How do public health expansions vary by income strata? Evidence from Illinois’ all kids program.
This paper examines how income levels affected the substitution of
public health insurance for private health coverage under expansions of
Children’s health encompasses the physical, mental, emotional, and social well-being of children from infancy through adolescence.
Insurance Program (
Building on a technique developed by Abadie and Gardeazabal (2003), I
estimate that among children whose family incomes are between 200% and
300% of the federal poverty level (
), 35% of those covered by SCHIP
would have retained private coverage in the absence of SCHIP.
Significant substitution also appears between 300% and 400% FPL, but
surprisingly I find evidence that the introduction of SCHIP caused an
increase in private health insurance coverage for those with family
incomes between 400% and 500% FPL.
The question of whether expansions in Medicaid coverage lead to
declines in private coverage has emerged as an important issue in the
health policy community. As early as 1996 (
Variant of curry2.
and Gruber 1996a),
economists noted that expanding eligibility for coverage does not
necessarily result in full coverage due to uneven program adoption. As
public health insurance coverage expands, some individuals with private
health insurance may choose to substitute that coverage with public
insurance instead. Public expansions therefore can “crowd out”
private coverage, offsetting many of the coverage gains expected in the
expansion populations. Stimulated by Cutler and Gruber’s (1996)
early work on the extent of crowd-out resulting from the Medicaid
expansions of the late 1980s and early 1990s, scholars have proceeded to
examine the extent of crowd-out in the subsequent State Children’s
Health Insurance Program (SCHIP) expansion that provided public health
insurance coverage to children with family incomes below 200% of the
federal poverty level (FPL). Depending on the research design and data
selected, these studies have produced very different results of how many
children taking up public health insurance would have had private health
insurance in the absence of a public option.
The extent of crowd-out from public insurance is of particular
interest in light of the passage of the Children’s Health Insurance
Reauthorization Act of 2009 and the Affordable Care Act (
) of 2010.
The Reauthorization Act expanded health insurance coverage to children
with family incomes below 300% FPL, while the Affordable Care Act
expanded Medicaid eligibility to all adults with incomes below 133% FPL.
In both cases, this will result in the expansion of public health
insurance to income thresholds that have not previously been covered.
Since earlier research on crowd-out addressed data at lower income
thresholds, the applicability of these results at higher income levels
is unclear. While the coverage expansions slated to take place in 2014
as a result of the Affordable Care Act make it somewhat unlikely that
public health insurance expansions will continue in the near term, the
trend over the past two decades has seen a continued
/pro·lif·er·a·tion/ () the reproduction or multiplication of similar forms, especially of cells.prolif´erativeprolif´erous
public health insurance options extended to children and the
Aid to Families with Dependent Children
Aid to Families with Dependent Children
n abbr (US) (= Aid to Families with Dependent Children) → ayuda a familias con hijos menores
Medicaid expansions of 1990 and 2010, and the SCHIP expansions of 1997
and 2009. Given this long-term trend, the impact of public health
insurance expansions is likely to return to the policy agenda at some
point in the future.
While the extent of crowd-out has been extensively studied for
low-income populations, an important limitation of earlier studies is
that they do not estimate substitution that would occur among higher
income groups. However, states vary considerably in the income
thresholds that are set as eligibility criteria for public insurance.
Moreover, there is a theoretical consensus that crowd-out rates are
likely to differ significantly depending on the income criteria
specified for eligibility. In the few instances where researchers have
examined income (notably Dubay and Kenney 1996 and Card and
Shore-Sheppard 2004), significant differences have been found in
coverage and crowd-out across different income levels. To what extent
did SCHIP reduce the number of uninsured children, and to what extent
did it crowd out private insurance coverage? And how do answers to these
questions vary by the income level targeted by the insurance expansion?
In examining the impact of further expansions of SCHIP, I focus my
analysis on the Illinois SCHIP. Illinois represents an ideal test case
for three reasons. First, Illinois is unique among the states in
extending SCHIP coverage to children of all income levels, while other
states continue to restrict SCHIP enrollment based on family income.
Illinois’ SCHIP was initially typical of many other public
insurance programs, providing health insurance to children living below
185% FPL. This program expanded dramatically in 2006 under then-Gov. Rod
Blagojevich’s All Kids program, which extended SCHIP to children of
all income levels. Illinois therefore allows examination of the effect
of public expansions at income levels not implemented elsewhere in the
U.S. Secondly, Illinois’ demographic characteristics and health
insurance coverage profile closely approximate the average values found
elsewhere in the U.S., a point demonstrated later in my analysis. Thus,
patterns found in Illinois might
1. to spread throughout the body, as when local disease becomes systemic.
2. to form a general principle; to reason inductively.
more broadly across the
country if SCHIP coverage is expanded elsewhere. Third, by examining
SCHIP at different income levels within the same state, I facilitate
comparisons across different income levels while holding constant
factors that vary across states (i.e., economic performance).
/cau·sal/ () pertaining to, involving, or indicating a cause.
relating to or emanating from cause.
from the Illinois experience is
complicated by the lack of an obvious counterfactual control region that
can approximate Illinois’ coverage levels in the absence of the All
Kids intervention. In this paper, I overcome this problem by using
“synthetic” controls, an idea demonstrated in Abadie and
Gardeazabal’s analysis of
political terrorism (Abadie and
Gardeazabal 2003), and extended in a recent paper by Abadie, Diamond,
and Hainmueller (2010). The key insight of this approach is that a
counterfactual synthetic control region can be constructed as a
of multiple “donor” units unaffected by
Illinois’ health insurance expansion. In this example, the donor
units are other American states that have not extended SCHIP coverage to
higher income levels. The synthetic unit is constructed such that its
relevant demographic characteristics and health insurance profile
closely resemble that of Illinois prior to the introduction of the All
Kids program in 2006.
The synthetic control unit thus allows
Illinois’ health insurance profile would look like after 2006 in
the absence of All Kids, and causal inference can be drawn by comparing
observations from the real Illinois to those of the synthetic unit in
the post-treatment period.
I begin this paper with a discussion of theories about the
relationship between crowd-out and income. While the existing literature
1. Having greatest ascendancy, importance, influence, authority, or force. See Synonyms at dominant.
argues that expanding public health insurance to higher
income populations leads to higher crowd-out, I argue that there are
theoretical reasons why this expectation may not always hold true. Next,
intr. & tr.v. sum·ma·rized, sum·ma·riz·ing, sum·ma·riz·es
To make a summary or make a summary of.
the extensive literature on crowd-out, much of which ignores
the impact of the targeted income level on coverage and crowd-out. In
particular, I discuss issues involved in applying models that are
standard in the crowd-out literature to expansions at higher income
levels. I then discuss the synthetic control methodology used in this
paper and how this procedure overcomes the problems raised in the
earlier literature. This methodology is then applied to examine how All
Kids affects insurance coverage and turnout in Illinois for children
living between 200% and 500% FPL. Among other findings, my estimates
suggest that there may actually be a surprising increase in private
coverage among children between 400% and 500% FPL following the
expansion of public health insurance. Further examination traces the
likely cause of this to a reduction of insurance prices for high-income
individuals in Illinois. I conclude with some thoughts on the policy
implications of my findings.
Dubay and Kenney (1996) were among the first researchers to
v. the·o·rized, the·o·riz·ing, the·o·riz·es
To formulate theories or a theory; speculate.
To propose a theory about.
a relationship between crowd-out and income. Reflecting the
general consensus among health economists, they argued that because
private insurance is often unavailable to lower income populations,
expansion of public insurance to higher income levels is likely to
increase participation in public programs among income groups where
private coverage is already available. Consistent with their
expectations, Dubay and Kenney found that crowd-out was much higher for
near-poor children (those living in families with incomes between 100%
and 133% FPL) than for poor children (those living below 100% FPL).
However, this research has not been extended to income levels currently
covered by SCHIP.
While plausible, the previous depiction ignores three
countervailing forces that are likely to
To moderate in force or intensity.
or even reverse
crowd-out as SCHIP expands to higher income levels. First, there may be
mark of Cain
God’s mark on Cain, a sign of his shame for fratricide. [O. T.: Genesis 4:15]
attached to SCHIP as a low-income
stigma imposes a private cost to insurance substitution that is likely
to be increasingly costly as an individual’s income level
increases. Beyond the stigma attached to welfare, SCHIP’s status as
a low-income entitlement is likely to produce
tr.v. mis·in·formed, mis·in·form·ing, mis·in·forms
To provide with incorrect information.
n.pl the conditions that define who may be entitled to dental benefits, when persons first become entitled to such benefits, and any provisions that determine how long an individual remains entitled to benefits.
