The polls: presidential referendum effects in the 2006 midterm elections.

Cohen, Jeffrey E.

Many commentators saw the 2006 midterm elections as a referendum on George W. Bush's presidency, especially his unpopular Iraq War policy. Presidential referendum effects may be highly likely given the degree of party polarization and the concomitant attachment of large numbers of voters to the Democratic and Republican parties. But local factors, especially incumbency effects, may also have affected the outcome of the 2006 midterm contests. Using state-level presidential approval polls, analysis finds that Bush's approval, an indicator of referendum voting, had little impact on voting in Senate contests, once applying controls for the ideological and partisan composition of state voting populations. In contrast, even with these controls, Bush's approval influenced voting in gubernatorial contests. But Iraq War battle deaths affected voting for senators, suggestive of another type of referendum effect. Battle deaths, however, did not affect the gubernatorial races.

**********

Democrats decisively defeated the Republicans in the 2006 midterm elections, unseating Republicans across the nation and for all types of offices up for grabs. Overall, the Democrats realized seat gains in both houses of Congress and the state governorships. For instance, Democrats gained 6 Senate seats out of 33 contested, increasing their representation in the upper chamber from 45 to 51, thereby assuming control of that chamber. Similarly, a gain of 30 seats led to Democratic control of the House, after Republican control that dated back to 1995. And 2006 saw Democrats control a majority of governorships, increasing from 22 to 28. In each of these cases, the Democrats entered the 2006 races as the minority and emerged with the majority. Democratic party fortunes also improved in the state legislatures. Prior to the midterm, the Democrats held 21 state houses, the Republicans 19, 9 were split, with 1, Nebraska, nonpartisan. Coming out of the midterm, Democratic control of state houses rose to 23, with a decline of 2 for the Republicans, while 9 still remained under split control.

Moreover, in all cases of U.S. Senate and gubernatorial races, the Republicans were unable to take any seats that the Democrats held going into the election. Another indication of the decisiveness of the Democratic victory was that almost all closely contested races fell into the Democratic column, such as the Senate seats in Virginia and North Dakota; Republicans could win only a smattering of close contests.

The consistent Democratic electoral showing in all regions and for all types of races led many postelection accounts to argue that national factors worked to the Democrats' advantage. In other words, the 2006 midterms, unlike many other midterms, was a national election, a referendum on President George W. Bush, especially his Iraq War policy. Yet local conditions might still have mattered; national affairs were not necessarily the whole story of the 2006 midterm.

For instance, local scandals, such as those associated with Ohio's Republican Governor Robert Taft, presumably affected the race for governor in that state, in which Democrat Ted Strickland beat Republican John Blackwell by a hefty 60 to 37 percent margin. This was in a state that had for the past several election cycles been the quintessential battleground state, with Republicans and Democrats usually in close contests, but with the Republicans often edging out the Democrats.

In addition, some Republicans running for election possessed a personal popularity that allowed them to buck the national tides in 2006. For example, Republican incumbent governor of California, Arnold Schwarzenegger, was highly popular. His popularity had recovered from dips of the past year, and he managed to win a comfortable election victory in a state that has been decidedly Democratic in recent years. As another example, Richard Lugar, Republican incumbent senator from Indiana, won approximately 90 percent of the vote; the Democrats did not field a candidate to run against him. Lugar faced only token opposition from the Libertarians.

This article asks what impact the presidency had on the midterm elections in 2006. Was the election's outcome a referendum on an unpopular president and his unpopular war policy in Iraq? At the time of the 2006 midterm, Bush's job approval ranged in the 35 to 40 percent level, according to the major national polls. (1) With regard to the Iraq War, CNN reported a 33 percent approval rating for the war in the first week of November 2006, and 53 percent of respondents to the USA Today/Gallup Poll of the same week called the Iraq War a mistake. Other questions from these firms, as well as polls results from other firms, all point to the same conclusion--the war in Iraq was unpopular with the public by sizeable margins. Did Bush's approval and public opinion on the war affect voting in the midterm contests?

To test for whether the 2006 midterm could be considered a referendum on the president, I use newly available state-level polls from SurveyUSA which provide us with presidential approval levels, as well as other important public opinion indicators such as state partisanship and ideological leanings, across all fifty states. SurveyUSA has been conducting such polls and making their results publicly available since mid 2005. Despite some limitations, discussed below, these survey data have been used to some effect in recent research (Jacobson 2006; Cohen 2006).

