Abstract
Although research on the link between health and political behavior at the individual level has flourished, there have been no systematic analyses regarding the policy consequences of health inequalities in political voice. Using a unique dataset that measures the health bias in voter turnout across the fifty states from 1996 to 2012, I find that state electorates that are disproportionately more representative of healthy citizens spend less on health and have less generous Medicaid programs. The negative relationship between the degree of health bias in state electorates and these outcomes remain after controlling for the degree of class bias in voter turnout. These findings have important implications for democratic theory and policy responsiveness, as well as our understanding of variations in population health and health policy across the American states.
There is growing evidence that healthy citizens are more likely to turn out than unhealthy citizens (e.g., Mattila et al. 2013; Schur et al. 2002), that the link between health and turnout develops in young adulthood (Ojeda and Pacheco 2019), and that health and wealth are independent inputs into the decision to vote (Pacheco and Fletcher 2015). 1 Healthy citizens also tend to identify with the Republican Party (Pacheco and Fletcher 2015; Schur and Adya 2013), are less likely to think that social policy is effective at improving public health (Robert and Booske 2011), and are less supportive of government involvement in healthcare (Schur and Adya 2012). In short, there is little doubt that “health and illness shape who we are politically” (Carpenter 2012, 303).
At the same time, that research on the link between health and political behavior at the individual level has flourished, there have been little systematic analyses regarding the political consequences of health inequalities in political voice. It is reasonable to think that if healthy people are more likely to turn out and have systematically different policy preferences, as suggested by previous research, then electoral results and the policies that are enacted may have a “health bias.” Yet, without empirical evidence looking directly at policies, it is difficult to make assertions about how health inequalities in political voice contribute to biases in the public policy process.
This paper tests the proposition that an electorate disproportionately representative of healthy citizens produces biased policies that disproportionately disadvantage unhealthy citizens. I do this using a unique dataset that combines measures of “health bias” in voter turnout with state spending on health and Medicaid from 1996 to 2012. I find that state electorates that are disproportionately representative of healthy citizens spend less money on health and have less generous Medicaid programs. The negative relationship between the degree of health bias in state electorates and these inferences remain after various robustness checks. These findings have important implications for democratic theory and policy responsiveness, as well as how we understand variations in health policy and population health across the American states.
Do Health Disparities in Turnout Lead to a Bias in Public Policy?
Political equality and democratically responsive government are cherished American values. Yet, political scientists have known for a long time that participatory inequalities are not randomly distributed, but systematically biased toward the more privileged citizens who are highly educated and wealthy (e.g., Lijphart 1997). Even more troubling is that class biases in the electorate have significant political consequences. There is mounting evidence that political officials are more responsive to the interests of wealthy citizens over those who are poor (Gilens 2012; Rigby and Wright 2011; Ura and Ellis 2008) suggesting that US policymaking is dominated by economic elites and business interests with little to no input by average citizens (Gilens and Page 2014). The policy consequences of class-bias in the US electorate extend to the fifty states as well. States tend to offer more generous welfare benefits when the poor vote in equal rates to the wealthy (e.g., Avery and Peffley 2005; Hill and Leighley 1992).
Although political scientists have traditionally focused on the sociological divisions of political power based on income, there are other politically relevant disparities. Paralleling class differences are disparities on the basis of gender (Schlozman et al. 2012) and race or ethnicity (e.g., Griffin and Newman 2007) indicating that inequalities above and beyond those associated with wealth translate into differential rates of policy responsiveness (Griffin and Newman 2005; Verba, Schlozman, and Brady 1995).
When applied to health, the obvious expectation is that individual-level differences in turnout across health status lead to aggregate level health disparities in policy responsiveness and, therefore, public policy. There are three reasons, however, to suspect that this expectation may not be empirically borne out. First, health and wealth are highly correlated at the individual (Lantz et al. 1998) and aggregate levels (Wilkinson 1996). Gross domestic product (GDP) is arguably the single most predictive determinant of health at the national level (Johns et al. 2013) and health disparities trend closely with income inequality (Kawachi 2000). Hence, any relationship that exists between health inequalities in political voice and public policy may be difficult to disentangle due to the class biases that exist in voter turnout.
