## Abstract

Context: In November 2017, Maine became the first state in the nation to vote on a key provision of the Affordable Care Act: the expansion of Medicaid.

Methods: This study merged official election results from localities across Maine with Census Bureau and American Hospital Association data to identify characteristics of areas that support Medicaid expansion.

Findings: Places with more bachelor's degree holders more often vote in favor, whereas those with more associate's degree graduates tend to vote against. Conditional on education rates, areas with more uninsured individuals who would qualify for expanded coverage tend to vote in favor, while those with more high-income individuals tend to vote against. Also conditional on education rates, greater hospital employment is associated with support for expansion, but the presence of other health professionals, whose incomes might decrease from expansion, is associated with less support.

Conclusions: Voting patterns are mostly consistent with economic self-interest, except for the sizable association of bachelor's degree holders with support for Medicaid expansion. Direct democracy can shift Medicaid policy: extrapolating to other states, the model predicts that hypothetical referenda would pass in 5 of the 18 states that had not yet expanded Medicaid at the time of Maine's vote.

The landmark 2012 US Supreme Court decision in National Federation of Independent Business (NFIB) v. Sebelius (567 U.S. 519 [2012]) upheld the 2010 Affordable Care Act (ACA) as federal law. Yet, it limited a key provision of the law that would have expanded health insurance coverage rates for adults in low-income households in all 50 states. Traditional Medicaid limits eligibility based on both financial resources and categorical rules (e.g., age, disability status, and family structure). As enacted, the ACA raised the income eligibility cutoff to 138% of the federal poverty level and removed Medicaid's categorical rules, thereby permitting nondisabled, nonelderly adults with no children to qualify for benefits. The NFIB v. Sebelius decision prevented the federal government from requiring states to expand Medicaid but allowed states to elect to do so individually.

Leaving the expansion decision to individual states created uncertainty about the timing and eventual scope of Medicaid expansion under the ACA (Rosenbaum and Westmoreland 2012). The first state expansions came into effect in January 2014. By November 6, 2017, 31 states and Washington, DC, had expanded their Medicaid programs, while 19 others had not. State expansion of Medicaid largely aligned with Democratic Party control of state government, but other factors, including medical providers' political power and the state's administrative capacity, also facilitated expansion (Hertel-Fernandez, Skocpol, and Lynch 2016; Jacobs and Callaghan 2013; Lanford and Quadagno 2016). States' decisions to expand Medicaid have had wide-ranging effects, including on insurance coverage and access to care (Courtemanche et al. 2017; Frean, Gruber, and Sommers 2017; Sommers et al. 2015), labor force participation (Duggan, Goda, and Jackson 2017; Gooptu et al. 2016), rates of uncompensated care at hospitals (Dranove, Garthwaite, and Ody 2016; Nikpay, Buchmueller, and Levy 2015), and payday borrowing (Allen et al. 2017).

Although researchers can estimate the costs and benefits of Medicaid expansion across various subgroups in the population, it is unclear whether these groups actually support or oppose expansion. Little is known about how voting for Medicaid expansion aligns with social and economic characteristics of individuals and local areas. Do those who gain financially from the expansion support it the most at the ballot box?

Opinion polls, such as those conducted by the Kaiser Family Foundation, reveal correlations between individual characteristics, such as race, party affiliation, and Medicaid receipt, and stated opinions about Medicaid expansion (Grogan and Park 2017a, 2017b).1 Unfortunately, existing surveys lack the ability to identify the opinions of health care providers. It is anyhow unclear how stated support for Medicaid expansion in a survey translates into actual voting behavior. Because survey responses carry no economic consequences, respondents may not invest in developing their opinion or respond truthfully (Brunner, Ross, and Washington 2011; Matsusaka and McCarty 2001). Furthermore, survey samples may not be representative of the voting population. To overcome these limitations, we instead examine voting data directly from a referendum on Medicaid expansion as a stand-alone issue.

On November 7, 2017, Maine became the first state in the nation to hold a popular vote on whether to expand Medicaid. In this historic referendum, 59% of votes were cast in favor of expansion. This vote provides a unique view into popular support for the Medicaid expansion when isolated from other political issues. Indeed, to date it is the only element of the ACA that has been evaluated through direct democracy. In this article, we analyze how citizens voted across localities in Maine.

Our analysis provides the first empirical evidence on the characteristics of populations that support Medicaid expansion as a stand-alone ballot issue. This enables us to provide new insight into factors that affect voting for Medicaid expansion. We identify key economic and demographic factors that correlate with localities in Maine voting in support of Medicaid expansion and assess how these factors align with financial self-interest. Finally, we use these factors to predict vote shares for hypothetical stand-alone ballot measures on Medicaid expansion in other states. Dissonance between state policy and these predicted voting outcomes point to a potential role for direct democracy in shaping the evolution of Medicaid policy across states.

## Background

### Politics of Medicaid Expansion

Reflecting the highly politicized nature of the ACA, political partisanship played a large role in states' decisions to expand Medicaid, with Democratic politicians generally favoring expansion and Republicans opposing it. Nevertheless, several Republican-led states have opted to expand their programs. Medicaid expansion was more extensive in states with more affluent residents, state administrative capacity, generous Medicaid coverage rules already in place before the ACA, and politically influential health care providers; it was less extensive in areas with higher levels of conservative ideology or racial resentment in the population (Hertel-Fernandez, Skocpol, and Lynch 2016; Jacobs and Callaghan 2013; Lanford and Quadagno 2016). When conservative politicians support expansion, they sometimes do so in combination with other program changes using a Medicaid waiver (Rosenbaum and Westmoreland 2012). This strategy enables them to frame the expansion as a “reform” to Medicaid rather than as an expansion of the existing program (Grogan, Singer, and Jones 2017).

These various political factors result in state Medicaid expansion policies that appear imperfectly aligned with popular opinion. Barrilleaux and Rainey (2014) found that increased public support for the ACA does not reduce gubernatorial opposition to Medicaid expansion. Grogan and Park (2017b) found that greater public support for Medicaid expansion among whites increases the probability of a state expanding Medicaid but that public support among nonwhites does not. The conflict is evident in nonexpansion states where large majorities of the population favor expanding Medicaid.2 Opinion polling itself may not represent constituents' preferences if their opinions are not fully formed (Matsusaka and McCarty 2001) or truly reported (Skibba 2016) or if low response rates yield unrepresentative samples (Skibba 2016). Nevertheless, the discrepancy between polling and policy suggests that statewide votes on Medicaid expansion might yield different policy choices in some states.