The next section of this article discusses the literature on subpresidential and midterm elections. Then I present the SurveyUSA polls used in this study; analysis of the senatorial and gubernatorial midterm election results follows. The conclusion raises some issues of interpretation, especially of detecting presidential referendum effects in a highly polarized political context.

Local and National Factors in Subpresidential and Midterm Elections

Due to the federal structure of our political system, a combination of local and national factors determines the outcomes of subpresidential elections. National tides may influence the outcome of subpresidential elections, but their impact may vary, in part a function of their intensity. On the other hand, local factors, acting like breakwaters, may soften the impact of these national tides on local elections.

Local factors that affect subpresidential elections are primarily structural in nature. For instance, voting is for candidates as opposed to parties. Voters must know the party of the local candidate and the president and then decide to reward or punish the candidate for being of the same party of the president. And some voters may refuse to tie local officials to the president in such a way, because the functional job duties of the local official are purely local (Carsey and Wright 1998a; Stein 1990).

Moreover, many local elections are not held at the same time as the presidential election, further divorcing the local contest from the presidency. When the president does not appear on the ballot at the same time as the local candidate does, voters may not connect the local election to the president. In part to insulate local politics from national tides and presidential association, many states in the mid-twentieth century changed the timing of their gubernatorial elections. The intention of this reform was to focus state citizenry on the needs of the state executive and keep presidential considerations from contaminating that decision. Similarly, staggered terms for office, such as the U.S. Senate, to some degree also insulate that body from presidential and national politics--only one third of the body is up for election at a time, which limits the number of seats that can change hands in any one election.

Further, office holders may build incumbency advantages that will also insulate them from the effects of national tides. Members of Congress have developed and exploited resources of office that help in that regard. Perhaps most critically, office holders, due in part to their own efforts, but also a function of their names appearing on the ballot, build a personal following and develop a personal image. This "personal vote" has helped many an incumbent, well-liked by his or her constituents, to survive national political storms, such as unpopular presidents, failed policies, and economic slowdowns (Ansolabehere, Snyder, and Stewart 2000; Herrera and Yawn 1999; Lockerbie 1999; Dawes and Bacot 1998; Swain, Borrelli, and Reed 1998; Erikson and Palfrey 1998; Levitt and Wolfram 1997; Cox and Katz 1996; King and Gelman 1991; Garand and Gross 1984; Cain, Ferejohn, and Fiorina 1984; Gross and Garand 1984; Campbell 1983; Cover and Brumberg 1982; Alford and Hibbing 1981; Cover 1977, 1980; Abramowitz 1975; Erikson 1971).

High local barriers to national tides are currently in place. Only eleven state governors are now elected in presidential years; thirty-six are elected during the congressional midterm and the remaining five during odd-number years. (2) Members of Congress still enjoy a substantial incumbency advantage. Over 90 percent of incumbents who run for reelection win, including 90 percent in 1994 and 95 percent in 2006, years when the party controlling the House lost its hold of that chamber. In those years, most seat change came from open contests which swung Republican in 1994 and Democrat in 2006.

Yet extreme polarization in the party system may be nationalizing subpresidential elections, as voters view politicians as members of a party team. Now, not only do the parties differ across almost every policy domain, but party differences coincide with ideological differences. Few liberal Republicans or conservative Democrats currently exist in the ranks of office holders or voters. Further, the number of self-identified Independents has dwindled and party identification appears as perhaps the strongest determinant of the individual vote choice (Bartels 2000; Jacobson 2005; Hetherington 2001).

Partisanship and ideology also strongly structure the public's opinion about the president and his policies. Presidential copartisans rarely disapprove of his job in office; identifiers of the opposition party rarely approve. The partisan gap in presidential approval is at its highest rates since poll questions have existed (Bond and Fleisher 2001; Jacobson 2006). Given this polarized environment, did presidential approval and the Iraq War play a role in the 2006 midterm election outcomes? These are the questions that I address in this article. (3)

Presidential Approval at the State Level: The SurveyUSA Polls

To test for presidential referendum effects in the 2006 Senate and gubernatorial races requires data on public attitudes toward the president, as well as on vote decisions and/or election outcomes. For instance, one can analyze individual vote choice by using data from national surveys, like the American National Election Studies, that ask voters about their attitudes toward the president and their vote choice. Such studies, however, cannot tell about factors affecting state-level outcomes, which are important because they determine seat distributions and party control in Congress. Moreover, major national surveys like these will not be publicly available for analysis for some time to come.

