Voter Mobilization and the Obama Victory. Tracy Osborn, Assistant Professor

Transcription

1 Voter Mobilization and the Obama Victory Tracy Osborn, Assistant Professor Department of Political Science University of Iowa 341 Schaeffer Hall Iowa City, Iowa Scott D. McClurg, Associate Professor Department of Political Science Southern Illinois University Mailcode 4501 Carbondale, IL Benjamin Knoll, Graduate Assistant Department of Political Science University of Iowa 341 Schaeffer Hall Iowa City, Iowa Abstract As with the other presidential elections from this decade, the 2008 election was followed by considerable speculation as to how new efforts to mobilize voters affected the eventual outcome. Although the conventional wisdom implies that Democrats benefit from higher turnout, previous research in political science demonstrates that such a conclusion applies to actual election results inconsistently. In this paper, we outline the difficulties involved with assessing turnout effects within a particular election and proceed to test the hypothesis that the Obama campaign benefited from higher turnout using three different methods. The evidence suggests that the Obama campaign benefited substantially from voter turnout, particularly in comparison to the Kerry campaign in 2004, yet they also were successful in changing the minds of already mobilized voters. Although various data difficulties suggest the presence of some bias in our estimates, the consistency of the results across tests supports the general conclusion. CITATION: Osborn, Tracy, Scott D. McClurg, and Benjamin R. Knoll Voter Mobilization and the Obama Victory. American Politics Research: forthcoming.

2 Introduction Among the many notable traits of the 2008 presidential election, one of the most discussed concerned the potential impact of increased voter turnout on the election. In both 2000 and 2004, representatives from all political walks devoted new attention to mobilizing voters to take advantage of technological innovation and changes in fundraising. In the 2000 election, these efforts were largely driven by the influx of soft money funds to political parties (Holbrook and McClurg 2005). These efforts were greatly extended in the next presidential election by groups like America Coming Together and the much discussed 72-Hour Task Force assembled by the Republican Party to bring conservative voters to the polls through phone calls and s. Given the significant rise in voter eligible turnout in the United States to 50% and 61% 1 in the 2000 and 2004 elections, campaigns focused even more effort on voter mobilization in The most prominent example of this, of course, was the much lauded Obama organization which created a national phone bank, stimulated local meetups, had an unprecedented individual-level fundraising effort, and had an extensive ground organization for personal contacts. Like the attention given to the 72 Hour Task Force in 2004, much media focus was given to the Obama campaign s innovations in targeted turnout, such as using texting to develop a campaign database of young voters and using prominent black politicians to target African American voters. And once again, the effort seemed to pay off with turnout up to 62%, its highest point since the late 1950s. 1 All turnout statistics in this paragraph are based on voter eligible turnout (McDonald and Popkin 2001) and can be found at 1

3 With all the attention parties and interest groups devoted to producing turnout in the past few elections, it would be easy to assume that these efforts and the concomitant fluctuations in turnout played an integral part in determining which candidate prevailed. To be sure, this kind of claim was made frequently in the wake of the Obama victory. The PBS headline on election night declared that Obama s presidential win was fueled by huge turnout 2, and news organizations noted Obama s intense campaigning in the week before the election in Virginia, North Carolina, and Indiana and high turnout in these states as explanations for his victories, implying a connection between the two. Such assertions are reminiscent of long standing conventional wisdoms about turnout which imply that when turnout is high that Democrats do better. Yet an examination of political science scholarship provides mixed evidence in support of this claim. Studies across a wide variety of legislative (DeNardo 1980), state (Nagel and McNulty 1996), and presidential (Radcliff 1994) elections provide what can at best be construed as inconsistent evidence for the conventional wisdom. For instance, DeNardo (1986) uses a formal model to demonstrate that there is no natural, persistent turnout effect for either political party and that the electoral effects of turnout depend a great deal on what happens from one election to the next. McClurg s (N.d) analysis of the 2004 election suggests that in thinking about the logical effects of turnout, Republicans actually have a persistent structural advantage because of a correlation between voting, turnout, and socioeconomic factors, and that it is inter-electoral fluctuations that produce differential turnout benefit effects

4 Drawing on insights from prior research, this paper tests the claim that the Obama victory was driven by turnout. We begin our examination by discussing the analytic difficulties involved in trying to determine the political consequences of voter turnout within any single election. Acknowledging the drawbacks of any particular approach, we outline three different methods for evaluating the effects of turnout in the 2008 election. Though the final estimates are lacking in precision, the consistency of our results across methods and models implies that in the context of this election, Democratic efforts at voter mobilization did in fact contribute to Obama s victory. However, Obama s victory was not only due to the effects of mobilization on turnout; changes in voter choice played a role as well. In the conclusion, we discuss the implications of this in light of theories about and empirical evaluations of the effects of mobilization. What Constitutes a Turnout Effect? The principal methodological problem confronting our study is how to measure a turnout effect. At the heart of the conventional wisdom is the plausibility of a key assumption fluctuation in voter turnout is disproportionately a function of two decisions made by low socioeconomic (SES) status voters. Since there is a strong positive relationship between SES and voter turnout at the individual level of analysis, there are fewer untapped votes among high status voters (who tend to vote Republican) than among the low status voters (who tend to vote Democratic). Consequently, aggregate fluctuations are more likely to be driven by the decisions of potential voters to 1) turn out in the first place, and therefore, 2) vote Democratic. Or so the logic goes. 3