The second reason that health disparities in turnout may have little impact on public policy is because, as Carpenter (2012) explains, health politics is different from class-based politics. Although health and class are both related to human identity, it is much easier to divide society along economic lines (e.g., rich/poor, business/labor) than health primarily because health is multidimensional with no clear line between those who are healthy and those who are sick (Carpenter 2012). To complicate matters, disease-specific associations differ in resources and organizational strength, regardless of the underlying public health threat (Armstrong et al. 2006) leading to vast disparities in mobilization opportunities for individuals with different diseases. Take the disability rights movement contrasted with the political activity of mental health advocates as one example. The disability rights movement is nearly a century old and efforts have abounded in recent years to make the political process more open to those with a physical disability, with a certain degree of success (Alvarez et al. 2012). Mental disability, however, has only recently been redefined to include depression (Benfer 2009) and although research on the economic and social consequences of depression has flourished (e.g., Greenberg et al. 2003), there is little electoral policy to make voting easier for people with mental disorders. Thus, although healthy citizens are more likely to vote, inequities in political participation may vary widely across different health conditions making a link to public policy difficult to uncover.
Finally, health status may not translate into differential policy responsiveness because the mechanism linking health to political power and influence is blurry at best. Money is the root of representational inequality causing Gilens (2012, 10) to describe it as the “mother’s milk” of politics. The rich are able to fund their preferred candidates that then disproportionately respond to their preferences over the preferences of middle- or low-income groups. The mechanism through which health influences differential responsiveness and public policy is less understood (see, for example, Pacheco and Fletcher 2015) particularly given that the mechanism through which health influences participation differs across health ailments (Ojeda and Pacheco 2019; Rahn and Gollust 2015).
To summarize, although the expectation is that health disparities in turnout lead to a health-bias in public policy, there are several reasons to suspect that this may not be the case. Health and wealth are highly correlated, health politics are fundamentally different from class politics, and the mechanism linking health to political power is underdeveloped. Nonetheless, if we are concerned about the broad policy consequences of differential turnout across health groups, statistical analyses must address policy directly. To this end, I take advantage of the policy variation across the fifty states to provide a more robust empirical test of whether electoral disparities across health lead to health biases in public policy.
Using the Fifty States to Test the Effect of Health Bias on Health Policy
The American states provide a unique unit of analysis to study the effect of health disparities in turnout on public policy because states vary widely with respect to health biases in their electorates, health indicators, and health policy. I consider state health policy because research demonstrates that preferences on government involvement in healthcare are inversely related to health status (Henderson and Hillygus 2011; Pacheco and Fletcher 2015; Robert and Booske 2011). This type of research design allows for the comparison of health policies in states where the participation gap between those in excellent and poor health is relatively small to states with large inequalities in participation across health. Such a research design contributes to the comparative study of state politics and policy as well. There has been considerable research on the determinants of state expenditures (Jacoby and Schneider 2009; Matsusaka 2004; McLendon, Hearn, and Deaton 2006; Pacheco 2013) as well as on the variations in health policy across the states (e.g., Bailey and Rom 2004; Pacheco and Boushey 2014; Shipan and Volden 2006; Volden 2006) including, more recently, the decision for states to expand Medicaid under the Affordable Care Act (Sommers and Epstein 2010) and healthcare reform (McDonough et al. 2008; Oberlander 2007; Pacheco and Maltby 2017). Yet, the relevance of health biases in the electorate for state health policy has largely been ignored.
There are two reasons why scholars have not incorporated measures of turnout across health status into studies of public policy. First, although research on the link between health and voter turnout has flourished in recent years, systematic comparative research on the impact of health on political behavior is in its infancy (Pacheco and Fletcher 2015). Thus, it is only recently that scholars of political behavior have realized that health may be associated with political voice and power. Second, although measures of electoral class bias are pervasive, no such measures exist for health disparities in turnout no doubt because health indicators are typically not included on existing surveys that ask about voter turnout and that have large enough samples to accurately describe the states. Below I demonstrate how multiple imputation (MI) techniques can be used to create measures of health disparities in turnout across the states over time.
Measuring Health Bias Using the Biobehavioral Risk Factor Surveillance Survey (BRFSS) and the Current Population Survey (CPS)
The first step to using MI to create measures of health disparities in turnout across the states and over time is to identify data sources. Datasets must be repeated cross-sectional with state identifiers and either a measure of voter turnout or health status. Additionally, the surveys should be large so that there is enough information to make inferences about the less populated states or the surveys should employ a research design such that estimates are representative of the states. Two such surveys that meet these criteria include the Voting and Registration Supplement for the Current Population Survey and the Biobehavioral Risk Factor Surveillance Survey.
