Balancing in the States,

Balancing in the States, 1978-2009 Michael A. Bailey Department of Government & Public Policy Institute Georgetown University Intercultural Center 681 Washington, DC 20057 (202) 687-6021 baileyma@georgetown.edu Elliott B. Fullmer Department of Government Georgetown University Intercultural Center 681 Washington, DC 20057 (202) 687-6130 elliott.fullmer@gmail.com AUTHOR INFORMATION: Michael A. Bailey is the Colonel William J. Walsh Professor of American Government in the Georgetown University Department of Government and the Georgetown Institute of Public Policy. His research covers Congress and the Supreme Court, methodology, welfare, trade, and inter-state policy competition. Elliott B. Fullmer is completing his Ph.D. in American Government at Georgetown University. His research interests include state politics, voting systems, congressional behavior, and the presidential nomination process. His dissertation examines the effects of early and absentee voting laws in the U.S. 1

Abstract Since the Civil War, the president s party has lost seats in the House of Representatives in all but three midterm election cycles. Many attribute this pattern to balancing by moderate voters who prefer a Democratic Congress when Republicans control the White House, and vice versa (Fiorina 1988, 1992, 1996; Alesina and Rosenthal 1995). Although a number of scholars have tested the balancing hypothesis, the debate remains unsettled. We argue that states provide an excellent way to further analyze the issue. Similar to the national government, states feature an executive and (in most cases) a bicameral legislature which comes up for election in both gubernatorial and state midterm years. If voters balance, we ought to observe such behavior in state elections when an executive s partisanship is known and a legislative choice is necessary. We examine state legislative elections from 1978-2009 and find evidence consistent with the balancing hypothesis. 2

One of the clearest patterns in U.S. politics is that the president s party fares poorly in midterm elections. In fact, since the Civil War the president has lost seats in all but three midterm cycles (1934, 1998, and 2002). Some scholars explain this phenomenon in terms of balancing, the idea that voters balance a president of one party by voting for members of Congress from the other party. In this view, voters expect that a Congress controlled by the opposite party of the president will moderate policies that emerge from the legislative process. For all the plausibility of the balancing hypothesis, however, doubts remain. The president s party has picked up seats in two of the last three midterms. In addition, research on the issue has yielded differing conclusions (Beck et. al 1992; Sigelman, Wahlbeck, and Buell 1997; Burden and Kimball 1998; Roscoe 2003; Lewis-Beck and Nadeau 2004). In this paper, we explore the issue in a new way. The balancing hypothesis is not specifically about Congress or the presidency, but rather about voters and the interaction of their choices in legislative and executive branch elections. Therefore, we examine state elections in order to vastly increase the sample size of cases in which voters have a chance to balance. We focus on state legislative elections during gubernatorial midterm years, as the balancing hypothesis suggests that voters should reward the party opposite of that of the sitting governor in such cases. We assess legislative outcomes in each chamber of U.S. states from 1978 to 2009 and find evidence that voters balance. In most specifications, we find that Democrats in a state legislature are at a disadvantage in midterm elections when there is an incumbent Democratic governor and vice versa. The balancing effect is typically around two or three percent of a chamber, but in some reasonable specifications is as large as a thirteen or fourteen percentage point difference in the proportion of a legislature that is Democratic. 3

Balancing: the evidence so far The balancing hypothesis was first introduced and tested as an explanation for split-ticket voting (supporting one party for president and another for Congress) during presidential election cycles. Fiorina (1988, 1992, and 1996) argued that many moderates engage in split-ticket voting in presidential election years out of a desire to have divided government. Slightly right-of-center voters who are inclined to support Republican presidential candidates may prefer a Democratic House member because a divided government headed by a Republican chief-executive will appear to represent them better than a unified Republican one. Conversely, slightly left-of-center voters who are inclined to support Democratic presidential candidates may prefer a Republican House member because a divided government headed by a Democratic chief-executive will appear to represent them better than a unified Democratic one. Analyzing the 1984 and 1988 presidential elections, Fiorina found some support for his hypothesis and argues that while most voters do not consciously balance, a preference for divided government influences the way a small segment of voters perceive candidates and ultimately make voting decisions. Additional evidence on this question has come in one of three forms: studies of ticket splitting, survey experiments and analysis of national election results. The most expansive literature addresses ticket splitting. This literature typically assesses the degree to which selfreported ideology and perceptions of the ideological positions of the two major parties relate to split-ticket voting. Those who perceive the parties as highly polarized and themselves as fairly moderate, the argument goes, should be more supportive of divided government and more likely to split their presidential and congressional tickets on Election Day. While some have found support for this idea through survey research (Smith et. al 1999; Lewis-Beck and Nadeau 2004; 4

Saunders et. al 2005), most have reported little or no evidence of such behavior (Beck et. al 1992; McAllister and Darcy 1992; Alvarez and Schousen 1993; Born 1994, 2000; Petrocik and Doherty 1996; Sigelman, Wahlbeck, and Buell 1997; Lacy 1998; Burden and Kimball 1998; Roscoe 2003). There have also been a number of survey analyses that focus more directly on balancing. Here again, we see no consensus as to whether voters balance. On the one hand are several papers that support the idea. Carsey and Layman (2004), for example, argue that the aforementioned ticket splitting findings should not represent the death knell for the balancing hypothesis. It is quite possible in their view that voters may possess balancing preferences and still engage in straight-ticket, rather than split-ticket, voting. In 1996, for example, most voters arguably knew that incumbent Bill Clinton was likely to be reelected. Interested in balancing the government, slightly right-of-center voters would be expected to vote for a Republican member of Congress. Knowing that Clinton would win, however, they could also vote their conscience for president by casting their respective ballots for Dole. Conversely, some may engage in either split or straight-ticket voting for reasons other than balancing. Carsey and Layman therefore focus simply on whether voters ideological locations and their perceptions of the two parties influence voter attitudes regarding partisan or divided control of government. The authors conduct a telephone survey of Illinois adults on the eve of the 2000 presidential election. In the survey, they ask If [George W.] Bush is elected president, which type of Congress would you prefer? They later ask the same question, substituting Bush for Vice-President Al Gore, the Democratic nominee. Carsey and Layman find that 17% of the sample supports a divided government under all conditions. These respondents prefer a Democratic Congress if Bush were elected, but a Republican one if Gore won. 5

