Voice and Balancing in US Congressional Elections

Abstract

If the median voter wrote the Constitution, every Tuesday would be Election Day. Consider the case of the United States: Halfway into a presidential term, congressional elections allow the people to adjust the course of federal policy. Two complementary mechanisms describe how this opportunity is embraced by centrists: a direct mechanism, which strengthens the out-party in Congress to “balance” the president’s policy impact, and an indirect mechanism, by which midterm voting serves to “voice” dissatisfaction as a signal to the president. A model of repeated elections unites the two mechanisms: whereas midterm balancing reacts to the preceding presidential election, midterm voice anticipates the following one. Using micro and macro data for all House elections from 1956 through 2018, I show that balancing and voice work hand in hand: it is those voters with both policy incentives who contribute most to the notorious “midterm loss,” and particularly so under circumstances that make balancing more necessary and voice more promising. Yet although policy-oriented behavior typically restrains dominant parties, it may also cushion the fall of unpopular administrations. Centrists can be creative.

A fundamental characteristic of modern democracy is the periodic election of officeholders for a limited term (Huntington 1991; Powell 2000; Schumpeter 1942). In a stable polity, voters and politicians can safely assume that elections will be repeated on a regular basis. This arrangement is often celebrated for fostering the precious virtue of “loser’s consent” (Anderson et al. 2005). Less apparent, yet perhaps equally important, is the ability of repeated elections to accommodate the political center. The reality of party polarization implies that moderate voters tend to find themselves without appropriate representation. Repeated elections give these voters the opportunity to have their voices heard and to seek a balance of different ideologies.

These general considerations are particularly relevant for the case of US midterm elections. The mere absence of a presidential race at midterm highlights the fact that these ballots follow the election of a president, and they will be followed by the election of another. I argue that the specifics of this situation—the ramifications of the last presidential election and the anticipation of the next—are crucial for voting behavior aiming at policy influence.

A focus on models of repeated elections brings together insights from as yet distinct literatures to demonstrate how voters use midterm elections to adjust the course of federal policy halfway into a presidential term. Such adjustment works through the interplay of two mechanisms:

1. 1. A direct mechanism, commonly known as “balancing,” by which voters react at midterm to the preceding presidential election by strengthening the party of the losing presidential candidate in Congress, so that the legislation produced by the federal government as a whole will be more moderate.

2. 2. An indirect mechanism, commonly known as “voice” or “signaling,” by which voters use midterm elections to send a message of dissatisfaction with policy to the president, on pain of defection in the following presidential election in case the incumbent fails to respond.

Because both balancing and voice typically hurt the incumbent administration, they contribute to the recurring phenomenon of the “midterm loss”: the almost universal pattern of the party controlling the presidency (the “in-party”) losing support to the other major party (the “out-party”). At the same time, the model also predicts patterns that may be less intuitive. Because policy-oriented behavior is attracted by circumstances that make balancing more necessary and voice more promising, it tends to spare the in-party in dark electoral times—or may even come to the aid of the president. Thus, thorough application of balancing/voice logic shows that conventional expectations of its political impact underestimate the rationality of the electorate.

Although balancing and voice are theoretically distinct, this does not mean that they are mutually exclusive. Quite the contrary—I argue theoretically and demonstrate empirically that policy-oriented behavior is pursued by those voters who have a balancing incentive and are most motivated to use voice. In essence, effective voice depends on the exit threat that accompanies balancing, and balancing is more valuable when combined with the opportunity to voice a preference for centrist policy. Modeling the two mechanisms in interaction yields a more powerful explanation of electoral behavior than either one alone.

This article proceeds as follows. In the next section, I develop my theory on the basis of relevant literatures on US elections, comparative behavior, and formal theory. A set of hypotheses is derived regarding individual voting behavior and contextual moderators. The third section describes the dataset used to test my hypotheses for all House elections from 1956 through 2018. The results section is split into two parts: the first shows on the micro level that a significant number of voters behave in line with joint balancing and voice incentives, and the second demonstrates on the macro level that this behavior is particularly frequent when the underlying incentives are most pronounced. The final section discusses implications of my findings and opportunities for future research.

Theory and Hypotheses

Most research on policy-oriented voting analyzes elections as isolated events in time. In the tradition of Downs (1957), voters are assumed to support the candidate or party whose policy agenda for the coming legislative period will produce outcomes closest to their own preferences. But Downs’s original theory was more versatile, acknowledging that the rational citizens populating the model have a past and a future. For example, they evaluate government policy by comparing the incumbent elected in the preceding election to changes the challenger would have made (39ff.), and they may cast strategic “blackmail” votes to influence the policy positions taken by other parties in future elections (131f.). ¹ However, these relatively rudimentary ideas went largely into hibernation in favor of the book’s wealth of more elaborate propositions.

During the eventual renaissance of research on the chronology of voting, midterm elections have served as a focal point. As the very label “midterm” expresses, these elections are subordinate to the surrounding presidential elections that define the “term.” This makes midterm elections particularly contingent on their place in the order of events. Voting behavior at midterm may be affected by reactions to the preceding presidential election, by the anticipation of the following one, or both.

The notorious midterm loss of the in-party is a case in point, because it is precisely defined relative to the preceding on-year election. Under the current party system, the party of the president has lost House seats in 38 of 41 midterm elections (the exceptions being 1934, 1998, and 2002). Theories trying to explain this phenomenon include an essentially “mechanical” effect of seat exposure (Oppenheimer, Stimson, and Waterman 1986), the withdrawal of presidential coattails at midterm (Campbell 1985), the regression to normal behavior in the course of “surge and decline” (Campbell 1960), a psychological tendency of voters to overrate negative information (Kernell 1977), and the midterm serving as a “referendum” on presidential performance (Tufte 1975).