A central argument for direct democracy is that allowing voters to initiate laws without going through their elected representatives can better align public policy with public opinion (e.g., Besley and Case 2003; Lascher, Hagen, and Rochlin 1996).3 A key feature of ballot initiatives, like the one we studied here, is that they allow voters to decide on a single issue in isolation. This achieves an “unbundling” of the various social and economic positions that each candidate for political office offers to voters (Besley and Coate 2000). Empirical studies that compare states with and without direct democracy provisions have found mixed evidence of greater policy alignment (for an early summary of this literature, see Besley and Case 2003: 47; see also Bowler and Donovan 2004; Lax and Phillips 2009). In theory, the ability to pursue a ballot initiative should in itself affect policy outcomes (Besley and Coate 2000; Gerber 1996). In practice, voters in several states make regular use of the initiative process to create policy (Matsusaka 2005; Matsusaka and McCarty 2001).

Maine is one of 24 states with direct democracy provisions that allow new statutes or constitutional amendments to be adopted through a ballot initiative process, and its politics made Medicaid expansion a natural question to pose to voters. Over the years, Governor Paul LePage vetoed five different bills passed by Maine's legislature to expand the program under the ACA, despite public opinion polling showing 68% support for expansion (Maine People's Resource Center n.d.). Medicaid expansion's peripheral status as a political issue likely contributed to this dissonance. The Maine electorate voted in 2017 to expand Medicaid, despite having reelected Governor LePage only three years earlier. Even after the Maine referendum passed, Governor LePage continued to obstruct its implementation.

### Maine Ballot Initiative

The Maine ballot initiative was the first binding vote in any state on a component of the ACA. After a petition for the initiative obtained the required number of signatures, Secretary of State Matthew Dunlap wrote the ballot question and determined that it would be positioned second on the ballot, following an initiative to allow slot machines or a casino in York County. The question was worded as follows: “Do you want Maine to expand Medicaid to provide healthcare coverage for qualified adults under age 65 with incomes at or below 138% of the federal poverty level, which in 2017 means $16,643 for a single person and$22,412 for a family of two?”

The original proposed question text asked about “expanding Medicaid to provide health insurance,” but a group of Republican lawmakers advocated for changing “insurance” to “taxpayer-funded health benefits” or “government-funded health benefits” (Thistle 2017b). The group also wanted to include an estimated annual cost in the question text. The final wording (stated above) referred to “healthcare coverage” and did not include estimates of direct costs, federal transfers, or cost savings from the expansion.

The proposed legislation that the question represented was titled “An Act to Enhance Access to Affordable Health Care.” As per direct democracy procedures in Maine, the legislature could amend the law after the vote and before its adoption, but the governor could not veto it. In advance of the vote, the secretary of state circulated a summary of the legislation, an estimate of its fiscal impact (annual costs of $54 million to Maine and$525 million to the federal government), and two public comments filed in support of the measure. No public comments were filed in opposition to it.

The organization leading the pro-expansion efforts, Mainers for Health Care, was a consortium of advocacy groups, including the Maine Center for Economic Policy, Maine Equal Justice Partners, Maine People's Alliance, Maine Voice's Network, and Planned Parenthood of Northern New England Maine Action Fund. The consortium spent nearly $2 million on the campaign. The primary opposition group was the Welfare to Work PAC, run by a former adviser to Governor LePage and formed in late August 2017 (Thistle 2017a); this group spent$385,000.

Echoing broader Medicaid politics (see Hertel-Fernandez, Skocpol, and Lynch 2016; Lanford and Quadagno 2016), the business community in Maine was divided on the issue. The National Federation of Independent Business of Maine opposed expansion, consistent with its parent organization's role as the plaintiff in the 2012 NFIB v. Sebelius lawsuit. But the Maine Small Business Coalition endorsed expansion, as did several health care interest groups, including the Maine Medical Association, the Maine Primary Care Association, and the Maine Hospital Association (Stone 2013; Wight 2017). The Maine Chamber of Commerce did not take a position.

Proponents of the initiative argued for the morality of expanding insurance coverage to low-income individuals; advertisements featured health care providers and others making personal appeals for clients or friends who would benefit from the expansion. The economic arguments in favor of expansion centered on obtaining federal subsidies for expansion (Ardoff 2017) and improving the financial status of hospitals (Peterson 2017). The opposition's arguments focused on the potential costs of the expansion to the state, citing Maine's experience with a 2002 Medicaid expansion and claiming, in an appeal to senior citizens, that Medicaid expansion would necessitate cutting funding for nursing homes.

## Data and Methods

We examined official election results of the November 2017 referendum from the Maine Bureau of Corporations, Elections, and Commissions, which reported nonzero vote totals for 495 localities. Our analysis consisted primarily of weighted least squares regressions.

The primary outcome of interest was the vote share in favor of Medicaid expansion. Vote shares reflect actual voting behavior that has real financial consequences for taxpayers, health care providers, and people who could potentially become eligible for Medicaid coverage. These real financial consequences distinguish votes on ballot initiatives from responses to survey questions, which suffer from potential biases from survey respondents misreporting their preferences, characteristics, or intentions to vote (Brunner, Ross, and Washington 2011; Matsusaka and McCarty 2001).

Explanatory variables were drawn from the US Census Bureau's American Community Survey 5-year estimates (ACS 2011–15), with three exceptions: we obtained information about hospital employment from the American Hospital Association's 2012 annual survey; voter registration in August 2017 from the Maine Bureau of Corporations, Elections, and Commissions; and the land area of each locality from the US Census Bureau's 2014 gazetteer files. We also examined the determinants of voter turnout, defined as the total number of votes cast in the referendum divided by the number of registered voters. After matching, we were left with a final sample of 491 observations. Summary statistics for all variables are listed in the  appendix. Regressions explaining voting outcomes were weighted by the total number of votes cast in each locality, and turnout regressions were weighted by the number of registered voters.

