Early coverage expansions under the Affordable Care Act and Supplemental Security Income participation

Pinka Chatterji University at Albany and NBER

Yue Li University at Albany

8/24/17

Keywords: Supplemental Security Income, SSI, disability, Affordable Care Act, ACA, Medicaid, health insurance

Acknowledgements and Notes: The authors thank participants at the 2016 ASHEcon meeting and the 2016 APPAM fall conference, particularly our discussants Ben Cook and Jody Schimmel Hyde, for helpful comments. The authors also thank seminar participants at the 2016 University at Albany Health Economics Summer Work Group for helpful comments and feedback. A working paper version of this paper is published as National Bureau of Economic Research Working Paper w22531.

*Correspondence to: Pinka Chatterji, Department of Economics, University at Albany, SUNY, Albany, NY 12222, USA. Telephone: 518-442-4746, Fax: 518-442-4736, E-mail: [email protected].

1

ABSTRACT We test whether early Affordable Care Act (ACA) Medicaid expansions in Connecticut (CT), Minnesota (MN), California (CA), and the District of Columbia (DC) affected Supplemental Security Income (SSI) receipt. We use a synthetic control approach, comparing the SSI receipt rate pre and post each early Medicaid expansion (“Early Expanders”) to the weighted receipt rate in states that expanded Medicaid in January 2014 (“Later Expanders”). In CT, the Medicaid expansion is associated a statistically significant, 7 percent reduction in SSI beneficiaries among individuals aged 18-64. For DC, MN and CA, we do not find consistent evidence that the Medicaid expansions affected disability-related outcomes.

2

1. Introduction The United States has an employment-based health insurance system for the working-age population, and this system has both advantages and disadvantages for people with disabilities (Gruber, 2000). The main advantage is the pooling of risks, as employers have the potential to bring together large groups of people for reasons unrelated to their health status. This pooling of risks reduces adverse selection and spreads administrative costs over large groups, lowering the price of insurance (Gruber & Madrian, 1993). This is important for individuals with disabilities who work and have access to employer-sponsored health insurance. One disadvantage of an employment-based system, however, is that individuals with disabilities who cannot work full-time may not have access to private health insurance through an employer. These individuals also are likely to face cost-related barriers and, until recently, were likely to encounter waiting periods and preexisting conditions exclusions in purchasing private insurance directly from an insurance company. Even when affordable private insurance plans are available, these plans typically have not covered all the services that individuals with disabilities need, such as therapy and long-term care (The Arc, 2012; Peele et al., 2002). For these reasons, some individuals with disabilities may turn to the federal disability programs – Supplemental Security Income (SSI) and Social Security Disability Insurance (DI) – not just for the income support these programs provide, but also because these programs in some cases confer public health insurance eligibility. The 2010 Patient Protection and Affordable Care Act (ACA) addressed some of the structural barriers that individuals with disabilities face in obtaining insurance coverage. In the private insurance market, the ACA built on prior state and federal reforms by eliminating insurance waiting periods and denials of coverage for preexisting conditions; banning annual and lifetime dollar limits on benefits; and making it illegal for insurance companies to rescind insurance coverage when a beneficiary becomes ill (Kaiser Family Foundation, 2012a). The ACA’s Health Insurance Marketplaces also are expected to

3

offer more affordable and comprehensive private coverage options for people with disabilities compared to what was available previously (The Arc, 2012). In the public health insurance market, the original goal of the ACA was to broaden Medicaid eligibility to include all adults with incomes below 138% of the Federal Poverty Level (FPL) by January 2014. Although this policy did not target adults with disabilities specifically, it may have opened an important new avenue to accessing public insurance among low-income, childless adults with disabilities, most of whom previously had no access to public health insurance programs other than through the federal disability programs. This is particularly true in states that previously had Medicaid programs with low income-eligibility thresholds, and Medicaid programs that previously did not cover any childless adults (Kennedy & Blodgett, 2014). In earlier Medicaid expansions that targeted lowincome, childless adults, the newly covered beneficiaries tended to resemble the SSI population in the sense that they had high levels of health needs, including high rates of behavioral health disorders (Sommers et al., 2010). Thus, it is likely that the ACA’s expansions, while not specifically targeted at individuals with disabilities, will tend to attract a large number of new beneficiaries who have debilitating and chronic health conditions. If Medicaid expansions opened a new pathway to public insurance for adults with disabilities, then these expansions may have reduced federal disability caseloads, if obtaining public insurance was a primary motivation for some individuals to apply for disability benefits. Similarly, if Medicaid expansions improve health among adults with disabilities, these individuals may be more likely to be able to continue working and stay off or exit the federal disability programs. On the other hand, if lowincome people with high levels of health need tend to migrate to or avoid moving away from states that expanded Medicaid, this increase in the low-income, relatively unhealthy population is likely to induce increases in the numbers of disability recipients. Because of these potentially counter-veiling effects, the net effect of Medicaid expansions on receipt of federal disability benefits remains an empirical question. 4

The proportion of the US population receiving disability benefits, and the costs of federal disability programs, have risen sharply in recent decades. One of the primary causes of these increases is the aging of the workforce.1 Liebman (2015) estimates that individuals aged 50 to 64 years old are four times more likely than those aged 20 to 49 to be participating in federal disability programs. With the backdrop of an aging US population, it is therefore highly policy-relevant to consider how public insurance coverage expansions under the ACA affect federal disability programs. In this paper, we test whether early ACA coverage expansions in Connecticut (CT), Minnesota (MN), California (CA), and the District of Columbia (DC) affected the number of SSI beneficiaries. We also examine the mechanisms that potentially link coverage expansions to SSI receipt, including health, insurance status, and population changes. We use a synthetic control approach, in which we examine the number of SSI beneficiaries before and after the early coverage expansion in each of the four states, utilizing a weighted combination of states that expanded Medicaid on 1/1/14 as a comparison group. For CT, the expansion is associated a statistically significant, 7 percent reduction in SSI beneficiaries aged 18-64, but for DC, MN, and CA, we do not find evidence that the early coverage expansions affected SSI receipt.