Lacy and Paolino (1998) reach a similar conclusion in their analysis. Echoing Downs (1957), they find that presidential vote choices depend more on the perceived distance between voter ideal points and expected policy outcomes under each candidate than on the distance between voter ideal points and candidate platforms. The authors analyze a 1996 Texas poll in which voters were asked to identify the positions of Bill Clinton and Bob Dole, along with the overall ideological position of the government if either was elected. While many respondents found Clinton to be left-of-center, a sizable number perceived that the government (given their expectation that Republicans would control Congress) would be decidedly more moderate under him than Dole. This expectation was found to aid Clinton s support level with many voters closer to Dole on the ideological spectrum. On the other hand, several papers find mixed or no survey evidence for balancing. Saunders et. al (2005) test the degree to which balancing is more prevalent in years featuring close, or unpredictable, presidential elections and find mixed evidence for balancing only in 1996. Geer et. al (2004) expose a sample of undergraduates to varying polling data concerning the upcoming congressional and presidential races on the eve of the 2000 presidential election. This allowed them to manipulate expectations concerning the victors of the looming elections. The authors did not find those who believed Democrats would control the White House to be more eager to elect a Republican Congress, or vice versa. The third approach to assessing the balancing hypothesis is to look at election outcomes. Alesina and Rosenthal (1995) argue that balancing is more likely to occur during midterms, as voters no longer face uncertainty regarding control of the executive branch. In presidential election years, voters are (at least somewhat) uncertain who will control the presidency and Congress, forcing moderate voters and leaners to hedge their bets by supporting the party 6

closest to their preferences in both races. In midterm elections, this uncertainty disappears, and voters can act to balance the federal government. This, Alesina and Rosenthal predict, should lead to losses for the president s party in midterm cycles. They study district-level election results in midterm elections from 1950 to 1986 (excluding years just after redistricting) and find that With a Democratic president at midterm, Republicans record a higher two-party vote share than they did in the previous on-year election (when the presidency was uncertain) in a large majority of congressional districts. With a Republican president at midterm, Democrats record a higher two-party vote share than they did in the previous on-year election (when the presidency was uncertain) in a large majority of congressional districts. The authors also provide an additional analysis in which they control for incumbency effects. They find that voters balance in two primary ways: 1) voters send the president s party a signal that the electorate wishes to moderate policies by diminishing the vote shares of some incumbents and 2) voters defeat the president s party in marginal open seat races, or those seats in which a party won with less than 55% of the vote share in the previous contest. The very plausibility of these results we know, of course, that the president s party tends to lose during midterms elections is also one of the idea s limitations. The hypothesis was derived in part based on the observation of midterm losses by the president s party. How can we be sure that balancing is not simply an ex-post rationalization of a well-known behavioral pattern? Balancing: a state-level analysis 7

States provide an ideal opportunity for assessing the balancing hypothesis as they feature voters making choices in separate elections for legislative and executive branch officeholders. Examining balancing in the states therefore provides a new venue for testing the hypothesis a venue that was not the source of the idea. In addition, states provide dramatically more observations and variation on the circumstances of interest. At the federal level, for any given election, the balancing hypothesis predicts movement in the same direction (away from the president s party in a midterm); at the state level, we can have numerous states with Democratic incumbents (where we would expect balancing to advantage Republicans in the legislature), other states with Republican incumbents (where Democrats would be advantaged in the legislature) and states with no incumbent (where neither party should have an advantage legislatively). This provides a much richer environment for testing the idea. Model With panel data, there are a number of different specifications that have been used in the literature. In particular, the literature has models with and without lagged values of dependent variables, with and without differencing of the dependent variable and with and without fixed effects. Our approach is to estimate models across the range of specification approaches; fortunately, the effects are generally consistent across approaches. One question in the literature is whether or not we should include a lagged dependent variable in the model. Oppenheimer, Stimson and Waterman (1986) argue that a lagged dependent variable is very important to include in a model of one party s legislative results, as this variable captures the exposure of the party. When the party has more seats than usual, it likely includes a number of marginal seats that it is more likely to lose in the future. 8