. For example, Haley and Kenney (1996)
found that while large numbers of people have heard of Medicaid and
SCHIP, significantly fewer people were aware that children could
participate in these programs without receiving welfare. In particular,
high-income groups may be more likely to lack familiarity with the
enrollment process required for public benefits, and specifically to
lack the knowledge that a public insurance program may be accessible to
Secondly, at higher income levels the quality of private insurance
offered through an employer is likely better than that offered in lower
income jobs. In deciding to substitute public for private insurance,
individuals must consider not only the quality of the public insurance
option, but also the quality of the private insurance plan they are
leaving. Even if SCHIP and Medicaid offer comparable coverage to a
private plan, they still may be a less attractive option than private
coverage. Compensation to care providers is typically lower in Medicaid
than private insurance, so providers are often less willing to see
publicly insured patients.
Finally, advocates of public insurance argue that the entry of a
new competitor in the high-income insurance market has the potential to
reduce private insurance prices. This is especially likely in highly
concentrated insurance markets like Illinois, where the two largest
HCSC Here Comes Santa Claus
Blue Cross Blue Shield and Coventry) account for 70% of
the market share. This level of market concentration is hardly unique to
Illinois; in over half of the states the two largest insurers had a
combined market share of 70% or more (Emmons, Guardado, and Kane 2011).
In particular, Bresnahan and Reiss (1991) found that most of the changes
in competitive conduct within oligopolistic markets occur with entry of
the second and third firms. Faced with lower costs, some high-income
individuals may choose to purchase a private health insurance plan for
their children that provides better coverage than SCHIP. While the
previously discussed mechanisms are only likely to mitigate crowd-out,
the effect of lower private insurance rates may actually produce
“crowd-in” that is, an increase in private coverage that
accompanies the increase in public coverage (
To summarize the previous discussion, there are multiple reasons
why one might expect to find a relationship between income and
crowd-out. However, these reasons point to different potential
relationships, casting doubt on whether crowd-out is positively,
negatively, or even non-monotonically related to income. While the
posited mechanisms differ considerably, collectively they provide strong
reasons that crowd-out rates calculated from low-income populations may
differ considerably from those in higher income populations. In the next
section, I review the crowd-out literature as applied to low-income
populations. While more comprehensive literature reviews on the large
crowd-out literature can be found elsewhere (i.e., Gruber and Simon 2008
Congressional Budget Office
2007), my review focuses primarily on
methodology and discusses why the application of previous estimation
techniques to high-income populations can produce biased estimates.
Any review of the crowd-out literature begins with Cutler and
Gruber (1996), who inspired a
Of considerable size; fairly large.
literature devoted to estimating
crowd-out effects. Using individual-level data from the Current
Population Survey (
), Cutler and Gruber examined crowd-out resulting
from the 1987-1992 Medicaid expansions by separately estimating the rate
at which public insurance was taken up, and the rate at which private
insurance was dropped, as people became eligible for public insurance.
These were estimated using two linear probability models:
Private Coverage = [[beta].sub.1] [Eligible.sub.i] +
[beta][X.sub.i] + [SIGMA] [[alpha].sub.s] [State.sub.i] + [SIGMA]
[[alpha].sub.i][Time.sub.i] + [[epsilon].sub.i] (1)
Public Coverage = [[beta].sub.2] [Eligible.sub.i] + [beta][X.sub.i]
+ [SIGMA] [[alpha].sub.s] [State.sub.i] + [SIGMA]
[[alpha].sub.i][Time.sub.i] + [[epsilon].sub.i] (2)
where Public Coverage and Private Coverage are
1. Divided or dividing into two parts or classifications.
2. Characterized by dichotomy.
variables indicating coverage type, Eligible is a dichotomous variable
indicating whether individual i is eligible for public health insurance,
X is a set of demographic controls, and State and Time are state and
variables. Since [[beta].sub.1] measures marginal take-up for
private insurance and is negatively signed, and [[beta].sub.2] measures
marginal take-up of public insurance, -[[beta].sub.1]/ [[beta].sub.2] is
an estimate of the fraction of individuals substituting public for
private coverage as eligibility expands.
Cutler and Gruber recognized that [[beta].sub.1] and [[beta].sub.2]
were unbiased estimates of marginal insurance take-up only if
eligibility for public health insurance was
to the private and
public coverage variables. To address this issue, they used a
“simulated instrument” for their measure of public insurance
eligibility following the work of Currie and Gruber (1996ab). Using
their simulated eligibility variable, Cutler and Gruber obtained
estimates of [[beta].sub.1] = -.074 and [[beta].sub.2] = .235, which
implies a crowd-out rate of 31%. [[beta].sub.2] also implies that 23.5%
of those eligible for coverage took up public insurance, while 27% of
those who were newly eligible for coverage were uninsured. This implies
that in the absence of crowd-out, the take-up rate among the uninsured
would have almost reached 90%.
Cutler and Gruber were subsequently challenged by a series of
papers, many of which used difference-in-differences (DID) estimators
(Ashenfelter and Card 1985). Under simple DID designs, outcomes are
observed for two groups over a pre-treatment (or intervention) period
and a post-treatment period. While neither group is exposed to treatment
during the pre-treatment period, in the post-treatment period one group
is exposed to the intervention while the control group is not.
Estimation of the treatment effect occurs by subtracting the average
change in the control group from that of the treatment group. This
design thus attempts to eliminate biases from comparisons over time in
the treatment group (i.e., insurance coverage trends over time unrelated
to the treatment), and to fix biases in comparing treatment and control
groups resulting from permanent differences between those groups.
A critical assumption underlying DID analysis is that unmeasured,
time-varying factors are assumed to have the same effect on treatment
and control group members.1 When this assumption fails, the estimates
are biased. This is particularly likely to be true in cases where the
treatment and control groups are drawn from different populations–a
situation that is true of every published DID design on crowd-out for
children I could find. For example, Dubay and Kenney (1996) conducted a
DID analysis comparing changes in insurance coverage of children
relative to adult men, an approach criticized for assuming there were no
other factors changing over time differentially for the two groups
(Cutler and Gruber 1997). Yazici and Kaestner (2000) and Blumberg,
Dubay, and Norton (2000) conducted DID analyses comparing changes in
insurance coverage of children who became eligible to those who still
1. Disqualified by law, rule, or provision:
for public health insurance, under the assumption that
time-varying factors did not differentially affect the eligible and
ineligible populations. Similarly, Hudson, Selden, and Banthin (2005)
estimated DID models using never-eligible children of different income
levels and married childless women of different income levels as their
control group. These DID studies typically estimated lower rates of
crowd-out than Cutler and Gruber, ranging from 5% to 15% for children in
poverty. However, it is known that such estimates are sensitive to
somewhat arbitrary changes to the control group, even when the choice of
data set is held constant. (2)
Recognizing the limitations of the original DID designs, Card and
Shore-Sheppard (2004) exploited Medicaid eligibility rules in a
n. pl. dis·con·ti·nu·i·ties
1. Lack of continuity, logical sequence, or cohesion.
2. A break or gap.
3. Geology A surface at which seismic wave velocities change.
design to identify the effect of two separate
Medicaid expansions on low-income children–the 1991 expansion to 100%
FPL and the 1990 expansion to children under 6 below 133% FPL. Following
the trend of lower estimates, Card and Shore-Sheppard did not find
statistically significant crowd-out in either Medicaid expansion. Card
and Shore-Sheppard also estimated that the expansion to 100% FPL
increased Medicaid coverage by 10% for children born just after the
cutoff birth date determining eligibility, while the expansion to 133%
FPL had no effect on health insurance coverage. The
100% and 133% FPL suggests that income can interact with coverage in
highly unusual ways, providing additional empirical support for the
kinds of unusual income dynamics that I demonstrate later in
Illinois’ All Kids program.
The trend toward lower crowd-out estimates using DID subsequently
led researchers to re-examine simulated eligibility models with mixed
results. Shore-Sheppard (2008) replicated the original Cutler-Gruber
models and found that the inclusion of age*year interaction variables
resulted in crowd-out estimates of zero and lower public insurance
take-up rates than reported earlier. This finding, confirmed by Ham and
Shore-Sheppard (2005) with different data, suggests that omitted trends
in insurance coverage by child age and state are
v. cor·re·lat·ed, cor·re·lat·ing, cor·re·lates
1. To put or bring into causal, complementary, parallel, or reciprocal relation.
expansions in eligibility. (3) However, similar models have continued to
find significant levels of crowd-out from SCHIP expansions. Using a
simulated eligibility design, Lo Sasso and Buchmueller (2004) found a
marginal take-up rate for public insurance of 8.1%, and crowd-out rates
between 18% and 50% depending on the exact specification used. A similar
paper by Gruber and Simon (2008) largely confirmed these results,
estimating a low take-up rate of 5.5% for public insurance accompanied
by a direct crowd-out rate of 30%.
