Local Unemployment and Voting for President: Uncovering Causal Mechanisms

Transcription

1 Local Unemployment and Voting for President: Uncovering Causal Mechanisms Taeyong Park Washington University in St. Louis Andrew Reeves Washington University in St. Louis Abstract How does local unemployment influence presidential elections? Some argue that, for voters, the state of the local economy is an afterthought to that of the national economy. Those who argue that local unemployment matters fall into two camps. Recent research finds that local unemployment is a partisan issue that benefits Democratic candidates because voters believe they are the party best equipped to deal with the issue. Alternatively, others have posited that the local economy provides information about the competency of the governing party. This view holds that the incumbent party, Democrat or Republican, will be punished when local unemployment is high. In this article, we investigate these two distinct mechanisms jointly. In an individual-level mediation analysis of the 2008 and 2012 presidential elections, we present evidence that both mechanisms are at work. Local unemployment bolsters support for Democratic presidential candidates, but, through its influence on views of the national economy, drives down support for the incumbent. Keywords: Local unemployment; Presidential elections; Retrospective voting; Issue-ownership voting; Causal mediation analysis

2 Local jobs are an oft-evoked electoral currency of U.S. presidents. From President Carter trumpeting the 450 new jobs created by a new steel plant in Perth Areboy, New Jersey in September of to President Obama touting the nearly 400 jobs created as a result of a solar energy center in Arcadia, Florida in October 2009, 2 presidents frequently campaign on local job creation in the communities they visit. Despite presidential rhetoric, existing studies reach differing and contradictory conclusions about the relationship between local employment and presidential voting. In this paper, we consider this relationship in the context of a causal inference framework and find that local employment growth influences presidential voting through direct and indirect mechanisms that are determined by both partisan reputations and economic evaluations of voters. 1 Local Jobs and Presidential Voting Despite presidential rhetoric, studies differ on both the nature and the extent to which voters hold presidents accountable for the local economy. One dominant view is that most citizens engage in national retrospective voting considering national economic conditions when voting for presidents and paying little attention to personal or local influence. Presidential support rises when macroeconomic conditions are good and falls when they are bad (Erikson 1989; Fair 1978; Hibbs 1989; Lewis-Beck and Stegmaier 2000; Norpoth 1984; Tufte 1978). These findings from aggregate studies are replicated at the individual level. Voters who view national economic conditions better are more likely to vote for incumbent presidents (Feldman 1982; 1984; Kinder and Kiewiet 1981; Lewis-Beck and Stegmaier 2000; Markus 1988). Voters set aside their personal and local experiences and engage in sociotropic voting by weighing the plight of the broader economy most heavily in voting for president. 3 The evidence for local retrospective voting considering local economic conditions when voting for president is smaller. The mechanisms are compelling and straightforward: personal Indeed, this theoretical view of voter accountability underpins foundational views of presidents as seekers of the national good against local passions of members of congress (Kriner and Reeves 2015). 1

3 observations and experiences such as daily drives to work or conversations with family, friends, and co-workers provide low-cost information about the local economy such as joblessness, rates of home foreclosures, or rising costs of goods. As voters look for low-cost information to make decisions about the quality of incumbent performance (Popkin 1994), the local economy may provide much guidance. Just as with national retrospective voting, local economic factors provide valence issues by which to judge the incumbent president or his party. Despite these straightforward mechanisms, there is little direct evidence that the local economy and in particular, local unemployment matters when it comes to presidential elections. While some early studies found evidence that local economic hardship was related to presidential voting (Gosnell and Colman 1940; Gosnell and Pearson 1941), more recent studies have found modest evidence or null effects. 4 In characterizing the literature on local economic voting, Eisenberg and Ketcham (2004) notes the relative unimportance of county level economic conditions (21) and confirm the lack of an effect of unemployment in their own study. In another study of the 2008 presidential election, Hill, Herron and Lewis (2010) finds that rising county-level unemployment over 2004 increased Obama s vote share by just one-tenth of one percentage point (42). In an analysis of presidential elections from 1988 to 2000, Kim, Elliott and Wang (2003) finds that local unemployment matters in only 1992, which they attribute to the strong third party showing of Ross Perot. 5 Though presidents rarely miss a chance to discuss the new jobs brought to a community from a new research institute or energy initiative, the evidence of voting in response to local unemployment is quite limited and characterized by null results. 6 Others have set aside a direct relationship between local employment and presidential election results and have instead examined attitudes toward a potential mediator: views of the national 4 See Cho and Gimpel (2009) for a notable exception with respect to presidential voting and Rogers (2014) with respect to MSA-level unemployment and approval of President Obama. 5 The definition of what is local is, of course, subjective. Our focus is county-level though others have found evidence of state economic context and presidential voting (Abrams and Butkiewicz 1995; Ansolabehere, Meredith and Snowberg 2014; Ebeid and Rodden 2006). 6 Though we focus here on unemployment, others have found evidence that voters hold presidents accountable for factors such as local income growth, local federal spending, local loan defaults, changes in wages, or local relief spending in response to natural disasters (Kriner and Reeves 2012; Gasper and Reeves 2011; Healy and Lenz 2016). 2