5 Some empirical support for these assumptions is found in research on how full turnout meaning what would happen if every citizen voted would influence U.S. elections. Studies in this vein begin by using individual-level survey data to estimate the relationship between demographic information and voting. The parameters from these models are then used to predict the number of votes each party would receive from different groups whose proportions are measured with aggregate information. The effect of full turnout is the difference between this estimate and the actual outcome. Almost without exception, these studies show that Democrats would receive more votes but not necessarily win more elections if everyone voted (Citrin et al. 2003; Hanjal and Trounstine 2005; Martinez and Gill 2006). Such results support the argument that there is greater latent support for Democrats than for Republicans, a key part of the conventional wisdom. Yet there is good reason to question the conventional wisdom. DeNardo (1986, 1980) shows that while it may be true that there is more latent, un-mobilized support for Democrats, the effects of this group on an election outcome depend upon their loyalty. So, if there is a large increase in voter turnout among low status voters, it might not benefit Democrats if those presumably Democrat-sympathetic voters do not cast their ballots for the party at the same rate as voters who do consistently turnout and vote Democratic. This means that whatever correlation exists between voter turnout and voter choice can be the product of two factors, voter mobilization and voter defection. Not only does this help explain inconsistent empirical results in observational studies (DeNardo 1986), but it also explains why the impact of full turnout is not consistent across all electoral jurisdictions or elections within jurisdictions. For our 4

6 purposes, it also raises important issues about how to analyze a turnout effect within a single election by showing that analysis of turnout effects within an election is complicated by inter-electoral shifts in both preferences and turnout. An additional complication in studying turnout effects on election outcomes revolves around the type of variation under examination. At issue here is trying to understand what the appropriate counterfactual is for understanding turnout effects on voting. The easiest way to see this is to consider the aforementioned empirical studies. Research that approaches this subject from what we have called the full turnout perspective is implicitly studying intra-electoral variation by comparing what did happen in the context of specific elections to what could have happened. Taking this approach one step further, McClurg (N.d.) argues that variation between electoral jurisdictions within a single election can yield important insights into these questions (also see Radcliff 1994). Contrasting this approach are observational studies of the correlation between turnout and elections which tend make comparisons over time (e.g., comparing presidential elections) or across different jurisdictions (e.g., comparing across legislative elections). The question that confronts us here is which type of variation synchronic or diachronic is the best underlying comparison for understanding a specific election, such as the 2008 presidential race. The natural inclination is to contrast the results of the election to the results in other, similar units. In the case of the 2008 presidential election, then, it might make the most sense to compare it to other presidential elections. Whether this is done using jurisdictions, such as states (Sides et al. 2008) or counties (McClurg N.d.) within an election or the election itself as the unit of analysis, the strength of 5

7 comparing across time is that what happened previously is often a very good predictor of what comes next. Yet despite the high levels of inertia that exist in elections because of demographic factors (i.e., the relationship between income and voting changes very slowly), this comparison is fraught with assumptions. The most important of these is the assumption that all change that occurs from one election to the next is a product of turnout. If we were to compare 2008 to 2004, for example, this seems to be a dubious assumption even if mobilization of more voters is an important explanation of the outcome, it remains plausible that there was shift in sentiment against the Republican party that also influences our interpretation of what happened in 2008 relative to Comparing variation within a single election carries with it similar strengths and weaknesses. On the one hand, each of the units under examination (e.g., counties) is comparable in the sense that the voters are making decisions in the same national electoral context. As such, an intra-electoral study is the mirror image of inter-electoral study in that the contextual sources of variability are easily controlled. But on the other hand, the other sources of variability in both turnout and voting become more difficult to pin down. To be sure, we can to some degree control for these by including demographic control variables at the aggregate level (e.g., median income). However, doing so with confidence requires the fairly strong assumption that the model is accurately specified in the cross-section and that the parameters that describe the relationship between the control variables, turnout, and voting remain constant across time. Implications for Understanding the 2008 Presidential Election 6