What these theories have in common is that they assign voters rather “sincere” roles, which tend to follow structural dynamics. In contrast, I believe that more strategic behavior is also at play. This does not mean that I seek to refute other theories. Quite the opposite—I will show that more sincere and more strategic voting behavior coexist in different parts of the electorate, and that the former may even serve as a stimulus to the latter.

A more strategic, policy-oriented basis for the midterm loss was first noted by Erikson (1988). After scrutinizing various theories, Erikson concluded that the empirical record is best described by a “presidential penalty” that is applied to the in-party in Congress to balance its advantage of controlling the executive. With rising interest in the politics of divided government, this proposition was elaborated in the context of a wider notion that voters support candidates of the out-party to prevent policy from tilting too far in the in-party’s direction (Alesina and Rosenthal 1995; Fiorina 1992). Further research produced a growing body of evidence: policy change triggers a backlash in party support (Bølstad 2012; Wlezien 1995), surprise about the outcome of a presidential election leads to balancing at midterm two years later (Scheve and Tomz 1999), voter motivations for balancing evolve throughout the midterm campaign (Bafumi, Erikson, and Wlezien 2010), the “presidential penalty” holds up when confronted with other theories (Knight 2017), and balancing behavior can also be found in various institutional environments outside the United States (Kedar 2009).

The basic intuition of midterm balancing is that centrist voters react to the preceding presidential election by weakening the presidential party in Congress. If this intuition is empirically correct, we should observe on the individual level that voters who find themselves “sandwiched” between the two parties are quick to abandon the in-party at midterm:

H1 (balancing): The midterm loss is particularly strong among voters whose policy preferences are torn between the two party platforms.

Midterm balancing describes a mechanism through which the outcome of a previous election affects the outcome of a current election. Of equal theoretical relevance is the other direction in which causality can go in a system of repeated elections; that is, from a future election to the current one. The mechanism there works through anticipation. In a series of formal models, scholars have shown how citizens may use their vote to signal preferences for future elections rather than to affect current policy (Castanheira 2003; Hummel 2011; Kselman and Niou 2011; Meirowitz and Shotts 2009; Meirowitz and Tucker 2007; Myatt 2017; Piketty 2000; Razin 2003). Among the many scenarios modeled in this literature, the closest one to my case of US midterm elections is Meirowitz and Tucker (2007), who consider a game of sequential parliamentary and presidential elections. Inspired by the 1996 reelection of Russian president Boris Yeltsin, they show formally how voters may have used the parliamentary election in the preceding year to send a message to the president by withdrawing support from the party backing his agenda.

Empirical evidence of “signaling” behavior in general comes from various electoral systems. ² Studies that best reflect the specifics of the US midterm context analyze elections to the European Parliament, the supranational legislature of the European Union (Lindstam 2019; Weber 2011). European Parliament elections generally fall during the electoral cycles of the EU member states, executive power is not at stake, turnout is low, and the parties controlling the national executives lose votes—just like the party controlling the US executive does at midterm (Franklin and Weber 2010; Reif and Schmitt 1980). Using pan-European survey data, Weber (2011) shows that the midterm loss in the EU is caused by voters who voice dissatisfaction with their national executives, and Lindstam (2019) adds that voters use their EU ballots to signal the salience of issues neglected in national politics. In the tradition of Hirschman’s (1970) “Exit, Voice, and Loyalty,” “voice” is modeled as a strategy to bring about performance improvement in one’s own party, in contrast to the “exit” strategy of looking for a new home in another party.

Translated to the US context, voice behavior would mean that voters use midterm elections to send a message of dissatisfaction to the president. These are voters who hope for improvement during the second half of the term. Adding authority to their hope, they use a current midterm election as a means to influence a future presidential race. As we will see, such a strategy is most attractive in a president’s first term, when a reelection bid is obvious.

Although voice behavior is inherently policy oriented, its focus can be broader than that of balancing. Hirschman (1970) defines the origin of voice-provoking dissatisfaction using the terms of supplier “performance” and product “quality.” In the electoral sphere, this may certainly refer not only to policy positions (as for balancing) but also to nonpositional factors.

Just like balancing, however, voice is not for everybody. The general optimism underlying this behavior needs some kind of source. Hirschman (1970, 77ff.) therefore expected voice to be most common among loyalists of an organization. In the words of a widely recognized exegesis, “loyalty does not normally mean a mere reluctance to leave a collectivity but rather a positive commitment to further its welfare by working for it, fighting for it and—where one thinks it has gone astray—seeking to change it. Thus, voice … is already built into the concept of loyalty.” (Barry 1974, 98). In a nutshell, loyalty functions as an “exit tax” and a “voice subsidy” (Gehlbach 2006).

If the intuition of voice as a recuperation mechanism is empirically correct, we should observe a disproportionate share of citizens who are generally loyal to the in-party shelving their allegiance at midterm. This is also what Weber (2011) finds in the EU context, where partisans turn against their own party around midterm precisely because they feel attached to it. Because of their attachment, they try to improve the performance of “their” party, rather than simply leaving it for good, as independents would. For the same reason, I expect the in-party in the United States to lose votes primarily among its most loyal supporters.

Importantly, voice resembles balancing in that the resulting vote reflects a policy preference, not a party preference. Voice is instrumental, not just expressive. Loyalists hope that the president’s response to their midterm signal will allow them to maintain their attachment in the long run, relieved of policy cross-pressure. Thus, although voice is driven by loyalty, it does not primarily express loyalty. Quite the opposite: voice even takes the form of exit, because loyalists abandon their party at midterm. This leads to my second hypothesis:

H2 (voice): The midterm loss is particularly strong among partisans of the president’s party.

To recapitulate, balancing and voice are two mechanisms that help policy-oriented voting travel through time in a system of repeated elections. Balancing aims at direct policy influence by compensating for the impact of a past election, whereas voice aims at indirect influence by anticipating the conditions of a future election.