Summarizing the literature, Cutler and Gruber’s estimates
suggest crowd-out rates of approximately 30% for the Medicaid expansions
of the early 1990s. However, estimates using alternative data and DID
have tended to find less crowd-out, and a
of the original
Cutler-Gruber model found no crowd-out once age*year interaction
variables were included. More recently, however, researchers applying
the Cutler-Gruber models to the Medicaid expansions of the late 1990s
have found lower rates of insurance take-up than before, accompanied by
crowd-out rates around 30%. While this crowd-out rate is consistent with
the original Cutler-Gruber estimates, the rates cannot be directly
compared because the early 1990s expansions primarily targeted children
living in poverty, while the late 1990s expansions primarily targeted
children with family incomes between 100% and 200% FPL.
Although the Cutler-Gruber models provide an attractive means to
estimate crowd-out for children below 200% FPL, their applicability to
higher income populations is limited. While Medicaid provides insurance
to some high-income children, these children are highly
/atyp·i·cal/ () irregular; not conformable to the type; in microbiology, applied specifically to strains of unusual type.
general population and are frequently
Health insurance A high-risk person without health care coverage through private insurance who falls outside the parameters of risks of standard health underwriting practices. See Underwriting.
by private means. (4)
Simulated eligibility models at higher income levels therefore may bias
upwardly estimates of public insurance take-up and bias downwardly
estimates of crowd-out because eligible high-income individuals are
effectively forced to stay in a public insurance plan. (5)
In this section, I present a new method to assess the effect of
SCHIP using Illinois’ All Kids program as the treatment unit of
interest. This exposition closely follows the description in Abadie and
Gardeazabal (2003) and Abadie, Diamond, and Hainmueller (2010), and
which is implemented subsequently in Abadie, Diamond, and Hainmueller
(2011). In this study, I observe J + 1 = 48 states, where the first
state (Illinois) is exposed to an intervention of interest (All Kids)
and the remaining J states are not exposed. Then in state i and time t,
[Y.sup.N.sub.it] represents the outcome that would be observed in the
absence of the intervention for units i = 1 , … , J 4- 1 and periods t
= 1 , … , T. Also, [Y.sup.I.sub.it] represents the outcome that would
be observed for unit i at time t if it were exposed to the intervention
in periods [T.sub.0] + 1 to T, where [T.sub.0] is the number of
preintervention periods and 1 [less than or equal to] [T.sub.0] < T.
In this application, T = 16 periods occur between 1995 and 2010;
[T.sub.0] = 12 represents the periods that occur prior to the All Kids
intervention in 2006.
Assuming the intervention has no effect on the outcome before the
implementation period t [member of ] 1 , …, [T.sub.0] for all units i
[member of] 1 , …, J + 1, the effect of exposure to the intervention
on unit i at time t is [[alpha].sub.it] = [Y.sup.I.sub.it] –
[Y.sup.N.sub.it] For untreated units (i.e., i[not equal]1) or treated
units before the intervention (i.e., i = 1 but t < [T.sub.0]), the
effect [[alpha].sub.it] =0. Then more generally, defining [D.sub.it] to
that takes the value of one if unit i is exposed
to the intervention at time t (which is true if i = 1 and t >
[T.sub.0]) and zero otherwise, the observed outcome for any unit i at
time t is [Y.sub.it] = [Y.sup.N.sub.it] + [[alpha].sub.it] [D.sub.it.]
Estimation of the effect on the exposed unit in each post-treatment time
period t > [T.sub.0] is given by [[alpha].sub.it] = [Y.sub.lt] =
[Y.sup.N.sub.1t,] with [Y.sub.1t] observed for all time periods. The
central issue of estimation therefore centers on how to estimate
[Y.sup.N.sub.1t,] the outcome that would be observed in Illinois in the
absence of the All Kids intervention.
Following Abadie, Diamond, and Hainmueller (2010), I approach this
problem by estimating [Y.sup.N.sub.1t] as a weighted combination of
other states, chosen to resemble the demographic characteristics and
health insurance profile of Illinois prior to the introduction of All
Kids. This “synthetic” Illinois provides estimates of
[Y.sup.N.sub.1t], the outcome that would be observed in the absence of
All Kids. Given J control states (i.e., states not treated with All
Kids), I estimate W = ([w.sub.1] , …, [W.sub.j])’, a (J x 1)
Of, relating to, or being a quantity that is either positive or zero.
Adj. 1. nonnegative – either positive or zero
weights that sum to 1. Each single weight
[w.sub.j] represents the weight of state j in the
/for·mu·la·tion/ () the act or product of formulating.
American Law Institute Formulation
synthetic Illinois, and each set of weights W produces a different
the choice of W, I let [X.sub.1] be a (K x 1) vector of
state-level demographic and health insurance coverage variables for
Illinois, and I let [X.sub.0] be a (K x J) matrix containing the same
variables for the J potential control states. Also, I let V be a
of nonnegative components representing the relative
importance of the different predictor variables. The vector of weights
W* that defines the combination of control regions best resembling
Illinois during the pre-treatment period is chosen such that
W* minimizes ([X.sub.1] – [X.sub.0]W)’V ([X.sub.1] – [X.sub.0]
W) subject to [w.sub.j] [greater than or equal to] 0 and [J.
n. the final argument of an attorney at the close of a trial in which he/she attempts to convince the judge and/or jury of the virtues of the client’s case. (See: closing argument)
over j=2] [w.sub.j] = 1. While V can be selected subjectively based on
previous knowledge about the relative importance of each predictor, in
this application I choose V such that the path of the synthetic outcome
variable [Y.sup.N.sub.it,] best approximates the path of the true
observed outcome variable [Y.sup.I.sub.it], during the pre-intervention
period 1 [less than or equal to] [T.sub.0] < T.
Model performance is evaluated in three ways. First, one can
examine how closely synthetic Illinois’ pre-treatment demographic
and health insurance coverage variables [X.sub.0] W approximates
[X.sub.1,] the corresponding set of observed variables in the real
Illinois. Second, one can also examine how closely the synthetic outcome
variable [Y.sup.N.sub.1t,] approximates the observed outcome variable
[Y.sup.I.sub.1t,] during the pre-intervention period. Finally, the
out-of-sample performance of the model can be evaluated using a
“placebo” treatment. To do so, I choose a placebo treatment
period [T.sub.p] that occurs before the true implementation period
[T.sub.0,] and refit the same model, substituting the chosen [T.sub.p]
for [T.sub.0.] Under this condition, note that a crucial difference is
that during the placebo “post-treatment” period [MATHEMATICAL
v. re·pro·duced, re·pro·duc·ing, re·pro·duc·es
1. To produce a counterpart, image, or copy of.
2. Biology To generate (offspring) by sexual or asexual means.
or a set of codes used to represent letters, numbers, a few symbols, and control characters. Originally designed for teletype operations, it has found wide application in computers.
] (6) I thus can further
the model by seeing how closely the estimated [MATHEMATICAL EXPRESSION
NOT REPRODUCIBLE IN ASCII] tracks the observed [MATHEMATICAL EXPRESSION
NOT REPRODUCIBLE IN ASCII] in the period between the placebo and actual
The synthetic control procedure enjoys many advantages over other
methods used to estimate effects from comparative case studies. Abadie,
Diamond, and Hainmueller (2010) note that the synthetic control
procedure generalizes the DID model discussed previously. Additionally,
the procedure produces useful estimates in models with time-varying
coefficients. (7) This is not true for traditional DID designs, which
allow for unobserved confounders but restrict the estimated effects of
those confounders to be constant in time. In contrast, the synthetic
control model allows the effects of
characteristics to vary with time, thus creating less model dependency.
Furthermore, the procedure directly addresses many of the issues
raised in considering how earlier techniques used to estimate crowd-out
could be applied to higher income populations. In contrast to the
earlier DID designs where the treatment and control groups may not have
been comparable (i.e., comparing children to adults), synthetic controls
have the advantage of forcing the researcher to demonstrate an affinity
between the treatment and control units, in the sense that the predictor
values of the synthetic control [X.sub.0] W closely approximate those of
the observed unit [X.sub.1] during the pre-treatment period. This
affinity maximizes the chance that time-varying factors have the same
effect on the treatment and control groups. In effect, the procedure
draws causal inference using exact matching, where the treatment unit is
matched not only to the observed donor cases but to all possible
combinations of those observed cases. Also, the synthetic control
procedure addresses the selection issues in the simulated eligibility
design, where a direct application of those procedures to higher income
populations will yield biased estimates because the treatment units are
not representative of the larger population at the income strata of
interest. Synthetic controls address this issue by allowing the entire
state of Illinois to be used as a treatment unit, because the All Kids
treatment covers all children
Without consideration of; regardless of.
their income levels or
Data for these results were obtained from the annual March
Supplement to the Current Population Survey (CPS) from the years
19952010. I used CPS data for this paper because the synthetic control
methodology requires samples designed to be representative within
states, a feature that is not true of the
Survey of Income and Program
or the Panel Study of Income Dynamics (
Bureau of Labor
2010). I begin this analysis using data from 1995 because
that was the first year of the CPS following its
tr.v. re·de·signed, re·de·sign·ing, re·de·signs
To make a revision in the appearance or function of.
to improve the
quality of its health insurance data (Swartz 1997). All state-level
estimates of the total, private, and public coverage variables are
Kalman smoothed to reduce measurement error.
