of these health insurance
changes for household spending on medical care and on health
insurance. Table 6 displays 2SLS estimates of the effect of an additional person becoming eligible for Medicaid from Eq. (1) on various
spending dependent variables.
Row 1 displays estimates of the effect on medical spending variables. The estimates imply that an additional person in a family
becoming eligible for Medicaid reduces the likelihood that a family has any medical spending by 2.7 percentage points. Much of
this decrease appears to come from families with positive quarterly
spending that is less than $100, which falls by about 2 percentage
points. I do not find evidence that Medicaid reduces the likelihood
that families have medical spending over $100 per quarter, nor do
I find any evidence that Medicaid has impacts on average medical
spending, though I should note that I cannot rule out meaningful
effects for either outcome.19
Row 2 of Table 6 considers the effect of an additional person becoming eligible for Medicaid on prescription drug spending
and displays some evidence that Medicaid reduces the likelihood
that families have positive but small expenditures on prescription
drugs. Row 3 considers the effect of Medicaid on dental spending.
I find no evidence that an additional family member being eligible
for Medicaid affects dental spending. This null result may arise with
dental spending because dental spending is more discretionary and
families might not visit the dentist without dental insurance.20
Row 4 of Table 6 considers how spending on health insurance
changes as a result of an additional person in a family becoming eligible for Medicaid. The estimates suggest that the likelihood that
families have any spending on health insurance falls by 4.3 percentage points when an additional family member becomes eligible for
Medicaid. This decrease in the number of families with any health
insurance spending does not appear to come from families that
were spending small amounts on health insurance coverage as the
percentage of families with positive spending less than $100 does
not change. Instead, all of the increase in families with no spending
appears to come from families that were spending over $100 per
18
In results available upon request, I have redone the SIPP and CPS analysis focusing exclusively on the parental expansions and using the same years of data as Busch
and Duchovny (2005) and Hamersma and Kim (2013) and find similar results as they
do, which suggests that the different time periods and expansions are why we find
different crowd-out results rather than differences in our approaches.
19
Estimating the effect on average medical spending is complicated by medical
spending’s severe right skew. Likely for this reason, other research (e.g. Finkelstein
et al., 2012) often does not consider average effects and instead considers indicators for any medical spending or for any medical bills. I have also considered other
methods, such as estimating quantile IV regression models, but these models do
not converge, likely due to the large number of fixed effects. Therefore, I report the
estimated effects on average spending but focus the discussion on indicators for
categories of medical spending because of the issues with studying average medical
spending.
20
Although all states provide preventive dental coverage for children, many states
only provide emergency dental services for adults. When I restrict the sample to
states that provide dental coverage to adults during this time period (as defined in
Decker and Lipton, 2015), I still find no evidence of an effect of an additional family
member becoming eligible for Medicaid on dental spending.
9
quarter. Average per quarter spending on health insurance coverage falls by about $47.
Medicaid allowing families to spend less on private health insurance coverage is important. A concern with Medicaid crowding out
private coverage is that Medicaid could potentially not be helping
low-income people as was originally intended. Instead, Medicaid
could mean employers have to pay less in fringe benefits. However,
these results suggest that low-income families that were relying on
employer-sponsored coverage were doing so at a large cost.21 Since
these estimates imply that the coverage that was crowded was high
quality, they also suggest that the reduced medical spending likely
comes from the previously uninsured.
Table 7 considers the effect of Medicaid eligibility on medical
debt. Medicaid has the potential to affect medical debt since families may not pay medical bills immediately. The CEX only asks about
debt in certain interviews, so the sample sizes are much smaller,
which results in noise in the estimates. I do not find strong evidence that Medicaid eligibility affects the likelihood of having any
debt. However, it is important to note that I cannot rule out sizable
impacts on many of the medical debt outcomes.
4.3. Robustness
I next verify the robustness of the key health insurance and
spending results to various data and estimation choices. Column
1 of Table 8 displays the original estimates with state-by-familytype indicators, year-by-family-type indicators, and demographic
and labor market controls.22
The main results use reported income and drop families with
people who worked but had no income from the sample. I now consider the sensitivity of the results to three alternatives. Column 2 of
Table 8 displays results that no longer drop families with workers
but no income. Column 3 displays results that use imputed income
rather than reported income when imputed income is available
beginning in 2004. Column 4 displays results that drop workers
who had their incomes imputed once the CEX began imputing data
as well as families with workers but no incomes. In all cases, the
results are qualitatively similar and generally maintain statistical
significance.
