Winning Elections with Unpopular Policies: Valence Advantage and Single-Party Dominance in Japan Open Access

The puzzle of LDP dominance has captured the attention of generations of scholars of comparative politics (e.g., Krauss and Pekkanen, 2011; Pempel, 1990; Scheiner, 2006; Thayer, 1969), in part because it continues to win elections despite a major electoral reform intended to facilitate two-party competition and a more frequent alternation in power.

In elections for the House of Representatives, multiple parties compete in a multidimensional issue space under mixed-member majoritarian (MMM) electoral rules combining single-member districts (SMDs) and closed-list proportional representation (PR). This system has been in place since the 1996 general election, replacing the single non-transferable vote (SNTV) system used from 1947 to 1993. The LDP has consistently won seat majorities (from substantial vote pluralities) under both systems, with the exception of three elections.

In the 1993 general election, the LDP lost its majority for the first time (but was still the plurality winner). An eight-party coalition, excluding the LDP and the Japanese Communist Party (JCP), introduced the MMM electoral system before breaking apart. The LDP reentered government in 1994, ruling in coalition with smaller parties — since 2003, exclusively with Komeito, a small religious party (Ehrhardt et al., 2014).⁵ Between 1996 and 2009, a predominantly two-party system gradually took shape in the SMD tier, with the opposition coalescing around the center-left Democratic Party of Japan (DPJ), and smaller parties surviving mostly in the PR tier.

The LDP lost to the DPJ in the 2009 general election, resulting in its second period out of power (and the first and only time it lost a plurality of votes). However, the DPJ struggled to handle the 2011 Fukushima nuclear crisis and other issues (Kushida and Lipscy, 2013). In the 2012 general election, the LDP-Komeito coalition regained power, and also won large majorities in the next three general elections in 2014, 2017, and 2021 (Pekkanen et al., 2013, 2016, 2018, 2023). The DPJ tried to rebrand but ultimately split apart before the 2017 election, and the opposition entered a period of reorganization. Many former DPJ members eventually joined the Constitutional Democratic Party (CDP), created in 2017; others joined the short-lived Party of Hope or the Democratic Party for the People (DPP).⁶ Another new party, the Japan Innovation Party (JIP), was formed in 2012 with a base of support in the Kinki region (especially Osaka).

In this period, the LDP appeared secure in its renewed dominance. But beginning in 2022, two major scandals emerged and began to weaken the party’s image. The first scandal involved the revelation of close financial ties between many LDP members and the Unification Church (following the assassination of former Prime Minister Shinzo Abe, which was motivated by the suspect’s grievance over his mother’s relationship to the organization). The second scandal involved unreported factional slush funds distributed to numerous LDP members, and resulted in Prime Minister Fumio Kishida’s decision not to seek another term as party leader in 2024. His successor, Prime Minister Shigeru Ishiba, called a snap election in October 2024 in an attempt to restore public trust (or minimize losses). Although the LDP still managed to win the most votes and seats, the LDP-Komeito coalition fell short of a majority, resulting in a minority government (McElwain et al., 2025).

Scholars have proposed several explanations for the LDP’s dominance over time. These include the fragmentation of the opposition (e.g., Christensen, 2000; Reed and Bolland, 1999; Scheiner et al., 2016), structural advantages in campaign laws and the strategic timing of elections (McClean, 2021; McElwain, 2008), and (since the 2000s) the LDP’s pre-electoral coalition with Komeito (Liff and Maeda, 2019).⁷ Two other explanations are closely connected to our theoretical framework and empirical tests: flexible policy adaptation and clientelism.