Another option is to use state-level results from exit polls, like those from Edison-Mitofsky, which conducted the major exit poll for the consortium of large national news organizations. Exit polls have the virtue of being conducted at the voting site, thereby minimizing voter recall problems. But issues arise in using presidential approval measures gleaned from the exit poll. Presidential approval and vote choice are measured simultaneously in exit polls and the possibility exists that vote choice will affect attitudes toward the president as voters psychologically justify or rationalize their vote choice. Another problem specific to the 2006 midterm is that Edison-Mitofsky did not conduct exit polls in all states holding contests that year, nor were its questions uniform across the states. (4)

SurveyUSA provides us with another data source for presidential approval and other opinions at the state level which overcome some of the limitations noted above. First, SurveyUSA has been conducting statewide polls that ask the presidential approval question every month, in every state, since May 2005. Thus, we can use a presidential approval reading taken prior to the election which provides an unambiguous temporal ordering of opinion and the vote. SurveyUSA coverage across all fifty states also ensures that no state with an election is left out of the analysis.

Another advantage of the SurveyUSA polls is the use of the same sample size per survey, 600, which generates the same sampling error across states. SurveyUSA estimates this to be [+ or -]4.1 percent. Like many survey organizations, SurveyUSA samples by using a random-digit telephone-dialing process, but also employs a "robot interviewing process." The robot poll technique, however, is somewhat controversial.

The robot interview begins with an actor who reads the questionnaire, which is recorded and used to interview all respondents. Respondent voice activation technology advances the survey from question to question. Although each respondent is in effect interviewed by the same interviewer, which mitigates interviewer variance, the effects of the recorded or robot interviewer on the survey response is not well understood. SurveyUSA indirectly addresses this issue on their Web page, where they demonstrate that their state-level results closely match the results of other survey firms and that they have a better than average track record in forecasting election outcomes from their data. In a recent paper, Gary Jacobson (2006) extensively tests the SurveyUSA data for reliability and validity and finds them quite comparable to other state-level polls. Thus, despite this reservation, SurveyUSA provides us with unique data to test the impact of state-level presidential approval on state-level election contests in 2006.

SurveyUSA's data were downloaded from their Web site (http://www.surveyusa.com). All SurveyUSA data used in this analysis are based on the state's population, not intended voters. In these state polls, SurveyUSA does not present breakdowns for likely voters.

To measure presidential approval, SurveyUSA uses this question: "Do you approve or disapprove of the job George W. Bush is doing as President?," which is identical to the Gallup job approval question, the basis of so much research. I use the September 2006 reading. To take into account "don't know" responses, I calculate a corrected Bush approval level: (Bush approval)/(Bush approval + Bush disapproval). This proportion is converted into a percentage by multiplying by 100. Across the fifty states, Bush's corrected approval stood at 41 percent, with a standard deviation of 8, a minimum of 23 percent, and a maximum of 61 percent.

Data and Hypotheses

Research on Senate and gubernatorial elections has identified a number of factors that affect those outcomes in those races. (5) Some are easily identified as either national or local, but several of these factors seem to mix effects from both levels. Another issue to bear in mind is that the small number of Senate and gubernatorial races in 2006 limits the number of variables that can be entered into the analysis. The dependent variable here is the Republican percentage of the two-party vote, defined as (Number of votes for the Republican candidate)/(Number of votes for the Republican candidate + Number of votes for the Democratic candidate). This proportion is turned into a percentage by multiplying by 100.

The most obvious local factor is incumbency (see the sources cited above), with the hypothesis that incumbents, everything else equal, will do better in election contests than challengers. However, as the dependent variable used here is the Republican percentage of the two-party vote, we have to identify whether the incumbent is a Democrat or a Republican. We expect that a Republican incumbent should positively affect the Republican vote total, while a Democratic incumbent should lower the vote that the Republican candidate receives in the contest. Thus, I use separate dummy variables indicating whether a Democrat or Republican incumbent was running for reelection. By distinguishing between Democratic and Republican incumbents, the incumbency variable may no longer be a purely local variable, inasmuch as partisanship has the national connotations to voters as discussed above. The criterion category in the analysis is states without an incumbent running.

Economic variables, such as incumbency as defined above, may also contain both national and local elements (Gomez and Wilson 2003; Nadeau and Lewis-Beck 2001; Carsey and Wright 1998a; Lowry, Alt, and Ferree 1998; Svoboda 1995; Niemi, Stanley, and Vogel 1995; Jacobson 1990; Chubb 1988). The literature on election outcomes finds strong effects associated with economic factors; the party in power usually suffers when the economy is doing poorly. Debate in that literature usually focuses on the relative electoral impact of different economic factors. Here I use the state's percent unemployed in September 2006, a measure of the local economy. (6) Yet, like incumbency as defined above, state unemployment is a consequence of national economic conditions and policies, as well as local ones. It is neither a purely national nor local factor.