4 economy, a central driver of presidential voting. For example, Reeves and Gimpel (2012) illustrates the considerable variability in local economic conditions and then demonstrates that local unemployment, gas prices, and home foreclosure rates are associated with evaluations of the national economy. 7 Relatedly, Ansolabehere, Meredith and Snowberg (2014) finds that higher reported national unemployment rates are associated with voters whose neighbors or friends suffer from unemployment. While there is modest empirical evidence of a relationship between local economic factors and presidential voting, other studies have found strong links between objective local economic factors and perceptions of the national economy. These studies raise the possibility that any effect of the local economy may be mediated by perceptions toward the national economy. When the local joblessness rises, voters perceptions of the national economy worsen and, lacking a proper estimation strategy, the true effect of the local economy may be difficult to discern. 8 Others argue that voters consider local economic factors not as valence issues, but as partisan issues and reward the party that is best equipped to address the most pressing local economic tribulation. We refer to this as local issue-ownership voting. For example, voters view Democrats (and parties of the left generally) as prioritizing the issue of unemployment while they view Republicans (and parties of the right) as more able to handle rising prices for consumer goods (e.g. Alesina, Roubini and Cohen 1997). Indeed, some previous research demonstrates that the Democratic party gains electoral advantage from higher local unemployment (Brunk and Gough 1983; Kim, Elliott and Wang 2003; Wright 2012). Brunk and Gough (1983) finds that President Carter, running as the Democratic incumbent, performed better in states with higher unemployment in the 1980 election whereas Reagan, the Republican challenger, was hurt by higher unemployment. Wright (2012) shows that county-level unemployment and the Democratic vote for president move together in the presidential elections from 1996 to While much of the previous literature has 7 Though see Books and Prysby (1999) and Rogers (2014). 8 Though we do not consider it here, others have examined the determinants of perceptions of the local economy itself. See Rogers (2016; 2014). 9 We can find a similar pattern in the study of presidential approval in relation to the national unemployment rate. Swank (1993) shows that the approval rate for Democratic presidents benefits from rising unemployment whilst their Republican counterparts are hurt by rising unemployment. Carlsen (2000) presents the same evidence. See also Powell and Whitten (1993) and Martinsson (2009) for the discussions of issue-ownership effects associated with economic 3

5 focused on local economic outcomes as valence issues where all voters prefer superior outcomes and will reward any incumbent for achieving them these studies argue that this is not the case. Most notably for unemployment, there is evidence that Democrats are actually rewarded for local joblessness because of their perceived capabilities at addressing the issue. To summarize, while the most common view may be that there is no relationship between local employment and presidential voting, there are three distinct explanations for how these phenomena are related. First, and related to the null-effects argument, is that local joblessness can influence perceptions of the national economy but its effect ends there. Second is the view that local unemployment is a valence issue and that voters reward incumbent presidents or their party s nominee when their communities see more jobs. A third perspective argues that voters view Democrats as owning the issue of local unemployment and therefore are moved to vote for Democratic presidential nominees at higher rates when their local communities are suffering job loss. In this paper, we seek to explain how these seemingly disparate phenomena may coexist by uncovering two countervailing components of the total effect of local unemployment. We consider that the effect of local unemployment on voting for president is mediated by attitudes toward the national economy. When local unemployment goes down, citizens in those communities have a more positive view of the national economy and the incumbent president benefits (e.g., Cho and Gimpel 2009). However, local issue ownership also plays a distinct and sometimes countervailing role. Once evaluations of the national economy are accounted for, Democratic presidents benefit from local joblessness while Republicans are punished (Wright 2012). Consequently, local joblessness may hurt Republicans through either mechanism that voters regard unemployment as a valence issue or that they regard it as a partisan issue. On the contrary, under Democratic presidents, the two mechanisms may cancel each other, meaning that a group of voters who view unemployment as a valence issue punish Democrats for local joblessness while other voters who view it as a partisan issue reward Democrats for local joblessness. To examine all of these mechanisms at once, we improve upon previous analyses in two voting in the cross-national and Swedish contexts, respectively. 4

6 ways. First, instead of relying on presidential votes aggregated at the county level, we utilize large individual-level datasets on presidential voting that include the county of residence for each respondent. Second, we consider the question in a causal inference framework using causal mediation analysis, which allows us to detect the direct effect of local unemployment as well as the indirect effect mediated through views of the national economy. Specifically, we exploit a potential outcomes framework to decompose the total effect of local unemployment on presidential voting into the indirect mediation effect and the direct effect, which are not entirely separable without using counterfactuals. In doing so, we examine 1) whether there is a significant effect of local unemployment driven by individual views of the national economy, and 2) once the effect driven by individual views of the national economy is isolated, whether the rest of the total effect, which is the direct effect, is conditioned by partisan reputations for unemployment. To preview the main findings, our individual-level analysis for the 2008 and 2012 presidential elections provides evidence that local retrospective voting was at work in both elections in a way that translates people s discontent about local job conditions into voting against the incumbent party. Yet this negative relationship between rising unemployment and support for the incumbent was offset by the countervailing force of local issue-ownership voting under the Democratic incumbent in In contrast, the negative relationship was reinforced by the negative effect of local issue-ownership voting under the Republican incumbent in Though further data collections will allow us to examine more election cases in the future, with the data presently at hand, we uncover how the role that local economic particulars influence the presidential vote is conditioned by the incumbent s party. 2 Local Jobs and Presidential Voting in a Potential Outcomes Framework In this section we develop a voting mechanism model to test the three different perspectives discussed in the previous section in one unified framework: 1) local joblessness can influence percep- 5