2. Background Medicaid is a joint federal and state-funded public health insurance program that, until recently, primarily covered low-income children and parents, the elderly, the blind, and people with disabilities. The ACA provided federal funding for states to expand Medicaid to adults with incomes below 138 percent of the FPL (Bitler & Zavodny, 2014). As of March 2016, 31 states and DC have adopted the ACA expansion of Medicaid, while the remaining states have not yet done so (Kaiser Family Foundation, 2016). As the ACA’s Medicaid expansions unfold in many states across the U.S., millions

Other causes include changes in the eligibility criteria, the changing composition of the labor force, and changes in incidences of various disabling conditions (Liebman, 2015). 1

5

of low-income adults are gaining eligibility for public health insurance for the first time. This major health policy change is likely to have implications for health insurance coverage, labor market outcomes, health outcomes, and financial outcomes among low-income people, and recent studies have started to document these effects (Leung & Mas, 2016; Courtemanche et al. 2017; Hu et al., 2016; Kaestner et al, 2015; Frean et al., 2016; Sommers et al. 2010, 2013, 2014, 2016). The ACA Medicaid expansions, however, also may have important effects on the number of people applying for and receiving benefits from SSI. Before the ACA, one main route through which low-income adults with disabilities could access public health insurance was through SSI. Other ways in which adults with disabilities potentially can access Medicaid include: (1) the State Supplemental Program (SSP), which augments the SSI program in most states (potentially covering some individuals with disabilities whose incomes are too high to be eligible for SSI) and may be linked to Medicaid; (2) state Medicaid programs for low-income individuals with disabilities and for the medically needy populations; (3) state buy-in programs for individuals with disabilities (see Wagner, 2015 for more details). The SSI program provides monthly cash benefits for individuals who are low-income and have a work-limiting disability. For most recipients, SSI is their only source of income (Center on Budget and Policy Priorities, 2014). The program is funded through general tax revenues, and only people with disabilities as well as low-income and low assets are eligible (Social Security Administration, 2006). SSI applicants must go through a disability determination process to verify that it is not possible for them to engage in substantial work (Center on Budget and Policy Priorities, 2014). As of June 2017, there were 4.8 million adult, nonelderly SSI recipients receiving an average monthly payment of $565 (Social Security Administration, 2017). Many states have linked Medicaid eligibility to SSI eligibility, at least to some extent. States fall into three categories in this regard. In the first category, called “1634 states,” SSI eligibility confers Medicaid eligibility; the two programs share the same application, and Medicaid eligibility starts in the 6

same month as SSI eligibility. Currently, 33 states and DC fall into the 1634 category. Note that CA and DC, two of our analysis states, are 1634 states. In the second category, called “SSI criteria states,” the SSI and Medicaid programs have the same eligibility criteria, but have separate application processes (AK, ID, KS, NE, NV, OR, and UT fall into this category). In the third category, termed “209(b) states,” the Medicaid eligibility rules and application process differ from that of the SSI program. At the present time, HI, IL, IN, MO, NH, ND, OH, OK, VA as well as two of our analysis states CT and MN fall into this category (Social Security Administration, 2017). In the 209b states, there is at least one eligibility criterion for Medicaid that is more restrictive than that of SSI (Social Security Administration, 2014).2 Medicaid expansions may affect SSI participation through several mechanisms. First, Medicaid expansions may reduce SSI caseloads if accessing public health insurance is a primary motivation for many existing and potential federal disability beneficiaries. The expansion of Medicaid to low-income, childless adults weakens the linkage between federal disability programs and public health insurance, at least for some individuals with disabilities who may choose to continue working and not apply for disability benefits if they can access Medicaid without doing so. Second, Medicaid expansions may reduce SSI caseloads if the expansions improve health, allowing for individuals with disabilities to continue working. For example, in the Oregon Medicaid Experiment, access to Medicaid was associated with improvements in mental health (Finkelstein et al. 2012); if these types of effects on health are substantial, they could lead to reductions in federal disability caseloads. We acknowledge, however, that these kinds of effects may not be large enough to affect SSI participation in the shortterm.

In CT and MN, the Medicaid income and asset limitations are more restrictive compared to those of the SSI program. See http://www.nolo.com/legal-encyclopedia/connecticut-disability-benefits-social-security-filinginsurance-options.html and http://www.nolo.com/legal-encyclopedia/minnesota-disability-benefits-socialsecurity-filing-insurance-options.html. 2

7

On the other hand, Medicaid expansions may affect SSI participation through their impact on migration decisions. If a state’s Medicaid expansion attracts new migrants to the state, or reduces outmigration from the state, there is likely to be an increase in the SSI caseload. Schwatrz & Sommers (2014) test whether there was in-migration associated with the Medicaid expansions that took place in Arizona and New York in 2001 and in Maine in 2002, and health care reform in Massachusetts in 20062007. They find no evidence that these policy changes were associated with migration among lowincome adults. We lack evidence regarding whether the more recent expansions under the ACA may have affected migration. There are large literatures examining the effects of prior Medicaid expansions and policy changes, as well as many studies focusing on the effects of policies related to the federal disability programs. Only a few recent studies, however, focus specifically on how expansions in access to Medicaid affect federal disability program participation. Maestas et al. (2014) test whether the 2006 state health care reform in Massachusetts, which reduced uninsured rates in the state by 48 percent, affected participation in federal disability programs. They use a difference-in-difference (DD) study design, comparing the change in the rate of federal disability applications before and after the policy change in Massachusetts to this same change before and after the policy change in other Northeast states. Notably, these authors can examine heterogeneous effects by county. They find that in counties with relatively high insurance coverage rates, state health care reform in MA was associated with an increase in federal disability applications, supporting the idea that Medicaid expansions reduced “employment lock” in populations likely to have employer-sponsored health insurance coverage. In counties with relatively low insurance coverage rates, however, the MA health care reform was associated with a decline in federal disability applications, suggesting that at least some individuals in low-insurance counties were seeking disability benefits primarily to obtain Medicaid coverage.

8

In a recent paper, Burns & Dague (2017) use DD methods to test whether state Medicaid expansions targeting low-income, childless adults between 2001 and 2013 are associated with SSI participation. This paper is notable in that it draws on comprehensive data regarding these expansions, including information on enrollment caps and waiting lists (Burns et al. 2017). The classification of states’ Medicaid expansions, however, differs from our paper, as we discuss below. Burns & Dague (2017) report that on average Medicaid expansions for childless adults are associated with a 7 percent reduction in SSI participation among childless adults. This finding supports the idea that some federal disability beneficiaries may seek benefits primarily for access to Medicaid. We build on these two recent studies by testing whether recent Medicaid expansions under the ACA are associated with federal disability caseloads and associated mechanisms. We focus on some of the earliest ACA Medicaid expansions, which took place before the broader ACA expansion of Medicaid in January 2014. Five states - CT, New Jersey (NJ), MN, Washington (WA), CA - and DC expanded their Medicaid programs in 2010 and 2011. These early expansions were an option for states under the ACA or states initiated the expansions using federal 1115 waivers. The early expansions differed from later ACA Medicaid expansions in 2014 in several ways. First, income eligibility under the early expansions could vary from 138% of the Federal Poverty Level (FPL), while the later Medicaid expansions applied to individuals with family income <=138% FPL. Second, the traditional state match rate (which varies by state) applied for the early expansions, while the Federal government matched later Medicaid expansions at 100%. Finally, all of the early ACA Medicaid expansions replaced or enhanced state or local health insurance programs that already had existed for low-income adults; this was not necessarily the case for the later expansions (Sommers et al., 2014). Nearly 950,000 individuals enrolled in Medicaid between the time of these early expansions and January 1, 2014 (Kaiser Family Foundation, 2015). In CT, MN, CA and DC, there is evidence that these early expansions resulted in a substantial increase in the number of individuals receiving public insurance (Sommers et 9