Balancing and voice can both be exercised in two ways: either by voting for the out-party at midterm or—in a milder version—by abstaining. Vote switching is more potent, because it effectively counts double (one vote less for the in-party and one more for the out-party), whereas abstention only withholds one vote from the in-party. Given that balancing and voice are both instrumental in essence, this would suggest that vote switching is the preferred means. However, abstention has the advantages that it carries no transaction cost and does not require loyalists to vote actively against their allegiance. Thus, the specific mix of behavior will be an empirical matter. ³

Importantly, although balancing and voice are based on different theories, their premises are kindred in practice. There is reason to believe that voice is more effective when balancing is strategically feasible and, vice versa, that balancing is more effective in conjunction with a motivation for voice.

From the perspective of voice, note that voters considering its use face a fundamental dilemma. Although their loyalty motivates them to fight performance decline at midterm, it also reduces their elasticity in the coming presidential election. Because of their attachment, partisans are less likely than independents to defect for policy reasons (Jessee 2010). Vote maximization then suggests that the in-party has an incentive to discount partisan concerns whenever they clash with other demands (Erikson and Romero 1990). To be effective, voice therefore needs to be accompanied by a credible threat of defection. As Hirschman already noted, “The chances for voice to function effectively as a recuperation mechanism are appreciably strengthened if voice is backed up by the threat of exit” (1970, 82).

The most formidable “threat of exit” is posed by voters who combine a motivation for voice with an incentive for balancing. If partisans are in fact not better represented by their own party than by the competitor, exit is entirely credible on policy grounds. Generally, then, “the threat of exit will typically be made by the loyalist—that is, by the member who cares—who leaves no stone unturned before he resigns himself to the painful decision to withdraw or switch” (Hirschman 1970, 83). ⁴

Now let us consider the same logic from the perspective of balancing. To recall, the aim of this strategy is purely instrumental in that centrist voters use their influence to achieve policy moderation. However, whereas the prototype of the balancing voter does not care which party is in charge of policy as long as the outcome is right, a balancing voter who also happens to be a partisan does have a long-term preference: they would ideally like to see moderate policy being implemented by the in-party. Such a voter may use the same midterm vote to balance the power of the presidency and to voice a preference for a centrist agenda to the president. In policy terms, this strategy “kills two birds with one stone”—direct influence in Congress and indirect influence in the next presidential election. Voice then reinforces the specific positional policy goal of balancing. When paired with a motivation for voice, a policy incentive is thus more likely to “tip” the balance toward the out-party, overcoming any nonpolicy reasons that may have worked in favor of the president in the preceding election.

Interaction of balancing and voice is also attractive because each addresses inherent paradoxes of rationality in the other. Balancing behavior may be considered rational only if House control is on a knife’s edge, and only by voters who are registered in competitive districts. Although various generic arguments suggest that these assumptions may be too strict, ⁵ voice gives additional support in two ways. First, it is a signal for the following presidential election, in which congressional districts do not apply (except for the district tiers in Maine and Nebraska). Because many fewer voters live in safe states than in safe districts, a midterm vote as a presidential signal makes sense more broadly. Second, balancing requires a somewhat permissive definition of rationality in general. All voters, whether they pursue balancing or not, know that their vote will not affect the outcome of the election. Much scholarship therefore assumes that voters behave “as if” an individual vote would be decisive. For this assumption to materialize, however, voting needs to come with some utility that is not exclusively policy based. One source of such utility is voice. As already developed in Hirschman (1976, 387), the loyalist will find the use of voice personally rewarding: it helps their party stay on track and defies a painful decision whether to leave for good.

The interaction of balancing and voice also addresses an inherent problem of voice. In their formal model of sequential parliamentary and presidential elections discussed earlier, Meirowitz and Tucker (2007) show that voters will only use an election as a voice mechanism if the legislature elected on that occasion is relatively unimportant. For US midterm elections this is apparently not the case, because the congressional offices filled at midterm are the same as in on-year elections. While one solution to this theoretical problem may be that voters do not perceive midterm elections as equally important, balancing bolsters voice more systematically: if the opportunity for voice is backed by a balancing incentive, then voters do not need to be concerned about “wasting” their direct influence at midterm. Quite the contrary—overlapping balancing and voice incentives allow defection in the belief that it will do the best for the party and achieve the best policy outcome.

Taken together, this theoretical reasoning suggests that the mechanisms of balancing and voice work hand in hand. If these considerations are empirically valid, we should find that the behavior of the crossed group defined by the concurrence of balancing and voice incentives is particularly exacting with the in-party at midterm.

H3 (balancing × voice): The midterm loss is exceptionally strong among partisans of the president’s party whose policy preferences are torn between the two party platforms.

The first three hypotheses describe my core expectations on the individual level. To verify that the observed behavior indeed reflects balancing and voice incentives, I am also interested in differences between elections. Exploring contextual variation will provide more confidence that the mechanisms behind the correlations are operating as theorized.

Regarding the balancing mechanism, while often stated in general terms, its underlying motivation is actually quite contingent. Voters will turn against the in-party if they expect a concentration of legislative and executive power that would move policy away from their ideal point. Thus, if the in-party is expected to do fairly well at midterm, the chance of united government increases, and it will seem more necessary to correct the imbalance. Vice versa, if the in-party is expected to suffer a bad defeat at midterm, divided government becomes more likely, and voters will refrain from additional balancing.

A similar effect was found by Erikson (2016) in on-year elections, where balancing occurs concurrently with the presidential race. While the congressional vote of a party increases with the coattails of that party’s presidential candidate, it decreases at the same time with the candidate’s perceived chances of winning the presidency. Erikson attributes this effect to “anticipatory balancing.” My fourth hypothesis generalizes this logic to “responsive balancing” in midterm elections:

H4 (context I): The midterm loss due to balancing increases the better the president’s party does in the preelection polls.