My analysis was restricted to children below age 19 with family
incomes between 200% and 500% FPL. Although these data are
/strat·i·fied/ () formed or arranged in layers.
Arranged in the form of layers or strata.
income, there is sufficient data in the CPS to draw inference. Across
the 16 years of data, I average N=358.9 (cy -62.1), N=301.9 (cy = 55.1),
and N=215.3 (cy = 37.4) observations each year, respectively, for the
three income strata of interest (200% to 300% FPL, 300% to 400% FPL, and
400% to 500% FPL). Summary statistics of the key variables of interest
in Illinois and the
U.S. are presented in Table 1.
In this section, I present an application of the synthetic control
method to Illinois’ All Kids program. I begin by describing the
application of this method to insurance coverage at the 400% to 500% FPL
population in greater detail, focusing on the diagnostics used to assess
the fit of my model. I then apply the same procedure to coverage between
200% and 400% FPL to determine the net increase in child health
insurance coverage resulting from All Kids. Next, I apply the same
procedure to determine changes in public versus private coverage across
the 200% to 500% FPL income level. These results facilitate subsequent
estimates of crowd-out. (8)
I begin by constructing a synthetic control for Illinois using CPS
data from 1995 2006, the pre-treatment period before the enactment of
All Kids. Recall that the objective is to approximate Illinois’
predictors of child health insurance coverage between 400% and 500% FPL
using a convex combination of donor states. For predictor variables, I
use the percentage of children covered with health insurance below the
federal poverty level, at 100% to 200% FPL, 200% to 300% FPL, 300% to
400% FPL, and 400% to 500% FPL, along with a standard set of demographic
controls including race, income, income distribution, unemployment rate,
and education. Donor states include all states except for Alaska,
Hawaii, and the
District of Columbia
federal district (2000 pop. 572,059, a 5.7% decrease in population since the 1990 census), 69 sq mi (179 sq km), on the east bank of the Potomac River, coextensive with the city of Washington, D.C. (the capital of the United States).
, though my results are robust to
the inclusion of those states and the District of Columbia as well. (9)
This control group allows estimates of children’s health insurance
coverage in the post-intervention period, 2007-2010, because they did
not enact public health expansions at that income level during the
Table 1 summarizes the predictor values for the real Illinois and
its synthetic counterpart, averaged across the 1995 2006
n the protocols required before beginning therapy, usually of a diagnostic nature; before treatment.
n See predetermination.
period. Comparing the predictor values in columns 2 and 3, the synthetic
control closely reproduces the values from Illinois during that period,
suggesting a strong fit for the synthetic unit. Column 4 shows the same
values calculated for the 47 states included in the donor pool, which
largely represents the predictor values across the rest of the U.S. In
comparing predictor values between Illinois and the U.S., we see some
minor differences, most notably a $3,005 difference in mean income. This
is significant because it suggests that Illinois is not particularly
unusual among states; hence the results of this study may potentially
generalize to the U.S. at large. (10) Also note that for all variables,
the synthetic control more closely approximates the predictor values of
Illinois than the nation at large, especially with regard to income.
This suggests that the synthetic unit serves as a better control unit
than simply using all other states as a control. Finally, the standard
deviations of the predictor variables for the 47 control states are
shown in column 5. The relatively large deviations for all variables
suggest that some combinations of control states will produce extremely
poor synthetic controls that fail to
1. To produce a counterpart, an image, or a copy of something.
2. To bring something to mind again.
3. To generate offspring by sexual or asexual means.
the predictor values
Table 2 explores the construction of the synthetic control unit
more closely. While the estimated weights defining the combination of
states used to construct the synthetic unit only approximately sum to 1
because of rounding, 94% of the total weight is accounted for by five
states: Delaware, Maryland, Michigan, Oregon, and
island, 15 mi (24 km) long and 5 mi (8 km) wide, S R.I., at the entrance to Narragansett Bay. It is the largest island in the state, with steep cliffs and excellent beaches.
weights can be used to generate the synthetic control, which then can be
used to estimate the effect of All Kids on Illinois’ child health
insurance coverage rates between 400% and 500% FPL.
Figure 1 displays the percentage of children with family incomes
between 400% and 500% FPL covered by either private or public health
insurance in Illinois from 1995 through 2010, along with its synthetic
counterpart. Prior to the introduction of the All Kids policy treatment,
synthetic Illinois largely replicates the slight downward trend in total
health insurance coverage that is actually observed in Illinois.
Combined with the high balance on all predictors shown previously in
Table 1, this suggests that synthetic Illinois provides a reasonable
1. the act or process of bringing into proximity or apposition.
2. a numerical value of limited accuracy.
to the coverage that would have been
observed between 2007 and 2010 in the absence of the All Kids expansion.
(12) However, coverage in real Illinois diverges from coverage in
synthetic Illinois after the 2006 policy intervention. The gap between
the observed and synthetic coverage represents the estimate of All
Kids’ effect on the total insurance rate at this income level.
While the gap is
during the pre-treatment period, the large
positive gap suggests that All Kids significantly increased health
insurance coverage among children of this income group.
[FIGURE 1 OMITTED]
To evaluate the statistical significance of this gap, I conduct a
test to determine how likely a gap of this magnitude is
likely to occur by chance under the
n theoretical assumption that a given therapy will have results not statistically different from another treatment.
. More specifically,
I generate the distribution of mean squared prediction errors (
under the null hypothesis when no intervention has occurred, and compare
this distribution to the observed MSPE of 9.03 from the post-treatment
observations. In the original model, I use 12 pre-treatment observations
between 1995 and 2006 to fit a model that is then used to estimate four
counterfactual cases in the 2007 2010 post-treatment period. To
, I instead use a random selection of eight of the 12
pretreatment observations to fit the model (i.e., observations from the
1995-2002 period). This model also allows me to generate four
counterfactual cases for the remaining four pre-treatment cases (i.e.,
for 2003-2006). Of the 495 possible permutations, I test 100 such
permutations at random, calculating the average out-of-sample MSPE of
the remaining four observations for each trial. None of the control
trials achieves such a large MSPE, and the inclusion of 100 trials in
this sampling distribution suggests that one can reject the possibility
at the [sigma] = .01 level of significance. This
procedure can be
1. Involving an entire organ, as when an epileptic seizure involves all parts of the brain.
2. Not specifically adapted to a particular environment or function; not specialized.
to apply the logic of the
1. A large clasp knife.
2. Sports A dive in the pike position, in which the diver straightens out to enter the water hands first.
generate standard errors (Tukey 1958). (13)
[FIGURE 2 OMITTED]
As an additional check on this hypothesis test, I implement the
placebo treatment discussed in the methods section, where I refit the
model using a fake treatment period timed before the actual treatment
takes place. I conduct this test by using a placebo treatment in the
year 2000. This test proceeds by fitting the model using only the first
six observations (1995 2000), and comparing the predicted values of that
model to the observed values in the period for the post-placebo period
prior to the actual treatment (2001 2006). These out-of-sample
predictions should closely approximate the observed values since the
real treatment in 2006 has not yet occurred. In contrast to the
jackknife, this procedure tests for the possibility that there is always
between the synthetic and observed results after even a fake
treatment, however unlikely that may be. Figure 2 displays the observed
and synthetic approximation of the placebo test. As expected, the
placebo treatment does not cause the true and synthetic coverage rates
between 2001 and 2006, providing additional evidence that the
coverage gaps in the post-treatment period shown earlier in Figure 1 are
unlikely to occur by chance.
I repeat this analysis for the 200% to 300% FPL and 300% to 400%
FPL income levels, but discard some control states that expanded SCHIP
above 200% FPL. Results between 100% and 200% FPL cannot be obtained
because there are no control states that did not enact an SCHIP
expansion at that income level; however, coverage and crowd-out at this
level have been
tr.v. an·a·lyzed, an·a·lyz·ing, an·a·lyz·es
1. To examine methodically by separating into parts and studying their interrelations.
2. Chemistry To make a chemical analysis of.
using the methods discussed in the literature
review. In both cases, the synthetic unit accurately reproduces the
observed coverage values during the pre-treatment period.14 However,
estimates in the post-treatment period differ substantially. Between
200% and 300% FPL, the synthetic unit suggests that insurance coverage
would have increased slightly in the post-treatment period even in the
absence of All Kids. Nevertheless, All Kids appears to have increased
health insurance coverage above and beyond the expected increase. At the
300% to 400% FPL level, however, the story is quite different. All Kids
appears to have had no net effect on health insurance coverage at this
income level, and the increase in coverage shown in the synthetic unit
at the 200% to 300% FPL level resulting from factors other than All Kids
also appears to be absent.