7 Figure 1: The Voting Mechanism Model and Predicted Effects of Local Job Conditions on Voting for the Incumbent Party (A) Model (B) Predictions National economic perceptions <Democratic presidents> National economic perceptions ( ) <Republican presidents> National economic perceptions ( ) Local unemployment Voting for incumbent Rising local unemployment (+) Voting for incumbent Rising local unemployment ( ) Voting for incumbent [Notes: (A) MODEL: The total effect of local unemployment is decomposed into two causal pathways. The top causal pathway reflects local retrospective voting by which local job conditions shape individual perceptions of the national economy which in turn influence presidential voting. The bottom direct pathway between local unemployment and vote choice represents alternative mechanisms not intervened by perceptions of the national economy. It is hypothesized that once the effect working through perceptions of the national economy is accounted for, the direct pathway reflects local issue-ownership voting. (B) PREDICTIONS: Regardless of which party is in office, the predicted effect of rising local unemployment on voting for the incumbent, driven by national economic perceptions, is negative. In contrast, the predicted effect of rising local unemployment operating through the bottom direct pathway is positive when Democrats hold the presidency and negative when Republican presidents are in office, which reflects issue-ownership voting.] tions of the national economy but its effect ends there; 2) local unemployment is a valence issue, and it not only shapes perceptions of the national economy but also affects presidential voting such that more local jobs lead to voting for incumbent presidents or their party s nominee; 3) local unemployment is a partisan issue, and voters vote for Democratic presidential nominees as their concern about local jobs increases. Panel (A) in Figure 1 graphically illustrates the voting mechanism model. In the model, the total effect of local unemployment on vote choice is decomposed into two separate causal pathways. One is a pathway through which local retrospective voting is at work via perceptions of the national economy. As in Reeves and Gimpel (2012), we hypothesize that perceptions of the national economy are influenced by local factors but we extend this argument by examining the extent to which local unemployment influences voting for president through perceptions of the national economy. We also consider a direct pathway reflecting local issue-ownership voting. We 6

8 hypothesize that once the effect working through perceptions of the national economy is isolated, the remaining effect is conditioned by the incumbent party s issue ownership. 10 Based on these hypotheses, we derive testable predictions. First, local retrospective voting via perceptions of the national economy predicts that rising local unemployment hurts the incumbent party, Democrat or Republican. The negative signs associated with the top causal pathways in panel (B) in Figure 1 indicate this prediction. That is, regardless of which party is in office, the predicted effect of rising local unemployment on voting for the incumbent, driven by subjective evaluations of the national economy, is negative. 11 Second, local issue-ownership voting predicts that the effect of job loss is positive under Democratic presidents and negative under Republican presidents. The direct pathway reflects this type of voting, where the effect working through subjective evaluations of the national economy remains fixed. This prediction is represented by the positive and negative signs associated with the bottom pathways in panel (B) in Figure 1. Finally, our model permits some adjudication of the previous null-effects finding by generating testable predictions for the total effect. The total effect of rising unemployment on presidential voting is the sum of the two effects working through the local retrospective voting mechanism and through the issue-ownership voting mechanism. Therefore, the total effect is predicted to be negative under Republican presidents for which the negative effects driven by the two different mechanisms reinforce each other. On the other hand, the total effect can be negative, positive, or null under Democratic presidents depending on the relative size of the two mechanisms working in countervailing ways. To test these predictions, we consider conducting a mediation analysis that makes it possible to analyze the indirect mediation effect working through the top causal pathway and the direct effect working through the bottom causal pathway in one unified statistical framework. The traditional 10 It would be ideal to specify the issue-ownership voting mechanism as a causal pathway via individual beliefs about partisan reputations for dealing with unemployment. However, testing such a mechanism is barely feasible due to limited data on individual beliefs about distinct partisan reputations available in nationally representative surveys. 11 More specifically, rising local unemployment ( ) makes subjective evaluations negative ( ), and negative evaluations ( ) decrease support for the incumbent ( ). Thus, the overall effect is predicted to be negative. 7