Note that the logic of H4 implies there are more sources of the midterm loss than balancing or voice. Such sources are described by various theories, discussed earlier, which see voters in more sincere roles. In my model, these forms of behavior produce a baseline expectation of in-party performance. More strategic, policy-oriented behavior in one part of the electorate then reacts to a baseline produced by more sincere behavior in other parts. In fact, it is conceivable that the baseline expectation turns so severely against the in-party that policy-oriented voters feel it necessary to balance in favor of the president. Determining if and when the tipping point is reached will require empirical calibration.

My second context hypothesis proposes that the motivation for voice, too, will vary between elections. In particular, voice will be most promising when the president faces reelection two years later, and therefore has a tangible incentive to listen to supporters. Voice then serves to remind the president of the need for the moderate vote to win a second term. Once a president has won that term, however, the strategy should become less effective. Given that voice is essentially meant to signal the possibility of defection, using it against a term-limited incumbent would not appeal to the president’s own reelection concerns. Still, to the degree that executive power relies on party backing (e.g., for legislation, appointments, legacy, or protection from impeachment), the party’s interest in future electability extends the argument to the second terms as well. Voters can use voice to make their party exert pressure on the president—just with less prospect of success than in first terms. Thus:

H5 (context II): The midterm loss due to voice is stronger in the president’s first term.

Finally, to conclude my theoretical expectations, note that the two context hypotheses (H4 and H5) describe the behavior of certain voter groups, not that of the electorate as a whole. To the degree that balancing (H1) and voice (H2) operate separately, H4 and H5 refer to the behavior of these two groups, respectively; to the degree that balancing and voice go hand in hand (H3), the contextual factors will moderate the behavior of the crossed group, which combines both incentives.

Data

I test my hypotheses using survey data for 32 congressional elections, from 1956 through 2018. The main source is the ANES Time Series Cumulative Data File, which includes most variables central to my analysis beginning in 1956. For the midterm elections of 2006–18, which were not covered by the ANES, I use data from the Cooperative Congressional Election Study (CCES). The pooled dataset contains 233,487 observations. I applied three types of weights to the data: sample weights (as provided by the ANES and CCES), political weights (so that aggregate turnout and vote choice reflect the official results), and year weights (so that each election has the same weight in the analysis).

To identify individuals with an incentive for ideological balancing, I rely on the liberal–conservative scales regularly used by the ANES/CCES. Respondents were asked to place themselves, as well as the two parties, on a seven-point scale from “extremely liberal” (1) to “extremely conservative” (7), with “moderate” as the middle value (4).

A respondent is assumed to have a clear balancing incentive when located at equal distance between the two parties. I label these cases “ideologically torn.” The measure is likely to be felt subjectively because it depends on respondent perceptions of party positions. ⁶ Other types defined by ideology are “leaners” for each party (i.e., those respondents who locate themselves closer to one party than the other) and the ideologically “inert” (i.e., those who failed to place themselves and/or the parties on the liberal–conservative scale or those who placed the parties in reverse order).

Liberal–conservative scales are not available for the elections of 1956–70. These values were imputed based on a model of micro and macro correlates of ideology. A detailed discussion of the imputation procedure is available in the online appendix.

Partisanship was measured using the standard ANES/CCES scale. I coded as partisans those respondents who readily identified themselves as Democrat or Republican or who indicated in response to a follow-up question that they lean toward one of the parties (as is recommended in the literature, as in Petrocik 2009). Respondents who did neither of these were coded as independents.

Table 1 shows a breakdown of all respondents in my data by ideology and partisanship. I also distinguish between on-year and midterm elections to see whether group sizes change through the electoral cycle. The in-party is the one that wins the presidency (for on-year elections) and holds it during the four-year term (for midterm elections).

Table 1 Group Sizes in the Sample

Notes: N = 233,487. Probit-based standard errors clustered by election year (N = 32).

** p < 0.01; * p < 0.05

To begin with ideology, table 1 shows that the share of ideologically torn respondents is about 6.3% of the sample, with little variation between on-year and midterm elections. Ideological indifference has remained an impressively stable feature of the electorate (the sample share did not decrease over the seven decades covered by the data; quite the opposite, its trend is weakly positive, r =.24). Torn voters thus have defied the larger shift toward polarization that occurred around them. While they have seen the two parties drift in opposite directions, they themselves have remained torn in the middle—perhaps a first sign of policy orientation.

Among the other groups defined by ideology, about half the sample consists of the ideologically inert (many of whom do not vote, as we see in the next section), with the remainder split between the two parties. There is a tendency for the out-party to gain close respondents from the inert at midterm, but this difference is not significant.

Turning to partisanship, it appears that the in-party loses some partisans to the independent camp at midterm. Part of the baseline midterm loss may thus be explained by a differential decline in partisanship in periods of low mobilization.

The last section of table 1 shows the crossed groups defined by ideology and partisanship. As can be seen, there is little change in the group of “In-party & Ideologically torn,” which is of particular interest to me. No significant differences are found, and the two distributions are statistically indistinguishable on the whole. ⁷

Overall, table 1 shows little change in group sizes, compared to much stability. This is a good starting position for my analysis of vote choice and turnout in the next section. My interpretation there assumes that the various groups consist of roughly the same potential voters at midterm as in the preceding on-year election. I see little reason to doubt this assumption, given that the typology is defined by fairly fundamental variables and the groups of interest to my hypotheses do not change much. ⁸ Apparently some uncertainty remains, given that the data are from repeated cross sections and not from a panel. But note that the assumption of continuity in the groups is more convenient than critical. Even if it is violated, the incentives for balancing and voice that voters face at midterm remain intact. The logic of my hypotheses implies that voters use midterm elections to further their current and future policy interests. Whether they personally had these same interests in the past is secondary.