al., 2014). In NJ and WA, however, restrictive eligibility requirements and/or lower take-up rates reduced the impact of the expansions (Sommers et al., 2013). Thus, we focus on early Medicaid expansions in CT, MN, CA and DC in this paper. The expansions that we study are somewhat different than those studied by Burns & Dague (2017), in that the ACA early expansions were built on previous state-run programs, transferring large numbers of beneficiaries from state-run programs to Medicaid, and there were improvements over the previously state-run programs, such as more comprehensive coverage and expanded provider networks (Sommers et al., 2013). For example, Burns & Dague (2017) consider the 2007 1115 waiver expansion in CA, initiated in 10 counties with a maximum income eligibility of 200% FPL (Burns et al. 2017). In this paper, we focus on the 2011 expansion in CA, which built on the 2007 expansion by enrolling more people (the 2007 expansion had an enrollment cap), and by covering a much broader array of health services (Sommers et al., 2016; Burns et al. 2017). Burns & Dague (2017) consider the 2008 expansion in CT, which had low enrollment and significant cost-sharing, while we consider the 2010 expansion, which offered close to full Medicaid benefits without cost-sharing (Burns et al. 2017). In DC, Burns & Dague (2017) examine a 2001 expansion with limited benefits, while we consider the 2010 expansion to full Medicaid benefits. Finally, in MN, we study a transfer from a fully state-funded public expansion (which existed throughout the study period in Burns & Dague, 2017) to a jointly funded 1115 waiver program in 2011 (Burns et al. 2017). Table 1 summarizes the details of the Medicaid expansions in the four states we study. The expansions took place in 2010 for DC and CT, and in 2011 for MN and CA. All four of these early expansions were based on pre-existing state insurance programs for low-income people, but there was considerable heterogeneity across the expansions; for this reason, it makes sense to examine each state’s expansion separately. In DC, the expansion included the population up to 200% of the FPL, while in CT and MN, the income thresholds for eligibility were 56% and 75% of the FPL, respectively (Sommers et 10

al., 2013). In CA, a new program called the Low Income Health Plan was created for low-income adults up to 200% of the FPL. Individual counties in CA, however, could choose whether or not to participate in the Medicaid expansion, and participating counties could choose their own income thresholds. Between July 2011 and March 2013, 52 of the 58 counties in CA chose to expand Medicaid (Golberstein et al., 2015). Sommers et al. (2014), using difference-in-difference methods with neighboring states as the comparison groups, finds that the early Medicaid expansions are associated with increases in Medicaid enrollments in CT and DC in the targeted low-income populations, particularly among childless adults with health limitations. Similarly, Sommers et al. (2016), using difference-in-difference methods with counties that did not expand Medicaid as the comparison group, find that the CA Medicaid expansion was associated with an increase in Medicaid coverage. We build on these studies by examining effects on a different set of outcomes – federal disability outcomes – and by using synthetic control methods, which perform better than difference-in-difference methods in our case. 3. Data Our main dependent variable is the percent of the state’s adult nonelderly population receiving SSI benefits, which is measured at the state-year level for the time period 2000-2013. This variable was constructed by dividing the number of SSI beneficiaries aged 18-64 in the state in December each year by the state-level population estimate for the age group 18-64 as of July 1 that year, multiplying by 100.3 SSI data come from the Social Security Administration (SSA 2000-2013). Population estimates for each year come from the US Census Bureau. We also examine several dependent variables that may be mechanisms linking coverage expansions to SSI participation. These variables are: percent of state’s low-income 18-64 year old

3

We also considered SSI applications/awards as outcomes, but found that the synthetic control method did not work well because these data fluctuate a lot from year to year. For these outcomes, we could not achieve a reasonable degree of matching between the synthetic control and the treatment state in the pre-intervention time period for all four states.

11

population that is uninsured, the percent with ESHI, and the percent with public insurance; and percent of state’s population 18 years old and older in fair or poor health. We define low-income as adults living in households with income below 200% of FPL. The cut-off of 200% is chosen because this is the highest income threshold used in the early expansions. The health insurance data are constructed from the American Community Survey (ACS), and are available for 2008-2013 only. The health data come from the Behavioral Risk Factor Surveillance System. Using data from the Census, we create a measure of population growth for each treatment state in each year. This measure is: the state’s 18-64 year old population in the current year divided by the 18-64 year old population in the year 2000, multiplied by 100. To create a measure of net migration of low-income individuals into a state, we calculate the number of low-income individuals migrating into the state (from other states and abroad) minus those migrating out of the state to other states (we do not have information on those who migrate abroad) from the ACS (2000-2013). We divide this number by the state population in 2000, and multiply by 100. The controls are: the percent of state population aged 45-54; the percent of state population aged 55-64; the percent of state population aged 25+ with a bachelor’s degree; the percent of state with income at or below the FPL; and the unemployment rate. These five control variables are constructed from the ACS based on the age group 18-64.4 4. Methods We use the synthetic control method, introduced by Abadie et al. (2010, 2015), to test whether early ACA coverage expansions affected SSI participation. In our context, the synthetic control method

The goal is to construct a synthetic control we can closely match to the treatment unit in terms of the predictor variables that have the most predictive power on the outcome. Selecting the best set of controls is a process of trial and error. In our case, we found that the match was best when we used a small set of controls since lagged values of the outcome variables were the best predictors. As a sensitivity check to see if our findings were sensitive to adding more controls, we tried adding the following covariates: percent female in state, percent married in state, percent white in state, and percent Latino in state. The results are qualitatively similar to those shown in the paper, but the estimated effect for CT becomes statistically insignificant. These results are available upon request. 4

