The Affordable Care Act is a landmark piece of social legislation with the potential to reshape health care in the United States. Its potential to reshape politics is also considerable, but existing scholarship suggests conflicting expectations about the law's policy feedbacks, especially given uneven state-level implementation. In this article I focus on the policy feedbacks of the law's Medicaid expansion on political participation, using district-level elections data for 2012 and 2014 US House races and cross-sectional survey data from 2014. I find that the increases in Medicaid enrollment associated with the expansion are related to considerably higher voter turnout and that this effect was likely due to both an increase in turnout for new beneficiaries and a backlash effect among conservative voters opposed to the law and its implementation. These results have important implications for our understanding of the ACA and of the impact of welfare state expansions on political participation, particularly in federalized systems.
The question of how welfare state expansions impact political participation is central to the policy feedback literature, but no simple answer has emerged. Welfare state expansions may make new beneficiaries more likely to participate in politics by providing a basis for political organization or by providing them with more resources and stability in their lives (Campbell 2003; Mettler 2002; Mettler and Soss 2004). Expansions may also increase participation among nonbeneficiaries by sparking a backlash or “thermostatic” reaction (Soroka and Wlezien 2010). Conversely, participation in some welfare programs—particularly those that stigmatize recipients and necessitate negative or degrading interactions with government—may decrease an individual's willingness to participate in politics (Michener 2017; Soss 1999, 2000).
The uneven state-level implementation of the ACA's expansion of Medicaid offers a research opportunity in this regard. Following a 2012 Supreme Court decision that found that the federal government could not threaten the withdrawal of existing Medicaid funding to compel states to implement the law's expansion of Medicaid eligibility to all adults up to 138 percent of the federal poverty line, twenty state governments chose not to proceed with implementing the expansion, while the remaining states eventually chose to do so. For most implementing states, the expansion officially began in 2014. As a result, the 2012 and 2014 congressional elections offer data points before and after a major discontinuity in Medicaid eligibility in some states but not others. In this study, I leverage this variation to determine the Medicaid expansion's short-term impact on voter turnout in 2014. I find that increases in Medicaid enrollment associated with Medicaid expansion are significantly correlated with higher voter turnout in 2014 US House elections (i.e., a smaller decline from the presidential-year turnout of 2012). This relationship is robust to a wide array of control variables and different measurement strategies. The mechanism behind this correlation is less clear—analysis of individual-level survey data offers evidence that the expansion may have increased participation and engagement among both beneficiaries and opponents of the policy.
Political Impact of the ACA
Passage of the ACA and Political Aftermath
Passed by razor-thin congressional majorities and signed into law by President Barack Obama in March of 2010, the Patient Protection and Affordable Care Act (ACA) is considered to be the most important enactment in US health policy since the creation of Medicare and Medicaid in 1965. Perhaps the most important component of the law was the expansion of Medicaid, a joint federal-state program that provides insurance to the poor and other “categorically eligible” groups, in addition to covering long-term care costs for many senior citizens. Prior to the ACA, Medicaid eligibility was very inconsistent from state to state. While categorically eligible groups such as people with disabilities and parents of young children were generally well covered, and almost all otherwise uninsured children were covered by the related Children's Health Insurance Program (CHIP), standards for determining eligibility among childless, able-bodied adults under the age of sixty-five varied widely. The ACA sought to standardize eligibility by establishing a minimum eligibility cutoff of 138 percent of the federal poverty level in all states. As such, the people who stood to benefit most from the expansion were able-bodied, childless adults between the ages of eighteen and sixty-four with annual incomes of about $16,000 or less (assuming a single-person household).
While the expansion of Medicaid was to come with generous federal matching funds,1 a number of conservative Republican politicians at the state level were resistant to the change. They opposed expanding an entitlement program such as Medicaid on ideological grounds and were wary of seeming to cooperate with President Obama, a Democrat and a polarizing figure, to implement an unpopular law. A group of twenty-six state attorneys general, most with the blessing of their governors, filed a lawsuit against the federal government arguing (among other things) that the law's requirement that states implement the expansion or lose their existing Medicaid funds was too coercive. The Supreme Court's eventual decision on this case, while upholding the law's controversial “individual mandate” to purchase health insurance, sided with the conservative states on the Medicaid expansion, effectively making the expansion optional. This set the stage for a number of high-stakes legislative battles at the state level. Ultimately, some Republican-controlled states (for example, Arizona and Michigan) did opt to cooperate with the expansion, though they typically did so by incorporating some conservative policy changes. Still, as of this writing, nineteen states have yet to accept the expansion.
In terms of politics, few legislative enactments in American history have had the immediate impact of the ACA. While most supposed political “game changers” fail to change much of anything in American politics (Sides and Vavreck 2013), the ACA was the rare political event where the empirics backed up the hype: the law was a political failure for Democrats, at least in the short term. Specifically, several empirical studies have found evidence that the ACA had a real and perhaps decisive impact on the 2010 midterm elections (before most of the law's major provisions went into effect), which saw the Democrats lose their House majority along with sixty-three House seats and six Senate seats (Brady, Fiorina, and Wilkins 2011; Jacobson 2011; Nyhan et al. 2012). Democratic and Republican politicians clearly understood the tilt of the playing field on this issue: Republican House and Senate campaign ads were three times as likely as Democratic ads to mention health care, even though Democrats were traditionally thought to “own” the issue (Fowler and Ridout 2010: 9).