Individual-Level Results

To test my first three hypotheses, I look at reported vote choice and electoral participation in the same groups defined by ideology and partisanship. Differences between on-year and midterm elections were tested for statistical significance using probit regression, with standard errors clustered by election year.

Vote Choice

Table 2 shows the resulting estimates for vote choice. The first line summarizes the official election results, followed by the three different ways of breaking down the sample to test H1, H2, and H3.

Table 2 Support for the President’s Party (Percentage of the Two-Party Vote)

^a Given for comparison. Note that the sample categories do not fully add up to the average midterm loss of 3.7% because the sample-based figures are also affected by group size (see table 1) and differential turnout (see table 3).

Notes: Boldface type indicates hypothesis tests. N = 155,561 (the two-party vote). Probit-based standard errors clustered by election year (N=32). ** p <0.01; * p <0.05.

To begin with ideology, I find some evidence that the ideologically torn contribute to the midterm loss, as expected by H1. The decline in support for the president’s party at midterm is largest among this group (−5.4%). However, the difference is not statistically significant, and other groups tilt in the same direction. ⁹

Regarding partisanship, the in-party loses somewhat among their own partisans (−0.9%), as expected by H2. But the difference is far from significant and, in fact, is smaller than that for out-partisans (−3.4%). In-party support among the latter group is apparently low in all elections, but it appears to be even lower at midterm—yet another aspect of the baseline loss.

Up to this point, evidence for my unconditional hypotheses is rather meager. Neither the shifting of support among the ideologically torn nor that among the president’s partisans can be confirmed with sufficient certainty. When looking at the results for the crossed groups, however, I do find strong support for my expectations. The midterm loss in the group of the ideologically torn who are also in-partisans is statistically significant and larger than in any other group (−9.7%, with the only other significant group at −4.8%). This supports H3, which expected an interaction of balancing and voice motivations. The group of “torn partisans” (the short form I use from now on) expresses this interaction because they have an incentive for balancing that is reinforced by an opportunity for voice or, vice versa, an opportunity for voice that is reinforced by an incentive for balancing.

Turnout

Table 3 repeats the empirical analysis with electoral participation as the outcome variable. The most basic finding is that turnout decreases significantly at midterm in all groups (as would be expected on the basis of the literature). Looking at relative differences, the drop is largest in the three groups singled out by my hypotheses: the ideologically torn (−16.4%), the president’s partisans (−12.5%), and the crossed group of the two, the torn partisans (−16.7%).

Table 3 Voter Turnout (Two-Party Vote Only)

^a Given for comparison. Note that the sample categories do not fully add up to the average turnout decline of 11.1% because the sample-based figures are also affected by group size (see table 1).

Notes: Boldface type indicates hypothesis tests. N = 233,487. Probit-based standard errors clustered by election year (N = 32). ** p < 0.01; * p < 0.05.

The relatively large decline in these three groups suggests that balancing and voice may indeed work to some degree through differential turnout. Rather than casting their vote for the out-party, voters may choose to withdraw support from the in-party by abstaining from voting. However, the evidence is weaker for abstention than for vote switching. Although the decline of turnout in the three groups of interest is strong and significant, the differences to the remaining sample do not quite reach conventional levels of significance. ¹⁰ A more prudent interpretation would be that the evidence is consistent with balancing–voice but not strong enough to distinguish these predictions. For example, the finding that in-partisans have higher turnout in on-year elections than out-partisans, but then shift to lower turnout at midterm, is familiar from other theories, such as “surge and decline.” Balancing–voice would explain why the pattern is most pronounced among the ideologically torn, who show larger turnout decline than their fellow in-partisans, even though they already begin with lower turnout in on-year elections. It may thus be that different groups abstain at midterm for different reasons. However, as the case stands, the evidence points more to a choice-based mechanism behind balancing–voice than to a turnout-based mechanism. We revisit this question in the next section.

Overall, the individual-level analysis produced mixed findings. Support for my unconditional hypotheses is limited, but combining the two in H3 unleashes their explanatory power. This suggests that the mechanisms behind balancing and voice are intimately related. At the same time, these inferences are somewhat uncertain, given that similar (if less pronounced) change also occurs in other voter groups. I therefore turn to my context analysis, which examines whether the frequency of balancing–voice varies in predictable ways with macro incentives for such behavior.

Context-Level Results

To appreciate the importance of context, consider figure 1, which shows the historical trajectory of in-party support among torn partisans. Note that this is not a proper time series, because torn partisans are always defined in relation to the party that holds the White House. The purpose of the graph is to explore any apparent patterns in the extent of balancing–voice, as expressed by the arrows, each of which connects an on-year election to the following midterm election. A downward-pointing arrow means that the in-party experienced midterm loss among its torn partisans. As can be seen, this happened in 12 of 16 cycles. These cases visibly confirm the conventional wisdom of a midterm verdict against the president. But the length of the arrows varies considerably, and there are four midterm elections in which torn partisans increased their support for the in-party: 1958, 1986, 2006, and 2010. Variation is also found in the initial level of in-party support (not only its change), suggesting that presidential years may see anticipatory balancing as well. Relatively poor on-year performance is occasionally followed by less trouble at midterm, and vice versa. However, the pattern is far from systematic and does not simply default to “regression to the mean”—nor is there an obvious time trend in the level or change of in-party support.

Figure 1 In-Party Support and Midterm Loss among Torn Partisans over Time

Hypothesis Tests

Overall, the trajectory of figure 1 shows quite some variance, which does not lend itself readily to interpretation. In other words, this is a formidable test for deductive theory—such as that expressed in my two context hypotheses regarding the moderating role of expected in-party performance and the presidential term. In particular, I proposed that balancing–voice will be more frequent the better the president’s party does in the preelection polls (H4) and in the president’s first term (H5). Finding support for these hypotheses will help demonstrate more conclusively that the observed behavior indeed reflects balancing and voice incentives.