For the main analysis, I generate the simulated measure of eligibility by applying the income eligibility criteria in each state-year
to the CEX sample for that year. I now consider two alternative
methods of constructing the instrument. As the March CPS income
measure has advantages over the CEX measure, such as a larger
sample size and more comparability over time (BLS, 2016), I next
construct the instrument and eligibility measure using March CPS
data rather than CEX data. Specifically, I use the March CPS to compute the average number of family members eligible for Medicaid
as well as the simulated instrument for each family type, year, and
state combination. I then merge these measures to the CEX on family type, year, and state and replicate the CEX analysis using average
eligibility and the simulated instrument from the March CPS.23
The results are shown in Column 5 of Table 8. Another alternative
method of constructing the instrument is to apply the income eligi-
21
Note that the size of the estimates do not rule out that employers could have
reduced health insurance expenses because of crowd-out. Any reduced health insurance expenditures for firms arising from crowd-out may be profit for firms or may
be passed onto workers in a variety of ways.
22
Although Tables 8 and 9 focus on select results for conciseness, the omitted
results are qualitatively similar to the main results presented in the paper as well
and are available from the author upon request.
23
This strategy follows Bronchetti (2014), who studies how the effect of Medicaid
eligibility on health care access differs for immigrants and natives using outcomes
and demographic characteristics from the National Health Interview Surveys but
the instrument and eligibility measure from the March CPS.
10
M. Dillender / Journal of Health Economics 53 (2017) 1–16
Table 6
The effect of Medicaid eligibility on health-related spending.
Spending categories
Any spending
Spending positive but less than $100 Spending at least $100
Means
2SLS estimates
Means
2SLS estimates
Means
2SLS estimates
Means
2SLS estimates
Medical
0.312
0.131
Dental
0.099
Health insurance
0.355
−0.008
(0.010)
0.001
(0.012)
−0.012
(0.007)
−0.043***
(0.014)
118
0.261
−0.020**
(0.009)
−0.019**
(0.009)
0.000
(0.005)
−0.000
(0.004)
0.181
Prescription
−0.027**
(0.013)
−0.018
(0.014)
−0.011
(0.008)
−0.043***
(0.014)
11
(9)
−6
(6)
−2
(5)
−47**
(19)
Additional controls
−0.027**
(0.011)
−0.016
(0.013)
−0.011
(0.008)
−0.043***
(0.012)
0.160
0.031
0.038
x
−0.019**
(0.008)
−0.019**
(0.009)
−0.000
(0.005)
−0.000
(0.005)
0.101
0.069
0.317
x
Average spending
−0.008
(0.010)
0.002
(0.011)
−0.011
(0.007)
−0.043***
(0.011)
x
39
40
223
11
(10)
−5
(5)
−2
(5)
−47***
(17)
x
Notes: The data come from the 2000 to 2014 CEX. The sample contains 30,752 observations from 14,024 families. Each cell is the coefficient on the number of people in
a family eligible for Medicaid from Eq. (1). All regressions control for family-type-year and family-type-state fixed effects. The second column of estimates under each
subheading also includes controls for family income as a percent of the FPL, the square of family income as a percent of the FPL, an indicator for the head of the household
being male, indicators for the race/ethnicity of the head of the household, the number of working adults in the family, the age of the head of the household, the annual state
unemployment rate, and the number of family members with high school degrees, with some college, and with bachelors’ degrees. Standard errors are clustered at the state
level and are shown in parentheses below the estimates.
**
Significance at 5%.
***
Significance at 1%.
Table 7
The effect of Medicaid eligibility on medical debt.
Dependent variables
Mean
Has medical debt
0.064
Number of medical debts
0.072
Amount of medical debt
379
Medical debt more than $100
0.057
Medical debt more than $1000
0.030
Medical debt more than $10,000
0.005
Additional controls
2SLS estimates
0.011
(0.014)
0.003
(0.013)
−119
(425)
0.015
(0.013)
0.014
(0.011)
0.003
(0.005)
0.010
(0.014)
0.002
(0.013)
−122
(414)
0.014
(0.012)
0.013
(0.011)
0.003
(0.004)
x
Notes: The data come from the 2000 to 2014 CEX. The sample contains 6049 families/observations. Each cell is the coefficient on the number of people in a family eligible
for Medicaid from Eq. (1). All regressions control for family-type-year and family-type-state fixed effects. The second column of estimates also includes controls for family
income as a percent of the FPL, the square of family income as a percent of the FPL, an indicator for the head of the household being male, indicators for the race/ethnicity of
the head of the household, the number of working adults in the family, the age of the head of the household, the annual state unemployment rate, and the number of family
members with high school degrees, with some college, and with bachelors’ degrees. Standard errors are clustered at the state level and are shown in parentheses below the
estimates.
bility criteria in each state-year to the entire sample for all years.24
The results are shown in column 6 of Table 8. With both alternative instrument definitions, the results are similar to the original
estimates.