On the other hand, Achen (2001) counsels caution when considering inclusion of lagged dependent variables. He derives the asymptotic bias of coefficients on the lagged dependent variable and other variables and shows that there can be considerable bias when there is serial autocorrelation and trending in the independent variable. In this case, the lagged dependent variable does not conduct itself like a decent, well-behaved proxy. Instead it is a kleptomaniac, picking up the effect, not only of excluded variables, but also of the included variables if they are sufficiently trended. As a result, the impact of the included substantive variables is reduced, sometimes to insignificance (Achen 2001). Our approach on this question is to report both types of models. In the models without the lagged dependent variable, we control for first-order autocorrelation. As we later discuss, including a lagged dependent variable seldom affects the results, but it does occasionally reduce both statistical significance and the magnitude of coefficients. A second question is whether or not to difference the data. Differenced data in a panel context is the results for a given year and state minus the previous results in that state. The alternative is a level model in which the dependent variable is simply the result for a given year. These models are absolutely equivalent when a lagged dependent variable is included (Allison 1990), so there is no need to concern ourselves with this question for those specifications. When the lagged dependent variable is not included, both differenced and level models are unbiased and consistent under standard assumptions; the models do differ in their robustness to certain assumption violations and we follow Allison s (1990) preference for differenced models (see also Wooldridge 2009). Another specification issue is whether or not to include fixed effects. We generally believe inclusion of these controls is prudent as they capture two classes of unmeasured factors that 9

could be correlated with the included variables and cause bias. First, we include year fixed effects; these effects control for all factors common to a given year across all states, including the national economy, national scandals or simply national political moods that may make some years better or worse for Democrats. Second, we include state-level fixed effects. These account for any influence on legislative outcomes that is fixed for a state across our time frame. For example, the political culture of a state may make it consistently more Democratic than one would predict based only on our measured variables. However, some argue that fixed effects have some kleptomaniac tendencies as well (Beck and Katz 2001). Hence, we also report results without fixed effects in order to provide a sense of how dependent our results are on the use of these models. In our study, we use data on the partisan composition of state legislatures before and after elections since 1978 (The Book of the States 1978-2009; Klarner 2009). Our core model is DemLeg st (DemGov 0 X 4 st 1 st st x GovMT ) (RepGov x GovMT ) (DemGov st 2 st st 3 st x NotGovMT ) st where DemLeg is the percentage point change in the percent of a legislative chamber that is Democratic, DemGov is a dummy variable for Democratic incumbent governors, GovMT is a dummy variable for a gubernatorial midterm election, RepGov is a dummy variable for Republican incumbent governors and X is a vector of control variables, described below. The excluded category in our formulation is a non-gubernatorial midterm with a Republican incumbent. We omit elections which take place while an independent is serving as governor. This is not a large enough number to merit another set of variables, and there is no obvious reason to include independents with one party or the other. The balancing hypothesis predicts that Democrats in legislatures do worse when there is a sitting Democratic governor ( 1 < 0 ) and better when there is a sitting Republican governor ( 2 > 10

0). Or, most precisely, we predict that 2 > 1. This means that we test for balancing not by interpreting the coefficients themselves, but rather their differences. In our tables reported below, the test of balancing will be an F-test reported at the bottom of each table. Following Alesina and Rosenthal (1995), we focus on state midterm races because voters have certainty regarding partisan control of the executive branch. We include several control variables in order to account for other influences on state legislative composition. First, we account for possible surge and decline trends in the electorate. This idea, advocated by Campbell and others, contends that the electorate differs between presidential and midterm election cycles (A. Campbell 1966; J. E. Campbell 1985, 1987, 1991, 1997; Born 1990). Presidential cycles are high-stimulus affairs and feature high levels of campaign activity and voter turnout. Conversely, national midterm cycles involve low-stimulus elections, and thus produce less interest and turnout (Campbell 1997). In midterm years, the electorate is comprised of core voters, or those highly interested in politics. In presidential years, however, these core voters are joined by peripheral voters, who are stimulated by the perceived importance and media attention inherent in presidential races. While core voters are likely to have stable partisan feelings and voting patterns, peripheral voters are often guided by short-term forces in particular campaigns. In presidential election years, the victorious candidate is elected largely with the help of peripheral voters, who in the process help members of his party perform well in congressional races. Two years later, in midterm cycles, these peripheral voters are absent and the core voters determine election outcomes. This inevitably brings losses for the president s party, as it is no longer helped by the tide of peripheral voters that supported it in the previous election (Campbell 1997). 11

If surge and decline indeed affects elections, we should expect it to impact state legislative races which are held concurrently with national races as well. In presidential election years, we expect to see a surge when the Democratic candidate does well: the in-state performance of the presidential candidate should positively affect the performance of state Democratic legislative candidates. In non-presidential election years in which the previous state legislative election was in a presidential year, the success of a Democratic presidential candidate in the previous election should hurt the party in the current election. Therefore, we include the Democratic share of the two-party presidential vote at the state level for the current year in presidential election years. In non-presidential years in which the previous state legislative election occurred during a presidential election, we include the Democratic share of the two-party presidential vote in the previous election. We control for state economic conditions with a measure of change in state unemployment. We use unemployment because it is highly politically salient and data on it is available for the entire span of years we cover (from the Bureau of Labor Statistics). We created two variables unemployment change x Democratic governor and unemployment change x Republican governor so that unemployment increases would be allowed to have different effects depending on which party controlled the governorship. We prefer using change variables as we feel they capture the mood of the electorate; for example, while the level of unemployment was higher in 1984 than in 2008, the trend was better in 1984 and we believe the politics of the times responded accordingly. Using levels of unemployment instead of changes, however, has little impact on the results. After we present a number of specifications based on the core model, we also report specifications with additional variables. The limitation of these variables is that they are not 12