12

involves generating a synthetic comparison group for each treatment state. The synthetic comparison group is a weighted average of states that did not expand Medicaid early that matches the pre-expansion treatment state’s characteristics as closely as possible. As Abadie et al. (2010, 2015) discuss, the advantages of the synthetic control approach over regression-based difference-in-difference methods include: (1) it provides clarity regarding the relative importance given to each member of the comparison group; (2) it makes explicit how well the treatment group’s outcomes and other characteristics match those of the comparison group during the pre-policy time period; and (3) it avoids extrapolation, since the weights assigned to each member of the comparison group are positive in sign and must sum to one. The heart of the synthetic control method lies in using a weighted average of comparison units, rather than a single comparison unit or a simple weighted combination of comparison units, as a better approximation of what would have happened to the treatment group in the absence of the treatment. The goal is to use longitudinal, pre-policy data to create a weighted average of non-treated units that best matches the treated unit during the pre-policy time period. Abadie et al. (2015) motivate this idea by considering a sample which includes J + 1 units (for example, states). The unit j = 1 is the treated unit (for example, an Early Expander state) while units j = 2 to j = J+1 are the potential comparison units, called the “donor pool.” The goal is to construct a synthetic control that proxies what would have happened in the treatment unit had the unit not been treated. Thus, Abadie et al. (2010, 2015) recommend that the donor pool be selected carefully such that it is: (1) restricted to units with characteristics similar to those of the treated unit; (2) restricted to units that have not experienced a similar intervention as the treated unit; and (3) restricted to units that have not experienced any large, idiosyncratic shocks to the outcome variable during the study period (Abadie et al. 2010, 2015). In this spirit, we limit the donor pool to the 15 Later Expanders that expanded Medicaid on January 1, 2014

13

(AR, IL, IA, KY, MD, NV, NJ, NM, ND, OH, OR, RI, WA, WV, WI).5 In these Later Expanders, we expect that the political environments in support of coverage expansions would be more similar to those of the Early Expanders; in this sense, this group makes the most appropriate comparison group for the synthetic control method. We exclude from the donor pool the states that had earlier, extensive reforms (MA, NY, AZ, CO, DE, HI,VT), as well as states that either expanded Medicaid after January 1, 2014 (AK, IN, MI, NH, PA, MT), or have chosen not to expand Medicaid as of January 2016 (AL, FL, GA, ID, KS, LA, ME, MS, MO, NE, NC, OK, SC, SD, TN, TX, UT, VA, WY).6 Assuming that all units are observed for time periods t = 1,….T, the number of pre-intervention time periods is T0, and the number of post-intervention time periods is T1, so that T = T0 + T1+1 (adding the year of the intervention). In our case, the data span 2000 to 2013; the pre-intervention period is 2000-2009 for CT and DC, while for CA and MN, the pre-intervention period is 2000-2010. Abadie et al. (2015) propose that a synthetic control can be represented by a (Jx1) vector of weights W = (w2, …….. wj+1)’ with 0 ≤ wj ≤ 1 for j = 2, …., J and w2 +…….+ wj+1 = 1. In choosing W, the goal is to match pre-intervention characteristics of the synthetic control group as closely as possible to the treatment group. If X1 is a (k x 1) vector containing pre-intervention characteristics of the treated unit, and X0 is a (k x J) matrix containing these same characteristics for the donor pool, then the optimal synthetic control W* is the value of W that minimizes: ∑

(

−

) Eq. 1

5

Although we tried to exclude states from the donor pool that experienced similar interventions, there may have been relevant changes to Medicaid during the pre-policy period in these 15 states; this is a limitation. To gauge the sensitivity of our findings to the choice of our donor pool, we re-estimated the main specification with two alternative choices for the donor pool: (1) donor pool includes all states that expanded Medicaid as of 1/1/17; and (2) donor pool includes all states. The alternative donor pools exclude the four treatment states and MA. We compared the fit in the pre-period, and the predicted response using the alternative donor pools. In three of the four states, the fit gets worse as we expand the donor pool to include more states. The general pattern of findings is the same when we use the alternative donor pools, although the effect for CT loses statistical significance if we expand the donor pool to states that expanded as of 1/1/17 and all states. These findings are available upon request. 6

14

In Equation 1 above, X1m represents the value of the mth variable (m= 1,…k) for the treated unit and X0m represents the (1 x J) vector containing the values of the mth variable for the donor pool units. The pre-intervention characteristics used in this analysis are the five control variables (see Section 3) and the outcome variable. The term υm captures the weight given to the mth characteristic. These weights are based on how well each characteristic predicts the outcome, with larger weights placed on characteristics that have more predictive power. Finally, letting Y1 be a vector of post-intervention outcomes for the treatment group and letting Y0 be a matrix of post-intervention outcomes for the donor pool, the synthetic control estimator of the effect of the treatment is the difference Y1 - Y0W*. Essentially, the synthetic control estimator involves comparing the outcome in the treated unit (which was exposed to the policy intervention) to the synthetic control (not exposed to the policy intervention) during the post-intervention period. Abadie et al. (2010), using a linear factor model, argue that matching on pre-intervention values of the outcome variables and drawing on a long time series of pre-intervention data control for the effects of unmeasured factors and for possible heterogeneity in the effects of measured and unmeasured factors on outcomes. Notably, unlike the standard difference-in-difference (DD) model, the linear factor model allows for time-varying unmeasured factors; this is an advantage of the synthetic control approach over the standard DD framework, which requires that any unmeasured factors be constant over time so that they can be “differenced-out.” The idea behind the synthetic control method is that if trajectories of the outcome variable and the synthetic control are similar over an extended period of time before the intervention, then the measured and unmeasured determinants of the outcomes, as well as the effects of those determinants on the outcomes, should be similar as well (Abadie et al., 2010, 2015). The success of the method hinges on how well the trend in outcomes in the synthetic control matches that of the treatment unit during the pre-intervention time period. One can gauge the quality of the matching by comparing the predictors of the outcome, as well as the trajectory of the outcome itself, between the 15

synthetic control and the treated unit during the pre-intervention period. The synthetic control method would not be appropriate in cases in which the treatment unit and the synthetic control do not match well in the pre-intervention time period. Following Abadie et al. (2010), we conduct inference in the following manner. First, we use the synthetic control method and compute treatment effects, as described above. Next, we conduct a falsification exercise for each Early Expander state. To do so, we take each of the 15 Later Expander states in the donor pool one by one, treat each of these states as if it were an Early Expander, and recalculate synthetic control estimates. The synthetic control method works better in terms of matching on pre-intervention characteristics for some states versus other states. That is, the mean square prediction error (MSPE) in the pre-intervention period is lower for some states, in which the matching worked well, versus other states, in which the matching did not work as well.7 Thus, what is relevant is the ratio of the MSPE during the post-intervention period to the MSPE during the pre-intervention period, since this ratio provides information regarding how closely the treatment group’s outcomes matched the synthetic control post-intervention, relative to how well the matching worked pre-intervention. A relatively high MSPE ratio indicates worse matching between the treated state and the synthetic control in the post-intervention period relative to what would be expected given the quality of the matching in the pre-intervention period, reflecting that there is an effect of the intervention on the outcome. The null hypothesis being tested is the MSPE ratio for the true treated state falls within the distribution of the MSPE ratios for the false treated states. The p-value is calculated based on the MSPE ratio for each treated state and for each state in the donor pool when it is falsely assigned as the treated

MSPE = 1/ for MN.