The main analysis did not weight the estimation using the CEX
weights, which are designed to make the CEX representative of
regions of the United States. In column 7, I weight the regressions
using the CEX weights. Again, the results are similar regardless of
whether or not weights are used.
The CEX contains multiple observations per family. In column
8, I consider the robustness of the results to collapsing the data to
one observation per family by producing quarterly means of the
outcomes. I use eligibility and characteristics in the most recent
interview, meaning eligibility and the controls are imperfectly measured. Despite the measurement error, the results remain similar.
24
Using a fixed sample each year to compute the instrument is valid since the year
coefficients control for national changes in income over time, but the results not
being robust to using a fixed sample across years would still be troubling. Examples
of studies that use a fixed sample across years include Gross and Notowidigdo (2011)
and Shore-Sheppard (2008), while examples of studies that use different samples
for each year include Bronchetti (2014) and Gruber and Simon (2008)
In the main analysis, I restricted the sample to families with
incomes under 200 percent of the FPL because most of the Medicaid expansions during the 2000s were for parents making under
200 percent of the FPL. Since the CEX does not contain nearly as
many observations as the CPS and SIPP, focusing narrowly on the
affected population makes the estimates more precise. Nevertheless, we would be concerned if the results change drastically with
slight sample alternations. In column 9 of Table 8, I broaden the
sample to include families under 250 percent of the FPL. Although
a few of the coefficients are no longer statistically significant at conventional levels, the point estimates are similar and the results are
qualitatively the same.
An alternative to using a simulated IV strategy is to include the
eligibility threshold as an independent variable in OLS regressions.
Instead of providing an estimate of the effect of a marginal person
becoming eligible for Medicaid, this approach provides an estimate
of the effect of a change in the income threshold. As Hamersma and
Kim (2013) explain, the estimate of the effect of the income threshold is appealing since the threshold is a major policy instrument
that legislators can control. Since most of the identifying variation comes from parental eligibility, I focus on the effects of the
parental eligibility threshold. The results are shown in column 9
and tell a similar story as the original estimates. An estimate with
Table 8
Robustness of select CEX results.
(1)
Original
estimates
(4)
(5)
Using CPS for
Dropping
families with
instrument
imputed incomes
(6)
Using fixed
sample for
instrument
(7)
Using CEX
weights
(8)
(9)
Collapsing to one Broaden sample
observation per to 250 percent
family
FPL
(10)
income threshold
as independent
variable
0.169***
(0.058)
−0.057***
(0.020)
−0.026*
(0.015)
−0.011
(0.012)
−0.015
(0.013)
16
(16)
−0.044***
(0.015)
−0.006
(0.006)
−0.039**
(0.015)
−41**
(19)
0.165***
(0.059)
−0.046***
(0.015)
−0.027**
(0.013)
−0.022**
(0.008)
−0.005
(0.010)
−1
(10)
−0.043***
(0.013)
−0.003
(0.005)
−0.040***
(0.012)
−28**
(11)
0.169***
(0.049)
−0.056**
(0.022)
−0.027**
(0.012)
−0.021**
(0.009)
−0.006
(0.010)
12
(10)
−0.047***
(0.012)
0.002
(0.005)
−0.048***
(0.012)
−50***
(17)
0.155***
(0.050)
−0.051*
(0.026)
−0.031**
(0.015)
−0.016**
(0.007)
−0.015
(0.012)
2
(11)
−0.045***
(0.015)
−0.001
(0.005)
−0.044***
(0.014)
−52***
(19)
0.138***
(0.050)
−0.038*
(0.021)
−0.025
(0.017)
−0.023*
(0.013)
−0.002
(0.013)
−4
(14)
−0.039***
(0.014)
0.006
(0.005)
−0.045***
(0.016)
−51***
(18)
0.175***
(0.045)
−0.057***
(0.016)
−0.025*
(0.014)
−0.017
(0.012)
−0.009
(0.009)
6
(11)
−0.049***
(0.010)
0.003
(0.005)
−0.052***
(0.009)
−53***
(17)
0.137***
(0.048)
−0.050**
(0.019)
−0.023**
(0.011)
−0.020**
(0.008)
−0.003
(0.009)
5
(10)
−0.037***
(0.011)
−0.001
(0.005)
−0.036***
(0.010)
−32***
(12)
0.146***
(0.043)
−0.045**
(0.018)
−0.017
(0.011)
−0.016**
(0.008)
−0.001
(0.009)
17*
(9)
−0.037***
(0.009)
0.001
(0.004)
−0.038***
(0.009)
−41***
(14)
0.159***
(0.046)
−0.027
(0.023)
−0.025
(0.015)
−0.011
(0.010)
−0.014
(0.013)
4
(12)
−0.031**
(0.014)
0.006
(0.005)
−0.038***
(0.014)
−38*
(19)
Notes: The data come from the 2000 to 2014 CEX. In columns 1, 5, 6, 7, and 10 the sample contains 30,752 observations from 14,024 families. In column 2, the sample contains 41,509 observations from 18,448 families. In column
3, the sample contains 28,722 observations from 13,179 families. In column 4, the sample contains 22,170 observations from 10,526 families. In column 8, the sample contains one observation for each of the 14,024 families in
the main sample. In column 9, the sample contains 38,322 observations from 16,904 families. Each cell in columns 1 through 9 is the coefficient on the number of people in a family eligible for Medicaid from Eq. (1). Each cell in