Notably, all of this occurred before the more high-profile provisions of the ACA (the Medicaid expansion, the private health insurance exchanges) went into effect in 2014. The 2010 backlash was a reaction to a political event more than a real policy. As the ACA has transitioned from a political abstraction to a real set of rules and programs affecting people's lives, it has the potential to affect politics in different ways via policy feedbacks.
Potential Policy Feedbacks of the ACA: Focusing on Participation
As valuable as this work is in helping us understand how the public responded to the passage of the ACA and its immediate aftermath, it is not sufficient for guiding our understanding of how the ACA might have influenced subsequent and future elections. By 2014 the ACA was not just a political issue but a set of real policies in many states and the focal point of fierce new policy debates in others. Understanding the political impacts of the ACA beyond the 2010 and 2012 elections requires us to consider the potential policy feedbacks of the law's various provisions, particularly the unevenly implemented Medicaid expansion.
Political scientists developed the “policy feedback” concept to better understand the ways in which, to quote Schattschneider (1935), “new policies create new politics.” This literature (for example, Campbell 2003; Mettler and Soss 2004; and Pierson 1993) reimagines the classic unidirectional model of democracy in which politics (public opinion, elections) is the cause and policy is the effect or outcome, allowing for the possibility that politics today are shaped in part by the policy outcomes of the past. In the present study, I focus on one potential policy feedback of the ACA: its effect on voter participation.
The question of how welfare state expansions affect political participation is particularly relevant in the United States, where voter turnout lags well behind that in other advanced democracies (DeSilver 2015) and where inequalities in participation across ethnic, economic, and age groups are pronounced, particularly in nonpresidential elections (File 2015; Leighley and Nagler 2014). Functionally, this means that congressional candidates face a considerably more conservative electorate in nonpresidential election years—one reason for congressional Democrats' very poor showings in the 2010 and 2014 midterms after they enjoyed great success in 2008 and 2012. In short, variation in political participation has real consequences for American politics and public policy.
The literature on policy feedbacks offers insights on how a welfare state expansion like the Medicaid expansion might affect voter turnout, though these insights do not all point toward the same expectations. Perhaps the most obvious is that a welfare state expansion may prompt mobilization among beneficiaries to support the incumbent party out of gratitude (see De La O 2013). A somewhat subtler variant of this argument holds that welfare state expansions make recipients more aware of their personal stake in political matters. Notably, this is the mechanism Clinton and Sances (2016) posit in a study parallel to this one, which finds a positive effect of Medicaid expansion on turnout in closely matched counties within expansion and nonexpansion states. While Jacobs and Mettler (2015) do find evidence that gaining coverage under the ACA modestly improved individuals' attitudes toward the law in a panel survey, I argue that it is unlikely that such conscious mobilization occurred in 2014, given the historic defeat suffered by congressional Democrats and the complex nature of the Medicaid expansion, which many beneficiaries may not understand as a federal program or part of the ACA.
Another mechanism through which policy can affect politics is by redistributing resources associated with participation—what Mettler (2002: 352–53) calls “resource effects” (also see Mettler and Soss 2004). The resource effect concept builds on the “resource model” of political participation (Brady, Verba, and Schlozman 1995), which holds that participation is dependent on a set of key resources, including time, money, and civic skills. Given that all three of these are correlated with income and education, the resource model offers a causal theory for the strong relationship between socioeconomic status and participation. Since public policies have the potential to change the distribution of these resources, they have the potential to change participation as well. For example, in her study of the long-term policy feedbacks of the GI bill of 1944 (which extended educational benefits to members of the armed services), Mettler (2002, 2005) finds that the bill's extension of higher education to the masses was a key component in spurring increased engagement and participation.
While it may be decades before any such long-term impacts of the ACA will be visible, there are possible short-term channels through which the law may already be affecting the resources important to participation.2 First, Pacheco and Fletcher (2015) find that better health is associated with a greater likelihood of voting among Americans,3 since poor health is a drain on resources (time, money, cognitive abilities). So, to the extent that gaining health coverage helps to improve health, the Medicaid expansion has the potential to increase turnout among new beneficiaries.