8 As the previous section makes clear, unraveling the political consequences of voter turnout within a single election is a challenging task. And, unfortunately, the most common methods used for this task all come up short on some standard of evaluation. Thus, we will examine the 2008 presidential election by combining careful reasoning with three methodological approaches. While this kind of examination will not yield precise estimates of causal effects, it has the advantage of beginning to provide boundary estimates of the relationship between turnout and voting under different assumptions, thereby helping us gather an overall understanding of the role of turnout in A Cross-County Analysis. Our first approach will be to examine the correlation between turnout and voting among counties within the 2008 election alone. Specifically, we will use an approach outlined by McClurg (N.d) to decompose voter turnout into two different components: 1) structural and 2) mobilization. The first of these is the level of turnout predicted by a county s demographic socioeconomic attributes, which he argues reflects the structural bias in turnout that naturally benefits Republicans, the party that is on average more responsive to high status voters (Bartels 2000). The second attribute is the deviation from actual turnout and the structural component, thus representing the proportion of voter turnout that is abnormal relative to a county s demographic makeup. In his analysis of the 2004 election, McClurg suggests that this is the part of the turnout function which can account for most intra-electoral change and thereby shed light on the unique impact of turnout across electoral jurisdictions in a single election. Although this method cannot entirely address questions about DeNardo s two parameters, we argue that the structural component should capture the average loyalty rate of 7

9 persistent voters across jurisdictions while the mobilization component should measure the average loyalty rate of peripheral voters. Our hypotheses for this test are straightforward. We expect that the effect of structural turnout on county-level voting patterns should benefit Republicans, suggesting they do better than Democrats in those areas where turnout should be high because of the connection between participation and socioeconomic status. For the mobilization component, we expect a positive effect will benefit Democrats, as it contributed to support for Obama. If it is negative, then it implies that Republicans did better in areas with higher-than-expected turnout and that Obama was hurt by spikes in turnout. A Cross-Electoral Jurisdictional Approach. Our second approach will be to examine changes in the correlation between voter turnout and election returns in U.S. counties from 2004 to In this case, our principal unit of comparison is a prior election. If we can assume that the only changes in counties that occur between elections are political, then studying the change in turnout to predict the change in vote choice will provide insight into the potential effects that turnout had in this past election. The drawback is that the assumption of constancy at the county-level is not an accurate description and, importantly, not constant across counties. If any demographic shifts in counties that occur from 2000 to 2008 are related to voting returns, this would constitute bias in a coefficient estimate of turnout estimates. The other potentially problematic 3 We use counties as our jurisdictional unit because they are the principal reporting unit for election returns in states and are therefore somewhat comparable across states and elections. Moreover, they are important units for campaigns as they represent parts of states, media markets, and places to visit. Nevertheless, there are methodological concerns that accompany a county-level analysis which we discuss below. 8

10 issue has to do with using 2004 as a baseline against which to judge the 2008 election. Not only is this (to some degree) an arbitrary unit against which to judge 2008, but we also cannot tell which of DeNardo s two parameters, voter mobilization or voter defection, might explain any relationship in the correlation between turnout and voting. The 2004 election is, however, the presidential election immediately prior to 2008 and the comparison most likely to minimize the problem of confounding the effects of demographic shifts with changes in turnout. The hypotheses guiding this part of the analysis are straightforward. If we observe that counties with higher turnout in 2008 than in 2004 were more likely to vote Democratic, it would imply that Obama benefited from patterns of voter turnout in this last election cycle when compared to John Kerry. If the relationship is non-existent or negative, it would imply that Obama won in spite of changes in voter turnout. While the first scenario does not distinguish between effects caused by changes in turnout and voter choice, the second scenario would suggest that Obama s victory was principally driven by a change in the candidate people chose instead of mobilization. The Full Turnout Method. Our third and final approach will rely on what we have previously called the full turnout method, which estimates turnout effects by comparing actual outcomes to what would have happened under conditions of full turnout. The strength of this method is that it provides a clear counterfactual under plausible conditions. The voter choice function is computed based on individual voters within a single election, thereby controlling for the context of the vote choice and accounting for both of the parameters that can produce the correlation between turnout and voting 9

11 (DeNardo 1986). Similarly, the turnout function is based on individual estimates of turnout and information on the composition of the population. The drawbacks are that using full turnout as a counterfactual for thinking about a turnout effect sheds little light relative to previous elections. For our purposes, this complicates our interpretation of the degree to which Obama benefitted from increased turnout relative to other elections. If we ignore the results of simulations from previous elections, this answer is relatively clear. If the percentage of votes received by Democrats is greater under conditions of full turnout, then they were disadvantaged by the patterns of voter mobilization in the 2008 election. This would in turn suggest that Obama benefited principally from a shift in people s voting behavior, rather than voter mobilization. If the percentage of votes received by Democrats is less under conditions of full turnout, it would imply that they were advantaged by voter mobilization and that shifts in voter choice had less of an impact on the final outcome. This interpretation is complicated if we consider previous research, which suggests that in nearly every election Democrats would do better under conditions of full turnout. This would imply that Democrats doing better might still indicate a mobilization effect, though one that is difficult to identify in practice. Mobilization from a Cross-Sectional Perspective: A Cross-County Analysis To begin, we examine turnout from the perspective of variation within our election of main interest: The goal here is to parse county turnout into twodifferent factors: structural and mobilization. Structural turnout is the predicted level of voter turnout in a county given its socioeconomic characteristics, while the mobilization 10