Because we studied a single referendum in a single state, our analysis was limited to identifying correlations in a simple cross section. The economic and social factors that we studied may affect voting patterns in Maine differently than they do in other states. Our analysis is based on observational, not experimental or quasi-experimental, sources of variation. The estimates should be interpreted as correlations, not causal effects. For example, the coefficient estimate associated with college education might reflect the selection of individuals who attend college in addition to (or instead of) the influence of a college education on an individual's voting preference. Our aggregated unit of analysis limited us to studying variation across localities rather than within them. It also reduced the number of observations in the analysis, though each observation represents many votes.

## Results

### Educational Attainment

The first result to emerge from our analysis was a striking and robust relationship between the Medicaid vote share and educational attainment. Figure 1 reveals the pairwise association. The share of the population with at least a bachelor's degree (x-axis) is plotted against the vote share for Medicaid expansion (y-axis), using a circle for each locality in the data, sized to reflect the total number of votes cast. College education has a strong, positive, and roughly linear association with voting in favor of Medicaid. The association's estimated slope is 0.65 and highly statistically significant (p < 0.001). Variation in education also explains a majority of the variation in Medicaid voting across localities (R2 = 0.54).

Although having a college degree is highly correlated with income, controlling for income did not diminish the estimated importance of education. Rather, when variables that measure the population shares in various annual household income categories ($25–50K,$50–75K, $75–100K,$100K+) were added to the model, the slope on college education increased to 0.79.

As shown in table 1, controlling for college education has an important effect on estimates for income categories. When education is not included in the model (column 1), middle-income households are associated with less support for expansion than are low-income households (the difference is statistically significant for households with incomes of $50–75K and$75–100K), but high-income households ($100K+) are associated with even stronger support (coefficient = 0.18; p < 0.001). Adding the control for college graduates to the model (column 2) reverses the sign on high-income households to a statistically significant −0.46 (p < 0.001), making the general pattern more monotonic. Relative to the lowest income groups (<$25K and $25K–50K, which are not statistically distinguishable), middle-income levels are associated with moderate opposition (coefficients = −0.20 and −0.18) and high incomes are associated with the strongest opposition. This reversal of high-income households' association with support from positive to negative by controlling for college education suggests there is a major split within the group between high-income individuals with college degrees who support expansion and high-income individuals with less education who oppose it. This could mean that more income on its own lowers support for redistributive programs, possibly because of the anticipated tax burden and the limited personal benefit,4 but that a college education mitigates this effect such that even college graduates with annual incomes above$100K favor expanding Medicaid.

This variation in support for Medicaid expansion across households in different income categories might also reflect the importance of income inequality. Political economy theory predicts that growing inequality bolsters popular support for redistributive economic policies (Ashok, Kuziemko, and Washington 2015). To examine this hypothesis, we measured income inequality in each locality using the Gini coefficient. Higher levels of local inequality are associated with greater support for Medicaid expansion, both in models with (coefficient = 0.35; p < 0.001) and without (coefficient = 0.66; p < 0.001) a control for the share of college graduates. Statewide or national levels of income inequality might also affect support for redistributive policies, but we are unable to examine their effect in our setting, because they do not vary within our sample.

Because college graduates are more likely to live in more densely populated areas, the regressions reported in table 1 and all subsequent tables control for the natural logarithm of population density. That variable is positively related to voting for Medicaid expansion (coefficient = 0.03; p < 0.001), consistent with the tendency of more urban areas to support public policies that increase redistribution. Excluding the variable for density makes the estimated impact of college education even larger.

All models also control for the areas' racial and ethnic composition using the population shares of blacks and Hispanics. In opinion polling, a respondent's race is among the strongest predictors of support for Medicaid expansion (Grogan and Park 2017a, 2017b).5 Because race is measured at the locality level in our analysis, it also captures the effects of the locality's racial diversity on policy preferences (Matsubayashi and Rocha 2012). Controlling for race has virtually no effect on the estimated impact of college education.

Column 3 of table 1 reports results for a model that also controls for party affiliation, another strong predictor of support for Medicaid expansion (Grogan and Park 2017a, 2017b). We controlled for party affiliation using the Republican share of registered voters. We excluded party affiliation from our main model because it is itself a political outcome, which may partly reflect voters' preferences about the ACA and Medicaid expansion. Controlling for party affiliation could therefore cause the relationship between college education and support for Medicaid expansion to be understated. Nevertheless, the results in column 3 show that the effect of college remains large and highly statistically significant when we control for party affiliation (coefficient = 0.70; p < 0.001).

The relationship between college education and preferences for Medicaid expansion has not been highlighted in previous research on state policy making or public opinion about Medicaid expansion. The finding builds on opinion polling reported by Grogan and Park (2017a: table 4), which estimated that support for Medicaid expansion is 2.4 times greater among the college educated than among high school dropouts. But that estimate is not statistically significant, and Grogan and Park emphasized the importance of race, political partisanship, and health insurance coverage (i.e., being uninsured, a Medicaid beneficiary, or a beneficiary's friend or family member), not educational attainment. Other studies of public opinion about Medicaid expansion, as well as the broader literature on public opinion about the ACA, its other components, and health care reform more generally, have similarly focused on explanatory factors other than educational attainment (see, e.g., Berk, Gaylin, and Schur 2006; Blendon et al. 2008; Blendon and Benson 2017; Brodie et al. 2010; Grande, Gollust, and Asch 2011; Sussman, Blendon, and Campbell 2009).

One potential explanation for the stronger, and more precisely estimated, relationship between college and voting for Medicaid is that college education is related both to holding a more favorable opinion about expanding Medicaid and to being more likely to vote. Sondheimer and Green (2010) and Mayer (2011) have shown that education increases voter turnout in federal elections. We examined the relevance of this effect in our setting by analyzing turnout in the referendum. The raw correlation between college education and voter turnout across localities in our sample is depicted in figure 2. The estimation results using turnout as the dependent variable are reported in column 4 of table 1. Voter turnout was greater in areas where a greater share of the population had a bachelor's degree. A bachelor's degree is associated with a 16 percentage point higher turnout, a 47% increase relative to the sample mean.