column 10 is the coefficient on the parental Medicaid eligibility threshold as a fraction of the FPL. All regressions control for family-type-year fixed effects, family-type-state fixed effects, family income as a percent of the FPL,
the square of family income as a percent of the FPL, an indicator for the head of the household being male, indicators for the race/ethnicity of the head of the household, the number of working adults in the family, the age of the
head of the household, the annual state unemployment rate, and the number of family members with high school degrees, with some college, and with bachelors’ degrees. Standard errors are clustered at the state level and are
shown in parentheses below the estimates.
*
Significance at 10%.
**
Significance at 5%.
***
Significance at 1%.
M. Dillender / Journal of Health Economics 53 (2017) 1–16
0.142***
(0.050)
Number of private health insurance plans
−0.055***
(0.020)
Any medical spending
−0.027**
(0.011)
Medical spending positive but less than $100 −0.019**
(0.008)
−0.008
Medical spending on at least $100
(0.010)
11
Average medical spending
(10)
−0.043***
Any health insurance spending
(0.012)
−0.000
Health insurance spending positive
(0.005)
But less than $100
Health insurance spending on at least $100 −0.043***
(0.011)
−47***
Average health insurance spending
(17)
Number of people receiving Medicaid
(2)
(3)
Keeping families Using imputed
with workers but income
no income
11
12
M. Dillender / Journal of Health Economics 53 (2017) 1–16
this approach can be interpreted as the effect of increasing the eligibility threshold by 100 percentage points. This approach implies
that increasing the threshold by 100 percentage points increases
the number of people receiving Medicaid in an average family by
0.175 and lowers the average number of private plans that a family
has by 0.057.
4.4. Revisiting assumption of parallel trends
An important assumption of the analysis is that states that alter
their Medicaid eligibility rules would have trended similarly to
states that did not if not for the changes to the eligibility rules. One
way to relax this assumption that has often been used in individuallevel analysis has been to take advantage of the fact that Medicaid
expansions do not always apply to all ages. For instance, many
CHIP expansions commonly set different eligibility rules for young
children and for older children. Researchers can include state-year
fixed effects and identify the effects of Medicaid eligibility as how
outcomes of children at the affected ages change after Medicaid
expansions relative to how outcomes change for children at other
ages within the state. Studies of parental eligibility do not typically
include state-year fixed effects, presumably because the expansions do not vary by age for adults. With family-level analysis,
including state-year fixed effects is further complicated by the fact
that parents and children of all ages share a household budget, so an
increase to parental eligibility increases the value of the instrument
for all observations in a state. In principle, this study could take
advantage of the fact that differences in family structure mean that
raising the parental eligibility threshold increases the number of
people eligible for Medicaid in a newly eligible two-parent household by two and the number of people eligible for Medicaid in a
newly eligible one-parent household by one. In practice, there is
not enough family-type variation within state-year cells in the CEX
to identify the effect of an additional person becoming eligible for
Medicaid in this way.
To adapt the strategy to identify the effect of Medicaid eligibility using within-state variation, I include wealthier families in
the sample since they would presumably not be directly affected
by Medicaid expansions. Specifically, for each state, I calculate the
highest Medicaid threshold during the time period and include all
families with incomes above that amount in the sample. Since Medicaid expansions should not affect wealthier families, I include the
simulated eligibility measure as a control and use the simulated
eligibility measure interacted with an indicator for being in the
original, low-income sample as the instrument. I allow the state and
year effects to vary for low-income and high-income families. With
this broadened sample, I can now supplement Eq. (1) with separate
year-by-state fixed effects for each family type, meaning that identification of the effect of an additional person becoming eligible for
Medicaid comes from how outcomes change for low-income families after a state expands Medicaid relative to how they change for
high-income families of the same type in the state. This approach
is similar in spirit to Hamersma and Kim (2013), who perform a
placebo test using a sample of high-income individuals.