12 component is the deviation between real turnout and its structural component. The logic of this definition is partly based on the notion that the effects of turnout which stem from its relationship with education and income should not be confused with the effects of unusually high or low turnout that are ground in election-specific factors. As noted, this approach requires some relatively strong assumptions about the counterfactuals to which we make comparisons, but has the advantage of controlling for electoral context. For this portion of the analysis, our primary dependent variable is the Democratic percent of the two party vote in all U.S. counties from 2008 to The independent variable of interest is voter turnout, measured as a percentage of voting age population in the 2000 census. This latter measurement comes with three potential drawbacks. First, we are unable to use a measure of turnout as a percentage of the voter eligible population (VEP). As shown by McDonald and Popkin (2001), using voting age population (VAP) in the denominator underestimates turnout rates generally and will particularly do so in those areas with high levels of foreign born residents and those convicted of a felony. 4 Second, there can be significant growth in a county from the 2000 census to both the 2004 and 2008 elections. 5 Third, using this measure of turnout gives us a measure that 4 For our purposes, the difference between using VEP and VAP creates a likely measurement error that would inflate the size of the turnout effect on the Democratic vote by exaggerating differences between low and high turnout counties. To address this problem, we estimated two alternative specifications. First, we re-estimated our models with state-level demographic variables to compare the coefficients using VAP and VEP measures. Second, we added a measure of foreign-born population in each county to correct to some degree for the effect of non-eligible immigrants on VAP. In each of these re-estimations, the small differences in coefficients are consistent with our expectations regarding measurement error, yet they still support the conclusions we draw from the existing models. 5 To investigate the size and impact of this form of measurement error, we re-estimated our model using state-level data from both the 2000 Census and the 2006 American Community Survey. These results, available upon request, suggest that we are likely 11

13 includes some implausible cases (e.g., turnout over 100% in a handful of counties). 6 In part this may reflect the population dynamics we just mentioned, though it could also be a function of reporting and record keeping errors. Each of these issues represent a case in which measurement error can potentially threaten the inferences we make by over or under-estimating the voter eligible population. As an initial step we estimate the effect of turnout on Obama vote share in 2008 using a multilevel approach. We estimated a multilevel mixed effects model to approximate the relationship between turnout and the Democratic percentage of the twoparty vote at the county level. There are fixed effects at the county level and a randomintercept for the states, meaning that variance is estimated both within and between states. This should produce efficient and unbiased estimates of the relationship. 7 The results, presented in Table 1, demonstrate that Obama did comparatively worse in those areas where turnout was relatively high. It is worth noting that this effect is relatively mild (a 1% decline in the Obama vote for every 10 percentage point increase in turnout) and that it is one-half the size of the coefficient McClurg (N.d) reports for Kerry in How will these results change when we separate turnout into its separate components? underestimating a mobilization effect. However, as the state is a different unit of analysis that it is subject to a form of aggregation bias, we only feel confident suggesting that whatever the nature of the county-level measure bias, it does not seem to be the driving force behind our estimates. 6 We account for this by re-estimating each of our models after trimming 5% from the top and bottom of the turnout distribution. In each case, the estimates are consistent with discussion in the text. 7 These models were also estimated with ordinary least squares regression with a separate variable for every state, clustering standard errors within states, and weighing the results by county voting age population. The estimate of turnout influence is not appreciably different at with a t-value of

14 McClurg (N.d.) estimates mobilization by computing the residuals from a regression that predicts turnout with demographics. In the interest of brevity, we skip this step and simply re-estimate our model by including the demographic variables directly into the model. 8 Displayed in the second column of Table 4, these results confirm two things. First, they demonstrate that there is a potentially confounding relationship between socioeconomic status, turnout, and voting, at least in the cross-section. Second, they show that once we control for this relationship that the mobilization component of turnout the part of the variance not explained by the demographic characteristics is positively related to Democratic success. How important is mobilization in the scheme of the overarching election? For every 10% points higher turnout is than we would expect given the demographic makeup of a county, we find that the Democratic share of the two-party vote increases by 1%. This is not enough evidence to suggest that this was the centerpiece of the Obama victory, particularly when we consider that roughly 80% of the counties had a mobilization score between -10% and 10%. Among those counties where the mobilization score is greater than 10%, only nine had a Democratic vote share that was between 50% and 51%. The influence of voter turnout benefited the Democrats, but was certainly not the deciding factor. A second question to consider is how these effects compare with preceding elections. Though we consider the question of inter-election comparison more completely below, we try to determine the comparability of the structural and mobilization components of turnout by comparing them for the 2000, 2004 and Taking this approach we get a coefficient for the effect of structural turnout on vote share of with a t-score of