My measure of expected in-party performance is the percentage of the two-party vote reported in the final Gallup poll published before each midterm election (Gallup 2014, 2016, 2018) ¹¹ ; the variable used is the difference between that percentage and the actual in-party vote in the preceding on-year election. Note that this is a conservative measure to test H4. Because the Gallup polls may include voters who had already committed to a balancing–voice strategy, the aggregate behavior of torn partisans will tend to be positively correlated with the polls. My hypothesis, however, expects a negative correlation.

The variable for the first presidential term is a simple dummy. As intended by the theoretical rationale of H5, first term means that the president may run for reelection, whereas second term means that this is not an option. There is one exception: President Johnson’s term beginning in 1965. Because Johnson had served less than half of President Kennedy’s term, he was not term limited under the Twenty-Second Amendment. Thus, for a proper test of H5, Johnson’s second term was coded as his first.

Dependent variables were equally measured on the election level. In particular, the first dependent variable is the change in the in-party’s percentage of the torn partisan vote between a midterm election and the preceding on-year election; in other words, the midterm loss among torn partisans. The second dependent variable is the change in the percentage of torn partisans turning out to vote for one of the major parties. Both dependent variables are continuous and can be simply estimated using OLS regression. Descriptive statistics of all variables are available in the online appendix.

The tests of H4 and H5 are quite challenging in terms of statistical power, because there are only 16 observations on the election level. Yet, as the results in table 4 show, both hypotheses are supported by highly significant estimates and impressive explanatory power in Model 1, the regression of in-party support. ¹²

Table 4 OLS Regression of Aggregate Behavior of Torn Partisans

^a Dependent variable in M1 and M2: change in the percentage of the two-party vote for the president’s party among torn partisans between on-year and midterm elections.

^b Dependent variable in M3: change in the percentage of torn partisans turning out to vote between on-year and midterm elections.

^c R-squared is not clearly defined for multiply imputed data. The first value listed is the mean R-squared of the separate regressions for each of the imputations. The second value is from a single regression using the mean value across the imputations as the dependent variable.

Robust standard errors in parentheses. ** p < 0.01; * p < 0.05.

The coefficients of the dummy for the first presidential term are negative and quite sizable. Support for the president’s party among torn partisans is down by 20.4%. This suggests that voice is indeed concentrated on contexts in which it makes sense to threaten the president with potential defection (H5). In contrast, voice appears to be much less pronounced in a president’s second term, where it has no real target. The volume of voice thus varies systematically with the strategic opportunity structure, suggesting that this behavior is indeed guided by quite rational considerations.

A similar conclusion is supported by the effect of expected in-party performance. As hypothesized, the midterm change of in-party support suggested by Gallup is negatively associated with actual change among torn partisans (H4). When the polls indicate that the in-party will do relatively well compared to the preceding on-year election, torn partisans withdraw their support. The effect is quite marked, with support for the president’s party in this group decreasing by 1.68% for each percentage point of Gallup’s estimate. In contrast, when the polls suggest a defeat for the in-party, defection among torn partisans vanishes. Overall, then, torn partisans anticipate the electoral fortunes of the in-party and work against the trend. When it seems that their intervention is needed to restore the balance between the parties, they side with the out-party; when the balance is already about to be restored through baseline decline, they follow their partisan allegiance.

Figure 2 visualizes the prediction of the two-party vote. As can be seen, actual change in levels of in-party support among torn partisans decreases with the overall change suggested by Gallup before the midterm election. It does so both in the president’s first and second term, but on different levels. For the first term the model predicts balancing–voice against the in-party for all cases but the most severe defeats, whereas for the second term it predicts such an effect only for expected in-party victories. ¹³

Figure 2 Prediction of Aggregate In-Party Support among Torn Partisans (with 95% CIs)

In general, the prediction is very accurate. For example, torn partisans strongly turned against their parties in 1998 and 2002, which are the only two cases in which Gallup suggested a sizable surge for the in-party. The case of 1998 is particularly remarkable, given that this was a sixth-year midterm election for which I expect the motivation for voice to be largely absent. But as the polls sensed that the Democrats might retake the House, the balancing motivation shined through, even though the president had already been reined in during four years of divided government. In contrast, in 2006 and 2010, two years that saw some of the most severe defeats of the in-party, torn partisans even came to the rescue. This illustrates how these voters anticipate the behavior of the electorate at large and then use their ballots to work in the opposite direction. Overall, policy-oriented voting contributed to the midterm loss in the majority of cases, but the degree depends on contextual factors—up to a point where torn partisans begin to balance in favor of the in-party.

The one election falling clearly outside the confidence intervals is easily interpreted in historical context. In 1974, torn partisans deserted the Republican Party in droves. In this midterm held just a few months after Watergate, the need for voice was arguably more apparent than usual—in addition to the fact that President Nixon’s resignation installed an incumbent whose electability had never been established.

Model 1 in table 4 does not say anything about the mechanism behind balancing–voice: Is it vote switching or differential turnout? This question is addressed in Model 2. With the same dependent variable, the predictor measures the change in the turnout of torn partisans from an on-year to a midterm election. The small, insignificant coefficient of this variable indicates that a decline in turnout does not systematically come with more midterm loss for the in-party.

The same conclusion results from Model 3, which tests H4 and H5 with the change in turnout as dependent variable. Neither predictor of this regression has a noteworthy or significant effect. Thus, unlike party support, turnout and abstention among torn partisans do not follow macro incentives for balancing–voice.

Overall, table 4 corroborates the impression from tables 2 and 3 that balancing-voice works through vote switching, rather than through differential abstention of in-party supporters. By directly supporting the out-party, torn partisans thus maximize the impact of their vote, which underscores the instrumental basis of their behavior.