The power of college education in explaining Medicaid vote shares naturally raises questions about what aspects of college are driving the positive effect. We investigate this in two ways. First, we considered effects of postsecondary schooling leading to an associate's rather than a bachelor's degree. Second, we tested for differences among graduates of 4-year colleges based on their subject of study. The results are reported in table 2.

Higher rates of postsecondary education in the form of associate's degrees have a large, negative, and significant association with support for Medicaid expansion; this is the opposite of what we found for bachelor's degrees. In our baseline model, an additional 1% of a locality in which voters' highest educational level is an associate's degree (rather than a high school diploma or less) is associated with a 0.70 percentage point lower vote share for Medicaid (column 1). When controlling for party affiliation (column 2), the estimate is smaller in magnitude but still highly statistically significant (coefficient = 0.38; p < 0.001). The opposite signs of the vote share associations for bachelor's and associate's degree holders may result from the preferences or economic circumstances of students who obtain 2-year rather than 4-year degrees, their instructional experiences (e.g., part-time vs. full-time), or the subjects they study. Associate's degree programs involve less coursework and tend to be more vocationally oriented.

There is also substantial variation in Medicaid support among localities with higher rates of 4-year college graduates based on the graduates' primary field of study. The ACS divides bachelor's degree holders based on the broad subject area of their (first) bachelor's degree. We included in our model the shares of the local population who have bachelor's degrees in each of the following groups: science and engineering; science- and engineering-related fields (e.g., nursing, architecture, and math teacher education); business; education; and arts, humanities, and others.

The different majors are associated with starkly different voting patterns. Undergraduate business majors appear more similar to 2-year college graduates than their peers with other bachelor's degrees. Localities with more business majors tend to have lower vote shares for Medicaid. The positive association with bachelor's degree holders overall is driven by substantial positive relationships for science and engineering, education, and arts and humanities majors. The largest association is 0.85–0.97 for arts and humanities majors (p < 0.001). Each additional person in an area with one of these degrees is associated with nearly one more vote for Medicaid expansion. The strength of this association could reflect robust preferences by those degree holders for Medicaid expansion. It may also reflect their being more likely to turn out to vote in the election.

We examined the question of turnout directly by again analyzing the total number of votes cast scaled by the number of registered voters. The results are reported in column 3 of table 2. We found a large and significant positive relationship between arts and humanities majors and turnout for the Medicaid vote (coefficient = 0.57; p < 0.001). We also found a significant increase in turnout associated with more business majors (coefficient = 0.44; p < 0.01), which suggests that turnout amplifies the impact of that group's preference against the expansion.

Higher rates of advanced degrees are also correlated with Medicaid voting. We repeated the analysis after breaking down bachelor's degree holders into those with no advanced degree, those with a master's degree, those with a professional degree, and those with a doctoral degree (data not shown). The model continued to control for the population share with an associate's degree, population density, and race. The estimates imply that bachelor's degree holders without advanced degrees are modestly more supportive of Medicaid expansion than are people without college degrees (coefficient of 0.21; p < 0.001), whereas those with master's and doctoral degrees are even more supportive (coefficients of 1.10 and 1.24, respectively; p < 0.001), and those with professional degrees are less supportive (coefficient = −0.62; p < 0.05). Further analysis suggests that the negative estimate for professional degrees is driven by health professionals, as the effect was no longer present after we controlled for the population share employed in health care diagnosis and treatment. We discuss health professionals in greater detail in the next section.

### Interest Groups

Four-year college graduates' strong implied support for Medicaid expansion is interesting in part because it does not obviously align with their economic self-interest. College graduates are less likely to be uninsured or to become eligible for Medicaid under the expansion.

Self-interest is, however, an apparent motivator for other populations. In table 3, we report results for other directly affected groups. First, we looked at the population that is most likely to gain insurance coverage from the proposed expansion of Medicaid: the uninsured with income below 138% of the poverty line. Consistent with self-interested voting, areas with higher population shares in this group are significantly more favorable toward Medicaid expansion (column 1; coefficient = 0.78; p < 0.001).

We also examined the votes of health care providers, who would benefit from their patients gaining expanded insurance coverage that increases demand for their services (Card, Dobkin, and Maestas 2008; Finkelstein et al. 2012, 2016; Taubman et al. 2014). The financial benefit is greater for hospitals, because expansion reduces uncompensated care (Nikpay 2016; Nikpay, Buchmueller, and Levy 2015, 2016). Unlike physicians' offices, hospitals are required (by the 1946 Hill-Burton Act and the 1986 Emergency Medical Treatment and Labor Act) to provide care to patients in emergency situations regardless of their ability to pay. Outside the hospital setting, however, the financial impact of Medicaid expansion on health care providers is less obvious. Medicaid has a relatively low reimbursement rate,6 and physicians may not see much financial gain from its expansion, particularly if some of the expansion comes from crowding out higher-paying private insurance (Frean, Gruber, and Sommers 2017; Kaestner et al. 2017).

Consistent with both incentives, we found that support for Medicaid expansion is positively associated with hospital employment (total full-time equivalent hospital employment as a share of the area population) and negatively associated with nonhospital providers.7 Like our finding for the uninsured, the positive association with hospital employment (coefficient = 0.31; p < 0.001) is consistent with self-interest, as hospitals expected to benefit from an expanded health insurance safety net. These votes align with the Maine Hospital Association's endorsement of the referendum. Although Maine's major physician and nurse associations also supported Medicaid expansion, the voting data suggest that their members were less likely to fall in line. The negative association with the population share working in health diagnosis and treatment8 (coefficient = −0.71; p < 0.001) is consistent with health care providers outside of hospitals expecting Medicaid expansion to make them worse off because of Medicaid's low reimbursement rate. The association is also negative but statistically insignificant for health technologists and technicians; it is positive and insignificant for health care support occupations.9

The model in table 3, column 1, includes controls for the shares of the population with associate's and bachelor's degrees. The other estimates should thus be interpreted as the associations with changes in those variables while keeping the shares of college graduates unchanged. Because workers in health diagnosis and treatment typically have bachelor's degrees, the negative estimate indicates that they are less likely to support Medicaid expansion than other 4-year college graduates, who tend to favor expansion (coefficient in this model = 0.57; p < 0.001). The other health care provider groups are drawn more from associate's degree holders, who are less likely to support the expansion (coefficient in this model = −0.61; p < 0.001).