15 elections. This effort is reported in Figure 1 and Table 2. In Figure 1, note that there is a close, linear relationship between the structural portions of turnout between 2004 and 2008 (r=0.96), suggesting the validity of the concept. Also note that there is a similarly close relationship between the mobilization measure in the two elections, which suggests that mobilization bleeds across electoral cycles. But importantly, the relationship is far less coincident as it is for the structural component (r=0.77). With this as a backdrop, we can now compare the structural and mobilization components to voting across the previous two elections in Table 2. In this first column of the table for 2004, the impact of structural turnout on voting behavior was negative (meaning higher status counties with higher turnout were less likely to vote Democratic) and substantively very significant (β=-0.210, s.e.=0.02). This effect is roughly twice as big as it is in This means that for every unit increase in turnout due to socioeconomic characteristics such as education or race, that the Democrats could expect twice as few votes in 2004 than in Similarly the mobilization effect in 2004 was approximately 60% greater, meaning that every additional increase in turnout due to mobilization yielded.42 more votes for Democrats in 2004 than in The third and fourth columns of Table 2 show the same analysis for Two things stand out here relative to Clearly, there is a stronger structural bias for Republicans in 2000 than in either 2004 or This implies that from 2000 to 2008, Democrats have succeeded in gaining more votes from what we might characterize as the regularly voting electorate. Such a result can be a function of either a decline in voter loyalty among Republicans or conversion of some voters to the Democratic side of the 14

16 ledger (DeNardo 1986; Finkel 1993). Second, there is no evidence of a mobilization effect from deviations in voter turnout for the 2000 election. Although we must buy the assumptions underlying this analysis to draw strong conclusions something we are remiss to do the pattern of results here does suggest that Obama s victory in 2008 was driven as much as changing the relationship between status and voting behavior as it was ground in voter mobilization. In comparison to the 2004 election, the Obama campaign did not benefit as much from deviations from normal turnout as it did a change in voter preferences. Yet relative to the Democratic campaign in 2000, these deviations in turnout played a significant role in Obama's success. And while we cannot rule out the possibility that some movement on the structural effect from 2004 to 2008 was because of voters newly mobilized in 2004, the implication is that turnout variations in 2008 were not entirely responsible for the Obama victory. Assessing Mobilization in Comparison to 2004: A Cross-Electoral Approach Our next approach to studying mobilization is to examine how differences in turnout from 2004 to 2008 are related to change in the Democratic vote between those elections. In the 3000-plus counties in the United States, there was notable change in the voting behavior of Americans, ranging from a decline of 20% for Democrats in Coal County, OK, to an increase of 18.5% in Clifton Forge City, VA. Similar variability is evident in voter turnout. Lake County, TN, experienced a decline of nearly 38% in turnout by our measure, while Clifton Forge City experienced an increase of nearly 38%. More generally, there is an average increase in both variables across all counties (2% for Democratic vote, 1% for turnout) with significant variability. 15

17 <Figure 2 about here> Figure 2 shows the scatterplot of these two variables with both a linear model and loess smoother superimposed. As this graph shows, there is significantly more variation among the counties when it comes to the Democratic vote. This suggests that, at a minimum, variation in voting outcomes at the county level can only be partly driven by changes in turnout. More importantly for our purposes, there is a clear positive and linear relationship between turnout and Democratic voting by these standards. While the relationship is not extremely strong, it does imply that Barack Obama did better in those areas where turnout was higher in comparison to This relationship does not change substantially if the outliers on the turnout distribution are trimmed. <Table 3 about here> To obtain a more precise estimate of this relationship, we estimated a multilevel mixed effects model to approximate the relationship between turnout change and vote change at the county level, similar to the model in the previous section. The results are reported in Table 3. Consistent with the hypothesis that the Obama campaign benefited in those places where turnout increased, we see a positive and statistically significant coefficient. The substantive effects are impressive, especially considering the range of the dependent variable. For every unit increase in turnout, there is approximately a.2 unit increase in voting Democratic. Thus, if turnout in 2008 was ten points higher in a county than in 2004, Barrack Obama s campaign could expect to out-gain John Kerry in the county by a little more than 2%. With change in voter turnout ranging from a decline of 38% to an increase of 37%, this can account for swings of nearly 16% across all counties on the continental U.S. The actual impact of voter turnout change is somewhat 16