The CPS is relatively straightforward and a mainstay in political science research. The voter turnout question asks respondents that “In any election some people are not able to vote because they are sick or busy or have some other reason, and others do not want to vote. Did you vote in the election held on Tuesday November 5?” to which respondents could answer yes, no, or I don’t know. This question has been asked every other year in November since 1976. The benefit of the CPS is its large sample sizes even for the less populated states. Sample sizes tend to range from about eight hundred to five thousand per state for any given year. Scholars use the CPS extensively to estimate levels of turnout in the nation and across the states as well as class bias in turnout over time (Hill and Leighley 1992; Rigby and Springer 2011; Wichowsky 2012). I use the information available for each election year from the CPS 1996–2012 to align with the available data in the BRFSS.
The BRFSS is a telephone survey of health indicators conducted by the Centers for Disease Control (CDC) in all fifty states. Unlike other health surveys that provide national estimates, the BRFSS is a state-based survey conducted by state health departments. In 2011, more than five hundred thousand interviews were conducted across the states and other geographic areas making it the largest telephone survey in the world (BRFSS Data User Guide 2013).
Although the BRFSS asks numerous questions that capture physical and mental health, risk behaviors, and health indicators, I opt to use a measure of self-rated health status (SRHS) to differentiate individuals. Empirically, SRHS is correlated with objective measures of mortality (Jylhä 2009) and health conditions such as coronary heart disease, cancer, and physical functioning (Bjorner et al. 2005) as well as health service use (Angel and Gronfein 1988). Conceptually, researchers view SRHS as an enduring self-concept of general well-being (Boardman 2006). Some even claim that “an individual’s health status cannot be assessed without” SRHS and that this single item captures “an irreplaceable dimension of health status” (Idler and Benyamini 1997, 34). Most importantly, there is evidence that turnout is related to SRHS such that citizens who report being in excellent health vote at significantly higher rates than those who report being in poor health (Matilla et al. 2013; Pacheco and Fletcher 2015; Rahn and Gollust 2015). In the BRFSS, individuals are asked to rate their general health as excellent, very good, good, fair, or poor. I use the BRFSS from 1996 to 2012.
I retain demographic variables used in the MI and the voter turnout measure or SRHS as well as the year and state identifiers for each dataset. I then stack the datasets. The combined dataset consists of 733,237 respondents from the CPS with valid responses to the turnout question and demographics (63% report turning out) and almost three million respondents from the BRFSS with valid responses on SRHS and demographics (5% poor, 13% fair, 30% good, 33% very good, and 19% excellent). 2 That is to say that about 21 percent of the combined data is missing on the SRHS question, whereas 79 percent of the combined sample is missing on voter turnout. The next step is to use MI to impute responses to the missing questions.
MI is a general approach for handling unit and item non-response in sample surveys and has been applied with increasing frequency in the past two decades. An advantage of MI is that it models the uncertainty associated with the predictions directly by iterative estimations, effectively creating j number of imputed datasets (in this case, j = 5). The completed datasets differ in their predicted values of the missing variables, due to the random error associated with each prediction and, thus, reflect uncertainty levels. Analysts can apply statistical methods to each imputed dataset separately or use a simple procedure to combine the results across the j datasets (King et al. 2001). 3
When performing MI, the first step is to decide what variables to include in the models. As Honaker et al. (2009) suggest, it is crucial to include at least as much information as in the analysis model and because the model is predictive and not causal, it is defensible to use as many variables as possible. All respondents have valid answers on demographic characteristics that are associated with turnout (e.g., Brady, Verba, and Schlozman 1995) and SRHS (e.g., Lantz et al. 1998) including age (ranges from 18 to 99), gender (female = 1), race (black = 1, other = 1, white is omitted category), employment status (1 = employed), marital status (1 = married), and education (no high school degree, high school degree, some college, and college degree or more). All of these demographic variables are used to impute turnout or SRHS for each individual for each year. 4 State and year variables are also included to account for differences in an individual’s propensity to turn out across space and time. 5
Having a large dataset that includes both the turnout and health measures for each state and over time, I then calculate the percentage of respondents who voted with poor SRHS for each state and election year combination. Additionally, I calculate the percentage of respondents who voted with excellent SRHS. As suggested by Gelman, King, and Liu (1998), I weight aggregate measures using the appropriate survey weights provided by CPS and BRFSS.