The flexible policy adaptation argument is that the LDP operates as a catch-all party that changes its policies to meet voter demand, sometimes by co-opting the most popular policies of the opposition (e.g., Calder, 1989; Kohei et al., 1991; Muramatsu and Krauss, 1987; Pempel, 1982).⁸ This kind of argument implies that the LDP wins elections because it proposes popular policies or, at least, policies that are popular with enough voters to win, given the other factors just described. Conversely, it could suggest that the opposition parties’ alternative policy proposals are too extreme or unrealistic for voters to get behind (e.g., Kohno, 1997; Maeda, 2012). This kind of situation — dominant party flexibility vs. opposition extremism — would accord with some comparative theories of policy-positioning strategies in systems where one party has a valence advantage (e.g., Calvo and Murillo, 2019; Greene, 2007; Groseclose, 2001; Riker, 1976; Schofield and Sened, 2006). The clientelism argument can be considered generally as the LDP’s advantage in capturing organized votes. It posits that the LDP wins elections due to its control over government resources and ability to reward geographical regions and organized interests (such as agricultural industries) with redistributive benefits in exchange for their support (e.g., Catalinac, 2025; Horiuchi and Saito, 2010; Reed, 2022). The LDP’s control over fiscal spending can also improve its ability to recruit high-quality candidates (Scheiner, 2006). If the LDP’s core support is based on these kinds of transactional redistributive policies, then it could potentially secure enough votes to win even if it deviates from its own (non-organized) supporters’ preferences on policies that are not directly connected to redistribution. Targeted benefits and strong organizational ties can also function as sources of non-policy endowments for advantaged parties like the LDP (Calvo and Murillo, 2019).

While existing explanations are plausible, we argue that previous empirical tests have been insufficient. Some studies assume that a segment of voters (e.g., farmers) prefer certain types of policies (e.g., agricultural protectionism) without subjecting these assumptions to rigorous empirical analysis (e.g., Horiuchi and Saito, 2010). Others focus on particular policy instruments (such as fiscal transfers) that the LDP controls (e.g., Catalinac, 2025), without directly testing whether these policies actually motivate individual-level vote choices.

Furthermore, existing studies do not examine how voters’ choices connect to preferences for the actual policy proposals that parties present to voters as manifestos during election campaigns. Since 2003, when parties first began producing such manifestos, the media has often interpreted the LDP’s victories as a public endorsement of its policies.⁹ But these claims have no strong empirical foundation (Horiuchi et al., 2018). The existing arguments about LDP policy flexibility (and, conversely, opposition policy extremism) also focus exclusively on the SNTV period (Calder, 1989; Kohei et al., 1991; Kohno, 1997; Maeda, 2012; Muramatsu and Krauss, 1987; Pempel, 1982). Our approach in this study allows us to measure voters’ multidimensional policy preferences and empirically investigate how they are associated with vote choices in contemporary campaigns.

Our approach also improves upon the existing research on LDP dominance by situating the case within the general framework of spatial models of voting. Classical spatial models assume that utility-maximizing voters will support parties based on their policy preferences and that office-seeking parties will propose policies that attract a majority of the electorate (e.g., Downs, 1957). But as long as multiple issues are at stake, opposition parties should be able to propose an alternative set of policies to attract voters and challenge the incumbent’s status (McKelvey, 1976, 1979; Schofield, 1978). Although electoral competition in Japan is multidimensional, no other party has become a stable alternative to the LDP. This is particularly puzzling given that a large portion of voters — so-called “floating voters,” often more than half of the electorate — do not support any party (Tanaka and Martin, 2003).¹⁰ These voters might be expected to be persuadable on the basis of alternative parties’ policy appeals. If they are not, is it because valence considerations better explain their behavior?

To motivate our research design and subsequent analysis, we adopt a standard random utility model of discrete choice (Alvarez et al., 2000; McFadden, 1973). Voters indexed by i choose one of the J parties indexed by j in a closed-list PR contest.¹¹ A categorical variable Z_i ∈ {1, … , J } denotes this observed vote choice outcome. Voters choose the party that will yield the highest overall utility,

where U_ij is a voter i’s utility for party j. The utility U_ij may include policy and non-policy considerations. We use boldface to indicate a J × 1 vector, where each element corresponds to a party.

When inferring the unobserved utility, we treat U_i as a random variable, and we decompose its systematic component into our quantities of interest: a valence component and a policy bundle component,

where α_j is party j’s valence advantage, β is the weight voters put on policy in their voting decisions, W_ij is the utility voter i derives from party j’s policies, and ε_i represents the remaining random component. If α_j = 0 for all parties j, a voter’s decision would be based solely on the policy bundle offered, i.e., E[Z_i] = arg max{W_i1, W_i2, …, W_iJ }.