available for the entire time period of our analysis. One factor that could be very important is that not all governors are created equal. Some are very popular; some are unmitigated disasters. Hence, the extent to which voters feel the need to balance may depend on the popularity of the governor. Certainly the evidence is strong that executive popularity influences the executive s party s performance in legislative elections, both at the federal (Arcelus and Meltzer 1975; Bloom and Price 1975; Tufte 1975; Kernell 1977; Fiorina 1981; Kiewiet 1983; Abramowitz 1985) and state-level (King 2001). We therefore include a measure of gubernatorial approval in the weeks preceding an election (Beyle et. al 2009). Because this polling data is not universally available, there is a considerable drop off in sample size when we include these data, causing a concomitant drop in statistical power. Changes in state partisanship and ideology could also explain legislative elections. We already have a number of controls that are related to such changes, but we also report results in which we include changes in state net Democratic identifiers (Democrats minus Republicans) and state net liberals (liberals minus conservatives) using data from Wright, McIver and Erikson (2004). This data is based on an aggregation of CBS News/New York Times national polls to create large state-level samples. The data is only available through 2003, again causing a significant drop off in sample size when included in the model. Finally, redistricting could conceivably affect legislative outcomes and be correlated with our main variable of interest, gubernatorial control. We therefore use data from Michael McDonald on partisan control of the redistricting process (see Bailey and McDonald 2006). There are two variables one is an indicator for when Democrats controlled the most recent redistricting process; the other is an indicator for when Republicans controlled the most recent redistricting process. Often these are both zero as neither party had complete control over the process. This 13

data is not available for all years, again causing a significant drop off in sample size when these variables are included. Results Table 1 presents the results for the main model for state assemblies from 1978 to 2009. Column 1 presents an extremely sparse model; the idea here is to see the relationship in its rawest form, relatively free of modeling assumptions. The F-test lines at the bottom indicate that the Democratic share of state assemblies is 2.6 percentage points lower when there is a midterm and a Democratic governor compared to when there is a midterm and a Republican governor. This difference is highly statistically significant (p = 0.001). <TABLE 1 ABOUT HERE> Of course, we wish to know if this result is robust to inclusion of controls for other determinants of state assembly elections. The next columns add increasingly more comprehensive controls. Column 2 adds variables for state unemployment, state-level presidential results and state and year fixed effects; the balancing effect is smaller, but the results are generally similar to column 1. Column 3 controls for first order autocorrelation. Some data are lost in the process, but the results do not change appreciably. Adding exposure in column 4 lowers the effect size to 1.4 percentage points and it remains statistically significant. <TABLE 2 ABOUT HERE> Table 2 presents results for the same specifications for state senates. The results are quite similar as for state assemblies. The sparse model in column 1 indicates that the difference between Democratic and Republican governors in gubernatorial midterms is about 2.4 percentage points and this difference is highly statistically significant (p = 0.002). Adding 14

independent variables and fixed effects in column 2 moderates the results a bit, but they are still significant. Correcting for autocorrelation in column 3 maintains an effect of 2.2 percentage points, despite a drop in sample size. As with the assembly results, the effect drops when the exposure variable is added in column 4, but it remains statistically significant. The results so far support the idea that voters balance, albeit somewhat modestly as the effect size is somewhere in the neighborhood of one to three percentage points. Translating this into terms of the U.S. House of Representatives, this would mean that a Democratic president would, all else equal, expect to lose between 4 and 13 House seats in a midterm. Applying the results to the context of an average-sized state assembly (about 110 seats), a Democratic governor would lose 1 to 3 seats. The results are statistically significant, but not breathtaking. We now wish to see if the results are robust to alternative specifications and if there are conditions under which balancing effects seem to be larger or smaller. In Table 3 we add additional controls for state assemblies. Columns 1 and 2 report results when gubernatorial approval ratings are included. The magnitude of the effect in column 1 (4.7 percentage points) is on the high end of anything we have seen. This effect is significant, as well. While the sample changes with the inclusion of the governor approval data, it appears that the changes in results are due to including the approval data; running the model on the 384 observations for which we have governor approval data, but not including the approval data (a model similar to column 3 of Table 1), yields a balancing effect of about 2 percentage points, nearly half of what we see in column 1. Column 2 adds a lagged dependent variable and the balancing effect is no longer significant. Although the results are often similar with and without a lagged dependent variable, when they do differ, they differ in this way: the balancing effect is smaller and sometimes insignificant when the lagged dependent variable is included. 15

In columns 3 and 4, we report results in which we include state-level partisanship and ideology variables. This sample has cut off all years after 2003 due to data availability, but it does not seem to change the results from what we saw in Table 1. The partisanship and ideology variables are not very significant, implying they do not explain legislative results over and above the other controls already in the model. <TABLE 3 ABOUT HERE> Columns 5 and 6 of Table 3 report models in which redistricting variables are included. The redistricting variables are available for only a subset of observations used in Table 1, but even with that reduction in sample size, we observe similar results as before. The balancing effect is between 2.5 and 3 percentage points and significant. <TABLE 4 ABOUT HERE> In Table 4 we conduct similar exercises for state senates. Columns 1 and 2 report results when gubernatorial approval ratings are included. The pattern is quite similar to what happened with state assemblies: the balancing effect is quite large in column 1 -- 5.9 percentage points and statistically significant. Again, it appears the gubernatorial variables are doing the work here, not the change in sample: running the model on data for which we have gubernatorial approval data but not including it yields a balancing effect of 1.5 percentage points, something similar to what we saw in Table 2. The balancing effect is relatively large in column 2, which adds a lagged a dependent variable, but is statistically insignificant. This is the same pattern we saw for state assemblies. Columns 3 and 4 report results in which we include state-level partisanship and ideology variables. The results are consistent with previous analyses: the balancing effect is in the 2 to 3 percentage point range and statistically significant. The partisanship and ideology variables are 16