7

∑!"

− ∑

∗



#. The MSPE =0.0002 for CT, 0.02 for DC, 0.004 for CA, and 0.006

16

state. The p-value is the number of states with a MSPE at least as high as the true treated state plus one (to include the treated state) divided by 16, which is the number of donor states plus the treated state.8 When analyzing the mechanisms, we use difference-in-difference (DD) methods because the health insurance data were only available for the time period 2008 onwards, making the pre-policy period too short to yield trustworthy synthetic control estimates. In Equation 2 below, yst, is the outcome variable for state s in year t. The independent variables include: an interaction between Early Expander and year in which the policy was implemented; an interaction between Early Expander and all years after the year of implementation (“Post”); a set of state-specific time-varying characteristics (Xst) – the five control variables described in Section 3; a set of year dummy variables; and state fixed effects. We consider each Early Expander state in a separate model. In Equation 2, $ reflects the anticipatory and first-year effects of the coverage expansion in a particular Early Expander state, while % captures the average impact from the second year and onward.

&' = )*+,-.+- + $(0.12&_045.+671' 7.1_8952797+- ) +%(0.12&_045.+671' :*,- ) +

'

; + 7.1 + <-.-7' + =' Eq. 2

We estimate Equation 2 using linear probability models weighted by state population from the 2000 Census, with Huber-White corrected standard errors adjusted for by clustering at the state level (Bertrand et al, 2004).

8

Limiting the donor pool to Later Expanders reduces power, since we have only 15 states for the falsification exercises. There appears to be a tradeoff between power and the quality of the match in the pre-period. For example, in CT, the MSPE for the pre-period is 84 percent smaller when only Later Expanders are included in the donor pool than the specification wherein we included all states.

17

5. Results First, we gauge whether the synthetic control method improved the quality of the matching during the pre-policy period. Table 2 shows pre-policy period means for each Early Expander state, the average of the 15 control states, and the synthetic control. The synthetic control substantially improves the quality of the matching. This is particularly true for the outcome variable in the pre-period because this variable is weighted heavily due to its importance as a predictor. In CT, the match is excellent for all of the characteristics. In the other three states, the match is good on all characteristics except for the percent with college degree and percent living in poverty in DC, which are considerably higher in DC versus its synthetic control. The match on variables was not good since DC has an unusual combination of a highly educated population but a relatively large fraction of the population living in poverty. Given the somewhat low quality of the matching for DC, we interpret the findings for this treatment state with caution.9 Figure 1 shows that the pre-period match between the synthetic control and the treated state is good for all states, and especially for CT.10 The synthetic control estimate of the effect of the ACA early coverage expansion on SSI receipt is the average difference between the synthetic control and the treatment state during the post-policy period. In CT and CA, there is a decline in the outcome relative to the synthetic control during the post policy period, suggesting that the expansion is associated with reductions in SSI recipients. In DC and MN, however, we see the opposite pattern; after the expansion took place, the trend in the percent of state receiving SSI is higher in the treatment state vs. the synthetic

Taking the average of the whole pre-policy time period may be problematic given the timing of the Great Recession. We tried breaking the per-period into two sub periods: before the Great Recession (2000-2006) and after the Great Recession (2007+) and using the average of the outcome variable for both sub-periods as predictors. Using this approach, we generate results that are qualitatively similar to those shown in the paper. Results available upon request. 10 Appendix Table 1 shows the synthetic control weights generated for each treatment state. In terms of the weights on the predictor variables, the most important predictor was the lagged outcome variable. Therefore, the composition of the synthetic control was largely based on which states had a similar proportion of people receiving SSI in the prior period as the treatment state. This is why the composition of the synthetic control is not necessarily based on characteristics such as geographic location and demographics. 9

18

control, suggesting that the expansion increased SSI receipt. We also note that in all four states there appears to be a slight divergence between the treated state and the synthetic control just before the expansion. This divergence may possibly reflect anticipatory effects. Table 3 displays the average of the difference in the mean SSI receipt between each treated state and its synthetic counterpart during the entire post-policy period. In CT, the expansion is associated with a marginally statistically significant, 0.11 percentage point reduction in SSI recipients, which is a 7 percent reduction at the pre-expansion mean. This magnitude is the same as what Burns & Dague (2017) report for SSI receipt (a 7 percent reduction), although these authors focus on a different set of expansions targeting low-income, childless adults. Similarly, in CA the expansion is associated with a 0.16 percentage point reduction in SSI recipients. In DC and MN, on the other hand, the expansions are associated with 0.50 and 0.24 percentage point increases in SSI recipients. In CA, DC, and MN, however, these associations are not statistically significant at conventional levels. Early ACA coverage expansions may reduce SSI receipt because now low-income individuals with medical conditions can access public insurance without receiving SSI benefits, or perhaps because having access to public insurance improves health. On the other hand, the expansion may have increased SSI recipients because the expansion changed migration patterns, attracting a less healthy lowincome population compared to the current population. In Table 4, we examine the plausibility of these mechanisms. Panels A-C in Table 4 show findings from DD models of insurance outcomes among lowincome, adults (percent uninsured, percent with ESHI, percent with public insurance), while Panels D-E show results from models of overall health among adults aged 18 and over (Panel D), models of population growth among adults aged 18-64 (Panel E), and models of net migration of low-income adults of ages 18-64 (Panel F). Each column in Table 4 shows models for a particular state – CT, DC, CA or MN.

19

In CT, the expansion was associated with a 1.1 percentage point increase in public health insurance coverage during the year when the expansion took place and a 4.7 percentage point increase in public coverage during the post-policy period. In CT, the expansion also was associated with reductions in ESHI during the year of implementation and the post-policy period, suggesting some degree of crowd-out. The expansion in CT was associated with declines in the uninsured rate, but these associations were not statistically different from zero. In the other three states, most of the associations between coverage expansions and health insurance outcomes were not statistically different from zero, with the exception of the DC expansion being associated with a statistically significant, 5 percentage point decline in ESHI (Table 4, Panels A-C). In all four states, the coverage expansions are associated with reduced probability of being in fair or poor health; these effects are large in magnitude, representing 9-12 percent reductions in fair/poor health at the sample means. These effects are statistically significant in the year of implementation only in CT and DC, but they are statistically significant in both the year of implementation and the postpolicy period in CA and MN. Although our health measure is somewhat crude and measured among all adults (not just low-income adults), this finding is consistent with the idea that improved health may be a mechanism leading from the coverage expansion in CT to the decline in SSI recipients. Finally, we examine DD estimates of the effects of the expansions on population growth and netmigration of low-income adults. In CT, the expansion is associated with a statistically significant increase in net migration of low-income adults during the year of implementation only. In CA, we observe that the expansion is associated with population growth as well as an increase in net migration of low-income adults. These findings are consistent with the idea that the expansion led to increased migration of low-income adults to these states, increasing SSI beneficiaries. We emphasize, however, that the association between the expansion and SSI receipt was not statistically significant in CA, although the sign was negative. In DC, there is no statistically significant association between the 20

expansion and population growth and migration. In MN, the pattern of findings is hard to interpret since the expansion is associated with increased population, but decreased net migration of low-income adults during the post-policy period. These changes in population growth in MN may be related to the coincident oil boom that was occurring in a neighboring state (ND) and attracting young workers.