18 milder, as 90% of US counties experienced change in voter turnout that ranged between a decline of 5% and an increase of 9%. We gain a greater understanding of these results by comparing them to the second column of Table 3, which replicates these analyses for the change from 2000 to In these results, the change in turnout from 2000 to 2004 again has a positive and statistically significant coefficient. However, the effect is smaller than in the 2004 to 2008 comparison; for each unit increase in turnout, there is a slightly less than.1 unit increase in voting Democratic. These results demonstrate that Obama received a better bump from turnout than Kerry did in Two final things should be noted about the impact of voter turnout on the Obama campaign as examined here. First, our comparison between the 2004 and 2008 elections based on 2000 census data may suffer from the drawback of demographic shifts caused by non-random population change across counties. That is, so-called growth counties with substantial demographic shifts, often the target of presidential campaigning, may drive some of the changes in turnout. To account for this problem, we consider two alternative data sources. First, we estimate our models using state-level data. Though the small-n in this data source renders the produced coefficients imprecise, the same pattern emerges in these results: Obama received more votes but not enough votes to influence the outcome. Second, we compare state-level information from the American Community Survey (ACS), which replicates much of the information in the 2000 census, to our statelevel census estimates. The pair-wise correlation coefficient for each variable from the ACS was over.90 for each variable from the census. These two comparisons suggest 17

19 that, at least at the state level, population dynamics over the eight years from do not change the nature of jurisdictions dramatically. While this is undoubtedly less true at the county level because it is a lower-level unit of aggregation, we believe this suggests the problem of demographic change is not significantly so severe as to alter the results of our analysis. In the final analysis these effects are mild. Roughly one-quarter of all U.S. counties had an increase in turnout that was sufficient for increasing the Democratic share of the two-party vote by one-percent. Those counties experienced an average change in Democratic vote share of 4%, with the overall vote share averaging just 43.5%. Indeed, only 13 counties in the data set experienced an increase in turnout that was 5% or greater where the Democratic % of the two party vote was between 0% and 1%. Although some of these were in battleground states, it does not seem a stretch to conclude that while Obama benefited from the surge in turnout, it was not the source of his Electoral College victory. 9 This effect is consistent with the results of the first section above. Assessing Mobilization Effects with Simulations: The Full Turnout Method Finally, we compare our conclusions drawn from the first two methods above to a replication of the method of simulating election results assuming conditions of full turnout used by Citrin, Schickler, and Sides (2003). The goal of these authors was to determine if election results would differ if everyone turned out to vote in a given 9 The 13 counties are Flagler, FL; Douglas and Newton in GA; Vanderburgh, IN; Olmsted, MN; Jefferson, MO; Gallatin, MT; Bladen, Caswell, and Richmond in NC; Barnwell, SC; and Chesapeake City and Staunton City in VA. 18

20 election. They examined Senate elections during the 1990s and concluded that Democrats would generally receive a small boost from universal turnout, but that it would change the outcome of only very close elections. Thus, they conclude that full turnout would have a minimal impact on the overall partisan balance of Congress. Relevant to presidential elections, these same authors simulated the outcome of state-by-state Electoral College results in the 1992, 1996, 2000, and 2004 presidential elections (see Sides, Schickler, and Citrin 2008). They concluded that Clinton would still have been elected in the 1990s, but that Gore and Kerry would have been elected in 2000 and 2004, respectively, due to the fact that those elections were much closer and depended on very small changes in only a handful of close swing states like Florida and Ohio. The simulation method employed by Citrin, et al. (2003) is fairly clear-cut. First, they estimated a binary logit model of partisan vote choice in a given Senate election using a variety of socioeconomic variables derived from state exit polls. They then applied these coefficients to model estimates using data from the Current Population Survey (administered by the U.S. Census Bureau) to predict partisan vote choice for individuals who did not vote in the given election. Finally, they calculated the simulated outcome under full turnout for the Democratic candidate by adding the estimated candidate choice among non-voters, weighted by the level of turnout in the election, with the actual distribution of voters in the election. The equation thus becomes: simulated full turnout = (actual % Democrat actual turnout %) + (predicted % Democratic (1-actual turnout %)). To simulate their method in predicting the results of the 2008 presidential election under full turnout, we replicated the same analysis using data from an early release of the 19