The final step is to create a measure of health disparities that reflects the relative turnout of healthy and unhealthy citizens for each state and in every election year. To do this, I take the ratio of turnout between the citizens in excellent health compared with those in poor health for each state/year. Relative values are calculated instead of absolute differences because previous literature generally finds that relative differences have greater political consequences. However, see the Supplemental Appendix for analyses of alternative measurement strategies. A value of 1 indicates equal representation across health status in the state electorate, values above 1 indicate bias favoring healthy citizens, and values below 1 indicate bias favoring unhealthy citizens. In the end, I have a panel dataset with 450 total observations (T = 9 election years, N = 50 states). As I explain below, the sample size decreases due to the methodological strategy.
Validation of MI Measures of Health Ratio
Before linking the health ratio measures to state policy, it is a useful exercise to validate the measures. I use the BRFSS special modules for measurement validation. Briefly, states can choose to include modules on a variety of public health topics in addition to the required core BRFSS questions. States occasionally add a module called “Social Context,” which includes a question about voting in the previous presidential election.
Table 1 shows which states included the Social Context module as well as the year administered and election year asked. The table shows, for instance, that respondents living in Alabama who completed the BRFSS in 2009 were asked, “Did you vote in the last presidential election? The November 2008 election between Barack Obama and John McCain?” to which they could answer “yes” or “no.” The BRFSS’s strength is its representative state samples. Hence, a BRFSS measure of health bias in voting using the states and election years in Table 1 should be representative of the state population and valid. One might argue that the BRFSS measures may merit a “golden standard” to compare with the MI measures.
Biobehavioral Risk Factor Surveillance Survey Social Context Module Information.
Table 2 shows direct comparisons of the various BRFSS measures and the MI measures. An eyeball test of Table 2 provides some validity check to the MI measures. The MI measures tend to underestimate the voter turnout rate of the high health group compared with the BRFSS. As a result, the MI health ratio has less variance compared with the BRFSS health ratio; the “partial pooling” result is consistent with Gelman, King, and Liu (1998) and an expected consequence of the MI procedure. Nonetheless, only two states (AK and HI) show ratios from the BRFSS that are in the opposite direction from the MI strategy. The correlation between the BRFSS vote for the high health group and the MI vote for the high health group is .84; the correlation between the BRFSS vote for the low health group and the MI vote for the low health group is .54; and the correlation between the BRFSS health ratio and the MI health ratio is .15 (N = 19). The correlations increase to .86, .76, and .34 when Alaska and Hawaii are excluded (N = 17). The results suggest the measures are moderately valid and, if anything, a conservative test of the hypotheses outlined in this paper.
Comparisons of Various BRFSS Measures and the MI Measures.
BRFSS = Biobehavioral Risk Factor Surveillance Survey; MI = multiple imputation.
Variations in Health Inequalities in Political Voice across States and Over Time
Figure 1 shows how levels of health inequalities in turnout vary across four select states in each region from 1996 to 2012. 6 As can be seen in Figure 1, there is both cross-sectional and temporal variations in health inequalities in political voice although there is more variance to be explained within states (77%) compared with between states (23%). The mean value across states and time is 1.04. Generally, the representation of healthy citizens has increased over time, although the slope associated with this increase varies across states. For instance, New York has experienced a steep increase in the representation of healthy citizens, particularly in the past few elections, whereas rates of representation in FL have remained relatively flat over the same period.
State variation in health disparities of turnout, 1996–2012.
Measuring State Health Policies and Other Variables
Now that I have a measure of health inequities in voting across the states, I explore the association of these inequalities with two state health policies. The first policy is a measure of overall direct health expenditures obtained from the State and Local Government Finance Data Query System provided by the Urban Institute’s Tax Policy Center. 7 The expenditure data are available for every state from 1996 to 2012 and are adjusted for inflation and state population (e.g., the measures are in constant 2011 dollars and per capita). Note that I take a log of per capita health spending for the regression analyses to account for the percentage changes (e.g., the growth rate) rather than changes in the absolute amount of spending. Analysis of variance (ANOVA) shows that the majority of variance in logged health spending is between states (66%) rather than within states (35%).