Whether Medicaid actually does improve physical health is debatable. A randomized controlled experiment embedded in an earlier state-level Medicaid expansion in Oregon (Baicker et al. 2013) found no solid evidence that gaining Medicaid coverage led to improvements in a pre-selected set of physical health indicators. Nonetheless, the same study found strong evidence that the program did improve mental health and reduce financial strain among beneficiaries even over the short term. Such changes in mental and financial well-being have the potential to increase political participation. Observational evidence from Gollust and Rahn (2015) supports this notion—in models primarily focused on the impact of chronic conditions on voter participation, they find a significant negative correlation between uninsurance and turnout in the 2009 Behavioral Risk Factor Surveillance System survey while controlling for numerous health indicators (including self-reported health) and a wide array of factors that might confound such a relationship (for example, income, education, race).4
Of course, a welfare state expansion has the potential to affect the political behavior of non-beneficiaries as well. In particular, it has the potential to mobilize opponents through a backlash effect. The idea of a backlash against a newly established policy is consistent with a thermostatic model of policy opinion and representation (e.g., Bendz 2015; Soroka and Wlezien 2010) in which voters react negatively to policy changes that go too far for their tastes. A backlash may be especially likely if the target population is socially constructed in a way that makes them seem undeserving, a key aspect of policy design according to the work of Schneider and Ingram (1993, 1997, 2005). As a sizable literature on welfare politics has found, there is a strong racial element to such perceptions, with attitudes about black Americans predicting attitudes about welfare programs (e.g., Gilens 1996, 1999; Lieberman 1998; Peffley, Hurwitz, and Sniderman 1997). The findings of Tesler (2012) and Henderson and Hillygus (2011) about the racial polarization of health care attitudes suggests that such a dynamic could play a role in the policy feedback of the Medicaid expansion. Jacobs and Mettler (2011: 928) also argue that views about the deservingness of Medicaid beneficiaries will play a role in the future politics of health care.5
In fact, contemporary news accounts suggest that this has been the case. In a November 2015 reported opinion piece for the New York Times on Republican electoral success in poor regions with large numbers of public beneficiaries, Alec MacGillis argues, based on his reporting, that resentment among “those who are a notch or two up the economic ladder” toward those on public assistance, including Medicaid, has fueled a voter backlash at the local and community level. MacGillis characterizes safety-net dependency as “the most visible manifestation of downward mobility in their declining towns” (MacGillis 2015). Similarly, a November 2015 article in the Lexington Herald-Leader notes that Matt Bevin, the newly elected Republican governor of Kentucky and a vocal opponent of the Medicaid expansion, garnered his highest vote totals in counties that also contained the greatest numbers of Medicaid recipients (Cheves 2015).
While the aforementioned arguments all point in the direction of increased participation in states that expanded Medicaid, another relevant line of scholarship points in the opposite direction. While Campbell (2003) and Mettler (2002, 2005) find that Social Security and the GI bill increased political engagement among beneficiaries, this was possible in part because both policies communicated the message that recipients were worthy and deserving members of society. In the terminology of Mettler (2002: 352–53), the policies conveyed positive “interpretive effects” to recipients. This is not always the case with social welfare policies in the United States. Soss's work on the political behavior of “welfare” recipients demonstrates that when programs stigmatize beneficiaries and treat them poorly (for example, by forcing them to comply with burdensome eligibility requirements), political engagement and participation tend to go down (Soss 1999, 2000). This could be because beneficiaries come to believe they do not deserve a voice in politics, or because their negative experiences with government bureaucrats make them cynical about politics in general, or both.
Should we expect this dynamic to play out with the Medicaid expansion? In a study of Medicaid's impact on political participation using pre-ACA survey data, Michener (2017) finds that Medicaid has a pronounced and statistically robust negative impact on political participation among beneficiaries. Michener's own unstructured interviews with beneficiaries lead her to conclude that the program conveys “disempowering” messages that discourage political participation. The implications of these findings for the ACA's Medicaid expansion are somewhat ambiguous, however. Michener also finds that the impact of Medicaid on participation varies considerably by state context and that a state-level expansion of Medicaid benefits sent a positive message to recipients that increased participation.
Theory and Hypotheses
Since the findings and insights of the existing policy feedback literature suggest conflicting expectations for the Medicaid expansion's impact on voter turnout, I test competing hypotheses in the present study.
First, since Medicaid has the potential to improve health and has been proven to improve stability and reduce financial strain for beneficiaries even in the short term (Baicker et al. 2013), it is reasonable to expect that voter turnout might increase in states that implemented the Medicaid expansion, since more potential voters should have the resources necessary to participate in politics. Second, Michener's (2017) findings about the positive effect of Medicaid benefits on participation in the context of benefit expansion also suggests increased participation among new beneficiaries.
Implementation of the Medicaid expansion will be associated with increased voter turnout among new beneficiaries
The literature on welfare state backlash and the thermostatic model of public opinion suggest that the Medicaid expansion also has the potential to increase participation among opponents of the expansion—in this case, Republicans and conservatives.
Implementation of the Medicaid expansion will be associated with increased voter turnout among Republicans, conservatives, and those opposed to the ACA
The work of Soss (1999, 2000) and Michener (2017) suggests that it is also possible for a welfare state expansion to drive down participation among new recipients. While Michener finds the opposite effect in the context of benefit expansion, it is not clear that this will apply to the ACA Medicaid expansion, especially since some states have adopted relatively stringent requirements for new beneficiaries. This leads to H3, a competing hypothesis to H1.
Implementation of the Medicaid expansion will be associated with decreased voter turnout among new beneficiaries
Election Analysis Data, Variables, and Models
Election Analysis Dependent Variable
The key dependent variable in my analysis is total voter turnout for all 435 US House races in 2014.6 In isolation, however, this information does not allow me to make meaningful inferences, given the heterogeneity of districts and my interest in the effects of Medicaid changes. Thus, I use turnout in the 2012 House elections as a baseline and analyze change in turnout in each district between the two elections.7
There are a number of advantages to using House election results as the dependent variable. This approach avoids some pitfalls of survey analysis: sampling error and bias (since I have the full universe of data), issues with self-reported voting, and obstacles to calculating the aggregated effects of survey respondents' reported behavior on actual elections. To accomplish a similar analysis with survey data would require a panel survey with a very large sample size and meaningful numbers of respondents in each congressional district. To my knowledge, no such panel survey exists for the time period of interest.