6. Discussion and Conclusions Given the size, cost, and changing demographic composition of the federal disability caseloads (Social Security Administration, 2016, Center on Budget and Policy Priorities, 2014; CBO, 2016), there has been concern that many individuals with medical conditions who could work instead seek SSI benefits to gain access to Medicaid. The early ACA expansions in CT, DC, CA and MN allow us to test whether broader access to public insurance is associated with receipt of SSI. If access to Medicaid is an important motivation for SSI enrollment, one would expect that expanding and strengthening public insurance would reduce disability caseloads. Since there are other potential pathways linking coverage expansions to disability receipt, however, the relationship between broader public health insurance eligibility and federal disability caseloads is not clear-cut. We examine the association between early coverage expansions and SSI receipt in four states using synthetic control methods recently introduced by Abadie et al. (2010). In CA, MN and DC, our findings do not indicate that early coverage expansions affected SSI receipt. In CT, however, the early coverage expansion was associated with a 7 percent reduction in SSI recipients among age group 18-64. Compared to the other three states, the pre-existing program in CT provided very limited provider network for eligible adults. Expanding the network could be potent for individuals with medical conditions, and hence substantially reduces their tendency to receive SSI. Although the mechanisms underlying this shift are hard to capture using secondary data, our DD findings show that the coverage expansion in CT was associated with an increase in public insurance coverage among low-income adults, and a reduction in the percent of adults in fair/poor health. These 21

results on mechanisms are consistent with coverage expansions reducing the size of the SSI population; however, the DD findings also indicate that CT’s coverage expansion increased in-migration of lowincome adults, which would tend to increase SSI receipt. The main lesson we can draw from these findings is that the effects of coverage expansions on federal disability outcomes are likely to be heterogeneous across states, depending on the characteristics of the pre-existing state and local programs. The more recent ACA Medicaid expansions initiated since 2014 differ from the Early Expander States in that the early expansions were all built on previous staterun programs; had income thresholds lower than 138% of FPL (in the case of CT and MN); and were not affected by the cross-state difference in Health Insurance Marketplaces (Sommers et al., 2013). On the other hand, these early coverage expansions took place during the same time period when two other early provisions of the ACA went into effect: (1) the prohibition of preexisting conditions exclusions; and (2) the dependent care mandate. Although these two provisions would be expected to affect the treated states and the synthetic control similarly, the backdrop of the early expansion period is an environment in which individuals with disabilities perhaps had new opportunities to obtain private insurance coverage. As data become available, researchers are starting to examine the effects of the more recent Medicaid expansions in 2014 on SSI receipt (Schimmel et al., 2016; Schmidt et al. 2017). This is particularly critical given that recent change of administration had led to a potential repeal of the ACA. Repeal may have unintended consequences for federal disability caseloads, and it is critical that these effects are considered when weighing different policy alternatives to the ACA.

22

References 1. Abadie, A., Diamond, A. & Hainmueller, J.. (2010). Synthetic control methods for comparative case studies: Estimating the effect of California’s tobacco control program. Journal of the American Statistical Society, 105(490): 493-505. 2. Abadie, A., Diamond, A. & Hainmueller, J.. (2015). Comparative politics and the synthetic control method. American Journal of Political Science, 59(2): 495-510. 3. The Arc of the United States (August 2012), The ACA: What disability advocates need to know. Accessed at: http://www.scdd.ca.gov/res/docs/pdf/Whats_New/The%20Affordable%20Care%20Act%20What %20Disability%20Advocates%20Need%20to%20Know.pdf, August 2016. 4. Bertrand, M., Duflo, E. & Mullainathan, S. (2004). How much should we trust differences in differences estimates? The Quarterly Journal of Eonomics, 119(1): 249-275. 5. Bitler, M. & Zavodny, M (May 2014). Medicaid: A review of the literature. National Bureau of Economic Research Working Paper 20169. 6. Burns, M., & Dague, L. (2017). The effect of expanding Medicaid eligibility on Supplemental Security Income program participation. Journal of Public Economics, 149, 20-34. 7. Burns ME, Dague L, and Kasper, M. Medicaid Waiver Dataset: Coverage for Childless Adults, 1996 – 2014. Version 1.0. University of Wisconsin-Madison, 2016. [machine-readable data file]. Madison, WI: Data and Information Services Center, University of Wisconsin-Madison [distributor], 2016. ; ( 10 August 2017). 8. Center on Budget and Policy Priorities, (February 2014). Introduction to the Supplemental Security Income (SSI) Program. Available at: http://www.cbpp.org/sites/default/files/atoms/files/1-10-11socsec.pdf. 9. Congressional Budget Office, US Congress (June 2016). Social Security Disability Insurance: Participation and Spending. Available at: https://www.cbo.gov/sites/default/files/114thcongress-2015-2016/reports/51443-SSDI_Participation_Spending.pdf. 10. Courtemanche, C., Marton, J., Ukert, B., Yelowitz, A. & Zapata, D. (Winter 2017). Early Impacts of the Affordable Care Act on Health Insurance Coverage in Medicaid Expansion and Non-expansion States. Journal of Policy Analysis and Management, 36(1), 178-210. 11. Finkelstein, A., Taubman, S., Wright, B., Bernstein, M., Gruber, J., Newhouse, J.P., Allen, H. & Baicker, K. (2012). The Oregon Health Insurance Experiment: Evidence from the First Year. Quarterly Journal of Economics 127(3): 1057-1106. 12. Frean , M., Gruber, J. & Sommers, B.D.. (April 2016) Premium subsidies, the mandate, and Medicaid expansion: Coverage effects of the Affordable Care Act. National Bureau of Economic Research Working Paper 22213. 23