insignificant, again implying they do not explain legislative results over and above the other controls already in the model. Columns 5 and 6 of Table 4 report models in which redistricting variables are included. Including these variables drops the sample size considerably. Column 5 reports a balancing effect of 2 percentage points, but it is not significant at conventional levels (p= 0.116); losing about half the data seems the most reasonable explanation here as the results are similar if we run the model on this subset of the data and do not include the redistricting data. In column 6, we include a lagged dependent variable and the balancing result completely disappears. Finally, we explore conditions under which balancing may be stronger or weaker. One possibility is that balancing at the state level may be stronger in non-presidential years as the intensity of a presidential election may so occupy voters that they pay less attention to potential benefits of balancing. Therefore, Table 5 presents results where the sample is limited to nonpresidential years. The first three columns present results for assemblies. The results are indeed stronger. Column 1 presents a model without lagged dependent variables (it is equivalent to column 3 of Table 1). The balancing effect is 7.2 percentage points (compared to 1.7 percentage points when both presidential and non-presidential years are included). Given the strong effect gubernatorial approval had on results earlier, we add approval data to the model in Column 2; the balancing effect shoots up to 13 percentage points. Column 3 includes a lagged dependent variable (making is comparable to column 4 of Table 1) and the effect falls, as it has generally done when adding lagged dependent variables, but the effect remains very large (9.4 percentage points) and is highly significant. <TABLE 5 ABOUT HERE> 17

We run the same specifications for state senates in columns 4 through 6, running a series of models based only on non-presidential election years. The balancing effect of 3.5 percentage points in column 4 is large (but not as large as for assemblies) and statistically significant. The effect skyrockets to 13.9 percentage points in column 5 and is highly significant. The effect falls back to earth a bit in column 6, reaching 6.4 percentage points, but is nonetheless on the high end of our findings. The results are not statistically significant, but some consideration must be made for the fact that we have less than 1/3 of the data as in our original state senate models from Table 2. Another possible wrinkle in the estimation is that the South could somehow be distinctive. In particular, the realignment pattern of the region over the last 30 years could have led to different patterns of balancing. Therefore, we run models excluding the eleven former Confederate states. Our results, however, are largely the same. For example, re-running the specifications in Table 1 without the South leads to an estimated balancing effect of between 1.6 and 2.5 percentage points, with p-values ranging from 0.008 to 0.032. These are highly significant even though the sample size is cut considerably. Re-running the tests in Table 2 without the South leads to an estimated effect of between 2.0 and 2.5 percentage points, with p-values spanning from 0.005 to 0.027. Again, these are highly significant despite the reduced number of cases. Conclusions The debate over balancing remains unsettled after nearly two decades of attention. The ticket splitting literature is divided, although there is perhaps more weight against balancing. The survey-based tests of balancing are also divided, though scholars are more likely to find evidence consistent with balancing. 18

We seek in this work to provide new evidence on the question. Our work is premised on the notion that the balancing hypothesis is essentially about voters and therefore it makes sense to look at voting behavior at the state level to see if there are signs that voters act in the manner predicted. In conducting our tests, we were able to call upon the relative abundance of state data over thirty years. How should one interpret our findings? Balancing skeptics could point to some insignificant results and a number of other results in which the effect is around one percent. Hence even if there is balancing, one percent (equivalent to four seats in the U.S. House or one seat in the U.S. Senate) is not going to dominate election outcomes. On the other hand, balancing proponents may reply that the balancing hypothesis survived even in the face of demanding models that included fixed effects and lagged dependent variables, elements that are often suspected of soaking up the effects of meaningful independent variables. In addition, in some quite reasonable models the balancing effect is large, often near or above three percent and as high as 14 percent. A three-percent balancing effect is equivalent to 13 U.S. House seats and four U.S. Senate seats, or (in an average-sized U.S. state legislature) three assembly seats and one senate seat. A 14- percent effect equates to 61 U.S. House seats and 14 U.S. Senate seats, or 15 state assembly seats and six state senate seats. In Pennsylvania, a larger state, the effect would be 28 assembly seats and seven senate seats. We come down somewhere in the middle. We believe that the evidence suggests balancing effects at the state level are real. The results are significant across a broad array of specifications, including many which are quite demanding of a relatively small data set. And there seem to be two factors that make the results particularly strong. First, balancing effects are clearer during presidential midterm elections when the intensity of national presidential elections does not 19

dominate the political environment. Second, balancing effects are clearer when we account for the popularity of the governor, as popular governors mitigate balancing and unpopular governors intensify it. This last point also opens interesting directions for future research on balancing. Suppose our conclusion is correct and voters do indeed balance. What should rational politicians do? One possibility is that incumbents would temper their policy goals in the run-up to a midterm election, especially if legislative control of a chamber is in question. To the extent that moderate voters respond to such tempering of policy, this could have several interesting implications. First, it could cause the observed effects to be smaller; for example, Democrats may moderate in anticipation of balancing and thereby do better than they would have done with a more aggressive agenda. Second, such anticipatory policy behavior could in and of itself be an interesting and important topic for further analysis. Finally, the micro-level determinants of voter balancing could be further probed. Who is it that balances? Is it sophisticated voters making rational calculations? Or is it unsophisticated voters going with the mood of the times? Our results here imply that this line of research merits continued attention. 20