Of course, there is a disadvantage to the analysis of elections data as well: such analysis quickly runs into the ecological fallacy problem when it attempts to make inferences about individual-level behavior using aggregate data. Correlations at the aggregate level could be driven by many possible mechanisms at the individual level, and an aggregate correlation or lack thereof may obscure meaningful variation in individual-level correlations. I address these shortcomings by including a companion analysis of cross-sectional survey data, discussed in the “Survey Analysis” subsection.
Election Analysis Independent Variables
The chief independent variables of interest capture Medicaid changes between 2012 and 2014. The simplest of these is a dummy variable indicating whether or not the state took part in the Medicaid expansion; this is the approach Clinton and Sances (2016) take.8 Since this is a state-level decision, this variable varies only at the state level. It also misses important sources of variation in Medicaid signups, including state outreach efforts, differences in previous and current eligibility requirements,9 secular changes in the economy,10 and the so-called woodwork effect in which previously eligible but uninsured individuals become aware of their eligibility due to the publicity surrounding the expansion, whether or not their state chose to expand Medicaid. Furthermore, focusing solely on the decision to implement the expansion fails to account for the fact that three states (California, Connecticut, and Minnesota) actually began early implementation of the expansion before the 2012 election. Thus, I also use the district-level change in the number of Medicaid beneficiaries as estimated by the US Census Bureau's American Community Survey (ACS).11,12 I consider the change in both the total number of Medicaid beneficiaries and the number of beneficiaries in what could be considered the “target demographic” for the expansion: adults aged eighteen to sixty-four with incomes that put them below 138 percent of the federal poverty line. The latter is a reasonable proxy of the number of signups directly related to the Medicaid expansion, though it does not capture woodwork effect signups among people outside that demographic and may include a number of woodwork effect signups within it.
For the election results analysis, I must control for a number of district-level factors that are known to affect voter turnout: the presence of races higher on the ballot (gubernatorial and senate elections), the competitiveness of the race (winning candidate's margin of victory and the absence of a major party challenger), and the presence of a major third-party challenger (measured in “other” party vote share). I also control for the total number of district residents with incomes that place them below 138 percent of the federal poverty line. Since voter turnout is lower for poorer Americans, and the number of poor residents is highly correlated with Medicaid enrollment, this is a crucial control variable. I also control for the presence of both Democratic and Republican incumbents, though the impact of incumbency on turnout is more ambiguous. Given the specification of the election models (see the following subsection), all of these variables are operationalized in terms of the change in their values between 2012 and 2014.13
Election Analysis Models
These analyses use a “first differences” approach, regressing the district-level change in voter turnout between 2012 and 2014 on the changes in the independent variables. Since these models analyze only change within districts, they eliminate concerns about heterogeneity between districts. Since the dependent variable is continuous, I use ordinary least squares (OLS) regression with robust standard errors clustered by state to allow for correlation of standard errors within states. Equation 1 summarizes the election models.
Medicaid change will be operationalized in three different ways, as detailed in the “Election Analysis Independent Variables” subsection. If H1 or H2 is correct, β1 should be positive and statistically significant, indicating that Medicaid expansion increases turnout. If H3 is correct, β1 should be negatively signed and statistically significant.
Survey Analysis Data, Variables, and Models
Survey Analysis Dependent Variable
To partially compensate for ecological fallacy concerns, I augment my analysis of elections data with an exploratory analysis of Common Content questions on the Cooperative Congressional Election Study (CCES). The CCES is a survey of over fifty thousand Americans, conducted every year, with pre- and postelection waves in midterm and presidential election years. The sample includes between 45 and 232 respondents per congressional district in 2014. While not a panel survey, the CCES is an ideal cross-sectional data resource in terms of sample size and geographic distribution. The Common Content also contains a wealth of demographic and political questions, though unfortunately it does not clearly identify Medicaid recipients or those who recently gained health coverage. I address this issue in the discussion of variables below. The key dependent variable in these analyses is self-reported voting.
Self-reported voting in the CCES is, as per usual, exaggerated. Seventy-three percent of respondents reported voting, though only 33.2 percent of the voting age population actually voted in 2014 (McDonald 2014). Part of this discrepancy may be due to sampling, but a large portion of it is surely due to misreporting.14 Importantly, Ansolabehere and Hersh (2012) find that misreporting is driven by social desirability bias and that this bias is strongest among voters with high levels of resources and engagement. In other words, the sorts of people who usually vote (wealthy, well educated, politically attentive) are the most ashamed of failing to vote and therefore the most likely to lie about it. Since this study largely focuses on low-income people eligible for Medicaid, who are less likely to feel this social pressure, misreporting is less of a problem. In any case, pairing the survey analysis with the election analysis guards against being misled by misreporting.