Local political context may also affect electoral outcomes. In particular, one would expect the Republican party to do better in places where many Republicans live rather than few. The same expectation exists for the distribution of conservatives versus liberals. Thus, I enter the distribution of partisan and ideological identifiers in the state. (7) Repeating the concern raised with regard to the partisan incumbency variables, inasmuch as party and ideological identification in the mass public is a national orientation rather than a local or regional one, this local context takes on a national element as well. Thus far, all the above control variables mix elements of local and national effects.

One variable, such as presidential approval, can be viewed as a purely national variable, although it varies across states: the per-capita number of Iraq War deaths from the state (Gartner, Segura, and Barratt 2004). (8) This is a national variable because it relates directly to a national policy, the Iraq War, a policy over which local office holders have no jurisdiction. The hypothesis here is that, as casualties mount, displeasure with the Republicans will mount, and the vote total for Republican candidates will decline. However, given that governors had no jurisdiction over this policy, whereas senators did, voters may not hold gubernatorial candidates to account for this consequence of that policy.

Finally, I control for the size of the state, a purely local factor. Research has shown that election contests tend to be narrower in larger states than smaller ones, due in part to the relative degree of population homogeneity in small versus large states (Carsey and Wright 1998b; Hibbing and Alford 1990; Lee and Oppenheimer 1997). Again, as the dependent variable is the Republican percentage of the two-party vote, I have to unfold this variable's effects on the vote total. I do this by entering two variables into the analysis, the log of population and the square of the log of population.

Analysis

The multivariate analysis proceeds as follows. First, I present the results of the eight control variables on the two-party vote for senator and governor. Then I add the presidential approval variable to the equation. Comparing across these two equations will allow me to identify the contribution of presidential approval over and above the effects of the other variables. Also in 2006, elections were held for thirty-three Senate and thirty-six governorships, but the analysis here uses only thirty-one of the Senate races because of a complicated three-way race in one state and lack of opposition in another. (9)

Effects of the Control Variables

Turning to the baseline models, that is, the equations without the presidential approval variable, results indicate for both senatorial and gubernatorial elections that the estimations account for a substantial amount of the variance in the two-party vote share, 83 percent in the senatorial case and 67 percent for gubernatorial races. Six of the eight variables are significant predictors in the Senate estimation, while only three emerge as significant predictors in the gubernatorial estimation.

For the Senate contests, presence of a Democratic incumbent, but not a Republican, affects the vote total, as do unemployment, battle deaths, the two population measures, and the percentage of Republican identifiers in the state. The percentage of conservatives in the state does not reach statistical significance, in part because of its strong correlation with the state's Republican percentage (.64, p = .000). In gubernatorial races, both types of incumbency status affect the election totals with the proper signs, as does the percentage of Republicans, but no other variable attains statistical significance. Comparatively, it appears that national factors weigh more heavily in the Senate races than they do in the gubernatorial contests, and I am able to explain with these variables a greater percentage of the variance in Senate than in gubernatorial races.

Presidential Referendum Effects in Multivariate Perspective

Senate Races, My primary question deals with the impact of presidential approval. With these controls, how does presidential approval fare as a predictor of the vote? Results in Table 1 report that, with controls, Bush's approval now does not affect the Senate outcomes, contrary to the significant effects in the bivariate analysis. Removing the two insignificant control variables has no impact on the Bush approval variable, which remains insignificant.

The Iraq War death variable is also a national variable, which, as expected, drags down the vote of Republican Senate candidates. Oddly, the battle death variable is not correlated to Bush's approval (r = .09, p = .51), but results indicate a populace highly responsive to battle deaths emanating from state of residence when voting for senator. Each additional death per 100,000 lessens the Republican vote total by nearly 14 points. States suffered on average 1.1 deaths per 100,000, ranging from a low of 0.55 to a high of 2.89, which produces a vote loss from low- to high-battle-death states of from 7.6 to 40.2 percent. These are massive effects on voting behavior. (10)

How could Republican Senate candidates survive, given such war aversion? Other factors worked in their favor, primarily the percentage of voters who identified with Republicans. Each 1 percent increment in Republican identifiers adds 0.84 percent to the Republican candidate vote total. The average state contains nearly 48 percent Republican identifiers, based upon the total of Democrats and Republicans. At this average level, Republican Senate candidates would garner about 40 percent of the vote as a base. If all other factors canceled out, only states with the highest battle-death totals would defeat the Republican candidate. Still, the extreme sensitivity to battle deaths in the Iraq War clearly cut into Republican vote totals and may have accounted for several Republican defeats and close contests. Thus, while presidential approval per se had little effect on the electoral performance of Republican Senate candidates, the Iraq War did. To a significant extent, the 2006 midterm Senate contests were a referendum on the president's Iraq War policy.