An additional indicator of the generosity of health policies includes total Medicaid personal health care spending per enrollee 1996–2012 in constant 2011 dollars obtained from the Centers for Medicare & Medicaid Services. Again, to account for the percentage changes (e.g., the growth rate) rather than changes in the absolute amount of spending, I take the log of Medicaid personal health care spending. ANOVA shows that the majority of variance in Medicaid spending is within states (57%) compared with between states (43%).
I rely on health expenditures for both methodological and theoretical reasons. Methodologically, these measures are relatively easy to obtain and available for all states during the specified timeframe. Theoretically, these measures better represent policy output, which is potentially influenced by voters as opposed to policy outcomes (e.g., the percentage of residents with health insurance), which are less directly manipulated by voters.
Control Variables
I include a number of control variables that may influence state health expenditures separate from health disparities in turnout. First, I include a measure of income vote bias as simply the ratio of the percentage of rich voting to the percentage of poor voting. More specifically, I use the CPS data to generate a set of state-year-specific aggregated participation measures for individuals making less than $12,500 (e.g., poor voting) and those making more than $75,000 (e.g., rich voting). Overall, disparities in turnout across income are much greater than those across health status; the mean level of income vote bias is 1.73 with a maximum value of 3.58. This is likely due to the methodological approach used to measure state health ratios. The correlation between the class bias measure and the health disparities measure is low (r = .08) suggesting that health inequalities differ from class-based biases in the vote. Again, however, this may be an artifact of the measurement strategy, which tends to underestimate the variance in health inequities of voting.
Policy preferences are also important to include as mass preferences toward spending, partisanship, and ideology all correlate with state expenditures (e.g. Kousser 2002; Pacheco 2013). There are, however, challenges to measuring state public opinion over time (see Pacheco 2013 for a description). Although scholars have made significant strides in developing state public opinion over time, data limitations still exist. For instance, measures of state partisanship, policy mood, and citizen ideology developed by Enns and Koch (2013) are unavailable past 2010, which is unfortunate since my dataset runs from 1996 to 2012. Given the majority of the variance in the health expenditure data is between states rather than within states and there are only nine time periods limiting the ability to make inferences about dynamics, I average the measures of state partisanship, policy mood, and citizen ideology for each state over time. More specifically, I include a measure of the percentage of state residents who are Democrat and the percentage of residents who are liberal for each state. For policy mood, higher values indicate a state that is relatively more liberal on policy outcomes. All of these variables are from Enns and Koch (2013). Note that these variables are time invariant and can only correlate with differences across the states in average levels of health and Medicaid spending.
Previous studies find that state expenditures are associated with the racial makeup of the state (Matsubayashi and Rocha 2012) as well as the percentage of residents with a college education. As a result, I include measures of the percentage of black residents and the percentage of residents with a college degree; both are obtained from the CPS and are time varying. Finally, it is important to include measures of institutional control because the expectation is that states under Democratic control spend more money on health and have more generous health programs. The democratic control variable is an additive scale of Democratic power in the legislature where a value of 1 = Democratic control of chambers; 0 = Republican control of both chambers; .5 = Democrats control one chamber, Republicans the other; .25 = Republican control of one chamber, split control of the other, and .75 = Democratic control of one chamber, split control of the other (see Klarner 2003). 8
Methodological Strategy
Given the nature of the data, there is complexity in estimating models that account for both unit heterogeneity and autocorrelation (see Beck and Katz 1995, 2011; De Boef and Keele 2008). Proper identification of the modeling strategy requires an understanding about the source of variation in the dependent variable as well as the limitations of the data. The majority of variance for both dependent variables is between states, suggesting that explaining dynamics is challenging. Additionally, because N (N = 50) is greater than T (T = 9), asymptotics exist in N and not in T (see Wooldridge for more information). As a result, I opt to focus primarily on comparing states to themselves to see how turnout disparities influence policy and largely ignore the temporal correlation structure of the errors. 9
The primary theoretical concern with panel data is that either systemic factors (e.g., national economic conditions) or unit-specific forces (e.g., state policy mood) that are not included in the model are influencing the dependent variable and causing omitted variable bias. Methodologically, systemic factors can be controlled away with variables that capture time trends (e.g., cubic spline or year dummies), while fixed unit effects (e.g., state dummies) account for unit heterogeneity. Yet, one consequence of including state fixed effects is that non-time-varying covariates are excluded from a model.