The policy component W_ij captures all issues that are spatial (in other words, positional). If movements in the space make one segment of voters better off and another segment of voters worse off, this counts as policy. For example, some voters might prefer a party that promotes the use of nuclear power, while others might prefer a party that promotes the deactivation of nuclear power plants — a key source of policy disagreement since the Fukushima crisis. The policies considered during the election aggregate to a value of W_ij, which represents a voter i’s preference for the bundle of policies that party j proposes.

If voters do not choose parties based on their most preferred policy bundle, what other aspects of parties do they consider? Existing theories of voting behavior introduce the concept of valence into the model to capture nonspatial preferences, and that is captured by the α here. Following Stokes (1963), valence is the component of a voter’s evaluation of a party that does not depend on the spatial distance between the party’s policies and the voter’s policy preferences (see also Ansolabehere and Snyder, 2000; Schofield, 2003, 2004, 2007). Scholars have labeled a host of qualities such as trust, credibility, and human capital as valence considerations, reasoning that voters should consider them to be universally valuable (i.e., not spatial/positional) and independent of policy (e.g., Adams et al., 2005; Clark, 2013; Green and Jennings, 2017; Schofield and Sened, 2006). In the case of Japan, valence might include a governing party’s perceived trustworthiness (absence of corruption), competence in handling the economy and national security, or leadership following the Fukushima crisis or the COVID-19 pandemic (e.g., Lukner and Sakaki, 2018; Pekkanen et al., 2023).

As noted earlier, the theoretical literature produces conflicting predictions about the behavior of a valence-advantaged party, and predictions get increasingly ambiguous with more than one issue dimension. A common theory predicts that a party (or candidate) with a valence advantage will take moderate positions, while disadvantaged parties will be forced to look for votes at the extremes (e.g., Calvo and Murillo, 2019; Greene, 2007; Groseclose, 2001; Riker, 1976; Schofield and Sened, 2006). In contrast, other theories and empirical tests suggest it is the low-valence candidates or parties that moderate more (e.g., Adams and Merrill, 2009; Buisseret and Van Weelden, 2022; Clark, 2013; Zur, 2021). Broadly speaking, models where the parties are uncertain about the position of the median voter produce the first result, whereas models where there is some uncertainty in the valence advantage produce the second result.¹² Our study provides an empirical answer to the question of balance between policy and valence in Japan’s multidimensional issue space. But three remarks are in order on how we construct our quantities of interest. First, our distinction between policy and valence follows standard formal models: essentially, valence is any systematic attribute that is not explained by voters’ preferences for positional issues in a multidimensional space. Defined this way, an attribute must be shared across all voters to be considered valence (i.e., α, the parameter that governs valence, is not indexed by voter i).

Second, it is reasonable in our setting to treat a party as having a single unified platform. This is because Japanese electoral competition is increasingly nationalized and based on party considerations rather than candidate considerations (Hamzawi, 2022; McElwain, 2012; Reed et al., 2012). In a more candidate-centered electoral system like the United States, defining valence as uniform within a party may be overly simplistic because candidates of the same party vary in their personal valence attributes (e.g., Stone and Simas, 2010). Furthermore, our primary analyses use the intended vote choices in the closed-list PR tier of the MMM electoral system, where voters choose between parties rather than candidates.

Third, our setup is agnostic to the number of dimensions in the policy space or the functional form that generates the policy utility W_ij. Classical spatial models often make assumptions about the issue space to generate precise predictions — for example, that policy dis-utility is the sum of squared distances in one or more dimensions (e.g., Enelow and Hinich, 1984; Kedar, 2005; MacDonald et al., 1991). Instead, we take a more flexible and data-driven approach to inferring preferences for policy bundles.