The results are shown in column 1 of Table 9. Overall, the results
are less precise when identifying the effect on an additional person
becoming eligible based on within state changes. The point estimate for the effect on Medicaid coverage is smaller, but the estimate
is statistically significantly different from zero and not statistically
different from the original estimate. Some of the other estimates
are no longer statistically significant from zero, but they are qualitatively similar to the original estimates and never statistically
different from them.
The results from using the within-state control group do not
provide evidence against the main empirical strategy but are not
precise enough to completely allay concerns about unobserved
trends. Another issue with this approach is that it can only account
for state-level trends that affect all families in a state. This strategy
would not account for trends that only low-income families experience. Given that Medicaid is targeted at low-income people, it is
not unreasonable to worry that different trends in health insurance
for low-income people may cause states to change their Medicaid
eligibility rules.
Another way to relax the parallel trends assumption is to supplement Eq. (1) with state-specific time trends for each state. If the
estimates change dramatically, we would be concerned that differential trends prior to changes in eligibility rules drive the results.
Estimates of Eq. (1) supplemented with linear state-specific time
trends are shown in the second column of Table 9. The estimates
are similar to the original estimates, which is reassuring that state
trends do not drive the results.
A more direct test of pre-existing trends is to examine if and
how the outcome variables change immediately prior to changes
in Medicaid eligibility. To do this, I supplement Eq. (1) with an indicator variable that is equal to one in the year before a state changes
its Medicaid thresholds. Significant coefficients on this indicator
variable would raise concerns about pre-existing trends. The coefficients on the indicator variable are shown in column 3 of Table 9 and
are statistically indistinguishable from zero in all cases. To restrict
the focus to larger changes, I next create an indicator equal to one
in the year before eligibility increases by more than 10 percentage
points of the FPL. Again, I find no evidence of pre-existing trends
that are not accounted for by the empirical strategy.
4.5. LATE estimates and estimated effects of treatment on the
treated
The main analysis focuses on estimating the effect of an additional person in a family being eligible for Medicaid rather than on
the effect of an additional person signing up for Medicaid coverage for two main reasons. First, Medicaid take-up is endogenous.
Sicker, more risk-averse, and more financially savvy families may
be more likely to take up Medicaid conditional on being eligible. Eligibility rules, on the other hand, are decided by policy. The
focus on the effect of eligibility means that the main estimates
are of the average intent-to-treat effect, which is a policy-relevant
parameter since policymakers can influence eligibility more easily than they can influence take-up. Second, it is possible that
additional people receiving Medicaid may have effects on the nonMedicaid population. For instance, if Medicaid expansions crowd
out private health insurance for sick people, then the Medicaid
expansions would lower health risks in the private market, which
could lower private health insurance premiums. Alternatively, if
Medicaid increases access to care, waiting times for non-Medicaid
patients may increase, which could lower spending on health care
by reducing access to the health care system. Medicaid expansions
having broader effects than just on those who sign up would bias
estimates that assume Medicaid only affects those who sign up for
Medicaid.
Despite the appeal of the intent-to-treat estimates, comparing
intent-to-treat estimates across studies can be difficult. To facilitate
comparisons of my estimates to other estimates and to provide an
alternative way to assess the plausibility of the estimates, I compute the average effect of the treatment on the treated, which is
equivalent to the local average treatment effect (LATE). As these
estimates require the critical assumption that Medicaid eligibility
only impacts those who take up Medicaid, which may be untenable,
they should be taken with caution.
To calculate the LATE estimates, I replace ELIG in Eq. (1) with
the number of people in a family receiving Medicaid and estimate
the model using 2SLS. As with the main analysis, the instrument
is the simulated measure of eligibility. Table 10 displays the LATE
M. Dillender / Journal of Health Economics 53 (2017) 1–16
13
Table 9
Considering the possibility of pre-existing trends.