Gubernatorial Races. Matters differ for gubernatorial races, in which national factors have less sway, but still a hint suggests that presidential approval might have had some marginal effect on gubernatorial election outcomes. Before turning to that variable, note that incumbency and a state's percentage of conservative identifiers had the most pronounced effect on these races.

Being an incumbent governor, by these estimates, would help a Republican by 13.5 points. Democratic incumbents similarly saw benefits, but at lower levels, about 7.5 points. For some reason, perhaps because of their relative political moderation, personal appeal, or ability to evade the Iraq War issue, Republican incumbent governors bucked the national tide against their party in the 2006 midterm. Some sense of the crossover appeal of Republican incumbents can be seen in the negative sign on the conservative identification variable. Each 1 percent decrease in conservative identifiers, which here equates with increases in the relative percentage of liberal identifiers, adds nearly 0.5 points to the Republican candidate's vote total. In other words, Republican candidates for governor did quite well in attracting votes as electorates became less conservative and more liberal. (11)

With the full complex of controls in place, the Bush approval variable falls just shy of statistical significance (p = .058 using robust regression). Eliminating the insignificant variables (12) allows the Bush approval variable to become decidedly significant, both statistically and substantively. The p value improves to .023, and the b is a healthy .57, indicating that each 1 percentage point gain in presidential approval is associated with a 0.37 percentage point gain for Republican gubernatorial candidates. This effect is substantial but might not be a boon to Republicans, as Bush's approval level was so weak across so many states. This result indicates that national factors seeped into the gubernatorial contests in 2006, but through a different route than we found for Senate races.

Conclusions

Results presented above show that the 2006 Senate and gubernatorial elections were in part referendums on the president, although the mechanism through which the referendum effects operated differed for the two offices. In the case of Senate elections, presidential approval had little effect on electoral outcomes, but Iraq War battle deaths suffered by a state's citizens did, and by a mighty amount. This is consistent with the idea that state electorates view senators as national policy makers, as their representative to a national policy-making institution, and that state electorates hold their senatorial candidates accountable to some degree for the local implications of national policies, in this case battle deaths. But voters also perceive the war as Bush's war, and in these results can be seen voters holding senators accountable for a presidential policy, too.

Presidential referendum effects in the 2006 gubernatorial races worked both more directly and diffusely than in the Senate case. Here the level of presidential approval in a state had clear implications for Republican candidates, who saw higher vote totals when the president was popular rather than unpopular. To a degree, voters may tie all members of the same party to the president in office, rewarding or punishing that party's candidates based upon assessment of the president's job performance. Such a reaction makes some sense in a highly polarized partisan context.

At the mass level, partisanship, and its strong covariate, ideological orientation, strongly predict attitudes toward the president. Jointly, with these SurveyUSA data, over 80 percent of the state-level variance in presidential approval is a function of the relative percentage of Republicans and conservatives in a state. (13) Thus, there is some penetration of presidential politics into gubernatorial elections through these mechanisms.

But governors are not completely captured by presidential-partisan-ideological effects. The Iraq War did not play a direct role in gubernatorial campaigns. Candidates for that office did not have to speak to the issue openly, because the governor's office possesses no policy responsibility for the war. And by not being in Washington and not having to go on record for or against the president's policies through roll-call voting and the like, Republican gubernatorial candidates could evade direct association with the president, other than the fact of being in the same party. Plus, they could stake out more moderate policies in an attempt to attract moderates, liberals, Independents, and Democrats, a strategy made necessary in blue states, where a surprisingly large number of Republicans hold the executive mansion.

References

Abramowitz, Alan I. 1975. Name familiarity, reputation, and the incumbency effect in a congressional election. Western Political Quarterly 28(4): 668-84.

Adams, Greg D., and Peverill Squire. 1997. Incumbent vulnerability and challenger emergence in Senate elections. Political Behavior 19(2): 97-111.

Alford, John R., and John R. Hibbing. 1981. Increased incumbency advantage in the House. Journal of Politics 43(4): 1042-61.

Ansolabehere, Stephen, James M. Snyder, Jr., and Charles Stewart III. 2000. Old voters, new voters, and the personal vote: Using redistricting to measure the incumbency advantage. American Journal of Political Science 44(1): 17-34.