We fielded nationwide surveys during the campaign periods for the House of Representatives elections on October 22, 2017; October 31, 2021; and October 27, 2024.¹³ We collected 6,065 responses in 2017 and 3,675 responses in 2021, using a Qualtrics panel. For the 2024 survey, we collected 4,095 responses using a PureSpectrum panel. For our surveys, we recruited a sample covering all 289 SMDs and a demographic composition matching official census statistics. Before the conjoint exercises, we asked respondents several basic demographic questions (gender, age, prefecture and SMD of residence, household income, and level of education). Following the conjoint exercises, we asked about ideology and partisanship, vote intention in the SMD and PR tiers, and support for the incumbent government, among other questions about political attitudes and trust in the parties and party leaders.

The LDP won a plurality of votes in all three elections — approximately 34% of PR votes in 2017 and 2021, and 27% in 2024. Combined with substantial pluralities in the SMD tier, these vote shares translated into seat majorities in 2017 (284 of 465 seats) and 2021 (259 seats), but fell short of a majority in 2024 (191 seats). The loss of 68 seats in 2024 was the second largest loss for the LDP in its 70-year history. The other main parties that competed in the elections were Komeito, the center-left CDP, the far-left JCP, and the centrist (or center-right) JIP, Party of Hope (in 2017), and DPP (in 2021 and 2024).¹⁴

The main feature of each survey is a multidimensional preference elicitation through a conjoint design. Respondents were shown a table containing two hypothetical party manifestos with randomized positions on the five (seven in 2024) most important policy issues discussed in the campaign. In the days preceding each campaign, we carefully followed the policy discussions in national daily newspapers, which often publish conjoint-like tables with summaries of the parties’ positions on major issues. After deciding on the main issues, we created descriptions of each party’s position (levels in the terminology of conjoint analysis) on each issue (attributes), using newspaper summaries and the actual text of published manifestos.¹⁵ Newspaper summaries are useful for our purposes because (1) they concisely and accurately describe the positions of the parties, and (2) voters are familiar with such summaries, increasing the ecological validity of our survey design.

The major issues included economic policy, security policy, and energy policy in all three elections, as well as constitutional revision (2017 and 2024), consumption tax policy (2017), diversity/social policy and COVID-19 policy (2021), and childcare policy, inflation policy, and political reform (2024). The positions were fully randomized in each profile, without cross-attribute or cross-profile constraints, so that less than 0.1% of the profiles we showed were populated entirely by the policies of a single party.¹⁶ This mitigates the risk that a respondent might associate a given policy profile with a specific party and base their choice on non-policy considerations connected to that party rather than the combination of policies presented.

For each pair of hypothetical manifestos, respondents were asked: “Imagine, hypothetically, that the following two parties were nominating candidates in this election. Which party would you support? Even if you are not entirely sure, please indicate which of the two you would be more inclined to support.”¹⁷ The respondents then registered their preference for one of the profiles, and this exercise was repeated 20 times in 2017 and 6 times in 2021 and 2024.¹⁸

The 2021 and 2024 surveys additionally included an experiment that randomly assigned party labels to a second set of conjoint exercises that were otherwise identical to the first set of exercises (i.e., without labels) each respondent considered. In this second set of exercises, one label (randomly assigned to a profile in each exercise) was always the LDP; the other label was always a different randomly assigned party.¹⁹ The label of the non-LDP party was randomly assigned for each task and for each respondent. This allows us to analyze within-respondent changes in the preference over conjoint choices between the two sets of exercises when a party label is exogenously added to a policy profile. The idea behind this design is that party labels provide low-cost informational shortcuts to voters regarding various considerations, including policy and valence (e.g., Downs, 1957; Kirkland and Coppock, 2018; Kobayashi and Yokoyama, 2018).

The party label treatment therefore potentially captures the combination of two components: (1) any policy -based support for the party that was not captured through the main set of issues included in the conjoint exercises (i.e., because the policy issue was not contested in campaign debates); and (2) voters’ non-policy -based support for the party, such as that arising from valence. Our experimental design facilitates the decomposition of these factors because the party-label effect is measured for each individual respondent. We regress the individual party-label effect on respondents’ attitudes related to policy and nonpolicy attributes of the LDP. As we will show, our decomposition suggests that most of the party-label effect is based on non-policy considerations, especially trust in the LDP.