Number of people receiving Medicaid
Number of private health insurance plans
Any medical spending
Medical spending positive but less than $100
Medical spending on at least $100
Average medical spending
Any health insurance spending
Health insurance spending positive but less than $100
Health insurance spending on at least $100
Average health insurance spending
Including within
state control
group
State-specific
time trends
Year before
eligibility change
Year before eligibility increase >10
percentage points
0.101**
(0.041)
−0.065
(0.039)
−0.010
(0.018)
−0.015
(0.013)
0.005
(0.013)
1
(25)
−0.035*
(0.018)
0.007
(0.009)
−0.042*
(0.021)
−66**
(31)
0.091*
(0.052)
−0.062***
(0.016)
−0.028**
(0.011)
−0.023**
(0.009)
−0.005
(0.009)
12
(11)
−0.039***
(0.012)
−0.003
(0.006)
−0.036***
(0.012)
−45***
(16)
0.013
(0.021)
0.004
(0.009)
0.007
(0.007)
−0.001
(0.005)
0.008
(0.007)
10
(11)
0.006
(0.009)
0.000
(0.004)
0.006
(0.007)
11
(11)
−0.016
(0.028)
0.014
(0.015)
−0.018
(0.012)
−0.009
(0.007)
−0.009
(0.009)
−5
(13)
0.006
(0.010)
−0.004
(0.004)
0.010
(0.009)
1
(14)
Notes: The data come from the 2000 to 2014 CEX. In column 1, the sample contains 75,549 observations from 29,875 families. In columns 2, 3, and 4, the sample contains
30,752 observations from 14,024 families. Each cell in columns 1 and 2 is the coefficient on the number of people in a family eligible for Medicaid from Eq. (1). Each cell
in column 3 is the coefficient on an indicator variable equal to one the year before Medicaid eligibility increases. Each cell in column 4 is the coefficient on an indicator
variable equal to one the year before Medicaid eligibility changes by more than 10 percentage points. All regressions in columns 2, 3 and 4 control for family-type-year fixed
effects, family-type-state fixed effects, family income as a percent of the FPL, the square of family income as a percent of the FPL, an indicator for the head of the household
being male, indicators for the race/ethnicity of the head of the household, the number of working adults in the family, the age of the head of the household, the annual
state unemployment rate, and the number of family members with high school degrees, with some college, and with bachelors’ degrees. All regressions in column 1 include
family-type-income-group-year, family-type-income-group-state, and family-type-year-state fixed effects rather than family-type-year and family-type-state fixed effects,
while all regressions in column 2 also control for state-specific linear time trends. Standard errors are clustered at the state level and are shown in parentheses below the
estimates.
*
Significance at 10%.
**
Significance at 5%.
***
Significance at 1%.
estimates for selected outcomes. The LATE estimates imply that an
additional person taking up Medicaid lowers the number of private
health insurance plans that a family pays for by 39 percent and
decreases the likelihood that a family has any medical spending
in a quarter by 19.3–20.4 percentage points. An additional person
taking up Medicaid reduces the likelihood that a family has any
spending on health insurance by 30.6–31.8 percentage points and
saves families $333–347 per quarter on health insurance.
The LATE can also be formulated as the effect of switching from
private coverage to Medicaid for the health insurance spending
variables. For this calculation, I replace ELIG in Eq. (1) with an indicator for whether or not the family pays for any private coverage.
These estimates require the even stronger assumption that Medicaid eligibility affects health insurance spending only for families
who experience crowd-out. The estimates suggest that Medicaid
saves families who experience crowd-out $1031–1071 on quarterly
health insurance spending.25
5. Discussion and conclusions
Medicaid has goals of expanding health care access and of reducing out-of-pocket expenditure risk for low-income families. While
many studies have explored the effect of Medicaid on health care
access, much less research has considered the effects on household
expenditures. In this paper, I examined how Medicaid eligibility
25
These results are not shown in Table 10. The first estimate ($1031) includes
demographic and labor market controls, while the second ($1071) does not. They
are statistically significant at the one-percent level with standard errors of 290 and
294.
affects families’ health insurance coverage and spending patterns
using CEX data.
My estimates suggest that an additional family member being
eligible for Medicaid reduces the likelihood that a family has any
medical spending by 2.7 percentage points. Although much of this
decrease appears to come from families who had small medical expenditures, I cannot rule out meaningful decreases in large
expenditure risks. These results are most comparable to Finkelstein
et al. (2012), who study the effect of the Oregon lottery on the likelihood that non-parents have any medical spending. In their study,
they calculate a LATE estimate for Medicaid coverage on the likelihood of having any medical spending in the past six months of
−0.200. The estimates from this paper suggest a LATE estimate for
an additional family member having Medicaid coverage on the likelihood of having any medical spending in the past three months
of −0.204 to −0.193. Our estimates are not directly comparable
because my estimates describe quarterly spending for a family,
while theirs describe six-month spending for adults without children. However, despite the different samples and methodologies,
the LATE estimates are similar, suggesting the results from the
Oregon health insurance experiment may generalize to the larger
population.