Variations and Robustness

Another strategy to explore the nature of policy-oriented behavior is to model complementary factors. While the R-squared of Model 1 does not leave much room, there are some differences that my general model ignored. In addition to the party affiliation of the president, torn partisans should be mindful of the specific policy constellation of the main actors in government. Such nuanced effects can be expressed with two complementary hypotheses, each of which moderates one of the logics of policy-oriented behavior. First, voice should be more sensitive when the president pursues more extreme policies. This derives directly from the purpose of voice to rein in the president’s ideological ambitions. Second, balancing should be more sensitive when the two parties are more polarized. Polarization amplifies the policy bias that torn partisans have to reckon with if they do not balance against the in-party.

To test these hypotheses, I rely on policy positions from the widely used DW-NOMINATE application, a scaling technique that places all members of Congress in a two-dimensional policy space based on their roll-call record (Lewis et al. 2020). The first dimension explains the large majority of the variance, and is commonly interpreted in terms of liberal-vs-conservative ideology. To measure party positions, I use the score of the median House member of each caucus. Moreover, DW-NOMINATE also places the president in a “common space” with Congress, based on legislative vetoes and other public positions.

Two variables are coded on the basis of the DW-NOMINATE scores: “Party polarization” is the distance between the two party positions on the [-1;1] scale. “President’s extremeness” is the distance of the president’s position from the midpoint (0).

Model 4 in table 5 shows the regression results when the two policy variables are added to the main model. As can be seen, neither variable has a significant effect, and there is not much gain in variance explained (compared to Model 1 in table 4). We can thus conclude that balancing–voice is not categorically limited to some particular context. At the same time, however, Model 4 is not an actual test of the expectation that policy-oriented behavior is more sensitive under certain conditions. For such a test, the two policy variables need to enter the model in interaction with the two main predictors. The expectation is that the effects of the main predictors (president’s first term and the Gallup estimate) will increase with party polarization and the president’s extremeness.

Table 5 Complementary Explanations of Aggregate In-Party Support by Torn Partisans

Note: Dependent variable is the change in the percentage of the two-party vote for the president’s party among torn partisans between on-year and midterm elections.

^a See table 4.

^b No score is available for President Trump at the time of this writing.

Robust standard errors in parentheses. **p < 0.01; *p < 0.05.

Regression results are shown in Model 5 and 6 of table 5. The coefficients of the multiplicative terms indicate that each of the policy variables interacts with one of the main predictors in the expected way: the first-term penalty increases with the president’s extremeness, and the strong-polls penalty increases with party polarization. These two interaction effects are large and negative, thus reinforcing the negative baseline effects.

Multiplicative interactions are a severe strain on a model with just 16 cases, and it is not surprising that significances appear weak. However, this does not mean that no statements can be made about the data (e.g., Brambor, Clark, and Golder 2006). To ease interpretation, marginal effects are visualized in figure 3.

Figure 3 Marginal Effects on In-Party Support, from M5 and M6 (with 95% CIs)

Figure 3a shows how the negative effect of the president’s first term increases over the range of the president’s extremeness in the data. As can be seen, presidential extremeness covers many different outcomes of midterm elections. A president with the maximum observed extremeness score is predicted to receive a first-term penalty of 33.5%—clearly larger than the average penalty of 20.4% (Model 1 in table 4). As expected, if the president abandons the ideological center, torn partisans voice their discontent at midterm. On the other end of the scale, a president with the minimum observed extremeness score is not predicted to receive much of a first-term penalty. In this case, the president remains fairly close to torn partisans, and there is little reason for voice—even if the widening confidence interval suggests more factors are at play.

Turning to Figure 3b, we see how the negative effect of Gallup’s midterm prediction increases over the range of party polarization in the data. Again, the scope of possible outcomes is comprehensive. For maximum observed polarization, the penalty for one additional percent in the polls is 3.26%—almost twice the average penalty of 1.68% (Model 1 in table 4). As expected, balancing the power of a rising in-party is considered particularly important under high polarization. As polarization decreases, this effect then diminishes toward the observed minimum. While the widening confidence interval again suggests more complexity, balancing becomes less necessary when the parties are more in agreement regarding policy, and torn partisans appear less concerned about the general trend.

Overall, two of the interactions thus work as expected (President’s extremeness × President’s first term, and Party polarization × Gallup estimate), whereas the other two are negligible: their coefficients are about 60% smaller than their respective counterparts, carry the wrong sign, and come with higher uncertainty (President’s extremeness × Gallup estimate, and Party polarization × President’s first term). This pattern harmonizes well with the larger theory, because it pairs the variables according to their target arena. Presidential term and extremeness both refer to the executive arena and thus condition voice, which is a signal to the president per se. Expected in-party performance and party polarization both refer to the legislative arena and thus condition balancing, which seeks direct policy impact in Congress. Overall, then, policy-oriented behavior appears to react in quite rational ways even to nuanced policy incentives.

The remaining two models of table 5 present robustness tests. Model 7 controls for possible contingency of balancing–voice on the preceding presidential election. First, to test for regression to the mean, it includes the lagged in-party vote of torn partisans. Second, to allow balancing in presidential and midterm years to be correlated, the model includes the final poll lead of the winning presidential candidate (Gallup, from Peters and Woolley 2020). The larger this lead, the more reason there was for anticipatory balancing, and the less balancing would be necessary in the following midterm election. However, the estimated coefficients of both variables are small and insignificant. This suggests that midterm elections really are a “new game”: torn partisans seek to balance the president’s power in the incoming Congress and to voice their concerns for the following presidential election. They do not somehow revisit their own behavior from the past. ¹⁴