Omitting the education variables from our regression model (column 2) results in a significant positive association for health diagnosis and treatment workers (coefficient = 0.62; p < 0.001) and significant negative associations for health technologists and technicians (coefficient = −1.10; p < 0.001) and for those in health care support occupations (coefficient = −0.89; p < 0.001).10 Taken together, these results appear to show that health diagnosis and treatment workers are more supportive of Medicaid expansion than the general population but less supportive than other 4-year college graduates; other health care workers are roughly as supportive as others with similar education.

Controlling for party affiliation leaves the estimates qualitatively unchanged but smaller in magnitude. Estimates from a model with this control, reported in table 3, column 3, can be compared to those from a model without it, reported in column 1. The magnitude of the coefficient estimate on the population eligible to gain insurance decreases only slightly (from 0.78 to 0.76), the one on hospital employment decreases moderately (from 0.31 to 0.26), and the one on health diagnosis and treatment workers decreases a lot (from −0.71 to −0.21). All three estimates remain statistically significant. The larger impact of these controls on the negative estimates than on the positive ones has multiple potential interpretations. For example, partisanship could be a bigger factor driving people to vote against expansion than to vote for it, or opposition to the ACA could be more likely to motivate moderates to register Republican than support for it motivates them not to.

## Projection to Other States

At the time of Maine's referendum, 18 other states had not yet expanded Medicaid under the ACA. Opinion polls in these states show popular support for expansion (KFF 2018), and advocates are campaigning for ballot initiatives in Idaho, Nebraska, and Utah. Yet it is unclear how these or similar initiatives would fare at the ballot box.

We used regression estimates from the vote for Medicaid expansion in Maine to project votes for expansion in other states, based on those states' economic and demographic characteristics. The resulting estimates predict how other states would vote in a referendum similar to Maine's but do not account for how differences in the context, framing, or timing of ballot initiatives in other states would affect their popularity. Maine is an unusual state in several dimensions. For example, it is more rural, has a smaller minority population, and has lower Medicaid reimbursement rates than average. Yet, it is not homogeneous. To the extent that the variation within Maine can inform variation elsewhere, a regression model estimated across localities in Maine can be used to predict vote shares in hypothetical referenda in other states. Because this assumption is impossible to verify, this exercise is fundamentally speculative in nature.

The goal of this regression is to predict total voting shares rather than to interpret the magnitude of individual coefficients, as we did in the regressions reported above. To increase the model's explanatory power, we included all of the variables (except voter registration) from the previous models. Instead of Republican voter registration, we used President Donald J. Trump's vote share in the 2016 election to capture party affiliation and voting preferences for the Obama agenda, including the ACA. We used Trump's vote share instead of voter party registration because of differences across states in the importance of party registration for selecting candidates for ballots. Altogether, these variables explain 92% of the variation in vote share across localities in Maine.

Figure 3 depicts the results of this analysis on a map of the United Sates. States are shaded based on their predicted vote share in a hypothetical Medicaid expansion referendum. States are shaded darker when their predicted share is greater than 50% (passing the referendum) and lighter when it is less than 50% (failure to pass). Within these categories, the shading is further differentiated on the basis of the status of Medicaid expansion policy, with darker shades for states that had not expanded Medicaid at the time of Maine's referendum.

Although we project the outcomes of hypothetical referenda in all states, not all states allow citizens to use direct democracy to make laws.11 For states without direct democracy, the projection results indicate how citizens would vote if they could. The projections allow us to identify states where the result of a hypothetical ballot initiative differs from the realized policy outcome. These differences point to a potential role for direct democracy in shaping Medicaid policy.

Among the states that had not expanded Medicaid at the time of the Maine referendum, our model predicts that five would vote in favor of an expansion in a statewide ballot. In decreasing order of predicted vote share, ranging from 57% to 51%, these are Utah, Texas, Wisconsin, Florida, and Virginia.12 Under current law, citizens in two of those five states (Utah and Florida) have access to direct ballot initiatives. A hypothetical initiative on Medicaid expansion is predicted to fail in the remaining 13 nonexpansion states, with predicted vote shares ranging from 46% (Idaho) to 29% (Mississippi).

We used a similar exercise to predict the outcome of expansion votes in states that have already expanded their Medicaid programs. For these states, we used the proportion of uninsured with incomes below 138% of the poverty line in 2012 (before Medicaid expanded) as the basis for the calculation. Perhaps unsurprisingly, the majority of these states are predicted to support expansion in a referendum, with vote shares ranging from 77% (Washington, DC) to 50.03% (Pennsylvania). However, the model predicts that a referendum to expand Medicaid would fail in 11 states that have already expanded. In decreasing order of predicted vote share ranging from 49% to 33%, these states are Michigan, Iowa, Alaska, Montana, Ohio, Indiana, Kentucky, Arkansas, North Dakota, West Virginia, and Louisiana. Six of these 11 states allow direct democracy ballot initiatives.

This last result should be interpreted carefully. The basis of our prediction is an empirical model of voting in 2017 in a state that had not yet expanded Medicaid under the ACA. This is not the same as predicting how residents of these 11 states would have voted at the time that their program was expanded; some of the reversals may reflect changing preferences rather than an adoption that constituents would not have voted for themselves. The status quo also likely affects voters' preferences. While our model predicts that majorities of voters in some expansion states would not approve a ballot initiative favoring Medicaid expansion, this does not imply that majorities would support a measure that takes away people's benefits by retracting their state's already adopted expansion.

## Conclusions

Maine's 2017 referendum on Medicaid expansion provides a unique opportunity to observe how people vote on Medicaid expansion as a stand-alone issue. Consistent with some votes reflecting self-interest, we found greater support for Medicaid expansion in areas where more of the population would gain insurance and where hospitals, which would benefit from a reduced need for uncompensated care, have a larger presence. At the same time, we found a robust positive association between the share of an area's population with at least a bachelor's degree and its support for Medicaid expansion, which appears to contradict financial self-interest. These correlations may reflect the causal impact of a 4-year college education on an individual's preference for Medicaid expansion and voting behavior, but they may also be driven in part by differences in the attitudes of individuals who self-select into college attendance and completion. Uncovering the mechanism that drives this association is an interesting area for future research.