This study is one of the first to consider the financial implications of the crowd-out of private coverage. The analysis provides
evidence that crowd-out occurs, which is a common but not
ubiquitous finding in this literature. An important and unique finding from this paper is that crowd-out saves families money on
health insurance spending. Assuming that only families who switch
from private coverage experience reduced health insurance expenditures implies that these families save $4124–4284 on health
insurance per year on average, which is similar to the average family contribution for employer-sponsored health insurance coverage
14
M. Dillender / Journal of Health Economics 53 (2017) 1–16
Table 10
LATE estimates of the effect of Medicaid.
Number of private health
insurance plans
Any medical spending
Medical spending positive but
less than $100
Medical spending of at least
$100
Average medical spending
Any health insurance spending
Health insurance spending
positive but less than $100
Health insurance spending of
at least $100
Average health insurance
spending
Additional controls
−0.394**
−0.387**
(0.184)
−0.204*
(0.105)
−0.146
(0.160)
−0.193**
(0.090)
−0.136*
(0.088)
−0.058
(0.079)
−0.058
(0.068)
79
(81)
−0.318**
(0.127)
−0.000
(0.064)
75
(79)
−0.306***
(0.113)
−0.002
(0.033)
−0.318**
(0.032)
−0.304**
(0.128)
−347**
(0.113)
−333**
(145)
(134)
x
Notes: The data come from the 2000 to 2014 CEX. The sample contains 30,752 observations from 14,024 families. Each cell is the LATE effect of an additional person in
the family enrolling in Medicaid. All regressions control for family-type-year and
family-type-state fixed effects. The second column of estimates also includes controls for family income as a percent of the FPL, the square of family income as a
percent of the FPL, an indicator for the head of the household being male, indicators
for the race/ethnicity of the head of the household, the number of working adults
in the family, the age of the head of the household, the annual state unemployment
rate, and the number of family members with high school degrees, with some college, and with bachelors’ degrees. Standard errors are clustered at the state level
and are shown in parentheses below the estimates.
*
Significance at 10%.
**
Significance at 5%.
***
Significance at 1%.
of $4316 in 2012 (Claxton et al., 2012). This crowded out insurance
being high value suggests that the medical spending effects are
likely driven by the previously uninsured.
Some have suggested that crowd-out is a reason to limit Medicaid eligibility (Cannon, 2005; United States Senate Republican
Policy Committee, 2015). But the results from this study highlight two important aspects about crowd-out arising from Medicaid
and about the spending effects of Medicaid that policymakers
should keep in mind. First, even though crowd-out occurs, Medicaid expansions still appear to lower medical spending for a
sizable portion of recipients, suggesting that the previously uninsured or underinsured experience welfare increases from Medicaid.
Second, the crowd-out from Medicaid allows families to spend
less on health insurance. As low-income families likely consume
where the marginal utility of consumption is high, this transfer
can result in overall welfare improvements. In short, Medicaid
realizes its intended effect of reducing medical expenditures for
low-income families, but focusing solely on Medicaid’s effect on
medical expenditures ignores a major part of Medicaid’s contribution to low-income families’ financial health.
Appendix A. The model for the conceptual framework
In this appendix, I provide a fuller discussion of the welfare
implications for the family from crowd-out. As described in Section
2, the welfare implications of crowd-out that arises from Medicaid expansions depend on the quality of private coverage that is
crowded out and the share the family was paying for the private
coverage. The following simple framework illustrates the possible
welfare changes that may be associated with crowd-out.
Suppose that a family has utility u(c) that is an increasing function of non-health consumption c. The family’s spending must
satisfy the budget constraint c = y − m − P, where y is income, m is
out-of-pocket medical expenditures, and P is the premium spent on
private health insurance. Both P and y are known in advance, while
m is a random variable with probability density function f(m) and
support [0, m̄].26 Suppose that private insurance is actuarially fair
and covers all expenditure risk, meaning that
m̄
P=
f (m)dm.
0
If the family purchases private insurance, its utility equals
u(y − ıP),
where ı ∈ [0,1] is the family’s portion of the premium.27
This simple framework allows for describing the utility available
to the family in each of the three pre-Medicaid eligibility cases from
Section 2. For simplicity, I assume the price and coverage for poor
health insurance are approximately zero, meaning that the family
would essentially be uninsured even though it shows up in survey
data sets as having private insurance. The utility for each original
case from Section 2 can be summarized as follows:
1. The family was paying very little for good coverage (ı=0): u(y)
2. The family was paying full price for good coverage (ı=1): u(y − P)
3. The family was paying very little for poor coverage (equivalent
to not purchasing insurance):
m̄
0
u(y − m)f (m)dm
Assuming that Medicaid covers all expenditure risks, then the
family’s utility is u(y) with Medicaid. Therefore, the utility gains
from switching to Medicaid are
1. The family was paying very little for good coverage:
u(y) − u(y) = 0
2. The family was paying full price for good coverage: u(y) − u(y − P)
3. The family was paying very little for poor coverage: u(y) −
m̄
0
u(y − m)f (m)dm
Utility being an increasing function of consumption means both
m̄
u(y) − u(y − P) and u(y) − 0 u(y − m)f (m)dm are positive. Thus, the
crowd-out of (2) and (3) make the family better off, while crowdout of (1) does not.