21 ANES 2008 Time Series Study. 10 This dataset included information on whether or not the individual respondent reported voting in the 2008 presidential election, which candidate the respondent voted for, as well as the relevant socioeconomic variables. This dataset also included partisan and religious variables which were not available to Citrin, et al. (2003), who explain that they would have incorporated partisanship or ideology had the relevant variables been included in the CPS data. 11 As with any survey of voter turnout, there was a 15.4% incidence of over-reporting on the part of respondents in the ANES data. While this is of course problematic, it is not overly different than the 13% incidence of over-reporting in the data analyzed by Citrin, et al. (2003). They argue, for a variety of reasons, that over-reporting does not significantly skew the results reported in their study (78). Table 4 reports the results of the logit equations predicting an individual-level vote for Obama in the 2008 election. Model 1 contains the same independent variables as employed by Citrin, et al. (2003) while Model 2 adds partisanship and religious conservatism. 12 The data for both models were weighted according to the sample 10 Because this is an advance release of the ANES study, the survey documentation recommends labeling the conclusions reached by analysis of its data as tentative, preliminary, or subject to change. 11 It is important to note that the major difference between our estimation of the Citron et al (2003) model is we cannot break the results down to the state level because we are using ANES data. This makes our estimation a sample of the nation rather than an estimation of effects on the Electoral College tally. 12 The partisanship variables are dummy variables for self-identified Democratic and Republican partisanship. The religious conservatism measure is a dummy variable coded 1 if the respondent reports that they believe that the Bible is the word of God and should be interpreted literally. This variable was added because previous research has shown that religion is playing an increasingly powerful role in determining individuallevel vote choice in the U.S., especially at the national level (see Green 2007, e.g.). The Biblical literalism question is used as a proxy for the modernist-traditionalist split in American religion, which has shown to be a more powerful predictor of vote choice than 20

22 weighting variables included in the ANES dataset. Table 5 then reports the predicted vote for Obama among non-voters using the logit coefficients from both models. 13 <Table 4 and Table 5 about here> To calculate the election results of full turnout under each circumstance (Models 1 and 2), we can apply the results from Table 4 to Citrin, Schickler, and Sides s (2003) simulation equation. In reality, Obama received 53.7% of the two-party vote in and 61.7% of the Voting Eligible Population turned out to vote. In the simulation of Model 1, Obama received 68.5% of the two-party vote among non-voters. Thus, (53.7) (61.7) + (68.5) (38.3) = Assuming full turnout, Obama would have increased his margin of victory by approximately six percentage points to about 59%. In the simulation of Model 2, which adds the explanatory variables of partisanship and religious conservatism, Obama would have received 71.5% of the two-party vote among nonvoters. Thus, (53.7) (61.7) + (71.5) (38.3) = 60.5, an increase of nearly seven percentage points. Under Citrin, Schickler, and Sides s (2003) full turnout simulation model, then, Obama would have won the 2008 presidential election with an expanded victory of approximately 60%. How does this 7-point gap compare to the difference between actual and simulated outcomes in previous elections? This question is somewhat tricky to answer, in no small part due to the fact that we have simulated national rather than state outcomes here. Nevertheless, we turn to Martinez and Gill s (2005) examination of either religious affiliation or frequency of church attendance (see Wuthnow 1998). Replications of the estimations here without Biblical literalism included yield a slightly higher (about one percent) vote share for Obama. 13 The logit equation being: P(vote Democratic) = exp(xβ)/1+exp(xβ). Xβ is the constant plus the sum of each variable multiplied by its coefficient. 14 President Obama received 52.9% of the overall vote, including minor party candidates. 21

23 turnout from 1960 to 2000 to put our results in context. They use National Election Studies data, but a different simulation method, to determine the relationship between different levels of turnout and the Democratic lean. Most relevant to our study, they find that Democrats would outpace Republicans by about 11% in 2000 their closest time point to our study compared to the real difference of approximately 1%. The fact that Obama would have benefited less from full turnout than Gore would have in 2000 implies that he benefited more from the choices made by the voters in the actual levels of turnout. While this data is consistent with that hypothesis, it is nevertheless important to recognize its limitations as a test of the turnout hypothesis in At a minimum, the evidence may simply reflect a decline in the pro-democrat bias of abstainers. This is consistent with the longer-term evidence that Martinez and Gill provide in their article. Discussion and Conclusion Our intention in this paper is to test the hypothesized relationship between turnout and Obama s victory in the analysis of the 2008 election. We do so with a particular eye to the persistent problem in political science research of how to measure the effect of changes in turnout on elections. Past studies of turnout s effect on election outcomes consistently grapple with the problem of separating increased turnout effects from changes in voter choice among those who turn out, and with the additional problem of extracting these effects from outcomes that in part are attributable to election-specific circumstances. Though we certainly do not claim to have fixed these measurement problems, by comparing methods of assessing turnout while considering the