In the 2024 survey, we further included an open-ended question asking respondents to explain the reasons behind their reported level of trust in the LDP in their own words. As the concept of trust is nebulous and can mean different things to different voters in different election contexts (e.g., Devine et al., 2024; Hetherington, 1998; Levi and Stoker, 2000; Lukner and Sakaki, 2018; Mayer et al., 1995; Parker, 1989), these responses help us to qualitatively explore and interpret the role of trust as a valence consideration in LDP support. While we pre-registered analysis plans for many of our analyses with the Open Science Framework for the 2021 and 2024 surveys, our pre-registration covers only part of the results we present, so our analyses as a whole should not be considered preregistered.²⁰

A common challenge in measuring voters’ preferences is that self-reported preferences are unreliable (or endogenous to partisanship) and may reflect a mix of attitudes toward policy and valence attributes (e.g., Tomz and van Houweling, 2008). For example, imagine asking a voter in the United States whether they agree or disagree that the Democratic Party is “too extreme in their positions.” Respondents may be uninformed about the policies Democrats propose (Fowler and Margolis, 2014), or they may use their own partisanship as a heuristic to make up their minds (Lenz, 2012).

This is where our conjoint design is particularly useful. This section describes how we develop a novel approach to modeling conjoint data with respondents’ demographics in order to estimate policy utility for any policy bundle within each demographic segment. We then examine whether the estimated policy utility for the parties’ actual policy bundles is associated with respondents’ reported party vote intentions.

We posit that respondents in our main conjoint experiment choose the policy bundle that yields the higher utility, without valence considerations. Intuitively, voters put positive weight on policy bundles that include issue positions they agree with, and this weight increases by the relative importance of the issue to each voter. Because these profiles contain no party labels, we can equate the preference over profiles as realizations of the respondent’s latent preference over policy (denoted by W_ij in the previous section) rather than simply a partisan preference.

We start by characterizing the preference over conjoint choices as a function of demographic and geographic covariates associated with the respondent, denoted by X_i, and the profiles presented to that respondent in the mth profile in task k, denoted T_ikm:

where the function of interest C : translates policy profiles T_ikm and voter demographics X_i into a measure of i’s utility for that conjoint choice. We use a tilde ( ) to denote that this utility is predicted from the survey responses, rather than directly observed.

To estimate the function C, we regress the respondent’s choice Y_ikm on the associated policy profiles. Our regression includes policy profiles T_ikm, respondent demographics X_i, pairwise interactions of X_i, and the pairwise interactions of both the policy profiles and those interacted demographics (T_ikmX_iX_i). The conjoint outcome Y_ikm is 1 if respondent i chooses the mth profile in task k, and 0 otherwise. This approach is similar to the computation of the standard AMCE but with two differences. First, we are interested in the fitted value of the function C, not in a particular coefficient within the function. The coefficients we estimate can be considered the importance (or weight) a voter puts on an issue when deciding between policy bundles. Second, while AMCEs are often estimated by running a regression with no interactions, our regression specification is more saturated by numerous interactions. The covariates we use are geographic region (one of eleven PR districts), age, education, income, employment status, gender, and selfreported ideology.²¹ These interactions are core to our methodological interest in allowing individuals with varying demographic characteristics to vary in their preferences.

Our approach is inspired by the marketing literature on market segmentation (Green and DeSarbo, 1979), where researchers are interested in identifying which segment of consumers are more likely to buy a product, not simply which attribute of a product is preferred by the average consumer. Our approach also has the flavor of structural estimation, in which the formal model is taken more literally and estimates of the parameters generate counterfactuals. In political science, conjoint designs are rarely used in this way. An exception is an analysis in Horiuchi et al. (2018, p. 15), who estimate the aggregate utility of each possible policy bundle across the entire sample and report how the parties’ actual policy manifestos rank in popularity (in aggregate). Bansak et al. (2021) also take a similar approach to study European voters’ support for austerity packages, fitting their model to the combination of policies that governments proposed.