Survey Analysis Independent Variables
While the CCES does not include questions specifically about Medicaid status, it does include a battery of questions about health insurance, including whether or not the respondent has insurance through a government program, offering the examples of Medicare or Medicaid. Other questions allow me to identify the two populations most likely to be covered by government programs other than Medicaid: senior citizens (almost universally covered by Medicare in terms of medical insurance, though Medicaid may cover their long-term care) and current or former military service people (many of whom receive coverage and care through the Departments of Defense and Veterans Affairs). Thus, a dummy variable indicating that the respondent receives insurance through a government program and is not a member of those two categories serves as a reasonable proxy for Medicaid coverage.15 Unfortunately, it is impossible to differentiate between longtime Medicaid recipients and those who recently started on the program. Instead, I interact this proxy with the percentage change in Medicaid enrollment at the congressional district level. The resulting interaction serves as an estimate of the probability that the individual recently signed up for Medicaid coverage.
Testing H2 requires identifying respondents likely to mobilize as part of a backlash effect against Medicaid expansion. Given the high degree of partisan polarization over the ACA (Smith 2015), the most obvious place to look for a backlash effect is among Republican voters. Since partisans on both sides of the aisle participate at greater rates than nonpartisans, I operationalize party identification as two dichotomous variables indicating identification with the Republican and Democratic parties rather than as a single scale. I interact these variables with Medicaid change measures in the models testing H2 (see following subsection). I also include both as control variables in the analyses testing the effect of new Medicaid receipt.
While I include the 2014 values for all of the district-level control variables referenced in the “Election Analysis Independent Variables” section, the cross-sectional survey analysis requires an additional set of individual-level control variables known to correlate with political participation, in addition to the party identification indicators. These include income (measured on a twelve-point ordinal scale), education (six-point ordinal scale), age, race (measured with a series of dummy variables16), and gender (dummy variable).
Survey Analysis Models
Since the dependent variable in the survey analyses (self-reported voting) is dichotomous, I use logistic regression with state-clustered standard errors. Equation 2 summarizes the models used to test H1 and H2.
If H1 is correct, β3 should be positive and statistically significant for at least some values of Medicaid change, which will be measured in terms of both the overall change in Medicaid enrollment and the change within the “target demographic,” as detailed in the “Election Analysis Independent Variables” subsection. H1 does not imply any prediction for the sign or significance of β1, since it predicts a positive effect for new recipients only; it is possible that Medicaid receipt depresses participation overall while increasing participation among new recipients in an expansionary context, per Michener (2017). If H3 is correct, β1 should be negative and statistically significant (indicating a negative effect of Medicaid receipt on participation), while β3 should be statistically insignificant (indicating that probable new recipients are no more likely to turn out than Medicaid recipients in general).
To test H2, I run a separate set of models including interactions of partisan identification with district-level Medicaid change, as summarized in Equation 3.
If H2 is correct, β3 should be positive and statistically significant for at least some values of Medicaid change, indicating that increased Medicaid enrollment triggered a backlash effect among Republican voters.
Election Analysis Results
I begin with a simple pooled analysis of the relationship between the number of Medicaid beneficiaries in a district and voter turnout in 2012 and 2014 US House races. As figure 1 demonstrates, there is a negative correlation between Medicaid beneficiaries (in this case, those in the age eighteen to sixty-four, <138% FPL demographic) and turnout in these elections, consistent with Michener's findings about the program's negative effect on participation.
Turning to the focal point of my analysis, change in both Medicaid eligibility and turnout presents a very different picture, as seen in figure 2. While voter turnout declined in all districts with contested House races, as expected given the change from a presidential to a midterm election, the decline was substantially smaller (i.e., turnout was higher) in districts with greater growth in Medicaid enrollment in the target demographic.17 As the figure also makes clear, this relationship appears to be driven by districts in states that chose to accept the Medicaid expansion.
Since the decision to implement the expansion and the number of new signups are likely correlated with a host of relevant district-level factors that also affect turnout, I regress the change in turnout on different indicators of expansion progress and district-level controls (omitted here and in other tables for space). The results of these models are displayed in table 1.
As the results of Model 1 make clear, a simple dummy variable indicating that the state in which the district is located implemented the expansion has essentially no predictive power for the change in turnout from 2012 to 2014.18 The result of Models 2 and 3 suggest that Model 1 misses important variations, however. When the change in Medicaid is operationalized as the change in enrollment, either overall or within the “target demographic,” there is a large and robust positive relationship with turnout. In fact, the coefficient estimates are so large (Model 3 actually estimates more than one additional voter per one Medicaid signup in the target demographic) that they suggest any positive effect of Medicaid change on turnout must necessarily involve citizens beyond those newly signed up for Medicaid.19
It is interesting and important that district-level variations in Medicaid enrollment do a better job of explaining turnout than does the statewide policy decision. This suggests that the effects of the Medicaid expansion may be a result of the actual downstream impacts of the program on people's lives or political activity rather than of the statewide political climate. To further ensure that the relationships I observe originate at the district level, and as a check against possible spatial correlation of errors, I also ran a series of spatial models (full results in appendix 2) incorporating geographic data from the US Census Bureau.20 Using both contiguity and inverse-distance approaches, I ran spatial autoregression models with autoregressive standard errors and spatial Durbin models based on Models 2 and 3 from table 1. In every case, the statistically significant relationship between Medicaid change and turnout change remained strongly statistically significant and changed little in terms of magnitude compared to the plain OLS models.21 In short, the evidence points toward a localized rather than statewide or regional relationship between Medicaid change and turnout.