9 linear structural equation modeling approach to a mediation analysis (e.g. Baron and Kenny 1986) is not directly applicable to nonlinear models with a discrete dependent variable (Imai, Keele and Tingley 2010; Imai et al. 2011). This is a critical limitation for our analysis because the nature of our dependent variable, vote choice, is binary. In contrast, the potential outcomes approach to causal mediation analysis specifies a general estimation algorithm that can be applied to nonlinear models. Moreover, this approach provides clear identification assumptions required to separate the indirect effect and the direct effect. For these reasons, we build on recent methodological developments of the causal mediation analysis based on the potential outcomes framework (Imai, Keele and Tingley 2010; Imai, Keele and Yamamoto 2010; Imai et al. 2011). The rationale for introducing the potential outcomes framework to a causal mediation analysis is to utilize counterfactual potential values. Specifically, for example, the direct effect should be an effect not intervened by subjective evaluations of the national economy. To estimate this effect, given the powers of omnipotence, we would hold subjective evaluations constant by turning off the indirect channel and make only local job conditions change. Yet, without such powers in reality, it is not possible to change local job conditions while holding subjective evaluations constant because as local job conditions move, subjective evaluations also move. The potential outcomes framework exploiting counterfactuals provides a solution. It allows us to estimate what would happen to presidential voting if 1) subjective economic evaluations changed with holding local job conditions constant or 2) if a local job condition changed with holding subjective evaluations constant. 12 Case 1) is associated with the indirect effect since it separates the indirect channel from the direct channel by extracting the effect only driven by subjective evaluations. Case 2) is associated with the direct effect, in that it switches off the indirect channel. 12 This conceptualization is from Becher and Donnelly (2013, 971). We acknowledge that the Becher and Donnelly study provided us with an important insight for applying the potential outcomes framework to our analysis. 8

10 3 Data and Methods We now turn to exam the influence of local unemployment on individual-level voting for president. As we described in the previous section, we aim to test three relationships: 1. An indirect effect of local unemployment through perceptions of the national economy that is negative for both Republican and Democratic incumbents. 2. A direct effect local unemployment that is negative for Republicans. 3. A direct effect local unemployment that is positive for Democrats. Our analysis relies on the 2008 and 2012 survey data from Cooperative Congressional Election Study (CCES) administered by YouGov/Polimetrix. Our analysis uses the full, nationally representative, stratified sample of 23,585 respondents in the 2008 survey and 37,909 respondents in the 2012 survey. These datasets identify the county of residence for each respondent so that we can match the survey respondents to their local economic statistics. Furthermore, the 2008 sample covers 2,335 counties, and the 2012 sample covers 2,497 counties. Without these large surveys encompassing numerous locales, it is difficult to test theories of locally contingent political behavior. Though more election cases will be examined with further data collections in future research, with the data presently at hand, we exploit different electoral circumstances with respect to the incumbent s party. Our dependent variable is whether the respondent reported voting for the incumbent (or his party) or the challenger. In 2008, a reported vote for Republican Senator John McCain reflects support for the incumbent, and in 2012 those who supported the incumbent supported Democratic President Barack Obama. The dependent variable takes 1 for a pro-incumbent vote, and 0 otherwise. The key dynamic that we are interested in is the effect of local unemployment on voting behavior. We obtain county-level unemployment statistics from the Local Area Unemployment Statistics (LAUS) provided by the Bureau of Labor Statistics Website 13 and match it to the CCES respon

11 dents using their county-level identifiers. Since our mediating variable is measured by a CCES survey item asking subjective evaluations of the nation s economy over the past year, we focus on the annual change in unemployment over the year preceding the presidential election contest. By doing so, our measurement of local unemployment corresponds to the data generating process of the mediating variable. Furthermore, credit or blame for the annual change is more plausibly related to the policies of the current government (Becher and Donnelly 2013). The level of unemployment is shaped by many structural factors that are embedded in each of different counties and are relatively persistent. Thus, voters are unlikely to attribute the level primarily to the current government. 14 According to our data, there is substantial variation in annual changes in unemployment across the country as Figure 2 depicts. Though local economies on average were plunging into a recession in 2008, there existed a great deal of variation. For instance, job losses between 2007 and 2008 were particularly severe in some counties such as Chambers County, Alabama (a 7.93 percentage points increase in the unemployment rate) and Lauderdale County, Tennessee (a 5.85 percentage points increase). By contrast, there are other counties in which the unemployment rate declined between 2007 and 2008 despite the nationwide economic turmoil. Such counties include Woods County, Oklahoma (a 3.08 percentage points decrease) and Wyoming County, West Virginia (a 1.41 percentage points decrease). Similarly, in 2012 local economies were recovering overall, but economic recovery was slow in some regions and fast in others. Moreover, a significant number of counties suffered job losses in spite of the nationwide trend of economic recovery. For example, residents in counties like Harlan and Letcher in Kentucky suffered job losses between 2011 and 2012 (a more than 2.9 percentage points increase). In contrast, the job conditions in Macon County and Clarke County in Alabama improved quite a bit during the same period of time (a more than 3.3 percentage points decrease) although the level of unemployment still remained at higher rates than the national average in each of these two counties. Given this substantial variation in our key explanatory variable, it is 14 As discussed below, we check for the robustness of the estimated effect of the annual change in unemployment. 10