However, unlike these two applications, we segment all voters into many more cells with multidimensional preferences. To the extent that Horiuchi et al. (2018) and Bansak et al. (2021) examine heterogeneity, they only focus on one variable: cabinet approval or left-right ideology. Our approach uses a high-dimensional set of demographics X_i to segment voters, generating a more individualized measure of policy utility. Other uses of conjoint analysis generally focus on estimating a quantity that applies to the whole population, whereas we generate individual utility estimates.²²

While our design is conducive to measuring policy utility, we must make two assumptions for identifying parameters. The first is that the set of policies displayed in the conjoint exercises (T) captures the relevant policies shaping voters’ preferences. In other words, we assume that a voter’s policy preference is consistently estimated by the policies we include in the conjoint exercises. When it comes to voting for a party, voters might consider other policies that are not included in party manifestos. Our approach hinges on these other policies not being so salient as to create measurement error in our estimation of policy preferences, and uncorrelated enough with the error term in Equation (2) to avoid biased estimation. Fortunately, direct questions about respondents’ policy priorities in 2021 and 2024 suggest that our chosen policies did not omit any major issue on voters’ minds.²³ Our secondary party-label experiment (in 2021 and 2024) also helps to shed light on the factors that might contribute to party support beyond the policy utilities we measure directly from the conjoint design, including unmeasured policies and various valence considerations.

A related concern with traditional conjoint analyses is that when choosing between hypothetical bundles of parties, respondents may infer something unintended from the profiles shown (Dafoe et al., 2018). For example, they may see that one party has an anti-US stance and infer that party’s stance against China. However, our main goal is to capture policy utility as a whole, rather than to attribute policy utilities to each of the policies shown. As long as these inferences are about policy, our estimates of W capture policy considerations as intended, regardless of whether they are directly shown or indirectly inferred.

The second assumption pertains to our estimation of the high-dimensional function. We assume that we enter individual demographic covariates into the model specification with sufficient granularity to estimate the function C accurately. Our regression model is highly interactive and saturated with pairwise combinations and triple interactions as outlined above, containing more than 10,000 possible coefficients. Because we are modeling the underlying preferences that generate a binary choice, it is natural to assume a logit model for C, i.e., C(T_ikm, X_i) = logit(Pr(Y_ikm = 1 | T_ikm, X_i)). We use a LASSO shrinkage estimator for the regression and generate out-of-sample predictions to avoid overfitting.²⁴

After estimating the conjoint choice model, we restrict our attention to the bundles in which the attribute levels are set to the actual policies that each party proposed during the campaign. Denote these J sets of policy bundles as t and t(j) as party j’s policy attributes. Then, for each respondent i, we generate J fitted values, using t(j) and the respondent’s demographics X_i,

where $W^ij$ represents the estimated preference for party j’s bundle of policies.

Using the estimates of $W~i$⁠, and each respondent’s reported vote intention across J possible parties measured directly in the survey, we then estimate the valence parameters and the policy parameter in Equation (2), by running a conditional logit (or McFadden’s multinominal logit). Unlike a logistic regression, the multinomial regression accounts for the rank ordering of more than two categories. The most common multinomial model assumes that the errors follow an extreme-value distribution (McFadden, 1973). This assumption allows us to specify the probability of party choice through a softmax transformation, that is, as a ratio of each choice’s overall utility scaled by an exponential function,

In this regression, α_j is a global mean that represents the difference in choice probabilities on the logit scale. To identify parameters, we fix the baseline level j to the LDP so that α_ldp = 0 and other intercepts are relative to the baseline. We estimate a single coefficient β representing the weight that policy utility contributes to a party vote choice. In the terminology of discrete choice modeling, W_ij is an alternative-specific rather than a respondent-specific variable because it varies by j in addition to i. This makes the multinomial regression into a conditional logit model, which is standard in studying vote choice in multiparty elections (Adams et al., 2005; Alvarez et al., 2000; Calvo and Murillo, 2019).