Table 3 summarizes the net effect of SCHIP on insurance coverage,
tabulating the net increase in health insurance coverage by year and
income strata. Following the logic outlined in the resampling test, I
estimate standard errors using the jackknife by calculating the
out-of-sample prediction errors for each observation in the pretreatment
period. Between 200% and 300% FPL, total coverage increased after four
years by 2.7 percentage points above what would have been expected in
the absence of any policy intervention. This increase covered 28.3% of
the uninsured population at the income level, so extension of coverage
to the uninsured is substantively large. Between 300% and 400% FPL,
increases in coverage were not statistically significant. (15) However,
between 400% and 500% FPL, the 3.36-percentage-point increase in net
coverage had an enormous effect, covering 49.9% of the uninsured. My
results therefore suggest that the marginal impact of public health
insurance on coverage does not change monotonically with the income
level of the expansion.
I extend the analysis presented previously, estimating the same
models separately for private and public health insurance coverage from
200% to 500% FPL with two changes. First, I replace the original
coverage control variables for private or public versions of those same
variables, as appropriate. Secondly, following Cutler and Gruber (1996)
and Gruber and Simon (2008), I treat cases in which an individual claims
to have both private and public coverage together as situations where an
individual is making a transition from private insurance to public
insurance. While there are generally few cases where an individual
claims both forms of insurance, this assumption will generally produce
higher estimates of crowd-out than the alternative of discarding the
Figure 3 plots observed and synthetic coverage rates for private
and public insurance between 1995 and 2010. Private insurance coverage
is shown along the top, while public insurance coverage appears on the
bottom. In five of the six cases, the synthetic control unit continues
to closely reproduce the corresponding coverage rates observed in
Illinois, providing evidence of good model fits. However, even without
looking at the post-treatment period, there is strong evidence pointing
to substitution in the pretreatment period; all income levels show some
decrease in private coverage accompanied by some increase in public
coverage during that time. Another notable trend in the data is that net
private and public coverage are strongly related to income–as income
increases, private coverage rates increase and public coverage rates
My estimates of public coverage rates highlight a limitation of the
approach because I am unable to construct a synthetic estimate for
public coverage between 400% and 500% FPL. This is not surprising, as
public health insurance coverage rates for high-income populations are
generally quite low,
v. fluc·tu·at·ed, fluc·tu·at·ing, fluc·tu·ates
1. To vary irregularly. See Synonyms at swing.
2. To rise and fall in or as if in waves; undulate.
between 2% to 4% in Illinois during the
pre-treatment period. In these cases, construction of a synthetic unit
can be very difficult because such units are typically interpolated as
convex combinations of states. Some of these donor states will have
higher public coverage rates than Illinois during the pre-treatment
period, but these are typically balanced by the inclusion of some donor
states with public coverage rates lower than those in Illinois. Since
Illinois’ public coverage rate is already quite low for this income
group, there are insufficient donor states with low public coverage that
are able to reproduce both Illinois’ public coverage pattern and
predictor values. Despite my inability to produce a synthetic
1. To duplicate, copy, reproduce, or repeat.
2. To reproduce or make an exact copy or copies of genetic material, a cell, or an organism.
A repetition of an experiment or a procedure.
of Illinois at the 400% to 500% FPL, the time series plot in Figure 3
clearly suggests that there is a significant take-up of public health
insurance following the treatment, since there is a large increase in
public coverage following the implementation of All Kids. (16)
[FIGURE 3 OMITTED]
The gaps in Figure 3 measure the difference between the observed
and synthetic private/public insurance coverage, which represents my
estimate of the effect of All Kids in the post-treatment period. At 200%
to 400% FPL, I get evidence consistent with the expectations of
crowd-out–a decrease in private coverage accompanied by an increase in
public coverage following the passage of All Kids. However, this pattern
diverges significantly at 400% to 500% FPL, where one can see an
increase in private coverage instead. This is likely accompanied by an
increase in public coverage as well, though the determination of this is
complicated by the lack of a synthetic Illinois to approximate the
Tables 4 and 5 present the effects numerically for public and
private insurance respectively during the entire post-treatment period.
(17) These values are needed to calculate crowd-out estimates in
relative (percentage) terms, but they are also of interest because they
also produce estimates of crowd-out in
as a proportion of
the entire population. Between 200% and 300% FPL, public coverage
increased by 6.54 percentage points in 2010, a large amount that is more
than 1.5 times larger than a standard deviation of child health
insurance coverage across the U.S. as reported earlier in Table 1.
Take-up of public insurance is lower at other income levels, varying
between 1.58 to 3.58 percentage points. In most cases, the observed
shifts are significant at standard levels. (18) The change in public and
private insurance rates also suggests that the full impact of the All
Kids intervention unfolds over many years-coverage rates for both types
of insurance still do not appear to have
v. sta·bi·lized, sta·bi·liz·ing, sta·bi·liz·es
1. To make stable or steadfast.
four years after the
intervention. The length of time that public health care expansions
require for full effect, and the measurement of crowd-out rates over
different periods of that expansion, may provide a partial explanation
why estimates of crowd-out vary so dramatically, even when the
estimation strategy remains constant. (19)
The crowd-out rate is defined as the fraction of children taking up
public health insurance who, in the absence of a public health insurance
option, would have taken up private health insurance instead. I
calculate it using the formula:
Crowd-out = -[DELTA]Private/[DELTA]Public
where [DELTA]Public and [DELTA]Private are the entries found in
Tables 4 and 5 respectively. Applying this equation, I calculate the
coverage and crowd-out rates shown in Table 6. Three results emerge from
this analysis. Between 200% and 300% FPL, crowd-out reached 35% by
2010–an estimate that is consistent with the magnitude of the effect
found by Cutler and Gruber (1996) for Medicaid below 200% FPL. Between
300% and 400% FPL, I observe no effect on net coverage, suggesting that
every marginal person taking up public health insurance would have
otherwise been covered by private health insurance. Finally, between
400% and 500% FPL, total coverage increased by 3.36 percentage points,
aided by a 2.62-percentage-point increase in private coverage shown
earlier in Table 5. Stated differently, we actually observe
“crowd-in”–an increase in private coverage rather than
substitution away from it. No standard errors are provided for crowd-out
estimates, so it is important to note the potential effect of this
n. 1) failure to perform an act agreed to, where there is a duty to an individual or the public to act (including omitting to take care) or is required by law. Such an omission may give rise to a lawsuit in the same way as a negligent or improper act.
on the interpretation of the results: as Hudson, Selden, and
Banthin (2005) have noted, standard errors for crowd-out estimates can
be much larger than the standard errors for the individual coefficients
used to estimate them.
As a robustness check, I also compare results derived from the
synthetic model against those using a simple DID model in Table 6. Using
individual-level CPS data from Illinois and its
1. One who lives near or next to another.
2. A person, place, or thing adjacent to or located near another.
3. A fellow human.
4. Used as a form of familiar address.
between 2003 and 2010, I estimate linear probability models of the form
Coveragei = [[??].sub.0] + [[??].sub.1], * IL + [[??].sub.2] Treatment +
[[??].sub.3] * Treatment *IL, where Coveragei is a binary variable
indicating whether individual i is covered by public or private health
insurance, IL is a dichotomous variable for an Illinois resident, and
Treatment indicates whether the
was surveyed after the
introduction of All Kids in 2006. (20) Each cell in columns 4 and 5
reports an estimate of [[??].sub.3], the estimated effect of All Kids on
public or private coverage at the specified income strata. The sum of
public and private coverage, shown in column 6, represents the DID
estimate of All Kids’ overall effect on health insurance coverage
in Illinois. Although there is significant uncertainty associated with
these estimates, the DID results are broadly consistent with the
patterns obtained via the synthetic model in suggesting that net
coverage increased between 200% and 300% FPL and between 400% and 500%
FPL. In particular, although the estimate for coverage between 400% and
500% FPL fails to achieve statistical significance, it is similar in
magnitude to the synthetic estimate, and the unusual increase in private
coverage at this income level that is estimated by the synthetic model
enjoys similar supporting evidence. Crowd-out estimates are 15% higher
between 200% and 300% FPL when derived using DID, though previous
caveats about the uncertainty surrounding such estimates also apply
Why Does Crowd-in Occur?
Estimates from the previous section suggest that at high-income
levels above 400% FPL, the expansion of public health insurance may
actually cause crowd-in–an increase in private coverage rather than
substitution away from it. Why might this occur? My key hypothesis is
that the entry of a new public competitor in the high-income insurance
market has the potential to reduce private insurance prices, or at least
slow their growth considerably. This is especially likely to be true in
insurance markets that are highly concentrated with few competitors–a
that is largely true of both Illinois and the U.S. at
large. Faced with lower costs, some high-income individuals may choose
to purchase a private health insurance plan for their children that
provides better coverage than SCHIP.