24 election, we can gain an understanding of whether increased turnout contributed to Obama s victory. We conclude that most likely, Obama did benefit by gaining votes from increased turnout compared to past elections, but that this turnout did not singularly hand him victory. Rather, voters also favored Obama at higher rates than expected, meaning it appears that Obama changed a few minds about vote choice instead of simply convincing more people predisposed to support him to turn out to vote. Comparing across methods, interestingly, our cross-county analysis within the 2008 election context suggests an increase in structural turnout should have actually slightly hurt Obama s vote share. Once we account for the mobilization component, however, these controls suggest there was some change in vote choice across what we might expect from the connection between status and vote choice, meaning Obama received a net benefit from increases in turnout because of those who chose to vote Democratic instead of (their perhaps baseline expectation of) Republican. In short, Obama did benefit from mobilization, but he also benefitted from voters changing their minds about vote choice. This outcome is corroborated by our county-level analysis comparing turnout and vote choice across 2004 and This analysis shows changes in turnout were not unidirectional; some counties experienced marked decline in turnout and in vote choice for the Democrat. In counties where turnout increased in 2008, Obama did make small gains in vote share, however, demonstrating some effect for mobilization. Finally, the full-turnout method suggests that indeed, had Obama (or all candidates for that matter) convinced everyone to vote, Obama would have increased his share of the popular vote substantially. Though we do determine in this piece the fate of the Electoral College 23

25 based on this national change in vote share, meaning we cannot comment fully on how this might have changed the overall outcome of the 2008 election, this is indeed more evidence that Obama would have received an even larger share of the vote had mobilization increased. Compared to the common wisdom about the outcome of the 2008 election, these results suggest we should not just think about Obama s victory as the outcome of his election team s success in getting those to vote who wouldn t otherwise. Though the full turnout model suggests that if Obama s team did convince typical non-voters to make the effort he would have won by a larger percentage, the county comparisons within the 2008 election suggest his team also made strides by convincing people who would typically vote anyway to vote Democrat. This outcome demonstrates the utility of intra-election jurisdictional comparison as a method to separate turnout from vote choice, as both DeNardo and McClurg suggest is useful. The unique benefit of understanding the structural component (the expected vote based on demographic characteristics) relative to the mobilization component (the deviation from this component based on turnout) is that we can see these components worked in opposite ways in the 2008 election. The underlying, or structural tendency in high SES counties was actually not to support Obama, but by accounting for the distinction between deviations in turnout and this tendency, we are able to see some suggestive evidence of Obama s ability to change minds as well as mobilize voters. More generally, these results suggest an interesting way to approach studying campaign dynamics. Can we separate the efforts of a campaign team to increase turnout and change minds? 24

26 We acknowledge we suffer from some of the same drawbacks found in other analyses of turnout and mobilization. Particularly, the baseline of a normal election and electorate remains elusive. Comparing 2008 to 2004 and 2000 is, we argue, probably a useful comparison, because the outcome of 2000 seemed to spur turnout to a high level replicated in However, are these three elections outliers from a longer series of American voter behavior or the proper baseline against which to judge future elections? Separating mobilization and turnout from voter choice estimates is a fruitful enterprise, but the relative importance of each component may also change across elections, especially compared to a low turnout context like a congressional midterm election or a less-exciting second term presidential election (for example, the Dole/Clinton contest of 1996). In short, work remains to be done to compare these two elements of voter mobilization in the 2008 election to the larger context of American elections and voter behavior. 25

27 Works Cited Bartels, Larry M Partisanship and Voting Behavior, American Journal of Political Science. 44(1): Citrin, Jack, Eric Schickler, and John Sides What if Everyone Voted? Simulating the Impact of Increased Turnout in Senate Elections. American Journal of Political Science 47(1): DeNardo, James Turnout and the Vote: The Joke s on the Democrats. American Political Science Review 74(2): DeNardo, James Does Heavy Turnout Help Democrats in Presidential Elections? American Political Science Review 80(4): Green, John C The faith factor: How religion influences American elections. Connecticut: Praeger. Hanjal, Zoltan & Jessica Trounstine Where Turnout Matters: The Consequences of Uneven Turnout in City Politics. Journal of Politics 67(2): Holbrook, Tom and Scott D. McClurg. Presidential Campaigns and the Mobilization of Core Supporters. American Journal of Political Science. 49(4): Martinez, Michael and Jeff Gill The Effects of Turnout on Partisan Outcomes in U.S. Presidential Elections. Journal of Politics. 67(4): McClurg, Scott D. N.d. The Political Consequences of Turnout in the 2004 Presidential Election: A County-Level Analysis. Unpublished paper. McDonald, Michael P. and Samuel L. Popkin The Myth of the Vanishing Voter. American Political Science Review. 95(4):

28 Nagel, John H. and John E. McNulty Partisan Effects of Voter Turnout in Senatorial and Gubernatorial Elections. American Political Science Review. 90(4): Radcliff, Benjamin Turnout and the Democratic Vote. American Politics Research. 22: Sides, John, Eric Schickler, and Jack Citrin If Everyone Had Voted, Would Bubba and Dubya Have Won? Presidential Studies Quarterly 38(3): Wuthnow, Robert The restructuring of American religion. New Jersey: Princeton University Press. 27