Atkeson, Lonna Rae, and Randall W. Partin. 1995. Economic and referendum voting: A comparison of gubernatorial and senatorial elections. American Political Science Review 89(1): 99-107.

--. 1998. Economic and referendum voting and the problem of data choice: A reply. American Journal of Political Science 42(3): 1003-07.

Bartels, Larry M. 2000. Partisanship and voting behavior, 1952-1996. American Journal of Political Science 44(1): 35-50.

Bond, Jon R., and Richard Fleisher. 2001. The polls: Partisanship and presidential performance evaluations. Presidential Studies Quarterly 31(3): 529-40.

Cain, Bruce E., John A. Ferejohn, and Morris P. Fiorina. 1984. The constituency service basis of the personal vote for U.S. representatives and British members of Parliament. American Political Science Review 78(1): 110-25.

Campbell, James E. 1983. The return of the incumbents: The nature of the incumbency advantage. Western Political Quarterly 36(3): 434-44.

Carsey, Thomas M., and Gerald C. Wright. 1998a. State and national factors in gubernatorial and senatorial elections. American Journal of Political Science 42(3): 994-1002.

--. 1998b. State and national factors in gubernatorial and senatorial elections: A rejoinder. American Journal of Political Science 42(3): 1008-11.

Chubb, John E. 1988. Institutions, the economy, and the dynamics of state elections. American Political Science Review 82(1): 133-54.

Cohen, Jeffrey E. 2006. The polls: The coalitional president from a public opinion perspective. Presidential Studies Quarterly 36(3): 541-50.

Cover, Albert D. 1977. One good term deserves another: The advantage of incumbency in congressional elections. American Journal of Political Science 21 (3): 523-41.

--. 1980. Contacting congressional constituents: Some patterns of perquisite use. American Journal of Political Science 24(1): 125-35.

--. 1986. Presidential evaluations and voting for Congress. American Journal of Political Science 30(4): 786-801.

Cover, Albert D., and Bruce S. Brumberg. 1982. Baby books and ballots: The impact of congressional mail on constituent opinion. American Political Science Review 76(2): 347-59.

Cox, Gary W., and Jonathan N. Katz. 1996. Why did the incumbency advantage in U.S. House elections grow? American Journal of Political Science 40(2): 478-97.

Dawes, Roy A., and A. Hunter Bacot. 1998. Electoral career patterns and incumbency advantage in the U.S. House of Representatives* Legislative Studies Quarterly 23(4): 575-83.

Erikson, Robert S. 1971. The advantage of incumbency in congressional elections. Polity 3(3): 395-405.

Erikson, Robert S., and Thomas R. Palfrey. 1998. Campaign spending and incumbency: An alternative simultaneous equations approach. Journal of Polities 60(2): 355-73.

Garand, James C., and Donald A. Gross. 1984. Changes in the vote margins for congressional candidates: A specification of historical trends. American Political Science Review 78(1): 17-30.

Gartner, Scott Sigmund, Gary M. Segura, and Bethany A. Barratt. 2004. War casualties, policy positions, and the fate of legislators. Political Research Quarterly 57(3): 467-77.

Gomez, Brad T., and J. Matthew Wilson. 2003. Causal attribution and economic voting in American congressional elections. Political Research Quarterly 56(3): 271-82.

Gronke, Paul, Jeffrey Koch, and J. Matthew Wilson. 2003. Follow the leader? Presidential approval, presidential support, and representatives' electoral fortunes. Journal of Politics 65(3): 785-808.

Gross, Donald A., and James C. Garand. 1984. The vanishing marginals, 1824-1980. Journal of Politics 46(1): 224-37.

Herrera, Richard, and Michael Yawn. 1999. The emergence of the personal vote. Journal of Politics 61(1): 136-50.

Hetherington, Marc J. 2001. Resurgent mass partisanship: The role of elite polarization. American Political Science Review 95 (3): 619-31.

Hibbing, John R., and John R. Alford. 1990. Constituency population and representation in the U.S. Senate. Legislative Studies Quarterly 15 (4): 581-98.

Highton, Benjamin. 2000. Senate elections in the United States, 1920-94. British Journal of Political Science 30(3): 483-506.

Jacobson, Gary C. 1990. Does the economy matter in midterm elections? American Journal of Political Science 34(2): 400-404.

--. 2005. Polarized politics and the 2004 congressional and presidential elections. Political Science Quarterly 120(2): 199-218.