Even without understanding the precise sources of each association, it is possible to use these associations to predict the outcomes of ballot initiatives on Medicaid expansion in other states. Our model predicts that ballot initiatives would pass in 5 of the 18 states that had not expanded Medicaid at the time of Maine's vote. Laws in two of these states—Utah and Florida—allow for ballot initiatives. These results suggest that, by allowing voters to isolate the issue of Medicaid expansion in a stand-alone ballot initiative, state-level direct democracy rights can lead to greater adoption of Medicaid expansion than is produced through elected state governments. Although direct democracy was originally adopted among Progressive Era reforms in the early twentieth century, in recent decades it has been used to limit taxation that supports the social safety net (Tolbert 2003). The outcomes of Maine's referendum and of our analysis indicate that direct democracy can nevertheless still be used to expand the social safety net.

## Acknowledgments

We thank Ariza Gusti and Kevin Hare for research assistance and Analia Schlosser and two anonymous referees for helpful comments.

## Notes

1.

Stated support for Medicaid expansion is also greater among people without health insurance and lower among those who believe that Medicaid stigmatizes recipients (Grogan and Park 2017a: table 4).

2.

The Kaiser Family Foundation's Health Tracking Poll asked residents of nonexpansion states if they thought their state should expand Medicaid or keep the current program rules. In July 2015, 61% favored expansion and 32% were opposed (KFF 2015). In February 2018, after three more states (Alaska, Montana, and Louisiana) expanded, 56% of the remaining nonexpansion states supported expansion and 37% were opposed (KFF 2018).

3.

A second argument is that direct democracy increases participation rates and voter turnout, but empirical support for this hypothesis has been limited (see, e.g., Lascher, Hagen, and Rochlin 1996; Tolbert and Smith 2005).

4.

Although Medicaid expansion in Maine would be funded largely by taxpayers in other states, Maine voters might worry that their expansion could lead other states to follow. They might also be concerned that federal contributions will decrease over time. The federal government, which provides 94% of expansion funding in 2018, is set to reduce funding to 90% from 2020 onward, and Congress could decide to reduce it further.

5.

Public support for the ACA overall may be affected by race and racial attitudes because of the law's strong association with President Obama. The racial divide in public support for health care reform increased by 20 percentage points between 1993–94 and 2009–10 (Tesler 2012), and experimental evidence shows that linking health care reform to President Obama rather than to President Clinton increases opposition among individuals with higher levels of racial resentment and implicit prejudice (Knowles, Lowrey, and Schaumberg 2010; Tesler 2012).

6.

Maine has a Medicaid-to-Medicare fee index for “all services” of only 0.64, which is lower than the national average of 0.72 (excluding Tennessee, where data are not available), 37 other states, and the District of Columbia (Zuckerman, Skopec, and Epstein 2017).

7.

Because of differences in data sources, the geographic basis for the occupation data on all health professionals is the worker's residence (which is also where they vote), but it is the hospital location for hospital employment.

8.

This Census occupational category is “health diagnosing and treating practitioners and other technical occupations” and includes physicians and surgeons, therapists, nurses, dentists, veterinarians, and chiropractors.

9.

The Census occupational category “healthcare support occupations” includes nursing, psychiatric, and home health aides; occupational and physical therapists and aides; massage therapists; dental assistants; and pharmacy assistants.

10.

In the model without education controls, the coefficient for the low-income uninsured population share remains positive but is smaller and not statistically significant. This is because the low-income uninsured population has lower average educational attainment, which is otherwise associated with opposition to Medicaid expansion; omitting education controls introduces a downward bias in the estimate for the uninsured.

11.

Twenty-one states have citizen-initiated state statutes: Alaska, Arizona, Arkansas, California, Colorado, Idaho, Maine, Massachusetts, Michigan, Missouri, Montana, Nebraska, Nevada, North Dakota, Ohio, Oklahoma, Oregon, South Dakota, Utah, Washington, and Wyoming. Three other states have citizen-initiated constitutional amendments: Florida, Illinois, and Mississippi. Maryland and New Mexico allow citizens to initiate a referendum to veto a new law but not to pass a new one.

12.

Virginia subsequently became an expansion state on June 7, 2018, when Governor Ralph Northam signed into law a state budget that included Medicaid expansion.