Before proceeding further, note that other researchers have also
uncovered evidence for crowd-in. In particular, there is growing
evidence that despite expanding eligibility for public health insurance,
the Massachusetts health reform simultaneously increased both public and
private health insurance coverage. Using data from a household survey,
Long (2008) reported that employer-based coverage and public insurance
each expanded by 2.9 percentage points, while Long, Stockley, and Yemane
(2009) found a 5.6-percentage-point increase in employer-sponsored
insurance accompanied by an 11.7-percentage-point increase in public
coverage, using a DID design with CPS data. These findings were largely
confirmed by Gabel et al. (2008), who found evidence of crowd-in using
employer surveys. Between 2007 and 2008, they found that the percentage
of firms offering health benefits increased from 73% to 79%, while the
percentage of firms offering coverage nationally was statistically
unchanged. Firms with 11 to 50 workers also increased coverage
significantly, from 88% to 92%, and were no more likely than firms
nationally to consider dropping coverage or restricting eligibility
(Gabel et al. 2008). These earlier results, coupled with the supporting
evidence that I present here, jointly point to the empirical
1. Seemingly or apparently valid, likely, or acceptable; credible:
2. Giving a deceptive impression of truth or reliability.
In this section, I use publicly available data from the Medical
Expenditure Panel Survey –Insurance Component (
) to investigate
this issue. (22) I tabulate the growth of total family premiums per
enrolled employee in the private sector between 2004 and 2008, covering
the two years before and after the introduction of All Kids, for both
Illinois and the U.S. at the fourth payroll
. I estimate the
effect of All Kids on premium growth as the difference in premium growth
between Illinois and the U.S. for the years 2006-2008, over and above
the difference in premium growth for the period 2004-2006. This analysis
allows me to control for
v. pre·ex·ist·ed, pre·ex·ist·ing, pre·ex·ists
To exist before (something); precede:
differences in premium growth that
jointly affect both Illinois and the U.S., but are unrelated to the All
My analysis is subject to two important caveats. First, note that
publicly available MEPS-IC data do not contain sufficient information to
calculate standard errors for my estimates. Secondly, the MEPS-IC does
not contain income data directly, but instead provides data on the wage
distribution of firms providing coverage. The fourth payroll quartile
thus refers to the quartile of the Illinois workforce in the group of
establishments with the highest average payroll, not to the top 25% of
income earners in Illinois. Nevertheless, the MEPS-IC remains the best
source of publicly available information on insurance premiums, and both
the direction and substantive magnitude of the changes I estimate are
largely consistent with expectations.
Table 7 presents my analysis of family premium
second column shows that premiums for the 2006-2008 period for employees
in high-payroll establishments grew 3.8% in Illinois after All Kids,
compared to 7.1% across the U.S. A simple estimate of the effect of All
Kids would be the difference between the two, or a 3.3% decrease that is
consistent with my expectation that All Kids reduced insurance premiums.
This simple calculation, however, does not account for the possibility
that Illinois may have had a different pre-existing trend in premiums
from the rest of the country. I examine this in the first column of
Table 7, which shows that Illinois’ premium rates were actually
growing 1.5% faster than the rest of the country. The differential
growth in Illinois versus the U.S. between 2006 and 2008, compared to
the premium growth rate between 2004 and 2006, yields -4.8% as my
estimate of the effect of All Kids on premiums in high-payroll
establishments. With an average family premium rate of $12,475 in 2008,
this effect amounts to $599 in reduced costs for a family in the fourth
payroll quartile in Illinois.
In the third and fourth rows of Table 7, I also tabulate premium
growth rates across the same periods for residents at all payroll
percentiles. While both Illinois and the U.S. show a substantial slowing
in the growth of premium rates after 2006, they still grow faster than
what was found earlier for high-payroll establishments in Illinois. The
reduction in premium rates caused by All Kids thus appears to have a
effect, focused primarily on premiums in high-payroll
establishments. This is consistent with my finding of crowd-in only at
high-income levels. Overall, I find evidence that All Kids lowered
family premiums considerably at high-payroll establishments, which may
partially explain why I observe crowd-in.
This paper examines the impact of SCHIP expansions on insurance
coverage across higher income populations. Health economists widely
believe that as eligibility levels for public health insurance programs
are expanded to cover higher incomes, the possibility of substitution
from private to public insurance rises because the expansions
increasingly target those with access to private insurance. However,
there are many reasons why this relationship may not be so simple–a
conclusion that previously drew empirical support from the work of Card
and Shore-Sheppard (2004). In examining Illinois’ All Kids program,
I find additional evidence that the relationship between income and
crowd-out is much more complex. More specifically, I find that for
children with family incomes between 200% and 300% FPL, the All Kids
program produced an increase in health insurance coverage of two
percentage points with 35% crowd-out. For children with family incomes
between 300% and 400% of FPL, All Kids produced no increase in overall
coverage with significant crowd-out. Finally, I find that for children
with family incomes between 400% and 500% FPL, SCHIP produced a
3.4-percentagepoint increase in coverage with crowd-in. These findings
are of substantive interest because there is clearly a desire by some
legislators to continue expanding SCHIP eligibility to higher income
Although comparisons at higher income levels are difficult to make
because of the
of published high-income studies, the 35%
crowd-out estimate at 200% to 300% FPL is largely in line with other
estimates. My estimate is similar to those at lower income levels found
by Cutler and Gruber (1996), and tends to be higher than estimates
derived using DID (see, for example, Dubay and Kenney 1996; Blumberg,
Dubay, and Norton 2000; Yazici and Kaestner 2000; Card and
Shore-Sheppard 2004). Most notably, the crowd-out estimate is consistent
with the Congressional Budget Office’s (
) projection of 33%
crowd-out for the Children’s Health Insurance Program
Reauthorization Act. The CBO projection is particularly important not
only because of its budgetary implications, but also because the
Reauthorization Act expanded coverage to 300% FPL; hence, it provides
the most comparable estimate by income level. This level of crowd-out is
seen by some as acceptable; in describing the projection of the
House-passed SCHIP reauthorization in 2007, former CBO Director Peter
Orszag stated that he “has not seen another plan that adds 5
million kids to SCHIP with a 33% crowd-out rate. This is pretty much as
good as it is going to get” (
BNA British North America
BNA Banco Nacional de Angola
Health Care Daily 2007).
Finally, my research also has implications for implementation of
the Affordable Care Act and its coverage expansions slated to take place
in 2014. At its core, the Affordable Care Act requires individuals not
already covered with health insurance to purchase private insurance or
pay a penalty. Moderate-income individuals will be eligible to meet this
requirement by purchasing private insurance on government-operated
health insurance exchanges at government-subsidized rates. While my
paper focuses on the crowding out of private insurance by public
insurance, the Affordable Care act may potentially produce some
crowd-out of unsubsidized private insurance by
tr.v. sub·si·dized, sub·si·diz·ing, sub·si·diz·es
1. To assist or support with a subsidy.
2. To secure the assistance of by granting a subsidy.
insurance in a manner similar to what I describe in this paper. My
results suggest that predicting future crowd-out rates resulting from
the ACA may be difficult, in part because subsidized insurance exchanges
will affect income strata much higher than what has typically been
observed in previous studies of Medicaid and SCHIP.
Abadie, A., A. Diamond, and J. Hainmueller. 2010. Synthetic Control
Methods for Comparative Case Studies: Estimating the Effect of
California’s Tobacco Control Program.
Journal of the American
–. 2011. Synth: An R Package for Synthetic Control Methods in
Comparative Case Studies. Journal of Statistical Software 42(13): 1-17.
Abadie, A., and J. Gardeazabal. 2003. The Economic Costs of
Conflict: A Case Study of the
, Basque Euzkadi, Span. País Vasco, comprising the provinces of Álava, Guipúzcoa, and Vizcaya (1990 pop.
. American Economic Review
Ashenfelter, O., and D. Card. 1985. Using the
Running in the direction of the long axis of the body or any of its parts.
Structure of Earnings to Estimate the Effect of Training Programs.
Review of Economics and Statistics 67(4):648-660.
Blumberg, L. J., L. Dubay, and S. A. Norton. 2000. Did the Medicaid
Expansions for Children
tr.v. dis·placed, dis·plac·ing, dis·plac·es
1. To move or shift from the usual place or position, especially to force to leave a homeland:
Private Insurance? An Analysis Using
. Journal of Health Economics 19(1):33-60.
BNA Health Care Daily. 2007. SCHIP: Governors, Health Officials,
Seek Withdrawal of
Rules Targeting Crowd-Out by SCHIP. August 31.
Bresnahan, T., and P. Reiss. 1991. Entry and Competition in
Concentrated Markets. Journal of Political Economy 9(5):977-1009.