The availability of actuarially fair insurance and the fam-
m̄
ily having a concave utility function mean that u(y) − 0 u(y −
m)f (m)dm > u(y) − u(y − P), which would imply that crowd-out
of (3) is more welfare improving than the crowd-out of (2). Of
course, these assumptions would also imply that low-income families always purchase insurance, which is clearly not the case.
26
This simple set up draws on Finkelstein and McKnight (2008), who consider the
welfare effects of Medicare.
27
The determinants of ı are beyond the scope of this paper. As described in Section
2, if workers value money spent on health insurance as much as they would value
the money itself, if employers can perfectly identify which people are eligible for
Medicaid, and if employers can perfectly and instantly adjust compensation, theory
predicts that employers will pass the costs of health insurance to employees in the
form of lower wages (Gruber, 1994; Summers, 1989). Therefore, Medicaid expansions have the potential to increase wages if employers no longer have to provide
workers with health insurance coverage. However, these assumptions may not hold
for a variety of reasons. For example, identifying Medicaid-eligible people is likely
difficult since employers would have to know total family income. Furthermore,
Finkelstein et al. (2015) find that the Oregon health insurance experiment’s Medicaid recipients only value Medicaid at 20 to 40 percent of Medicaid’s cost, suggesting
that many low-income workers may not fully value health insurance.
M. Dillender / Journal of Health Economics 53 (2017) 1–16
Families likely do not purchase health insurance because insurance
is not actuarially fair in reality, especially since some low-income
families may not pay emergency room medical bills, meaning they
do not bear all of the costs of their negative health shocks if they are
uninsured. The empirical result of Medicaid eligibility causing both
private coverage and health insurance spending to fall would imply
that crowd-out of (2) has occurred. The empirical result of Medicaid eligibility causing private coverage but not health insurance
spending to decrease, means crowd-out of (1) or (3) occurred.
Another point to note is that concavity means that
15
Table B.1
The effect of Medicaid eligibility on non-health spending, participation in other
programs, and labor market outcomes.
Dependent variables
Means
Panel A: other spending
Spending on food
1793
Spending on clothes
349
Spending on housing
3603
u(y − m)f (m)dm >
Spending on transportation
1831
u(y ) − 0 u(y − m)f (m)dm for y > y. In other words, the less
income a family has, the more welfare improving crowd-out of (2)
or (3) is for the family.
Spending on entertainment
419
Spending on education
156
Contributions to retirement
172
Spending on cigarettes
88
Spending on alcohol
41
Total spending
11,069
Panel B: other social programs
Receives SSDI
0.057
Receives SSI
0.058
Receives unemployment
0.056
Receives workers’ compensation
0.011
Receives food stamps
0.267
Receives welfare
0.057
Panel C: income and employment
Percent of federal poverty level
1.006
Number of people working
1.294
u(y) − u(y − P) > u(y ) − u(y − P) and u(y) −
m̄
m̄
0
Appendix B. The effect of Medicaid eligibility on
non-health spending, participation in other programs, and
labor market outcomes
The main manuscript focused on exploring the first-order effects
of Medicaid on household spending, but Medicaid has the potential
to affect household budgets in a variety of ways. I now consider
second-order effects of Medicaid eligibility on household budgets.
Through cost savings on medical care and health insurance,
Medicaid eligibility has the potential to lower total expenditures
or to increase non-health-related spending. As improved medical care access can lead to conversations with doctors that lead
families to consume less tobacco or alcohol and because smoking
cessation programs are part of Medicaid in many states, Medicaid
can also influence spending on categories that have health-related
consequences.
Panel A of Table B.1 displays estimates of the effect of an
additional person being eligible for Medicaid on various types
of expenditures and shows no evidence that Medicaid eligibility
affects other forms of spending. These results could indicate that
Medicaid eligibility has no impact on other forms of spending, but
they may also reflect that effects on other forms of spending are
too subtle to detect with these…
Purchase answer to see full
attachment

  
error: Content is protected !!