13. Golberstein, E., Gonzales, G. & Sommers, B.D. (2015). California’s early ACA expansion increased coverage and reduced out-of-pocket spending for the state’s low-income population. Health Affairs 34(10): 1688-1694. 14. Gruber, J. Health insurance and the labor market. 2000. In: Handbook of Health Economics, vol. 1A (Culyer, A.J & Newhouse, J.P. Eds.). Elsevier: Amsterdam. 15. Gruber, J. & Madrian, B.C. (September 1993) Limited insurance portability and job mobility: The effects of public policy on job lock. National Bureau of Economic Research Working Paper 4479. Cambridge, MA. 16. Hu, L., Katesner, R., Mazumder, B., Miller, S., & Wong A. The effects of the Patient Protection and Affordable Care Act Medicaid expansions on financial well-being. National Bureau of Economic Research Working Paper 22170, April 2016. 17. Kaestner, R., Garrett, B., Gangopadhyaya, A., Fleming, C (2015). Effects of ACA Medicaid Expansions on Health Insurance Coverage and Labor Supply. NBER Working Paper 21836. 18. Kaiser Family Foundation (September 2012a) Focus on Health Reform. Health insurance market reforms: Pre-existing condition exclusions. 19. Kaiser Family Foundation (April 2, 2012b). States getting a jump start on health reform’s Medicaid expansion. 20. Kaiser Family Foundation, March 20, 2015. Medicaid enrollment and the Affordable Care Act. 21. Kaiser Family Foundation, March 14, 2016, Status of state action on the Medicaid expansion decision. Available at: http://kff.org/health-reform/state-indicator/state-activity-aroundexpanding-medicaid-under-the-affordable-care-act/ 22. Kennedy, J. & Blodgett, E. (September 20, 2012). Health insurance-motivated disability enrollment and the ACA. New England Journal of Medicine, 367:e16. 23. Leung, P. & Mas, A. (2016). Employment effects of the ACA Medicaid expansions. NBER Working Paper 22540. 24. Liebman, J. B. (2015). Understanding the increase in disability insurance benefit receipt in the United States. The Journal of Economic Perspectives, 29(2), 123-149. 25. Maestas, N., Mullen, K.J. & Strand, A.. (2014). Disability insurance and health insurance reform: Evidence from Massachusetts. American Economic Review: Papers and Proceedings, 104(5): 329-335. 26. Peele, P. S., J. R. Lave, K. J. Kelleher. 2002. “Exclusions and Limitations in Children’s Behavioral Health Care Coverage.” Psychiatric Services. 33:591-594. 27. Schimmel, J.H., Anand, P., Colby, M., Hula, L. & O’Leary, P. (2016). 24

“The Impact of ACA Medicaid Expansions on Applications to Federal Disability Programs.” Presentation at the 2016 Disability Consortium Annual Meeting, Washington, DC. 28. Schmidt, L., Shore-Sheppard, L. & Watson, T. (2017) The Impact of the ACA Medicaid Expansion on Disability Program Participation. Unpublished manuscript. 29. Schwartz, A.L. & Sommers, B.D. (2014). Moving for Medicaid? Recent eligibility expansions did not induce migration from other states. Health Affairs, 33(1): 88-94. 30. Social Security Administration. (2000-2013). Annual statistical report on the social security disability insurance program. Available at: https://www.ssa.gov/policy/docs/statcomps/di_asr/index.html. 31. Social Security Administration, 2014, Medicaid and the Supplemental Security Income (SSI) Program, Available at: https://secure.ssa.gov/poms.nsf/lnx/0501715010. 32. Social Security Administration, 2017, https://www.ssa.gov/disabilityresearch/wi/medicaid.htm. 33. Social Security Administration, (August 2006). Trends in the Social Security and Supplemental Security Income Disability Programs, https://www.ssa.gov/policy/docs/chartbooks/disability_trends/. 34. Social Security Administration, (April 2016), Annual statistical supplement to the Social Security Bulletin, 2015. Available at: https://www.ssa.gov/policy/docs/statcomps/supplement/index.html. 35. Social Security Administration, (August 2017), Monthly statistical snapshot, June 2017. Available at: https://www.ssa.gov/policy/docs/quickfacts/stat_snapshot/#table3. 36. Sommers, B.D., Arnston, E., Kenney, G.M. & Epstein, A.M.. (2013) Lessons from early Medicaid expansions under health reform: Interviews with Medicaid officials. Medicare & Medicaid Research Review, 3(4): E1-E19. 37. Sommers, B.D., Chua, K., Kenney, G.M., Long, S.K. & McMorrow, S.. (2016) California’s early coverage expansion under the Affordable Care Act: A county-level analysis. Health Services Research, 51(3): 825-845. 38. Sommers, S.A., Hamblin, A., Verdier, J.M. & Byrd, V.L.H. (2010). Covering low-income childless adults in Medicaid: Experiences from selected states. Center for Health Care Strategies Policy Brief, August 2010. 39. Sommers, B.D., Kenney, G.M. & Epstein, A.M. (2014). New evidence on the Affordable Care Act: Coverage impacts of early Medicaid expansions. Health Affairs, 33(1): 78-87. 40. Steven Ruggles, Katie Genadek, Ronald Goeken, Josiah Grover, and Matthew Sobek. Integrated Public Use Microdata Series: Version 6.0 [dataset]. Minneapolis: University of Minnesota, 2015. http://doi.org/10.18128/D010.V6.0. 25

41. Wagner, K.L. (2015). Medicaid expansions for the working age disabled: Revisiting the crowdout of private health insurance. Journal of Health Economics, 40, 69-82. 42. Yelowitz, A. (1996). Why did the SSI-disabled program grow so much? Disentangling the effect of Medicaid. Journal of Health Economics, 17, 321-349.

26

Table 1: Early Medicaid Expansions in California, Connecticut, District of Columbia and Minnesota for Low-Income Adults State

Start Date

Pre Expansion Eligibility Cutoffs and State or Local Insurance Programs

Post Expansion Eligibility Cutoffs

CA

July 1, 2011

Varies by county, up to 250% FPL without asset test under the Health Care Coverage Initiative

Varies by county, up to 200% of FPL without asset test

CT

April 1, 2010

56% of FPL with asset test under the State Administered General Assistant

DC

July 1, 2010

200% of FPL without asset test under DC Healthcare Alliance

MN

March 1, 2011

75% of FPL with asset test under General Assistance Medical Care

Post Expansion Significant Changes in Benefits

Transfer from Previous State Program 59,000

1. Expanded provider network 2. Out-of-network Emergency Room and poststabilization coverage for those<133% FPL 3. Comprehensive insurance in place of coverage 56% of FPL 1. Expanded provider 45,000 without network asset test 2. Enhanced coverage for medical transportation 3. Coverage for long-term care, home health, and skilled nursing facility 200% of 1. Expanded mental health 34,000 FPL without coverage asset test 2. Expanded pharmacy benefits 75% of FPL 1. New benefit of non77,000 without emergency medical asset test transport 2. New coverage for longterm care 3. Elimination of inpatient cost sharing and annual inpatient limit

Source: Sommers, B.D., Arnston, E., Kenney, G.M. & Epstein, A.M.. (2013) Lessons from early Medicaid expansions under health reform: Interviews with Medicaid officials. Medicare & Medicaid Research Review, 3(4): E1-E19.