References Abramowitz, Alan I. 1985. Economic Conditions, Presidential Popularity, and Voting Behavior in Midterm Congressional Elections. Journal of Politics 47 (February): 31-43. Achen, Chris. 2001. Why Lagged Dependent Variables Can Suppress Explanatory Power of Other Independent Variables. Presented at American Political Science Association. Alesina, Alberto, and Howard Rosenthal. 1995. Partisan Politics, Divided Government, and the Economy. Cambridge: Cambridge University Press. Allison, Paul. 1990. Change Scores as Dependent Variables in Regression Analysis. Sociological Methodology 20: 93-114. Alvarez, R. Michael, and Matthew M. Schousen. 1993. Policy Moderation or Conflicting Expectations? Testing the Intentional model of Split-Ticket Voting. American Politics Quarterly 21 (October): 410-38. Arcelus, Francisco, and Allan H. Meltzer. 1975. The Effects of Aggregate Economic Variables on Congressional Elections. American Political Science Review 69 (December): 232-39. Bailey, Michael and Michael McDonald. 2006. Ideological Gerrymandering: Redistricting Effects on State Congressional Delegations. Paper presented at Midwest Political Science Association Annual Meetings, April 20-23, Chicago, IL. Beck, Paul Allen, Lawrence Baum, Aage R. Clausen, and Charles E. Smith, Jr. 1992. Patterns and Sources of Ticket Splitting in Subpresidential Voting. American Political Science Review 86: 916 928. Beck, Nathaniel and Jonathan N. Katz. 2001. Throwing out the Baby with the Bath Water: A 21

Comment on Green, Kim, and Yoon. International Organization, Vol. 55, No. 2 (Spring): 487-495. Beyle, Thad, and Richard Niemi, Lee Sigelman. 2009. Job Approval Ratings Database, 1958-2009. From www.unc.edu/%7ebeyle/jars.html, updated November 2009. Bloom, Howard S., and H. Douglas Price. 1975. Voter Response to Short-Run Economic Conditions: the Asymmetric Effect of Prosperity and Recession. American Political Science Review 69 (December): 1240-54. Born, Richard. 1990. Strategic Politicians and Unresponsive Voters. American Political Science Review 80 (June): 599-612. Born, Richard. 1994. Split-Ticket Voters, Divided Government, and Fiorina's Policy- Balancing Model. Legislative Studies Quarterly 19 (February): 95-115. Born, Richard. 2000. Policy-Balancing Models and the Split-Ticket Voter, 1972-1996. American Politics Quarterly 28 (2): 131-62. Burden, Barry C., and David C. Kimball. 1998. A New Approach to the Study of Ticket Splitting. American Political Science Review 92 (September): 533-44. Campbell, Angus. 1966. Elections and the Political Order. New York: Wiley, 1966. Campbell, James E. 1985. Explaining Presidential Losses in Midterm Congressional Elections. Journal of Politics 47 (November): 1140-57. Campbell, James E. 1987. The Revised Theory of Surge and Decline. American Journal of Political Science 31 (November): 965-79. Campbell, James E. 1991. The Presidential Surge and its Midterm Decline in Congressional Elections: 1868-1988. Journal of Politics 53 (May): 477-87. Campbell, James E. 1997. The Presidential Pulse of Congressional Elections. Kentucky: 22

The University Press of Kentucky. Carsey, Thomas M., Geoffrey C. Layman. 2004. Policy Balancing and Preferences for Party Control of Government. Political Research Quarterly 57 (4): 541-550. Downs, Anthony. 1957. An Economic Theory of Democracy. New York: Harper and Row. Fiorina, Morris P. 1981. Retrospective Voting in American National Elections. New Haven: Yale University Press. Fiorina, Morris P. 1988. The Reagan Years: Turning to the Right or Groping toward the Middle? in The Resurgence of Conservatism in Anglo-American Democracies (Edited by Cooper, Barry, Allan Kornberg, and William Mishler). Durham: Duke University Press. Fiorina, Morris P. 1992. Divided Government. New York: Macmillan Publishing Company. Fiorina, Morris P. 1996. Divided Government (2 nd ed.). Needham Heights: Allyn and Bacon. Geer, John G., Amy Carter, James McHenry, Ryan Teten, and Jennifer Hoef. 2004. Experimenting with the Balancing Hypothesis. Political Psychology 25 (1): 49-63. Kernell, Samuel. 1977. Presidential Popularity and Negative Voting: An Alternative Explanation of the Midterm Congressional Decline of the President s Party. American Political Science Review 71 (March): 44-66. Kiewiet, D. Roderick. 1983. Macroeconomics and Micropolitics. Chicago: University of Chicago Press. King, James D. 2001. Incumbent Popularity and Vote Choice in Gubernatorial Elections. The Journal of Politics 63 (May): 585-597. 23

Klarner, Carl. 2009. State Partisan Balance 1959 to 2007. State Politics and Policy Data Resource. From http://www.ipsr.ku.edu/sppq/journal_datasets/klarner.shtml, updated February 2009. Lacy, Dean. 1998. Back from Intermission: The 1994 Elections and the Return to Divided Government. In Herbert E Weisberg and Samuel C. Patterson, eds., Great Theatre: The American Congress in the 1990s. New York: Cambridge University Press. Lacy, Dean, and Philip Paolino. 1998. Downsian Voting and the Separation of Powers. American Journal of Political Science 42 (October): 1180-99. Lewis-Beck, Michael S. and Richard Nadeau. 2004. Split-Ticket Voting: The Effects of Cognitive Madisonianism. The Journal of Politics 66 (February): 97-112. McAllister, lan, and Robert Darcy. 1992. Sources of Split-Ticket Voting in the 1988 American Elections. Political Studies 40 (4): 695-712. Oppenheimer, Bruce, James Stimson and Richard Waterman. 1986. Interpreting U.S. Congressional Elections: The Exposure Thesis. Legislative Studies Quarterly 11, 2 (May): 227-247. Petrocik, John R., and Joseph Doherty. 1996. The Road to Divided Government: Paved without Intention. In Peter F. Galderisi, Roberta Q. Herzberg, and Peter McNamara, eds., Divided Government: Change, Uncertainty, and the Constitutional Order. Lanham, MD: Rowman & Littlefield. Roscoe, Douglas D. 2003. The Choosers or the Choices? Voter Characteristics and the Structure of Electoral Competition as Explanations for Ticket Splitting. The Journal of Politics 65 (November): 1147-1164. Saunders, Kyle L, Alan Abramowitz, and Jonathan Williamson. 2005. A New Kind of 24