Bureau of Labor Statistics. 2010. Comparison of SIPP with Other
Surveys. http://www.census. gov/sipp/vs.html Accessed February 1, 2010.
Card, D., and L. Shore-Sheppard. 2004. Using
1. interrupted; intermittent; marked by breaks.
2. discrete; separate.
3. lacking logical order or coherence.
Eligibility Rules to Identify the Effects of the Federal Medicaid
Expansions on Low Income Children. Journal of Health Economics
Chang, C. F. 2007. Evolution of TennCare Yields Valuable Lessons.
Managed Care 11:45-49.
Currie, J., and J. Gruber. 1996a. Health Insurance Eligibility,
Utilization of Medical Care, and Child Health.
Quarterly Journal of
–. 1996b. Saving Babies: The Efficacy and Cost of Recent Changes
in the Medicaid Eligibility of Pregnant Women. Journal of Political
Economy 104(6): 1263-1296.
Cutler, D., and J. Gruber. 1996. Does Public Health Insurance Crowd
Out Private Insurance? Quarterly Journal of Economics 111(2):391-430.
–. 1997. Medicaid and Private Insurance:
Evidence and Implications. Health Affairs 16(1):194-200.
Dubay, L., and G. Kenney. 1996. The Effects of Medicaid Expansions
on Insurance Coverage of Children. Special Education for Students with
Disabilities 6(1): 152-161.
Emmons, D., J. Guardado, and C. Kane. 2011. Competition in Health
Insurance: A Comprehensive Study of U.S. Markets, 2911 Update. Chicago:
American Medical Association.
Gabel, J., H. Whitmore, and J. Pickreign. 2008. Report from
Massachusetts: Employers Largely Support Health Care Reform, and Few
Signs of Crowd-Out Appear. Health Affairs 27(1):13-23.
Gabel, J., H. Whitmore, J. Pickreign, W. Sellheim, K. Shova, and V.
Bassett. 2008. After the Mandates: Massachusetts Employers Continue to
Support Health Reform as More Firms Offer Coverage. Health Affairs
Gruber, J. 2008. Covering the Uninsured in the
officially United States of America, republic (2005 est. pop. 295,734,000), 3,539,227 sq mi (9,166,598 sq km), North America. The United States is the world’s third largest country in population and the fourth largest country in area.
Journal of Economic Literature 46(3):571-606.
Gruber, J., and K. Simon. 2008. Crowd-out 10 Years Later: Have
Recent Public Insurance Expansions Crowded Out Private Health Insurance?
Journal of Health Economics 27(3):201-217.
Hacker, J. 2007a. Health Care for America. Accessed January 11,
2007. http://www.shared prosperity.org/bp180.html
–. 2007b. The Case for Public Plan Choice in National Health
Reform. Accessed February 1, 2010.
Haley, J., and G. Kenney. 1996. Coverage Gaps for Medicaid-Eligible
Children in the Wake of Federal Welfare Reform. Inquiry 40(2): 158 168.
Ham, J., and L. Shore-Sheppard. 2005. The Effect of Medicaid
Expansions for Low-Income Children on Medicaid Participation and Private
Insurance Coverage: Evidence from the SIPP. Journal of Public Economics
Hudson, J., T. Selden, and J. Banthin. 2005. The Impact of SCHIP on
Insurance Coverage of Children. Inquiry 42(3):232-254.
Long, S. 2008. On the Road to Universal Coverage: Impacts of Reform
in Massachusetts at One Year. Health Affairs 27(4): 270-284.
Long, S., K. Stockley, and A. Yemane. 2009. Another Look at the
Impacts of Health Reform in Massachusetts: Evidence Using New Data and a
Stronger Model. American Economic Review 99(2):508 511.
Lo Sasso, A., and T. Buchmueller. 2004. The Effect of the State
Children’s Health Insurance Program on Health Insurance Coverage.
Journal of Health Economics 23(5): 1059-1082.
Congressional Budget Office (CBO). 2007. The State Children’s
Health Insurance Program. Accessed February 1, 2010. Washington, D.C.:
Congressional Budget Office. http:// www.cbo.gov/publication/18638
Shore-Sheppard, L. 2008. Stemming the Tide? The Effect of Expanding
Medicaid Eligibility On Health Insurance Coverage. B.E. Journal of
Economic Analysis and Policy 8(2):45-49.
Swartz, K. 1997. Changes in the 1995 Current Population Survey and
Estimates of Health Insurance Coverage. Inquiry 34(1):70-79.
Tukey, J. 1958. Bias and Confidence in Not Quite Large Samples.
1. A chronological record of the events of successive years.
2. A descriptive account or record; a history:
of Mathematical Statistics 29(2):614.
Yazici, E., and R. Kaestner. 2000. Medicaid Expansions and the
Crowding Out of Private Health Insurance. Inquiry 37(1):23-32.
(1) Note that time-varying characteristics can be added to DID
models, and many researchers (e.g., Hudson, Selden, and Banthin 2005;
Shore-Sheppard, 2008) have acknowledged the need to do this.
(2) A prominent example is Hudson, Selden, and Banthin (2005), who
compared SCHIP crowd-out estimates using a difference-in-trends approach
with data from the 19962002 Medical Expenditure Panel Survey. Using
children with family incomes between 300% and 400% FPL as a control
group produced a crowd-out estimate of 56%, while using children with
family incomes between 400% and 600% FPL produced a crowd-out estimate
(3) Take-up and crowd-out rates were estimated to be 13.6% and 0%
(4) A notable example is Tennessee’s TennCare Medicaid
program, which insures more than 500,000 individuals who are uninsurable
by private insurance due to pre-existing conditions (Chang 2007).
(5) A similar point was made by Hudson, Selden, and Banthin (2005),
who argue that high-income families have different insurance options
than children targeted by SCHIP and Medicaid. Comparing Cutler-Gruber
style estimates using all children compared to children in families
between 100% and 300% FPL, they found that using the smaller sample
substantially reduced estimates of net coverage effects from public
(6) These results, of course, will not hold for t > To, since
the real treatment will have occurred.
(7) Proof of this claim under standard conditions is shown in
Appendix B of Abadie, Diamond, and Hainmueller (2010).
(8) For most SCHIP programs, stratifying data in this manner is
potentially problematic because it may be
/en·dog·e·nous/ () produced within or caused by factors within the organism.
1. Originating or produced within an organism, tissue, or cell.
with respect to
income (i.e., individuals might alter their income to become eligible
for public insurance). However, this is unlikely to be true in Illinois,
since there are no income-eligibility levels in the All Kids program.
(9) I remove Alaska and Hawaii because the FPL in these states is
different from those of the 48 contiguous states.
(10) Gruber (2008), for example, has argued that the health reform
introduced by former Gov.
in Massachusetts would be unlikely
to produce near-universal coverage if implemented in other states
because of its unusually high-income and coverage rates.
(11) Note that although Illinois is similar to the average of the
other states, the large standard deviations also suggest that Illinois
cannot be generalized to many other individual states with
characteristics that deviate unusually from the country at large.
(12) As an additional test
tr.v. val·i·dat·ed, val·i·dat·ing, val·i·dates
1. To declare or make legally valid.
2. To mark with an indication of official sanction.
the value of the synthetic
control unit, I calculated the
mean squared prediction error
pretreatment period using each donor state as the sole control unit.
Only one state, Utah, produced a pre-treatment MSPE lower than the
synthetic unit, and a quick check shows that Utah has a set of predictor
variables that differ significantly from those of Illinois. Hence, no
single control state can produce a set of coverage patterns and
covariates comparable to that of the synthetic unit.
(13) Under the jackknife, I discard one pre-treatment observation
and fit a model using the remaining 11 pre-treatment observations, then
calculate the deviation between the predicted value and the
v. dis·card·ed, dis·card·ing, dis·cards
1. To throw away; reject.
a. To throw out (a playing card) from one’s hand.
out-of-sample value. This can be done 12 times, using each of my 12
pre-treatment observations as testing data. The standard deviation of
this distribution represents my approximation of the standard error.
(14) Predictor values are also well reproduced; however, to save
tr.v. o·mit·ted, o·mit·ting, o·mits
1. To fail to include or mention; leave out:
a. To pass over; neglect.
weight and predictor value tables for the remaining models
in the paper. Plots comparable to Figure 1 are also omitted but
available upon request.
(15) The 2010 observation is marginally significant at [alpha] =.
1, but this is likely due to chance. As an additional check, I evaluated
the statistical significance of the average effect over the entire
post-treatment period from 300% to 400% FPL of -.21% by generating
sampling distributions over four observations as done earlier during the
resampling test for 400% to 500% FPL, and estimate a standard error of
.28% (t=.75) for full-period effects. Hence, I reject the possibility of
a coverage effect across the entire post-treatment period at the 300% to
400% FPL income level at standard levels of significance testing. In
repeating this test across all estimates (i.e., full, private, and
public coverage rates at all income levels), I typically find that the
standard errors for the full period are 50% to 100% smaller than those
estimated for single observations alone. Standard errors averaged over
four post-treatment observations are available upon request.