Notably, these turnout effects do not seem to have impacted partisan vote shares in the 2014 House election. Analyses using the Democratic candidate's share of the two-party vote as the dependent variable yields null results, regardless of the model specification. This suggests that the increased number of voters did not systematically favor one party over the other, perhaps because there were processes in play that mobilized voters on both sides of the aisle.
Survey Analysis Results
I begin by attempting to isolate the effects of changes in Medicaid on those who actually received coverage. Table 2 displays logit models of self-reported turnout with an interaction of the Medicaid proxy with the percentage change in Medicaid enrollment at the district level as the key independent variables. I also include the full array of district and individual level controls, though they are omitted here for space.
As the negative sign and strong statistical significance of the Medicaid proxy constituent term demonstrates, Medicaid recipients in general are less likely to report having voted, even when controlling for income. The positive and statistically significant coefficients on the interaction terms suggest that this is less likely to be the case for new Medicaid recipients in 2014, consistent with H1. Figure 3 illustrates this pattern graphically, using marginal predictions from Model 2 in table 3.22 Medicaid recipients with a higher probability of being newly covered in 2014 were more likely to report having voted, all else equal. In fact, for some values of the percentage change in Medicaid enrollment, the model predicts that this effect may erase Medicaid's overall negative impact on turnout.
Next, I test to see if the apparent effect of enrollment changes has a partisan dimension. Table 3 displays the results of models of self-reported voting, with the key independent variables being the interaction of partisan identification with the two indicators of district-level Medicaid enrollment growth. The excluded group in these models is respondents that did not voluntarily identify with either party. The results for the interaction terms are more ambiguous—only the interaction of Democratic partisanship with Medicaid change achieves conventional statistical significance (at the p < .10 level) and only in Model 2. Given that this is a categorical by continuous interaction, however, it is not possible to definitively determine statistical significance using a table alone, as Brambor, Clark, and Golder (2006) demonstrate.
To properly interpret the effect of the interaction term, I plot the predicted probability of voting by partisanship across the range of the change variable in figure 4, which shows the marginal effects of the interaction from Model 2 in table 3. While there is no evidence that Medicaid enrollment increased turnout among independents, it does appear that greater enrollment increases caused both Democrats and Republicans to widen their participation advantage over independents by several percentage points.
The apparent positive impact on turnout for Republicans is consistent with an anti-Medicaid backlash (H2), since most Republicans are strongly opposed to the ACA (though it should be noted that this effect appears to be somewhat sensitive to measurement) while the increased turnout among Democrats may be the result of the effect of the program itself (H1), since low-income people and Medicaid recipients are more likely to identify as Democrats than Republicans.23
Race and Possible Heterogeneous Effects
All of the above models and results are based on the assumption that the effect of Medicaid change on voter turnout is uniform across individuals, districts, and states. This may not be the case. In this section, I briefly consider the possibility that these effects may vary for different racial groups and contexts. The full results of the models discussed here are displayed in appendix 4, along with explanations of the additional variables used.24
There are a number of levels and pathways through which race might condition the effect of Medicaid on voter turnout. First, the impact of Medicaid enrollment on turnout may vary by the recipient's race. Michener (2017) documents that Medicaid may carry a racialized stigma for blacks in at least some contexts, which could depress participation. Moreover, it is possible that gaining insurance through Medicaid may not be enough to overcome the barriers to participation (e.g., language barriers) that depress turnout among Hispanics. Racial diversity at the district level may also have an effect—Michener finds that participation is lower among Medicaid enrollees in states with large numbers of black Medicaid enrollees due to the racial stigmatization of the program. On the other hand, the racialization of social welfare programs in diverse areas also has the potential to magnify any backlash effects, since racial stereotypes may lead potential voters to view beneficiaries as less deserving than they otherwise would (Gilens 1996, 1999; Lieberman 1998; Peffley, Hurwitz, and Sniderman 1997).
While neither of my data sources is sufficient to definitively document dynamics such as these, preliminary analyses do offer some suggestive results. In versions of the election models that consider district-level racial diversity, the effect of Medicaid change on turnout appears to be greatest in the least diverse (that is, “whitest”) districts (table A4-1). This suggests the racialized backlash effect discussed above is unlikely but lends credence to the idea that signing up for Medicaid had different effects on different types of people. To explore this further, I analyze versions of the survey analysis models interacting respondent race variables with the Medicaid proxy/Medicaid change interaction (table A4-2). While I find no evidence that the positive effect of new Medicaid receipt differed for black respondents relative to white respondents, it does appear that the effect was lessened and perhaps negated entirely for Hispanic respondents (figure A4-1). I also consider versions of the partisan interaction models displayed in table 3, incorporating triple interactions of Republican party identification, Medicaid change, and indicators for district racial diversity (table A4-3). If there were a racialized backlash effect, we would expect to see greater Republican turnout in more diverse districts with large jumps in Medicaid signups, but there is no evidence of such a pattern in these results.