--. 2006. The polls: Polarized opinion in the states: Partisan differences in approval ratings of governors, senators, and George W. Bush. Presidential Studies Quarterly 36(4): 732-57.

King, James D. 2001. Incumbent popularity and vote choice in gubernatorial elections. Journal of Politics 63(2): 585-97.

King, Gary, and Andrew Gelman. 1991. Systemic consequences of incumbency advantage in U.S. House elections. American Journal of Political Science 35(1): 110-38.

Kone, Susan L., and Richard E Winters. 1993. Taxes and voting: Electoral retribution in the American states. Journal of Politics 55(1): 22-40.

Lee, Frances E., and Bruce I. Oppenheimer. 1997. Senate apportionment: Competitiveness and partisan advantage. Legislative Studies Quarterly 22(1): 3-24.

Levitt, Steven D., and Catherine D. Wolfram. 1997. Decomposing the sources of incumbency advantage in the U.S. House. Legislative Studies Quarterly 22(1): 45-60.

Leyden, Kevin M., and Stephen A. Borrelli. 1995. The effect of state economic conditions on gubernatorial elections: Does unified government make a difference? Political Research Quarterly 48(2): 275-90.

Lockerbie, Brad. 1999. The partisan component to the incumbency advantage: 1956-1996. Political Research Quarterly 52(3): 631-46.

Lowry, Robert C., James E. Alt, and Karen E. Ferree. 1998. Fiscal policy outcomes and electoral accountability in American states. American Political Science Review 92(4): 759-74.

Marra, Robin E, and Charles W. Ostrom, Jr. 1989. Explaining seat change in the U.S. House of Representatives, 1950-86. American Journal of Political Science 33(3): 541-69.

Nadeau, Richard, and Michael S. Lewis-Beck. 2001. National economic voting in U.S. presidential elections. Journal of Politics 63(1): 159-81.

Niemi, Richard G., Harold W. Stanley, and Ronald J. Vogel. 1995. State economies and state taxes: Do voters hold governors accountable? American Journal of Political Science 39(4): 936-57.

Rudalevige, Andrew. 2001. Revisiting midterm loss. Referendum theory and [US] state data. American Politics Research 29(1): 25-46.

Simon, Dennis M. 1989. Presidents, governors, and electoral accountability. Journal of Polities 51(2): 286-304.

Simon, Dennis M., Charles W. Ostrom, Jr., and Robin E Marra. 1991. The president, referendum voting, and subnational elections in the United States. American Political Science Review 85(4): 1177-92.

Squire, Peverill, and Christina Fastnow. 1994. Comparing gubernatorial and senatorial elections. Political Research Quarterly 47(3): 705-20.

Stein, Robert M. 1990. Economic voting for governor and U.S. senator: The electoral consequences of federalism. Journal of Politics 52(1): 29-53.

Svoboda, Craig J. 1995. Retrospective voting in gubernatorial elections: 1982 and 1986. Political Research Quarterly 48(1): 135-50.

Swain, John W., Stephen A. Borrelli, and Brian C. Reed. 1998. Partisan consequences of the post-1990 redistricting for the U.S. House of Representatives. Political Research Quarterly 51(4): 945-67.

(1.) These are the percent approval ratings from national surveys in the first week of November 2006: FOX News/Opinion Dynamics, 38 percent; CNN, 35 percent; USA Today/Gallup, 38 percent; ABC News/Washington Post, 40 percent; Pew, 41 percent; Newsweek, 35 percent; and Time, 37 percent.

(2.) The total adds to fifty-two because two states, New Hampshire and Vermont, have a two-year term of office for their governors.

(3.) A considerable literature has looked at presidential referendum effects in subpresidential elections. See Atkeson and Partin (1998); Simon, Ostrom, and Marra (1991); Marra and Ostrom (1989); Cover (1986); Rudalevige (2001); Gronke, Koch, and Wilson (2003); and Carsey and Wright (1998a).

(4.) While loss of several states might not compromise analysis much if the missing states were not selected for random reasons, such is not the case for the 2006 Edison-Mitofsky exit poll. States were not polled because of lack of interest, for instance, states lacking close contests or important personalities running for office. The Edison-Mitofsky 2006 exit poll results can be found on several news media Web sites. I used CNN's site, http://www.cnn.com/ELECTION/2OO6/pages/results/states/US/H/00/ epolls.0.html.

(5.) The literature on senatorial and gubernatorial elections is huge. For representative analyses on the Senate, see Highton (2000), Carsey and Wright (1998a), Adams and Squire (1997), and Atkeson and Partin (1995). On gubernatorial elections, see King (2001); Carsey and Wright (1998a); Leyden and Borrelli (1995); Svoboda (1995); Niemi, Stanley, and Vogel (1995); Atkeson and Partin (1995); Squire and Fasmow (1994); Kone and Winters (1993); Stein (1990); and Simon (1989).