(16) I also note that my inability to generate a public coverage
estimate for 400% to 500% FPL does not
tr.v. in·val·i·dat·ed, in·val·i·dat·ing, in·val·i·dates
To make invalid; nullify.
estimates of total
coverage at that income level, shown earlier in Figure 1. For example,
both the Survey of Income and Program Participation and the Panel Study
of Income Dynamics are not representative within states, but are
nationally representative survey samples commonly used in health
(17) For public coverage between 400% and 500% FPL, I present
estimates assuming coverage stayed constant at 2006 levels after
treatment. Most likely, this estimate of the magnitude of the effect of
coverage is overestimated, as the
The curve described by a body moving through space, as of a meteor through the atmosphere, a planet around the Sun, a projectile fired from a gun, or a rocket in flight.
of public coverage shown in
Figure 3 was already trending upwards even before the introduction of
(18) With the exception of 2010, the private coverage reductions
estimated for 200% to 300% FPL are not individually significant.
However, in testing the joint significance of the average change over
the full post-treatment period, I find that the average post-treatment
estimate of -1.31 percentage points is marginally significant (t=-1.88).
(19) For example, Cutler and Gruber (1996) estimated early
crowd-out rates around 31% between 1987 and1992, while Gruber and Simon
(2008) obtained crowd-out estimates around 60% using similar models for
the 1996-2002 period.
(20) Neighboring states for the control group include Indiana,
Iowa, Kentucky, Missouri, and Wisconsin. Only data from these states and
Illinois were used in the DID regressions.
(21) Although DID using adjacent states as a control group yields
similar estimates to the synthetic model, it is worth noting that the
predictors of child health insurance coverage shown in Table 1 in the
five adjacent states deviate somewhat from Illinois. In general, the
averages of these five states have slightly higher rates of pre-2006
health insurance coverage, moderately lower rates of education,
significantly fewer minorities, and lower income than Illinois.
(22) Data for these analyses comes from Medical Expenditure Panel
Survey Insurance Component, tables VII.D.1, 2004, 2006, 2008, available
at http://www.meps.ahrq.govl mepsweb/data_stats/MEPSnetlC.jsp.
James Lo, Ph.D., is a research fellow at the Center for the Study
of Political Economy of Reforms at the
University of Mannheim
Germany. Email Dr. Lo at: email@example.com
Table 1. Predictors of child health insurance coverage, for children with family incomes at 400% to 500% FPL Illinois Variable Real Synthetic Child health coverage, below FPL (%) 80.60 80.89 Child health coverage, 100%-200% FPL (%) 81.03 81.17 Child health coverage, 200%-300% FPL (%) 89.50 89.42 Child health coverage, 300%-400% FPL (%) 92.62 92.54 Child health coverage, 400%-500% FPL (%) 94.25 94.22 Percent white 80.2 82.8 Mean income, U.S. dollars 36,434.81 36,335.81 Standard deviation, income 41,722.77 41,067.37 Unemployment rate (%) 5.63 5.62 Percent college educated 49.47 48.78 Control states Standard Variable Mean deviation Child health coverage, below FPL (%) 78.28 6.33 Child health coverage, 100%-200% FPL (%) 79.84 5.60 Child health coverage, 200%-300% FPL (%) 86.96 4.06 Child health coverage, 300%-400% FPL (%) 91.25 2.88 Child health coverage, 400%-500% FPL (%) 93.13 2.43 Percent white 85.1 10.0 Mean income, U.S. dollars 33,330.29 4,037.54 Standard deviation, income 38,049.24 4,973.45 Unemployment rate (%) 5.21 0.83 Percent college educated 47.30 4.98 Notes: All variables are averaged for the 1995-2006 period. Forty-seven states were included in the donor state pool, excluding Alaska, Hawaii, Illinois, and the District of Columbia. Table 2. State weights in synthetic Illinois State Weight State Weight Alabama 0.02 Montana 0 Alaska -- Nebraska 0 Arizona 0 Nevada 0 Arkansas 0 New Hampshire 0 California 0 New Jersey 0 Colorado 0 New Mexico 0 Connecticut 0 New York 0 Delaware 0.16 North Carolina 0 District of Columbia -- North Dakota 0 Florida 0 Ohio 0 Georgia 0 Oklahoma 0 Hawaii -- Oregon 0.25 Idaho 0 Pennsylvania 0 Indiana 0 Rhode Island 0.22 Iowa 0 South Carolina 0 Kansas 0 South Dakota 0 Kentucky 0 Tennessee 0 Louisiana 0 Texas 0 Maine 0 Utah 0 Maryland 0.19 Vermont 0 Massachusetts 0 Virginia 0 Michigan 0.12 Washington 0 Minnesota 0 West Virginia 0 Mississippi 0 Wisconsin 0 Missouri 0 Wyoming 0 Notes: All weights were rounded to two significant digits, so they may not sum exactly to 1 as presented here. The results suggest that a synthetic Illinois is best fitted using largely a convex combination of Delaware, Maryland, Michigan, Oregon, and Rhode Island. Table 3. Summary of percentage point changes in total insurance coverage in Illinois after All Kids expansion Changes (percentage point) FPL 2007 2008 2009 2010 200%-300% 2.06 2.49 2.40 2.70 (0.64) (0.64) (0.64) (0.64) 300%-400% 0.13 -0.08 -0.25 -0.63 (0.35) (0.35) (0.35) (0.35) 400%-500% 2.11 3.05 3.31 3.36 (0.29) (0.29) (0.29) (0.29) Notes: Jackknife standard errors are in parentheses. Numbers represent percentage point increases in net coverage by year, calculated as the difference in coverage between the observed and synthetic control. Table 4. Summary of changes in public insurance coverage in Illinois after All Kids expansion Changes (percentage point) FPL 2007 2008 2009 2010 200%-300% 2.57 4.72 5.94 6.54 (1.68) (1.68) (1.68) (1.68) 300%-400% 0.29 1.61 2.43 3.59 (0.66) (0.66) (0.66) (0.66) 400%-500% 0.69 1.31 1.64 1.58 (--) (--) (--) (--) Notes: Jackknife standard errors are in parentheses. Numbers represent percentage point increases in public coverage by year, calculated as the difference in coverage between the observed and synthetic control. Estimates for 400% and 500% FPL are derived naively by assuming that 2006 coverage levels remain constant. Table 5. Summary of changes in private insurance coverage in Illinois after All Kids expansion Changes (percentage point) FPL 2007 2008 2009 2010 200%-300% 0.43 -1.27 -2.13 -2.27 (1.24) (1.24) (1.24) (1.24) 300%-400% -0.35 -1.76 -2.52 -3.72 (0.68) (0.68) (0.68) (0.68) 400%-500% 0.67 1.48 1.96 2.62 (0.78) (0.78) (0.78) (0.78) Notes: Jackknife standard errors are in parentheses. Numbers represent percentage point increases in private coverage by year, calculated as the difference in coverage between the observed and synthetic control. Table 6. Summary of crowd-out estimates in Illinois after All Kids expansion Synthetic estimates (%) Using DID approach (%) [DELTA] [DELTA] [DELTA] FPL Coverage Crowd-out Public Private 200%-300% 2.7 34.8 11.5 -5.68 (0.64) (1.42) (1.67) 300%-400% -0.63 -100 3.03 -2.12 (0.35) (1.23) (1.51) 400%-500% 3.36 Crowd-in 0.76 2.34 (0.29) (1.21) (1.57) Using DID approach (%) [DELTA] FPL Coverage Crowd-out 200%-300% 5.78 49.6 (2.19) 300%-400% 0.9 70 (1.94) 400%-500% 3.10 Crowd-in (1.99) Notes: Standard errors are in parentheses. For synthetic results, change in coverage is calculated as percentage point differences between observed and synthetic unit in 2010. Crowd-out calculated as -[DELTA]private/[DELTA]public. Each private and public coverage estimate comes from a separate DID (difference in differences) regression, using data from 2003 to 2010, and states adjacent to Illinois as a control group. Net coverage results under DID are the sum of the changes in private and public coverage. Results between the synthetic estimate and DID are largely consistent with each other. Table 7. MEPS-IC growth rates in premiums for families by payroll quartile Change in growth (%) 2006-2008 2004-2006 2006-2008 minus 2004-2006 Fourth payroll quartile Illinois 16.6 3.8 -12.8 U.S. 15.1 7.1 -8 Difference 1.5 -3.3 -4.8 All payroll percentiles Illinois 13.7 7.0 -6.7 U.S. 13.7 8.1 -5.6 Notes: Standard errors are not available. The difference-in-differences analysis presented here suggests that All Kids reduced insurance premiums for residents in the fourth payroll quartile by 4.8%.