12 Figure 2: Distribution of Election Year Change in the Annual Average Unemployment Rate at the County Level, 2008 and Median = Median = 0.88 Frequency Annual change in the unemployment rate (%) [Notes: Each data point corresponds to a local job condition associated with an individual survey respondent. The dotted vertical line indicates the median value of annual changes in the unemployment rate. Local economies on average were plunging into a recession in 2008 and recovering in Still, there is substantial variation in annual changes in unemployment across the country.] meaningful to examine whether and how such variation explains presidential voting. The mediating variable is subjective economic evaluations of the national economy. We use a CCES survey item that measures respondents retrospective evaluations. This variable is a fivepoint scale ordered categorical variable. The survey question is worded Would you say that over the past year the nation s economy has...? Response options are coded as (1) gotten much better ; (2) gotten better ; (3) stayed about the same ; (4) gotten worse ; (5) gotten much worse. Though theoretical motivations lead us to focus on changes in local unemployment as our treatment variable, there are other local economic variables that may drive people s view of the economy and/or vote choice. Failing to control for alternative measures might produce omitted variable bias. Therefore, we account for three additional county-level economic variables: the level 11

13 of unemployment, gasoline prices, and home foreclosure rates. We provide further details about these variables in the online appendix (See A1). Moreover, we employ several alternative models to check for the robustness of the analysis results to unobserved heterogeneity across counties or across states. These alternative models yield virtually identical estimation results and do not change our primary inferences. Further details about these models and their estimation results are presented in the online appendix (See A4). Using the data described above, we aim to estimate two quantities: 1) the indirect mediation effect what would happen to presidential voting if subjective economic evaluations changed with holding local job conditions constant to test the predictions derived from local retrospective voting; 2) the direct effect what would happen to voting if an local job condition changed with holding subjective evaluations constant to test the predictions derived from local issue-ownership voting. As discussed earlier, neither of these two quantities is observable because subjective evaluations and objective job conditions move together and they are not entirely separable. The potential outcomes framework makes it possible to estimate these unobservable effects by means of simulated potential values. 15 The estimation procedure consists of two steps. The first step fits two regression models: the model for the mediating variable, retrospective evaluations, and the model for the outcome variable, vote choice. And then, the second step computes the mediation effect and the direct effect using simulated potential values that are generated from the two regression models. Detailed 15 Two assumptions are required. The first assumption is that a local job condition is not affected by potential outcomes for vote choice or subjective evaluations of the economy given the observed pretreatment covariates. The second assumption is that a subjective economic evaluation is independent of potential outcomes for vote choice given the observed pretreatment covariates and the observed values for local unemployment. These two assumptions are based on the sequential ignorability assumption proposed by Imai, Keele and Yamamoto (2010). We account for demographic and socioeconomic characteristics and political predispositions as pretreatment covariates to reduce the possibility of violating these assumptions. In particular, it is important to account for political factors such as partisanship and ideology, which could confound the causality between economic evaluations and vote choice (Anderson, Mendes and Tverdova 2004; Evans and Andersen 2006; Evans and Pickup 2010; Gerber and Huber 2010; Wlezien, Franklin and Twiggs 1997). Additionally, we conduct a sensitivity analysis to evaluate how much our results would be influenced by a possible violation of the assumptions. Detailed descriptions of the pretreatment covariates and sensitivity analysis results are in the online appendix (See A2 and A5). 12

14 explanations of this estimation procedure are in the online appendix. 16 Yet it is worth explaining here the basic idea behind the procedure. We begin by fitting the model for the mediating variable. Using the regression estimates, we simulate predicted values of retrospective evaluations associated with two different local unemployment rates holding all other variables constant such as Eval i (c x i ) and Eval i (c x i ), where i is an individual respondent, Eval i (Job i x i ) is the model of retrospective evaluations as a function of job conditions given covariates x, and c and c are two different realized values for Job i. In the same vein, we fit the model for the outcome to generate 1) predicted values of presidential voting associated with two different realized values for local job conditions holding all other variables constant and 2) predicted values of presidential voting associated with two different retrospective evaluations holding all other variables constant as follows: 1) Vote i (c, Eval i (c x i ) x i ) and Vote i (c, Eval i (c x i ) x i ), 2) Vote i (c, Eval i (c x i ) x i ) and Vote i (c, Eval i (c x i ) x i ), where Vote i (Job i, Eval i (Job i x i ) x i ) is the model of presidential voting as a function of unemployment and retrospective evaluations given covariates x, and c and c are two different realized values for local employment conditions. Thereafter, using these predicted potential values we 16 See A3. Also, see the general estimation algorithms for the mediation analysis that can be found in Imai, Keele and Tingley (2010, Appendix D). Furthermore, see Becher and Donnelly (2013, 971-4) that provides an example of applying the causal mediation analysis based on the potential outcomes framework to the study of economic voting. 13