27

New Enrollm ents (2012) 440,000

36,000

10,000

7,000

Table 2: Quality of Matching Early Expander States to Synthetic Controls Average Synthetic CA Average Synthetic CT Control of of Control Controls Controls % with college 0.302 0.290 0.292 0.373 0.288 0.345 degree % aged 45-54 0.218 0.235 0.229 0.248 0.235 0.239 % aged 55-64 0.148 0.169 0.161 0.173 0.166 0.168 % unemployed 0.078 0.071 0.077 0.061 0.067 0.068 % living in 0.131 0.122 0.130 0.086 0.119 0.094 poverty % receiving SSI 2.582 2.187 2.580 1.530 2.166 1.532 DC Average Synthetic MN Average Synthetic Control of Control of Controls Controls % with college 0.483 0.288 0.263 0.331 0.290 0.309 degree % aged 45-54 0.199 0.235 0.230 0.238 0.235 0.235 % aged 55-64 0.152 0.166 0.163 0.160 0.169 0.165 % unemployed 0.089 0.067 0.073 0.054 0.071 0.049 % living in 0.174 0.119 0.140 0.093 0.122 0.116 poverty 3.205 1.448 2.187 1.448 % receiving SSI 3.210 2.166 Note: Each covariate is the average of that variable during the pre-policy period for the state. The pre-policy period is 20002009 for CT and DC, and is 2000-2010 for CA and MN. The controls are the 15 states that expanded Medicaid on 1/1/14: AR, IL, IA, KY, MD, NV, NJ, NM, ND, OH, OR, RI, WA, WV and WI. The synthetic control is weighted combination of the 15 control states, with the weights shown in Appendix Table 1.

28

Table 3: Summary of Synthetic Control Findings

SSI recipients

CT

DC

CA

MN

-0.111* (0.063)

0.500 (0.438)

-0.163 (0.625)

0.237 (0.438)

Note: Table reports estimates of the average post-policy differences in SSI receipt between the treated state and the synthetic control states. The interpretation of the estimated coefficient is a percentage point change. The p-value is in parentheses. The p-value is calculated based on the ratio of the MSPE in the post-policy period vs. the MSPE in the pre-policy period for each treated state and for each state in the donor pool when it is falsely assigned as the treated state. The p-value is the number of states with a MSPE at least as high as the true treated state plus one (for the treated state) divided by 16, which is the number of donor states plus the treated state. * indicates statistically significant from zero at the 0.10 level.

29

Table 4: Early Medicaid Expansions and Mechanisms - DD Estimates State: Treat X Year Implemented Treat X Post Pre-policy period means R2 Treat X Year Implemented Treat X Post Pre-policy period means R2

Treat X Year Implemented Treat X Post Pre-policy period means R2 Treat X Year Implemented Treat X Post Pre-policy period means R2 Treat X Year Implemented Treat X Post Pre-policy period means R2

Treat X Year Implemented

30

CT DC CA MN Panel A: Percent uninsured among low-income adults -0.92 -2.00 -0.30 0.17 (0.44) (0.69) (0.76) (0.81) -1.49 -0.73 -0.02 -0.27 (0.36) (0.91) (0.99) (0.60) 26.92 12.99 42.82 24.53 0.928 0.929 0.944 0.938 Panel B: Percent with ESHI among low-income adults -1.35*** -1.46 0.53 0.09 (0.00) (0.31) (0.17) (0.79) -2.48*** -5.06** 0.38 -0.28 (0.00) (0.03) (0.38) (0.32) 32.18 35.33 23.75 33.85 0.946 0.946 0.973 0.952 Panel C: Percent with public coverage among low-income adults 1.13 0.22 0.05 -0.22 (0.39) (0.97) (0.96) (0.78) 4.65** 2.91 0.09 0.45 (0.02) (0.67) (0.93) (0.44) 35.25 44.55 26.54 34.69 0.87 0.851 0.853 0.856 Panel D: Percent in fair/poor health -1.22** -1.53* -1.52*** -1.00*** (0.01) (0.08) (0.01) (0.01) -0.06 -1.53 -2.10*** -1.03** (0.90) (0.18) (0.00) (0.01) 11.89 12.33 17.43 10.75 0.922 0.924 0.898 0.933 Panel E: Total population growth (relative to 2000) -0.03 -1.55 1.41 2.87*** (0.98) (0.79) (0.21) (0.01) 0.10 3.18 3.11*** 3.26** (0.95) (0.66) (0.01) (0.04) 103.42 102.86 106.72 105.85 0.852 0.851 0.892 0.847 Panel F: Net migration of low-income adults (relative to 2000) 0.48*** 0.38 0.16 -0.07 (0.00) (0.31) (0.16) (0.55)

Treat X Post Pre-policy period means R2

0.05 (0.49) 0.02 0.654

-0.72 (0.13) 0.95 0.648

0.22** (0.04) 0.27 0.604

-0.37*** (0.00) 0.07 0.650

Note: The table shows estimated coefficients and its p-values (in parentheses) of Equation 2. Robust standard errors clustered on state. Models include state fixed effects and year effects, and the five control variables described in Section 3 – estimated coefficients on these covariates are not shown. Data span 2008-2013 for Panels A-C (N=96) and 2000-2013 (N=224) for Panels D-F. *** indicates statistically significant from zero at the 0.01 level. ** indicates statistically significant from zero at the 0.05 level. * indicates statistically significant from zero at the 0.10 level.

31

Figure 1: Synthetic Control Estimates of Effects of Medicaid Expansion on SSI recipients

% points 2 3 1

1

% points 2 3

4

CT

4

CA

2005

2010

2015

2005

DC

MN

2010

2015

2010

2015

% points 2 3 1

% points 2 3

4

Year

1 2000

2005

2010 Year

2015

2000

2005 Year

Treated

32

2000

Year

4

2000

Synthetic control

Appendix Table 1: Synthetic Control Weights

Arkansas Illinois Iowa Kentucky Maryland Nevada New Jersey New Mexico North Dakota Ohio Oregon Rhode Island Washington West Virginia Wisconsin

33

CT 0.000 0.000 0.000 0.000 0.000 0.015 0.759 0.000

DC 0.000 0.519 0.000 0.481 0.000 0.000 0.000 0.000

CA 0.000 0.769 0.000 0.231 0.000 0.000 0.000 0.000

MN 0.000 0.152 0.000 0.000 0.000 0.000 0.195 0.000

0.000 0.000

0.000 0.652

0.000 0.000 0.226 0.000 0.000 0.000

0.000 0.000 0.000 0.000 0.000 0.000

0.000 0.000 0.000 0.000

0.000 0.000 0.000 0.000

0.000 0.000

0.000 0.000