## References

References
AHA (American Hospital Association)
.
2019
. “
AHA Annual Survey Database
Allen
Heidi
,
Swanson
Ashley
,
Wang
Jialan
, and
Gross
Tal
.
2017
. “
Early Medicaid Expansion Associated with Reduced Payday Borrowing in California
.”
Health Affairs
36
, no.
10
:
1769
76
.
Ardoff
Arthur Barry
.
2017
. “
Medicaid Expansion Would Be a ‘Win-Win’ for Maine
.”
Bangor Daily News
,
September
13
.
Ashok
Vivekinan
,
Kuziemko
Ilyana
, and
Washington
Ebonya
.
2015
. “
Support for Redistribution in an Age of Rising Inequality: New Stylized Facts and Some Tentative Explanations
.”
Brookings Papers on Economic Activity
, Spring,
367
405
.
Barrilleaux
Charles
, and
Rainey
Carlisle
.
2014
. “
The Politics of Need: Examining Governors' Decisions to Oppose the ‘Obamacare’ Medicaid Expansion
.”
State Politics and Policy Quarterly
14
, no.
4
:
437
60
.
Berk
Marc
,
Gaylin
Daniel
, and
Schur
Claudia
.
2006
. “
Exploring the Public's Views on the Health Care System: A National Survey on the Issues and Options
.”
Health Affairs
25
, no.
6
:
w596
w606
.
Besley
Timothy
, and
Case
Anne
.
2003
. “
Political Institutions and Policy Choices: Evidence from the United States
.”
Journal of Economic Literature
41
, no.
1
:
7
73
.
Besley
Timothy
, and
Coate
Stephen
.
2000
. “
Issue Unbundling via Citizens' Initiatives
.” National Bureau of Economic Research Working Paper No. w8036.
Cambridge, MA
:
National Bureau of Economic Research
.
Blendon
Robert J.
,
Altman
Drew E.
,
Benson
John M.
,
Brodie
Mollyann
,
Buhr
Tami
,
Deane
Claudia
, and
Buscho
Sasha
.
2008
. “
Voters and Health Reform in the 2008 Presidential Election
.”
New England Journal of Medicine
359
, no.
19
:
2050
61
.
Blendon
Robert
, and
Benson
John
.
2017
. “
Public Opinion about the Future of the Affordable Care Act
.”
New England Journal of Medicine
377
, no.
9
:
e12
.
Bowler
Shaun
, and
Donovan
Todd
.
2004
. “
Measuring the Effect of Direct Democracy on State Policy: Not All Initiatives Are Created Equal
.”
State Politics and Policy Quarterly
4
, no.
3
:
345
63
.
Brodie
Mollyann
,
Altman
Drew
,
Deane
Claudia
,
Buscho
Sasha
, and
Hamel
Elizabeth
.
2010
. “
Liking the Pieces, Not the Package: Contradictions in Public Opinion during Health Reform
.”
Health Affairs
29
, no.
6
:
1125
30
.
Brunner
Eric
,
Ross
Stephen
, and
Washington
Ebonya
.
2011
. “
Economics and Policy Preferences: Causal Evidence of the Impact of Economic Conditions on Support for Redistribution and Other Ballot Proposals
.”
Review of Economics and Statistics
93
, no.
3
:
888
906
.
Card
David
,
Dobkin
Carlos
, and
Maestas
Nicole
.
2008
. “
The Impact of Nearly Universal Insurance Coverage on Health Care Utilization: Evidence from Medicare
.”
American Economic Review
98
, no.
5
:
2242
58
.
Courtemanche
Charles
,
Marton
James
,
Ukert
Benjamin
,
Yelowitz
Aaron
, and
Zapata
Daniela
.
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
, no.
1
:
178
210
.
Dranove
David
,
Garthwaite
Craig
, and
Ody
Christopher
.
2016
. “
Uncompensated Care Decreased at Hospitals in Medicaid Expansion States but Not at Hospitals in Nonexpansion States
.”
Health Affairs
35
, no.
8
:
1471
79
.
Duggan
Mark
,
Goda
Gopi
, and
Jackson
Emilie
.
2017
. “
The Effects of the Affordable Care Act on Health Insurance Coverage and Labor Market Outcomes
.” National Bureau of Economic Research Working Paper No. 23607.
Cambridge, MA
:
National Bureau of Economic Research
.
Finkelstein
Amy
,
Taubman
Sarah
,
Allen
Heidi
,
Wright
Bill
, and
Baicker
Katherine
.
2016
. “
Effect of Medicaid Coverage on ED Use—Further Evidence from Oregon's Experiment
.”
New England Journal of Medicine
375
, no.
16
:
1505
7
.
Finkelstein
Amy
,
Taubman
Sarah
,
Wright
Bill
,
Bernstein
Mira
,
Gruber
Jonathan
,
Newhouse
Joseph
,
Allen
Heidi
,
Baicker
Katherine
, and
the Oregon Health Study Group
.
2012
. “
The Oregon Health Insurance Experiment: Evidence from the First Year
.”
Quarterly Journal of Economics
127
, no.
3
:
1057
1106
.
Frean
Molly
,
Gruber
Jonathan
, and
Sommers
Benjamn
.
2017
. “
Premium Subsidies, the Mandate, and Medicaid Expansion: Coverage Effects of the Affordable Care Act
.”
Journal of Health Economics
53
:
72
86
.
Gerber
Elisabeth R.
1996
. “
Legislative Response to the Threat of the Popular Initiative
.”
American Journal of Political Science
40
:
99
128
.
Gooptu
Angshuman
,
Moriya
Asako
,
Simon
Kosali
, and
Sommers
Benjamin
.
2016
. “
Medicaid Expansion Did Not Result in Significant Employment Changes or Job Reductions in 2014
.”
Health Affairs
35
, no.
1
:
111
18
.
Grande
David
,
Gollust
Sarah
, and
Asch
David
.
2011
. “
Polling Analysis: Public Support for Health Reform Was Broader than Reported and Depended on How Proposals Were Framed
.”
Health Affairs
30
, no.
7
:
1242
49
.
Grogan
Colleen M.
, and
Park
Sunggeun Ethan
.
2017a
. “
The Politics of Medicaid: Most Americans Are Connected to the Program, Support Its Expansion, and Do Not View It as Stigmatizing
.”
Milbank Quarterly
95
, no.
4
:
749
82
.
Grogan
Colleen M.
, and
Park
Sunggeun
.
2017b
. “
The Racial Divide in State Medicaid Expansions
.”
Journal of Health Politics, Policy and Law
42
, no.
3
:
539
72
.
Grogan
Colleen M.
,
Singer
Phillip M.
, and
Jones
David K.
2017
. “
Rhetoric and Reform in Waiver States
.”
Journal of Health Politics, Policy and Law
42
, no.
2
:
247
84
.
Hertel-Fernandez
Alexander
,
Skocpol
Theda
, and
Lynch
Daniel
.
2016
. “
Business Associations, Conservative Networks, and the Ongoing Republican War over Medicaid Expansion
.”
Journal of Health Politics, Policy and Law
41
, no.
2
:
239
86
.
Jacobs