15 estimate the mediation effect (ME) and the direct effect (DE) as follows: ME = Vote i (c, Eval i (c x i ) x i ) Vote i (c, Eval i (c x i ) x i ), (1) DE = Vote i (c, Eval i (c x i ) x i ) Vote i (c, Eval i (c x i ) x i ). (2) The estimated mediation effect (1) is associated with how much presidential voting would change if retrospective evaluations change, i.e. Eval i (c x i ) Eval i (c x i ), while the effect of local job conditions remains constant at c. The estimated direct effect (2) is associated with how much presidential voting would change if local job conditions vary, i.e. Vote i (c,...) Vote i (c,...), while a retrospective evaluation remains constant at Eval i (c x i ). The total effect (T E) is simply the sum of the mediation effect and the direct effect. By repeating this procedure for varying combinations of c and c, we can compute multiple sets of estimated ME, DE, and T E. In our analysis, we exploit nine combinations of (c, c ) to estimate the effects of a one percentage point change in local unemployment. 17 A one percentage point change in local unemployment represents a reasonable amount of change, which is about 0.7 standard deviations in our 2008 data and about 0.8 standard deviations in our 2012 data. Taking the average over the nine multiple sets of estimated ME, DE, and T E produces the average mediation effect (AME), the average direct effect (ADE), and the average total effect (AT E). Finally, we repeat drawing the regression estimates from their estimated sampling distribution 200 times to compute statistical uncertainty about the average effects (King, Tomz and Wittenberg 2000). The respective means of these 200 sets for AME, ADE, and AT E and their statistical uncertainty are the primary results that we present in the following section. 17 These nine combinations are illustrated as follows: (c, c ) = {(-2.5, -1.5), (-2.0, -1.0), (-1.5, -0.5), (-1.0, 0), (-0.5, 0.5), (0, 1.0), (0.5, 1.5), (1.0, 2.0), (1.5, 2.5)}. Note that every combination exploits a one percentage point change in local unemployment by forcing the difference between c and c to be one. 14

16 4 Results We consider how an increase in local unemployment affects voting for the incumbent through views of the national economy, which is the AME, as well as directly through other mechanisms, the ADE, as well as those two effects added together, which is the AT E. In Figure 3, each point is an estimated effect of a one percentage point increase in the annual average unemployment rate on the likelihood of voting for the incumbent party. Each horizontal line indicates uncertainty about the estimated effect at a 95% confidence level. The left panel depicts estimated effects on voting for John McCain in the 2008 presidential election while the right panel depicts estimated effects on voting for Barack Obama in The estimated AM E is statistically distinguishable from zero in both elections. The point estimates corresponding to the AME in the top row in Figure 3 indicate that a one percentage point increase in unemployment within a county, decreases the likelihood of voting for the incumbent party by about three-tenths of a percent in 2008 and by about six-tenths of a percent in 2012 by making voters think the national economy is getting worse. This finding is consistent with our hypothesis that the indirect effect is negative for both Republican and Democratic incumbents. In other words, this finding demonstrates that local retrospective voting were at work in both elections. To interpret the AME in a more intuitive way, suppose, for example, a county suffers a one percentage point increase in unemployment between 2011 and Suppose further that we isolate a group of citizens whose voting is driven by their views of the national economy from all other citizens whose voting is not driven by their views of the national economy. The estimated AME for 2012, 0.6%, means that if we only consider the group of citizens whose views of the 18 See online appendix A4 that includes four tables presenting the regression estimates from the models of retrospective evaluations and the models of vote choice. Table 1 and Table 2 in the appendix demonstrate that an increase in county-level unemployment leads to negative retrospective evaluations of the national economy in both the 2008 and 2012 elections, which is robust to several alternative model specifications. Table 3 and Table 4 in the appendix show that as people see the economy more negatively (positively), they tend to become less (more) likely to vote for the incumbent party in both elections, which is robust to several alternative model specifications. From these regression estimates, we may make a preliminary inference that rising local unemployment forms negative individual retrospective evaluations of the national economy which in turn decrease support for the incumbent party. In the main text, beyond this preliminary inference, we present the estimates of AME, ADE, and AT E as illustrated by Figure 3. 15

17 Figure 3: Effect of Changes in the County-level Unemployment Rate on Individual Vote Choice for the Incumbent Party, 2008 and AME ADE ATE 2% % The likelihood of voting for the incumbent party [Notes: The estimated average mediation effect (AM E), average direct effect (ADE), and average total effect (AT E) are presented. These effects are generated by a one percentage point increase in the annual average of county-level unemployment rate over the year preceding the election. The left panel depicts results for the 2008 election, and the right panel depicts the 2012 results. The AM E reflects the voting mechanism that voters evaluations of the national economy intermediate between a change in unemployment and vote choice. The ADE reflects all other alternative mechanisms. The AT E is the sum of the AME and ADE. Each point is the estimate, and each horizontal line is equivalent to a 95% confidence interval representing uncertainty about the estimated effects.] national economy drive their vote choice, a one percentage point increase in unemployment would decrease the likelihood of voting for Obama by about 0.6% in that county. Furthermore, this effect translates into about 1.2% vote swing from Obama to Romney. 19 To help put this magnitude in context, consider that Kriner and Reeves (2012) finds that increasing federal spending in a county by 80% results in swing of 1.1% toward the incumbent. 19 If Obama s share goes down by 0.6%, then Romney s share goes up by that amount resulting in a swing of 1.2%. As the AM E is calculated by taking the average over individual respondents predicted vote choice, the AM E indicates the difference between expected vote shares of the incumbent party (Becher and Donnelly 2013, 973). Therefore, a one percentage point increase in unemployment decreases individual respondents predicted voting for Obama, which translates into decreasing the expected vote share of Obama, by about 0.6%. 16