This characterization of the choice probabilities induces a property known as the Independence of Irrelevant Alternatives (IIA). It implies that the ratio of the choice probabilities for any two options j and j^t, U_j/U_j^l , is constant regardless of the valuations of the other options. This is a key mathematical property of Equation (5) that makes models of categorical choice coherent (Schofield, 2004), but it may not reflect the choice probabilities in some real-world settings (Alvarez and Nagler, 1998). Other estimation methods relax the IIA assumption but inevitably introduce other, similarly untestable assumptions.²⁵ Given these difficulties, we present estimates from the most commonly used conditional logit model.

We acknowledge two limitations in our estimation strategy. First, we can only account for uncertainty in Equation (5), while our overall approach technically encompasses two sources of uncertainty: the first comes from estimating policy utility W_i in Equation (4), and the second from estimating party vote choice in Equation (5). The prediction uncertainty in W_i is not included in our estimations from the conditional logit model (Equation (5)). Obtaining proper prediction uncertainties is known to be difficult even with a fully Bayesian LASSO (Kyung et al., 2010, p. 376). If there is considerable measurement error in W_i, that means our estimates of policy voting would be biased toward zero. Nevertheless, we do not expect the uncertainty in our estimated policy utilities to be any worse than the uncertainty in self-reported left-right ideological positions used in traditional studies of spatial voting.

Second, our method characterizing vote choice is a simple separation, assigning anything not explained by policy W_ij to the valence parameter and mean-zero noise. Non-policy considerations, such as a party’s perceived trustworthiness or competence, could be important components of the valence parameter, which is otherwise black-boxed in Equation (5). Later, we will attempt to decompose the valence effect into the subcomponents hypothesized in the literature. A challenge here, however, is that interpreting each component causally is not possible without strong independence assumptions. This problem is analogous to the challenge of mediation analysis. Trust could be endogenous to policy — e.g., respondents might trust a party more if their policies are congruent (Orr et al., 2023). Therefore, we only feature policy utility in our baseline model specification.

We begin by showing the distributions of our estimates of voters’ utility from each party’s policy bundle. We then focus on the association between this policy utility measure W_i, and reported vote intentions. The estimated coefficient (β) represents the contribution of policy, and the intercepts of the conditional logit regression represent the contribution of valence. In our subsequent analysis using the party-label experiment, we find that this contribution of the valence term warrants a causal interpretation.

Figure 1 shows the distributions of the estimated utilities for policy bundles in 2017, 2021, and 2024. For each year, we rescale each respondent’s policy utility into a Z-score such that a one-unit increase represents a one-standarddeviation increase from the average respondent’s preference for the LDP’s policy bundle. The mean of the LDP’s policy utility is 0 by design because of this standardization. The means of the other parties’ policy utilities are noted in the figure. Parties are ordered approximately according to the conventional left-right scale (Miwa, 2015; Proksch et al., 2011), with the JCP on the left (at the bottom in the figure) and the LDP on the right (at the top in the figure).²⁶ In the 2017 and 2021 elections, we estimate that the modal voter derived less utility from the LDP’s policies than almost all other parties, despite the party winning comfortable majorities.²⁷ In 2021, the JIP, a center-right regional party (strongest in the Kinki region) that campaigns on reformist policies, enjoyed the highest average utility: on average 0.88 standard deviations higher than the LDP’s policy bundle. It also had one of the largest leads (1.12 standard deviations) ahead of the LDP in 2017. Two other new centrist parties, Party of Hope (in 2017) and DPP (in 2021), similarly show higher average utilities than the LDP. The center-left CDP and far-left JCP follow, with less popularity than the JIP but still more than the LDP. For over half of the respondents in both of these years, the LDP’s policy bundle gave the lowest utility among all other alternatives.

In the 2024 election, in contrast, Komeito’s policy bundle was the most popular, followed by the JIP. Meanwhile, the LDP’s policy bundle was no longer the least popular, and the leftist parties (JCP and CDP) proposed policy bundles that enjoyed the lowest average utility among the respondents. These changes in the estimated policy utilities are notable, since the LDP (and Komeito) lost votes and seats in 2024, while the CDP gained the most votes and seats. If policy preferences were purely driving vote choice, in other words, our 2024 estimates would be hard to reconcile with the election result.²⁸

Our estimates for policy bundle preferences correlate roughly in the way one would expect: voters with a high utility for a left-wing (right-wing) party also have higher utilities for other left-wing (right-wing) parties, and voters who place themselves to the left (right) on a traditional left-right ideological scale have higher utilities for left-wing (right-wing) parties.²⁹ This is reassuring evidence that our proposed methodological approach meaningfully captures policy preferences while also capturing more within-demographic and within-party variation than traditional measures.