To account for systemic factors, I include a linear and squared measure of election year. 10 I also include a dummy variable indicating midterm election years. All time-varying covariates are included as a lag because the theoretical story is that health inequities in turnout lead to policy consequences in the future. All models use ordinary least squares (OLS) with standard errors clustered by state. 11 Unless otherwise noted, all continuous predictors are re-parameterized to range from 0 to 1 to ease statistical interpretation. The sample size decreases from 450 to 392 in these analyses because of the lagged covariates. In addition, Nebraska is excluded from the analyses because it has a non-partisan legislature and is missing on the unified Democrat score. Results are shown in Table 3.
OLS Model Predicting State Expenditures on Health and Medicaid, 1996–2012 (N = 392).
All continuous predictors range from 0 to 1; thus, the coefficients represent that maximum effect. Robust standard errors are shown in parentheses. OLS = ordinary least squares.
p < .10. **p < .05. ***p < .01 using a one-tailed significance test.
Results: Health Bias and State Health Policies
The expectation that states with higher levels of health bias in turnout spend less on health compared with those with low levels of health bias is borne out empirically. As shown in Table 3, health bias in turnout is associated with less spending in direct health expenditures as well as on Medicaid in the subsequent year, controlling for other important covariates. Substantively, the model predicts that a change in health bias from the minimum value to the maximum value results in a 21.5 percent decrease in direct health spending the following year. 12 It is important to note that this effect is statistically significant at the .10 level with a two-tailed test. 13 As shown in the second model, a change in health bias from the minimum value to the maximum value results in a 20 percent decrease in Medicaid payments per enrollee the following year. States tend to spend more on direct health during midterm years, while Medicaid spending is unaffected whether it is a midterm year. The model also shows that racial diversity, the level of education, and partisan control of government are associated with Medicaid spending. 14
Using Federal Election Turnout to Model State Outcomes?
Recall that the measure of health bias is derived from the CPS, which measures voter turnout primarily in federal election years. This begs the question why should federal election turnout be used as a measure of state election turnout? As shown in the Supplemental Appendix (see Tables S1–S3), many states do not have any state elections during even years. In fact, almost all states have state elections for their “senate” or “house” in odd years. Ideally, then, I would want a measure of the ratio turnout score in the odd years to correspond to state elections. Unfortunately, the CPS is only available in even years, thus the ratio measure is also only available during even years.
One solution would be to restrict the analyses to only those state/election years, which include a gubernatorial election. Essentially, I drop the states that do not hold gubernatorial elections in 1998, 2002, 2006, and 2010. Doing this significantly decreases the sample size; nonetheless, inferences regarding health expenditures are largely unchanged and, if anything, strengthened (see Table S4 in the Supplemental Appendix). However, the coefficient on the health ratio measure fails to reach statistical significant for Medicaid expenditures.
A better solution is to restructure the data so that the key covariate is the turnout ratio in the most recent gubernatorial election. This results in DE, IN, MO, MT, NC, ND, UT, VT, WA, and WV being included in the analyses, but this means that not all state turnout and spending variables are measured at the same time. The sample size is reduced again, but not as much as if I exclude these states. The inferences, however, again, are largely unchanged (see Table S5 in the Supplemental Appendix). The health ratio turnout measure continues to be negative and statistically significant for state health expenditures in these models, but fails to reach statistical significance for state Medicaid spending.
A final solution includes a measure of the year in which state held gubernatorial elections and interacts this with the turnout ratio. This solution is more parsimonious and preserves the sample size. As shown in Table S6 in the Supplemental Appendix, the inferences from the limited sample analyses are largely unchanged with this inclusion (again, the states with odd-year elections and Nebraska were dropped from the sample).
For healthcare expenditures, the interaction between lagged gubernatorial elections and lagged health bias ratio is negative and significant at the .10 level suggesting that federal turnout itself does not underlie this relationship. For Medicaid expenditures, the coefficient on the health bias ratio measure is significant and negative (i.e. health bias has a negative relationship in non-gubernatorial years), but the interaction fails to reach statistical significance. This is more suggestive of the idea that federal turnout might account for the relationship we see between Medicaid spending and health bias, which makes sense given that Medicaid is partially federally funded.