Lawrence
, and
Callaghan
Timothy
.
2013
. “
Why States Expand Medicaid: Party, Resources, and History
.”
Journal of Health Politics, Policy and Law
38
, no.
5
:
1023
50
.
Kaestner
Robert
,
Garrett
Bowen
,
Anuj
, and
Fleming
Caitlyn
.
2017
. “
Effects of ACA Medicaid Expansions on Health Insurance Coverage and Labor Supply
.”
Journal of Policy Analysis and Management
36
, no.
3
:
608
42
.
KFF (Kaiser Family Foundation)
.
2015
. “
Medicare and Medicaid at Fifty
.” Publication 8755-T.
Menlo Park, CA
:
KFF
.
KFF (Kaiser Family Foundation)
.
2018
. “
Kaiser Health Tracking Poll—February 2018: Health Care and the 2018 Midterms, Attitudes towards Proposed Changes to Medicaid
.” Publication 9161-T.
Menlo Park, CA
:
KFF
.
Knowles
Eric D.
,
Lowery
Brian S.
, and
Schaumberg
Rebecca L.
2010
. “
Racial Prejudice Predicts Opposition to Obama and His Health Care Reform Plan
.”
Journal of Experimental Social Psychology
46
, no.
2
:
420
23
.
Lanford
Daniel
, and
Jill
.
2016
. “
Implementing ObamaCare: The Politics of Medicaid Expansion under the Affordable Care Act of 2010
.”
Sociological Perspectives
59
, no.
3
:
619
39
.
Lascher
Edward L.
Jr.
,
Hagen
Michael G.
, and
Rochlin
Steven A.
1996
. “
Gun behind the Door: Ballot Initiatives, State Policies, and Public Opinion
.”
Journal of Politics
58
:
760
75
.
Lax
Jeffrey R.
, and
Phillips
Justin H.
2009
. “
Gay Rights in the States: Public Opinion and Policy Responsiveness
.”
American Political Science Review
103
, no.
3
:
367
86
.
Maine People's Resource Center
. n.d.
Maine People's Resource Center Public Opinion Survey: March 31–April 3, 2013
. mprc.me/research/0413_health_results.pdf.
Matsubayashi
Tetsuya
, and
Rocha
Rene R.
2012
. “
Racial Diversity and Public Policy in the States
.”
Political Research Quarterly
65
, no.
3
:
600
614
.
Matsusaka
John G.
2005
. “
Direct Democracy Works
.”
Journal of Economic Perspectives
19
, no.
2
:
185
206
.
Matsusaka
John G.
, and
McCarty
Nolan M.
2001
. “
Political Resource Allocation: Benefits and Costs of Voter Initiatives
.”
Journal of Law, Economics, and Organization
17
, no.
2
:
413
48
.
Mayer
Alexander K.
2011
. “
Does Education Increase Political Participation?
Journal of Politics
73
, no.
3
:
633
45
.
Nikpay
Sayeh
.
2016
. “
The Relationship between Uncompensated Care and Hospital Financial Position: Implications of the ACA Medicaid Expansion for Hospital Operating Margins
.”
Journal of Health Care Finance
43
, no.
2
:
72
89
.
Nikpay
Sayeh
,
Buchmueller
Thomas
, and
Levy
Helen
.
2015
. “
Early Medicaid Expansion in Connecticut Stemmed the Growth in Hospital Uncompensated Care
.”
Health Affairs
34
, no.
7
:
1170
79
.
Nikpay
Sayeh
,
Buchmueller
Thomas
, and
Levy
Helen
.
2016
. “
Affordable Care Act Medicaid Expansion Reduced Uninsured Hospital Stays in 2014
.”
Health Affairs
35
, no.
1
:
106
10
.
Peterson
Robert
.
2017
. “
Medicaid Expansion Is Critical to the Survival of Maine's Rural Hospitals
.”
Bangor Daily News
,
October
31
.
Rosenbaum
Sara
, and
Westmoreland
Timothy
.
2012
. “
The Supreme Court's Surprising Decision on the Medicaid Expansion: How Will the Federal Government and States Proceed?
Health Affairs
31
, no.
8
:
1663
72
.
Skibba
Ramin
.
2016
. “
The Polling Crisis: How to Tell What People Really Think
.”
Nature
538
, no.
7625
:
304
6
.
Sommers
Benjamin
,
Gunja
Munira
,
Finegold
Kenneth
, and
Musco
Thomas
.
2015
. “
Changes in Self-Reported Insurance Coverage, Access to Care, and Health under the Affordable Care Act
.”
JAMA
314
, no.
4
:
366
74
.
Sondheimer
Rachel Milstein
, and
Green
Donald P.
2010
. “
Using Experiments to Estimate the Effects of Education on Voter Turnout
.”
American Journal of Political Science
54
, no.
1
:
174
89
.
Stone
Matthew
.
2013
. “
Maine Hospital Association Supports Medicaid Expansion
.”
Bangor Daily News
,
March
27
.
Sussman
Tara
,
Blendon
Robert
, and
Campbell
Andrea
.
2009
. “
Will Americans Support the Individual Mandate?
Health Affairs
28
, no.
3
:
w501
w509
.
Taubman
Sarah
,
Allen
Heidi
,
Wright
Bill
,
Baicker
Katherine
,
Finkelstein
Amy
, and
the Oregon Health Study Group
.
2014
. “
Medicaid Increases Emergency Department Use: Evidence from Oregon's Health Insurance Experiment
.”
Science
343
, no.
6168
:
263
68
.
Tesler
Michael
.
2012
. “
The Spillover of Racialization into Health Care: How President Obama Polarized Public Opinion by Racial Attitudes and Race
.”
American Journal of Political Science
56
, no.
3
:
690
704
.
Thistle
Scott
.
2017a
. “
Opponents to Medicaid Expansion in Maine Form Political Action Committee
.”
Portland Press Herald
,
August
25
.
Thistle
Scott
.
2017b
. “
Republican Group Asks to Strike ‘Insurance’ from Medicaid Expansion Ballot Wording
.”
Portland Press Herald
,
August
22
.
Tolbert
Caroline J.
2003
. “
Direct Democracy and Institutional Realignment in the American States
.”
Political Science Quarterly
118
, no.
3
:
467
89
.
Tolbert
Caroline J.
, and
Smith
Daniel A.
2005
. “
The Educative Effects of Ballot Initiatives on Voter Turnout
.”
American Politics Research
33
, no.
2
:
283
309
.
US Census Bureau
.
2018
. “
American Community Survey Data
.” www.census.gov/programs-surveys/acs/data.html.
US Census Bureau
.
2015
. “
2014 US Gazetteer Files
.” www.census.gov/geo/maps-data/data/gazetteer2014.html.
Wight
Patty
.
2017
. “
Maine Medical Association Announces Support for Ballot Initiative Expanding Medicaid
.”
Maine Things Considered
,
October
11
.
Zuckerman
Stephen
,
Skopec
Laura
, and
Epstein
Marni
.
2017
.
Medicaid Physician Fees after the ACA Primary Care Fee Bump
.
Washington, DC
:
Urban Institute
.