Balancing Act: Electoral Certainty and Ticket-Splitting in the 1996 and 2000 Elections. Political Research Quarterly 58 (March): 69-78. Sigelman, Lee, Paul J. Wahlbeck, and Emmett H. Buell, Jr. 1997. Vote Choice and the Preference for Divided Government: Lessons of 1992. American Journal of Political Science 41 (July): 879-94. Smith, Charles E., Robert D. Brown, John M. Bruce, and L. Marvin Overby. 1999. Party Balancing and Voting for Congress in the 1996 National Election. American Journal of Political Science 43 (July): 737-64. The Book of the States. Various years. Lexington, KY: Council of State Government. Tufte, Edward R. 1975. Determinants of the Outcomes of Midterm Congressional Election. American Political Science Review 69 (September): 812-26. U.S. Department of Labor. Bureau of Labor Statistics. Unemployment Rates by State. From http://data.bls.gov/map/servlet/map.servlet.maptoolservlet?survey=la&map=stat e&seasonal=s, updated November 2009. Wooldridge, Jeffrey. 2009. Introductory Econometrics: A Modern Approach, 4 th edition. South- Western. Wright, Gerald C., John P. McIver and Robert S. Erikson. 2004. Aggregated CBS News/New York Times national polls [electronic file] php.indiana.edu/~wright1/cbs7603_pct.zip. 25

Table 1: Predicting Change in Democratic Proportion of State Assemblies (1) (2) (3) (4) Balancing Dem. gov. x Gov. midterm -0.016* -0.003 0.001-0.001 (2.18) (0.44) (0.07) (0.21) Rep. gov. x Gov. midterm 0.01 0.016* 0.018* 0.013+ (1.31) (2.19) (2.51) (1.88) Dem. gov. x Not gov. midterm -0.004 0.005 0.005 0.006 (0.58) (0.72) (0.76) (0.99) Surge and Decline Lagged Dem presidential vote share 0.034 0.014 0.202** (0.58) (0.22) (3.40) Current Dem presidential vote share 0.105+ 0.094 0.270** (1.77) (1.53) (4.55) Economic Unemployment change x Dem. gov. -0.005-0.007-0.007 (0.23) (0.30) (0.37) Unemployment change x Rep. gov. -0.029-0.028-0.031 (1.12) (1.05) (1.25) Other Pre-election seat share -0.265** (8.53) Constant -0.003 0.025-0.047 0.067 (0.68) (0.37) (1.52) (1.04) Year fixed effects No Yes Yes Yes State fixed effects No Yes Yes Yes AR1 model No No Yes No N 740 740 691 740 R 2 0.02 0.32 0.31 0.39 Rep gov midterm - Dem gov midterm 0.026 0.019 0.017 0.014 F-test: Rep gov midterm = Dem gov midterm 11.4 7.4 6.0 4.7 (p=0.001) (p=0.007) (p=0.014) (p=0.031) Note: Dependent variable is change in Democratic proportion of state assembly t-statistics are in parentheses + p 0.10; * p 0.05; ** p 0.01 (all two-tailed) 26

Table 2: Predicting Change in Democratic Proportion of State Senates (1) (2) (3) (4) Balancing Dem. gov. x Gov. midterm -0.023** -0.015+ -0.017+ -0.012 (3.02) (1.70) (1.95) (1.56) Rep. gov. x Gov. midterm 0.001 0.005 0.005 0.001 (0.12) (0.64) (0.57) (0.10) Dem. gov. x Not gov. midterm -0.013+ -0.003-0.006-0.001 (1.77) (0.40) (0.83) (0.15) Surge and Decline Lagged Dem presidential vote share 0.005 0.008 0.05 (0.14) (0.22) (1.64) Current Dem presidential vote share 0.093* 0.103* 0.169** (1.98) (2.07) (3.79) Economic Unemployment change x Dem. gov. -0.016-0.016-0.013 (0.54) (0.50) (0.47) Unemployment change x Rep. gov. -0.02-0.018-0.01 (0.65) (0.57) (0.34) Other Pre-election seat share -0.274** (9.78) Constant 0.001 0.095-0.034 0.165* (0.24) (1.12) (0.41) (2.10) Year fixed effects No Yes Yes Yes State fixed effects No Yes Yes Yes AR1 model No No Yes No N 689 689 640 689 R 2 0.02 0.18 0.18 0.29 Rep gov midterm - Dem gov midterm 0.024 0.020 0.022 0.013 F-test: Rep gov midterm = Dem gov midterm 9.4 6.2 7.0 3.1 (p=0.002) (p=0.013) (p=0.008) (p=0.079) Note: Dependent variable is change in Democratic proportion of state senate. t-statistics are in parentheses + p 0.10; * p 0.05; ** p 0.01 (all two-tailed) 27