Finally, Model 8 includes a set of standard macro indicators—the overall in-party vote in the preceding presidential year to model the effect of withdrawn coattails, presidential job approval (Gallup, from Peters and Woolley 2020) to test for spillover on the congressional vote, and the differential between the in- and out-party on Gallup’s “keeping the country prosperous” question (Gallup 2020) to cover valence. This model is essentially a “placebo” test, because its independent variables are generally used to predict vote shares in the electorate at large. Torn partisans in particular, however, are expected to focus on policy moderation. As can be seen in table 5, this expectation is confirmed: none of the additional predictors has a significant effect, and additional variance explained is relatively little for three extra variables (compared to Model 1 in table 4; the adjusted R-squared even declines). However, this does not mean that torn partisans ignore all these fundamental factors—just that they do not matter directly. Presidential performance, withdrawn coattails, and party valence all affect the expected performance of the in-party (as reflected in a multiple correlation of .63 with the Gallup prediction), and torn partisans then react to that expectation. Thus, the “fundamentals” matter indirectly, through their impact on other voters. Focusing on expected performance allows torn partisans to exercise balancing and voice with high confidence that their behavior will work in the desired direction.

Conclusion

Midterm elections have many faces. One of them is an opportunity to adjust the course of federal policy halfway into a presidential term. My findings show that centrist voters who share incentives for policy-oriented behavior typically turn against the in-party at midterm, and the frequency of this behavior varies systematically with the size of the incentives. A measurable impact requires the concurrence of two mechanisms of policy influence: a direct mechanism, which strengthens the out-party in Congress to “balance” the president’s policy impact, and an indirect mechanism, by which midterm voting serves to “voice” dissatisfaction as a signal to the president. The theoretical relation of these mechanisms, as well as their empirical interaction, may help refine our understanding of US congressional elections as parts of a sequence and contribute a new analytical lens to electoral research in general.

My analysis also suggests a reason why explanations based on policy-oriented voting are not more prominent in the midterm literature. The rationality of balancing and voice implies that these mechanisms tend to hurt the in-party in situations where conventional wisdom would not necessarily expect it. For example, voice is used primarily in the first presidential term, where it can affect future executive performance. In the second term—notwithstanding its “six-year itch”—voice behavior is rationally muted. Similarly, the negative impact of balancing is correlated positively with the in-party’s expected performance. I have argued that this is a rational reaction of “torn partisans” trying to achieve a centrist policy outcome. Thorough instrumental interpretation is key, because balancing–voice really is a means to an end. In fact, if this end is better served by giving the president a helping hand, my analysis shows that torn partisans do just that. To isolate these mechanisms, it was necessary to study the individual level. On the macro level, the impact of balancing–voice tends to be absorbed by the larger trends that this behavior seeks to curb.

In the context of the literature, this interpretation means that other theories of the midterm loss definitely have their place. In fact, my finding that torn partisans generally vote against the trend implies a source of variation in party support that is not balancing/voice. Policy-oriented voting naturally coexists and interacts with other motivations to produce the overall cycle of electoral support. While this perspective may frustrate the search for “grand” theory, it does help bridge the gaps between different modeling traditions.

A task left for future research is to explore the prevalence of balancing–voice relative to other motivations. By relying on widely available and well-studied measures of ideology and partisanship, my study maximized historical coverage and analytical precision. But one may well think of other incentives that are contingent to a degree on which party holds the presidency. For example, voters may care about moderate policy on specific issues, even if they do not think of them as ideologically constrained (e.g., Ansolabehere, Rodden, and Snyder 2008). Alternatively, voters may produce moderation across issues by demanding different policies from each party (e.g., Petrocik 1996). Another frontier faces nonpositional factors. Voice, with its focus on “performance” and “quality” (Hirschman 1970), is inherently open to matters of policy implementation, pledge fulfillment, or fiscal prudence. Balancing, too, extends beyond the spatial model if voters are “cognitive Madisonians” who fear the tyranny of concentrated power (e.g., Lewis-Beck and Nadeau 2004). Overall, the motivation to counter the electoral trend may turn out to be compatible with a group size considerably larger than my conservative estimate.

Other opportunities lie in comparative politics. The theorizing of my study liberally integrated comparative ideas into Americanist work. It is therefore only logical to consider the implications of the American findings for elections in other countries. For example, counter-trend voting may even be found in systems where a presidential penalty is not apparent at all—such as in France, where legislative elections are held just weeks after a presidential ballot and voters tend to strengthen the party of the newly elected president. My theory suggests that balancing–voice may be at work even under these circumstances, but that torn partisans often fail to outweigh a larger effect of the president’s “honeymoon.” As presidential races in France and elsewhere become more complex, policy-oriented voting may have more immediate consequences for government control. Whether these consequences will favor moderate outcomes is debatable. Comparative work reminds us that policy-oriented behavior can equally call for more extreme target marks (Kedar 2009; also Grofman 1985), and preferences for moderation and polarization may even interact to enhance the policy content of the vote (Weber 2015).

Balancing–voice also has implications for public polling. Although the Gallup polls used for my analysis were generally quite accurate, their accuracy varied systematically with the reported level of in-party support: the higher this level, the more the poll tended to overpredict the actual result, and the lower the level, the more it tended to underpredict. ¹⁵ My theory offers an explanation for this pattern, because certain voters react to the polls precisely to prevent them from becoming true—possibly contributing to the common decline in poll leads toward Election Day (e.g., Campbell 1996). This is the opposite of the well-known “bandwagon effect,” which has been the main hypothesis in research on feedback effects of the polls on public opinion (for an overview, see Hardmeier 2008).

In more general terms, my findings suggest that part of the notorious stalemate in Washington is due to a positive popular preference for policy moderation. While one may lament the consequences of this behavior for the clout of the federal government, my analysis advises against laying the blame on an ill-informed or uncoordinated electorate. At a time of strong antagonism in American politics, this conclusion may even give some increasingly rare reason to be optimistic regarding the rationality of electoral democracy.