(6.) The September figures are used because they were the latest that were available at the time of the election. We also experimented with the change in the unemployment level from August to September 2006, but results found that the variable always failed to reach statistical significance and thus is deleted from the presentation.

(7.) These estimates come from the SurveyUSA September 2006 polls.

(8.) These data come from the Iraq Coalition Casualty Count available from http://www.icasualties.org/oif/default.aspx. Unfortunately, the site only presents battle death totals, but not wounded totals, by state.

(9.) In Connecticut, incumbent Joe Lieberman ran as an Independent, having lost the Democratic primary to Ned Lamont. The Republican senatorial candidate, Alan Schlesinger, was never a viable candidate in this three-way race. The dependent variable, the Republican share of the two-party vote, makes little sense here. Schlesinger garnered a mere 10 percent of the vote in the general election, receiving little election support from the Republican establishment, which tacitly, if not overtly, supported Lieberman's candidacy. The Republican establishment, including the Bush administration, figured that Lieberman had a better chance of winning than Schlesinger and that he would continue to support the administration's Iraq policy. And in the Indiana Senate race, Republican incumbent Richard Lugar did not face a Democratic challenger and received only token opposition from Libertarian candidate Steve Osborn. Thus, Connecticut and Indiana are not included in the Senate analysis.

(10.) Although I can only speculate here as to why the effect is so large based upon the battle death figures, it is likely that the number of those wounded and the number killed in battle are highly correlated at the state level. Over twenty thousand U.S. military personnel had been wounded by the time of the midterm elections, about seven times the rate of battle deaths. The public may be reacting to the total number of those killed and wounded. If so, the regression results would be the same as presented here, but the conversion metric of casualties to votes would differ. In either case, the results demonstrate a public highly sensitive to casualty rates.

(11.) Even when the state's Republican percentage is dropped from the estimation, this negative sign remains intact.

(12.) Now the percentage of conservatives becomes significant, while the formerly significant variable, the percentage of Republicans, fails into statistical insignificance. The high collinearity between the two is the underlying reason for this switch.

(13.) The actual regression results are Corrected Bush approval = -.05 + .48 (Conservative %, p = .000) + .34 (Republican %, p = .000), [R.sup.2] = .82. Recall, the Conservative % is calculated as (Conservative %)/(Conservative % + Liberal %), and Republican % equals (Republican %)/ (Republican % + Democratic %). No other variable used in this analysis has a significant effect on Bush's approval.

JEFFREY E. COHEN

Fordham University

Jeffrey E. Cohen, professor of political science at Fordham University, is the author of several books, including Presidential Responsiveness and Public Policy-Making, as well as articles in numerous journals, including the American Political Science Review. American Journal of Political Science, and Journal of Politics.

AUTHOR'S NOTE: I want to thank Costas Panagopoulos for his helpful comments on an earlier version of this feature.

TABLE 1

Robust Regression Effects of Corrected Bush Approval on the Republican
Percentage of the Two-Party Vote in Senate and Gubernatorial Races,
2006
 Senate

 b P* b P *

Republican incumbent 3.83 .23 3.55 .25
Democratic incumbent -11.25 .01 -11.39 .01
Unemployment 3.67 .003 3.97 .001
Iraq deaths per capita -13.92 .001 -13.55 .003
Ln stare population -63.64 .001 -66.17 .001
Ln state population-sq. 1.99 .002 2.07 .001
Republican % .84 .001 .81 .01
Conservative % -.22 .12 -.35 .11
Bush approval xx .21 .35
Constant 527.40 .001 544.73 .001
[R.sup.2] .83 .83
n 31 31

 Governor

 h P * b P *

Republican incumbent 14.22 .000 13.5 .000
Democratic incumbent -7.66 .01 -7.47 .01
Unemployment .37 .39 .97 .21
Iraq deaths per capita 3.82 .17 5.68 .09
Ln stare population 32.07 .20 35.1 .19
Ln state population-sq. -1.03 .21 -1.11 .20
Republican % 0.4 .02 .15 .28
Conservative % -0.23 .08 -.49 .01
Bush approval xx .64 .06
Constant -215.78 .24 -244.97 .21
[R.sup.2] .67 .70
n 36 36

* One-tailed test.

Source: SurveyUSA, midterm election results, and data collected by
author. See text for details.