Table 3: Predicting Change in Democratic Proportion of State Assemblies (1) (2) (3) (4) (5) (6) Balancing Dem. gov. x Gov. midterm -0.035-0.016 0.004 0.001 0.000-0.006 (1.58) (0.85) (0.55) (0.14) (0.02) (0.58) Rep. gov. x Gov. midterm 0.012 0.011 0.019* 0.015* 0.030** 0.019+ (1.19) (1.19) (2.52) (2.20) (2.95) (1.74) Dem. gov. x Not gov. midterm -0.022-0.006 0.004 0.005 0.009 0.01 (1.01) (0.33) (0.52) (0.80) (0.84) (1.06) Surge and Decline Lagged Dem presidential vote share -0.051 0.017-0.053 0.185** 0.053 0.264** (0.46) (0.18) (0.79) (3.12) (0.55) (3.11) Current Dem presidential vote share -0.024 0.101 0.042 0.247** 0.129 0.314** (0.23) (1.07) (0.61) (3.94) (1.26) (3.52) Economic Unemployment change x Dem. gov. -0.017-0.029-0.053+ -0.027-0.017* -0.02 (1.22) (0.77) (1.78) (1.03) (2.12) (0.74) Unemployment change x Rep. gov. -0.032-0.029-0.050+ -0.042-0.024-0.025 (0.85) (0.80) (1.73) (1.64) (0.64) (0.68) Other Pre-election seat share -0.181** -0.374** -0.268** (3.84) (10.59) (5.96) Dem. Gov approval 0.018* 0.01 (2.41) (0.93) Rep. Gov approval -0.034-0.015 (0.92) (0.49) Change in net Democrats 0.00 0.00 (0.05) (1.28) Change in net liberals -0.001+ -0.001+ (1.81) (1.81) Dem. control redistricting 0.011 0.020+ (0.72) (1.90) Rep. control redistricting 0.004-0.016 (0.24) (1.31) Constant -0.161-0.059 0.037 0.108+ -0.161-1.158 (1.34) (0.28) (0.49) (1.83) (1.11) (0.73) Year fixed effects Yes Yes Yes Yes Yes Yes State fixed effects Yes Yes Yes Yes Yes Yes AR1 model Yes No Yes No Yes No N 384 433 528 575 358 407 R 2 0.34 0.40 0.35 0.46 0.32 0.39 Rep gov midterm - Dem gov midterm 0.047 0.027 0.015 0.014 0.030 0.025 F-test: Rep gov midterm = Dem gov midterm 4.3 2.1 3.3 4.1 7.1 5.6 (p=0.039) (p=0.151) (p=0.071) (p=0.042) (p=0.008) (p=0.019) Note: Dependent variable is change in Democratic proportion of state assembly t-statistics are in parentheses + p 0.10; * p 0.05; ** p 0.01 (all two-tailed) 28

Table 4: Predicting Change in Democratic Proportion of State Senates (1) (2) (3) (4) (5) (6) Balancing Dem. gov. x Gov. midterm -0.063+ -0.043-0.023* -0.018* -0.013-0.012 (1.85) (1.62) (2.11) (1.97) (1.05) (1.16) Rep. gov. x Gov. midterm -0.004-0.004 0.002 0.001 0.007-0.006 (0.35) (0.35) (0.16) (0.08) (0.61) (0.54) Dem. gov. x Not gov. midterm -0.051-0.029-0.009-0.005-0.001 0.005 (1.49) (1.11) (0.97) (0.62) (0.11) (0.53) Surge and Decline Lagged Dem presidential vote share -0.013 0.022 0.007 0.045 0.057 0.115** (0.26) (0.58) (0.17) (1.32) (0.97) (2.86) Current Dem presidential vote share 0.158+ 0.228** 0.086 0.116* 0.144+ 0.231** (1.95) (3.42) (1.39) (2.18) (1.84) (3.81) Economic Unemployment change x Dem. gov. -0.015-0.006-0.052-0.038-0.035-0.034 (0.27) (0.14) (1.30) (1.16) (0.83) (0.97) Unemployment change x Rep. gov. 0.008 0.011-0.025-0.002-0.022-0.024 (0.17) (0.27) (0.63) (0.07) (0.49) (0.63) Other Pre-election seat share -0.370** -0.364** -0.299** (8.86) (10.38) (7.74) Dem. gov. approval 0.035 0.023 (0.79) (0.66) Rep. gov. approval -0.046-0.031 (1.07) (0.97) Change in net Democrats 0.00 0.00 (0.73) (0.10) Change in net liberals 0.00 0.00 (0.53) (0.04) Dem. control redistricting 0.017 0.042** (1.06) (3.89) Rep. control redistricting 0.003-0.011 (0.16) (0.90) Constant -0.154 0.229+ 0.032 0.212** -0.093 0.049 (1.45) (1.89) (0.30) (2.61) (0.60) (0.73) Year fixed effects Yes Yes Yes Yes Yes Yes State fixed effects Yes Yes Yes Yes Yes Yes AR1 model Yes No Yes No Yes No N 350 399 488 535 328 377 R 2 0.31 0.42 0.18 0.33 0.14 0.30 Rep gov midterm - Dem gov midterm 0.059 0.039 0.025 0.019 0.020 0.006 F-test: Rep gov midterm = Dem gov midterm 2.9 2.2 5.3 4.5 2.5 0.3 (p=0.090) (p=0.138) (p=0.023) (p=0.035) (p=0.116) (p=0.559) Note: Dependent variable is change in Democratic proportion of state senate. t-statistics are in parentheses + p 0.10; * p 0.05; ** p 0.01 (all two-tailed) 29