18 On the other hand, we also find a substantively large direct effect that is unmediated by retrospective evaluations of the national economy. The point estimates corresponding to the ADE in Figure 3 show that a one percentage point increase in unemployment within a county decreases the likelihood of voting for McCain by about 0.7% in 2008 and increases the likelihood of voting for Obama by about 1% in 2012 through the direct causal pathway. This finding is consistent with our hypothesis that the direct effect is negative for Republicans and it is positive for Democrats, which provides evidence for the presence of local issue-ownership voting. For intuitive interpretations of the ADE, suppose again a county suffers a one percentage point increase in unemployment between 2011 and As in the case of the AME, continue to suppose that we divide the citizens in that county into two groups: first is a group of citizens whose voting is driven by their perceptions of the national economy and second is a group of citizens whose voting is not intervened by their views of the national economy. The ADE is associated with the latter group of people whereas the AME is associated with the former group as we discussed above. Thus, given that we take into consideration only the latter group of citizens, the positive estimated ADE, 1%, means that a one percentage point increase in unemployment in that county increases the likelihood of voting for Obama by about 1%. This translates into about 2% vote swing from Romney to Obama in the same manner described earlier. In contrast, the negative estimated ADE in 2008 reveals that a one percentage point increase in unemployment between 2007 and 2008 decreases the likelihood of voting for McCain by about 0.7% given the same group citizens under consideration. This translates into about 1.4% vote swing from McCain to Obama. Finally, consider the total effect of changes in local unemployment (AT E) which is the sum of the AME and ADE. According to the point estimates corresponding to the AT E in the bottom row in Figure 3, a one percentage point increase in local unemployment decreases the likelihood of voting for McCain by about 1% in 2008, and the same amount of increase has no substantial effect in The total effect of a 1% decrease in 2008 translates into about 2% vote swing from McCain to Obama. 17

19 Our individual-level analysis allows us to answer questions we have been unable to answer. Previous research on local unemployment and voting is characterized by disparate findings. Some studies find no significant effects, others support retrospective voting, and still others support issueownership voting (Abrams and Butkiewicz 1995; Brunk and Gough 1983; Eisenberg and Ketcham 2004; Hill, Herron and Lewis 2010; Kim, Elliott and Wang 2003; Wright 2012). Our analyses provide an explanation that helps reconcile these divergent findings. In 2008, rising local unemployment hurt McCain through both the direct and indirect causal pathways. In other words, the negative effect of rising unemployment working through retrospective voting was reinforced by the negative effect of issue-ownership voting. In 2012, on the contrary, the direct effect and the indirect effects of local joblessness cancelled each other. The negative effect of rising unemployment working through retrospective voting was offset by the countervailing force of issue-ownership voting. As a result, local unemployment appears to have no effect on the election outcome in terms of the total effect. However, this does not mean that voters did not hold presidents accountable for local jobs because the valence mechanism of retrospective voting was still at work. Based on these findings, we argue that local unemployment is a multifaceted issue influenced by both the valence and partisan mechanisms. This argument complements Wright s (2012) county-level evidence that unemployment is a partisan issue for voters, not a valence issue (699). The estimated total effect in our analysis is consistent with Wright s finding. Yet our individuallevel causal mediation analysis allows us to uncover the two components of the total effect that are found to be at work in both 2008 and Our multidimensional-issue perspective provides an explanation of why Democratic presidents have incentive to bring unemployment under control. Even if Democrats could benefit electorally from rising unemployment because of those voters who translate rising local unemployment into voting for the incumbent Democratic party, they still have incentive to attract another type of voters who credit the incumbent government for keeping unemployment low. Before we conclude, we discuss a supplementary analysis briefly. In addition to our main analysis, we investigated the role that political interests and partisanship might play in moder- 18

20 ating our estimated AM E, ADE, and AT E. Previous research reports that economic voting requires some level of political expertise and ordinary citizens vary greatly on this dimension (e.g. Delli Carpini and Keeter 1996; Gomez and Wilson 2001; Goren 1997; Krause 1997). Partisanship could also play an important role in the sense that the extent to which local conditions influence economic evaluations may be moderated by individual partisan orientations (Reeves and Gimpel 2012). Building on these previous literatures, we considered whether the mechanisms we had uncovered varied by political interests or partisanship. In this supplementary analysis, we find no conditioning effect of political interests or partisan predispositions. First, we find that the estimated effect of the change in local unemployment on subjective evaluations of the economy is not conditioned by political interests or partisanship. Specifically, the interaction of different levels of news interests and the change in local unemployment is practically zero. Similarly, the interaction of different partisan groups and the change in local unemployment has a statistically null impact. Moreover, we conducted a causal mediation analysis employing the interaction terms associated with political interests and partisan identification. We find no evidence that the estimated AME, ADE, and AT E vary by either political interests or partisan predispositions Conclusion This article reconciles disparate perspectives on the relationship between local joblessness and presidential voting. We demonstrate that local retrospective voting and local issue-ownership voting were at work in a reinforcing way in 2008 and in a countervailing way in Increases in local unemployment decreased the probability of voting for the incumbent party by way of influencing perceptions of the national economy in both elections. Through an alternative mechanism, increases in local unemployment were associated with decreased support for McCain in 2008 and increased support for Obama in These findings imply that local unemployment 20 Further details of this analysis are in the online appendix A6. 19