Placebo Effects
A primary threat to the empirical analyses is that the measure of health disparities in turnout is picking up some other unobserved confounding characteristic that is related to state expenditures. One way to assuage this concern is to conduct parallel analyses that look for unexpected effects on other outcomes. The logic here is that if health disparities in turnout are related to non–health-related policy outcomes, then that would suggest that something other than health disparities are driving the outcomes observed. On the contrary, if health bias in turnout is related to health outcomes, but not others, then the empirical analyses and arguments presented in this paper are bolstered.
Other outcomes that I use to test for placebo effects include state expenditures toward K-12 education, corrections, and welfare. I chose these three outcomes because states that tend to spend a lot on health also tend to spend on these issues (Jacoby and Schneider 2009). Collectively these issues tap into a state’s propensity to provide particularized benefits for their residents. If results show that health bias in turnout is unrelated to these spending variables, then that provides an extra layer of empirical evidence that supports the argument. Similar to the previous analyses, each spending variable is logged to account for percent change. Models are identical to what was presented previously. Results are shown in Table 4.
OLS Model Predicting Placebo Effects on State Expenditures on Education, Corrections, and Welfare, 1996–2012.
All continuous predictors range from 0 to 1; thus the coefficients represent that maximum effect. Robust standard errors are shown in parentheses. OLS = ordinary least squares.
p < .10. **p < .05. ***p < .01 using a one-tailed significance test.
As shown in Table 4, health bias in turnout is unrelated to state expenditures toward education and corrections. 15 On the contrary, as shown in the third model, health disparities are related to state expenditures on welfare. Substantively, the model predicts that a change in health bias from the minimum value to the maximum value results in an 18.5 percent decrease in state welfare spending the following year. It is also the case that the model suggests a negative association between income bias and state spending on welfare. A change in income bias from the minimum value to the maximum value results in a 60 percent decrease in welfare spending the following year. Although health disparities in turnout are related to welfare spending, the evidence that income biases matter more and not at all for health spending provides further empirical support for the argument that health inequalities of political voice are separate and independent from class-based inequalities.
Conclusion
The question of whether health disparities in political participation have policy consequences is central to issues of population health, representation, and health policymaking. Its importance is underscored by the burgeoning scholarly research on the link between health and political behavior at the individual level. The empirical evidence offered here suggests that health disparities in voice have important policy consequences. State electorates that are disproportionately more representative of healthy citizens are less likely to spend money on health and have less generous Medicaid programs. Many of these relationships are robust to alternative model specifications and remain significant after controlling for class biases in turnout.
Yet, I must emphasize that this finding may only hold within the current range of the data. Advances in medicine and technology coupled with an aging global population and new understandings of disease and the body have pushed the relevance of health into discussions of politics (Carpenter 2012). And, while on and off the political agenda since at least the Progressive Era, the failure of Clinton’s Health Security Plan and the passage of Obama’s Affordable Care Act brought health care reform—and health—back into the political spotlight. This suggests that the impact of health on political behavior and the policy consequences of health disparities in political power may not hold in previous time periods.
Nonetheless, I have provided evidence that who participates is critical to the formulation of health policies across the states. These results largely collaborate with previous research on the policy consequences of class bias in turnout, yet they also point to the importance of exploring the effects of unequal participation across a wider range of divisions. Although political scientists have traditionally focused on the sociological divisions of political power based on income, other disparities, such as those on the basis of gender, race or ethnicity, age, education, and health, exist. As importantly, these other kinds of electoral disparities matter for political outcomes, including the health policies examined here.
Finally, the results suggest that unequal representation across health status may coincide with growing levels of income inequality. Much has been written on the political consequences of rising income inequality with many scholars concerned that differential policy responsiveness across wealth will be exacerbated in the future. A substantial number of studies, however, show that income inequality is a determinant of population health. In particular, lower income inequality is related to higher standards of population health (e.g., Neckerman 2004). The results presented here suggest that the political consequences of increasing income inequality are not restricted to the poor, but also the sick.
Footnotes
Declaration of Conflicting Interests
The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The author(s) received no financial support for the research, authorship, and/or publication of this article.
Notes
Supplemental Material
Supplemental materials and replication materials for this article are available with the manuscript on the Political Research Quarterly (PRQ) website.