This article offers an intriguing first look at the impacts of a landmark welfare state expansion on political participation in a country where both have historically lagged behind the norm. While the results are not definitive and it is not clear what the long-term impacts might be, the present study does offer a number of preliminary conclusions.
First, the increased enrollment in Medicaid brought on by the ACA, in those states that chose to implement the law, was associated with higher levels of voter participation in 2014 House races relative to 2012 (that is, with a reduction in the size of the usual midterm drop-off in turnout). In contrast to Clinton and Sances (2016), who find a significant effect of policy implementation measured as a dummy variable in carefully selected counties, I find an effect only for actual changes in Medicaid enrollment. Second, there is considerable evidence that at least part of this increase in turnout was the result of new Medicaid recipients turning up to the polls, despite the fact that Medicaid enrollment in general is associated with depressed participation. The significance of the district-level measurements of enrollment change, as opposed to the statewide indicator of policy change or enrollment change in nearby districts (see appendix 2), suggest that this correlation is tied to actual signups at the local level rather than statewide or regional political dynamics. Furthermore, probable new Medicaid recipients were more likely to report voting in the 2014 CCES than those that were less likely to be new recipients. These findings buttress the findings of Michener (2017) on the conditional impacts of Medicaid on political participation. The program's overall negative impact on participation is visible in these data, consistent with the arguments of both Michener and Soss (1999, 2000) that some welfare programs decrease political engagement. In a period of expansion, however, it appears that such patterns can be reversed, as Michener argues.
The third conclusion, more tentative than the first two, is that there may have been a backlash associated with implementation of the expansion, as contemporary journalistic accounts have argued (e.g., MacGillis 2015). The effect sizes in the election models are simply too large (in one model, more than one additional voter per new Medicaid recipient) to reflect increased participation by new Medicaid recipients alone. Additionally the partisan interaction models of self-reported voting found that partisans on both sides of the aisle were more likely to vote in districts with higher enrollment, though these effects were not statistically significant for all values of the Medicaid change variables.
These patterns beg the question about the causal mechanisms that link changes in Medicaid enrollment to voter turnout. While the present study does not definitively establish causal mechanisms, it does offer some guideposts for future study. First, for reasons cited above, these findings imply that the new recipients themselves are a significant part of the increased turnout, suggesting some sort of individual-level process is in play. Clinton and Sances (2016) suggest that the effect operates through conscious mobilization, though I argue that this is unlikely given that 2014 was such a poor election for Democrats and that there was no discernible effect of Medicaid enrollment change on partisan vote share. I argue that a resource effect is more likely: Medicaid transfers resources to recipients that make their lives more stable, making them more likely to participate in politics. Nevertheless, more study is needed to validate this as the mechanism of interest.
It is important to remember that the overall effect of Medicaid receipt on political participation is negative in both of the datasets analyzed here, consistent with the findings of Michener (2017). This suggests that, if there are resource effects in play, something must counteract them for those in traditional Medicaid. Michener argues that this “something” is welfare stigma, which is less of a problem in the more welcoming context of benefit expansion. Additional survey research targeting Medicaid beneficiaries will be necessary to more definitively establish the role of stigma in mediating Medicaid receipt's effects on participation. This future research should also consider that effects may differ across racial and ethnic groups—my subgroup analysis suggests that there may not have been a positive effect of Medicaid change on turnout for Hispanics, for example.
Equally important for future study is establishing the precise mechanisms for welfare state backlash effects. Soss and Schram (2007) argue that, for most of the public, welfare programs are “visible” but “distant”—they do not directly affect most Americans' everyday lives, so symbolism and discourse shape attitudes more than do real policy changes. This study, by contrast, finds some evidence for localized backlash effects driven by real policy changes (see also MacGillis 2015 and Cheves 2015). In the case of the Medicaid expansion, the relevant policy changes may have been more “proximate” for Medicaid opponents—perhaps because they noticed and resented the new benefits enjoyed by some of their poorer neighbors. Notably, while race has played a central role in American welfare politics for decades, and research has shown that the ACA has become a racialized issue (Henderson and Hillygus 2011, Tesler 2012) at the national level, I find no evidence of a racially based backlash here. Taken together, these findings suggest a different mechanism for welfare state backlash than the traditional story of symbolic welfare politics.
Future work on the ACA's political impact should also continue to track these effects over time. It will be especially interesting from the standpoint of welfare state feedbacks to analyze whether this apparent bump in turnout fades or even reverses over time, as recipients continue to interact with the Medicaid systems in their state. In particular, it is likely that the more onerous requirements for eligibility imposed in moderate and conservative states will send the sorts of demoralizing messages that Soss and Michener find to be corrosive to participation. It is also possible that the returns to participation from increased emotional and financial stability, and perhaps even improved physical health, will increase over time.
I would like to acknowledge guidance, comments, and feedback from the following persons: Daniel Lee, Andrea L. Campbell, Frederick J. Boehmke, the attendees of the Citizens in Changing Welfare States workshop in Gothenburg, Sweden, the editor (Colleen Grogan), and two anonymous reviewers. I would also like to acknowledge Jenefer Jedele for her research assistance.
The author acknowledges support from a Robert Wood Johnson Foundation Scholars in Health Policy Research Fellowship while conducting research for this article.