21 is a multidimensional issue. Some voters translate rising local unemployment into voting against the incumbent regardless of the incumbent s party. Yet others consider parties reputations for handling unemployment to make a voting decision. By jointly taking into account these two distinct mechanisms, we can better understand how the local economy affects presidential accountability. Consistent with arguments related to national retrospective voting, local unemployment was not influential in the 2012 election in terms of the total effect. However, we find that underneath the total effect, both local retrospective voting and local issue-ownership were influential yet countervailing forces. Thus, it is incorrect to conclude that voters do not hold presidents accountable for the local jobs. While Democratic presidential candidates benefit from rising local unemployment as the issue-ownership voting mechanism suggests, the presence of the local retrospective voting mechanism explains why Democrats still work to bring local unemployment under control. There is still much to investigate. Though we provide evidence from 2008 and 2012, we should examine earlier and later elections. One limitation, though, is that it is often difficult or impossible to obtain local geographic information about survey respondents. Second, future research should consider the influence of the local economy for other office holders such as governors or members of Congress. Third, it is worth investigating the issue-ownership voting mechanism in a direct way that employs a mediator measured by people s distinct beliefs about how competently each party deals with unemployment. While currently there is no nationally representative survey that provides such data, once they are available, it will be straightforward to operationalize and test the local issue-ownership voting mechanism by exploiting the research design that we introduce in this paper. Finally, further research should also focus on how the relationships we examine vary across countries. Building on previous findings that, across advanced industrial democracies, parties of the left tend to benefit from rising or high unemployment (Carlsen 2000; Martinsson 2009; Powell and Whitten 1993), it will be fruitful to uncover, beyond the U.S. context, how these findings of issue-ownership voting can be reconciled with the conventional retrospective economic voting theory. Indeed, Becher and Donnelly s (2013) analysis employing national surveys from 20

22 18 countries between 1979 and 2001 finds a positive direct effect of rising unemployment on the vote share of the party of the incumbent chief executive, which is unmediated by voters national economic perceptions. Our findings suggest that the positive direct effect may be most acute among incumbent parties of the left. 21

23 References Abrams, Burton A and James L Butkiewicz The Influence of State-level Economic Conditions on the 1992 US Presidential Election. Public Choice 85(1-2):1 10. Alesina, Alberto, Nouriel Roubini and Gerald D. Cohen Political Cycles and the Macroeconomy. MIT Press. Anderson, Christopher J, Silvia M Mendes and Yuliya V Tverdova Endogenous Economic Voting: Evidence from the 1997 British Election. Electoral Studies 23(4): Ansolabehere, Stephen, Marc Meredith and Erik Snowberg Mecro-Economic Voting: Local Information and Micro-Perceptions of the Macro-Economy. Economics and Politics 26(3): Baron, Reuben M and David A Kenny The Moderator-Mediator Variable Distinction in Social Psychological Research: Conceptual, Strategic, and Statistical Considerations. Journal of Personality and Social Psychology 51(6): Becher, Michael and Michael Donnelly Economic Performance, Individual Evaluations, and the Vote: Investigating the Causal Mechanism. The Journal of Politics 75(04): Books, John and Charles Prysby Contextual Effects on Retrospective Economic Evaluations the Impact of the State and Local Economy. Political Behavior 21(1):1 16. Brunk, Gregory G and Paul A Gough State Economic Conditions and the 1980 Presidential Election. Presidential Studies Quarterly 13(1): Carlsen, Fredrik Unemployment, Inflation and Government Popularity: Are There Partisan Effects? Electoral Studies 19(2): Cho, Wendy K. Tam and James G. Gimpel Presidential Voting and the Local Variability of Economic Hardship. The Forum 7(1). Delli Carpini, Michael X and Scott Keeter What Americans Know about Politics and Why It Matters. New Haven, CT: Yale University Press. Ebeid, Michael and Jonathan Rodden Economic Geography and Economic Voting: Evidence from the US States. British Journal of Political Science 36(3): Eisenberg, Daniel and Jonathan Ketcham Economic Voting in U.S. Presidential Elections: Who Blames Whom for What. Topics in Economic Analysis and Policy 4(1):1 23. Erikson, Robert S Economic Conditions and the Presidential Vote. American Political Science Review 83(02): Evans, Geoffrey and Mark Pickup Reversing the Causal Arrow: The Political Conditioning of Economic Perceptions in the US Presidential Election Cycle. The Journal of Politics 72(04):