1. Prior to the ACA, wealthier states (in terms of per capita income) split Medicaid costs 50–50 with the federal government, while poorer states paid considerably less (for example, the federal government pays 74.2 percent of the cost in Mississippi). Importantly, the split for new Medicaid recipients covered under the ACA's expansion of the program is much more generous—the federal government pays 100 percent of the cost for the first three years, gradually ramping down to 90 percent thereafter.
2. One such channel is increased opportunities to register to vote, which the law dictates government must offer during interactions with citizens like Medicaid signups (though see Sneed 2015). Unfortunately, the data sources used in the present analysis are insufficient to rigorously explore this possible mechanism. The Cooperative Congressional Election Study does include a question about registration, but the self-reports of registration appear unrealistically high. Logit models using self-reported registration as a dependent variable show no evidence that Medicaid changes correlate with this measure.
4. In appendix 3, I consider the possibility that other ACA policies intended to reduce the ranks of the uninsured, including tax credits to directly purchase insurance on the exchanges, may have had similar effects. I find no such effects, consistent with expectations I lay out in the text of appendix 3. (Readers who wish to access the appendices mentioned in this article will find them, as well as replication data and do files for all models, at my website: faculty.missouri.edu/∼haselswerdtj/.)
7. The decennial reapportionment and redistricting that took place after 2010 makes it impossible to make comparisons with the previous midterm election in 2010. Using a 2010 baseline would require using a lower level of aggregation (e.g., the county or precinct level) where estimates of Medicaid enrollment become less reliable and, in some cases, unavailable. Furthermore, using the 2012 baseline minimizes the time between the pre and post measurements.
8. This variable is coded 1 only for states that began implementation of the expansion in 2014 or earlier. As expected, this variable is strongly and positively correlated with the Medicaid change statistics that follow.
9. For some states, the Medicaid expansion represents a drastic expansion of eligibility relative to the status quo, while in others the change is less pronounced. Furthermore, even if all states accept the expansion of Medicaid, considerable heterogeneity will remain in terms of eligibility requirements, with some states exceeding the 138 percent of FPL “floor” and others implementing stricter eligibility requirements via waivers.
10. Since it is a means-tested program, we should expect Medicaid signups to be influenced by economic indicators like the poverty rate.
11. I rely on census data because, as of this writing, administrative data through the Medicaid State Information System is unavailable. One drawback of the census data is the known problem of underreporting of Medicaid receipt, which may understate enrollment by 22.6 percent to 31.7 percent (State Health Access Data Assistance Center 2008: 11). For this reason, I substitute the change in the ACS estimate of the total number of “uninsured” for the Medicaid estimates as a robustness check in appendix 3. Research suggests that measurements of uninsurance are less affected by this problem, since Medicaid recipients are more likely to misreport that they have some other type of coverage than that they have no coverage (State Health Access Data Assistance Center 2008: 11).
12. The ACS insurance question asks about current-point-in-time insurance status, with seven response categories. Importantly, the wording of the ACS question on health insurance status did not change between 2012 and 2014, despite changes in other relevant federal surveys (see State Health Access Data Assistance Center 2014).
13. This means that the dichotomous variables are transformed into trichotomous variables with values of −1, 0, or 1.
14. The degree of misreporting seems to be similar to that found by Keeter, Igelnik, and Weisel (2016) in their analysis of a Pew Research Center survey from the 2014 midterm elections.
15. This variable flags 14 percent of the sample as likely Medicaid recipients, very close to the national incidence of about 13 percent. Moreover, as we would expect, those identified by this variable are much more clustered at the bottom of the income scale than other respondents: 73.3 percent of those identified by the variable report household incomes of less than $39,999, compared to only 31.2 percent for other respondents. While it is possible that the variable may misidentify those who directly purchase insurance via ACA exchanges as Medicaid recipients, since some respondents may interpret the exchanges as a “government program,” these numbers suggest this is not a major problem for this analysis.
16. These variables identify respondents in the following categories: black or African American, Hispanic, Asian, Native American, Middle Eastern, mixed race, or other. Non-Hispanic whites make up the excluded category.
17. The pattern is similar for total Medicaid enrollment change as well.
18. Dropping the three states that implemented expansion prior to 2012 (California, Connecticut, and Minnesota) does not change this finding.
19. Control variables largely perform as expected in these and other models, though I omit them here to conserve space. Full results tables can be viewed online in appendix 1.
20. See appendix 2 online.
21. There is evidence of spatial correlation in errors and the dependent variable, and some inconsistent evidence of spillover effects of Medicaid change in nearby districts, but the key findings presented here remain remarkably robust.
22. All marginal prediction graphs were created using Stata's margins and marginsplot commands. All control variables are allowed to assume their actual values.
23. While this study does not deal directly with policy attitudes, the CCES does contain a battery of three questions related to the Affordable Care Act that form a good scale of ACA-related opinion (alpha = .79). A preliminary analysis of this scale suggests that opposition to the ACA was significantly lower among those in districts with greater increases in Medicaid enrollment. As an independent variable, however, this scale fails to achieve statistical significance in a logit model of self-reported voting.
24. Appendix 4 (online) also includes subgroup analyses based on economic inequality and the implementation of a state-based exchange. Since the results from these analyses were ambiguous, I do not discuss them here.