Despite the lower popularity of the LDP’s policy bundles, respondents prefer the LDP for their party vote choice by wide margins. We use the PR vote as the outcome variable for this analysis for three reasons. First, the PR tier is fully contested by all parties, allowing us to fit the conditional logit model with fewer IIA assumptions. Second, there is less incentive for strategic voting (not voting for one’s favorite party because the vote would be wasted) in the PR tier because district magnitudes are large and every vote matters. Third, it is the most likely place to find policy voting, if it exists, since voters choose a party rather than a candidate. That said, existing research on Japanese voting behavior finds that contests in SMDs are also increasingly based on party considerations more so than candidate considerations (Hamzawi, 2022; Reed et al., 2012). We find consistent results for respondents’ SMD vote intentions, but caveat that this estimation is complicated by the uneven entry of parties across SMDs.³⁰

Table 1 classifies respondents by the party bundle that gave them the highest policy utility (in rows) and the party they reported intending to vote for in the PR tier (in columns), pooling across all three election surveys. The numbers in each matrix show the row percentages: the sum of these in each row is 100%. The leftmost column gives the number of respondents who most preferred each policy’s bundle. We dropped respondents who planned to abstain from voting or vote for a fringe party (both roughly 8% of the sample, totaling a sixth of the sample). If respondents’ policy preferences perfectly corresponded to their vote choices, all observations would fall in the diagonal set of bolded cells. However, this is not what we find. Among respondents who derived the highest policy utility from an opposition party bundle, anywhere from 25% to 63% (in 2017), 15% to 47% (in 2021), and 22% to 55% (in 2024) reported an intention to vote for the LDP (Table C.6).

For example, the first row in Table 1 shows that among the 2,053 respondents who most preferred the far-left JCP’s policy bundle in any of the elections, 22% reported an intention to vote for the LDP. About 29% reported an intention to vote for the other leftist party, the CDP. Only 16% reported an intention to vote for the JCP. The patterns for the JIP are also noteworthy. Across the three elections, more respondents preferred the JIP’s policies than any other party’s bundle (3,208). Nevertheless, only 16% of these respondents reported an intention to vote for the JIP. In contrast, nearly three times as many of them intended to vote for the LDP. The bottom row of Table 1 sums across all three elections. While only about 7% of respondents preferred the LDP’s policy bundle in any election (687 of 10,720 respondents), 38% nevertheless intended to vote for the LDP.

In sum, Table 1 shows two things simultaneously. First, some voters do vote as if they were policy-utility-maximizers, but others do not. Second, across segments, the LDP captured a substantial amount of votes, even among those for whom the LDP’s policy bundle was not the most preferred.

A conditional logit regression following Equation (5) better captures the weight voters put on policy versus valence. Rather than focusing only on the highest-scoring bundle as in Table 1, it accounts for the rank ordering of the preferences for all policy bundles. We regress the categorical variable of each respondent’s intended party vote choice on the matrix of policy bundle utilities, with each possible choice linked to a policy utility for that specific party that varies at the respondent level. Figure 2 visualizes the results of the regression as fitted probabilities. For each respondent, we can obtain $Pr^(Zi=j)$for each party j by taking the vector of observed policy utilities, and combining it with the posterior means of the estimated coefficients following Equation (5).

The first noteworthy pattern in Figure 2 is the positive association between policy utility and probability of vote choice. This indicates the positive estimate of β, i.e., the contribution of policy utility to the intended party vote choice. The estimated coefficient on policy, β, is 0.66 for the 2017 election and 0.52 for the 2021 election. In other words, a one-standard-deviation increase in the policy preference for party j is associated with a 0.66 or 0.52 increase in U_ij in these years.³¹ The coefficient is statistically significant, with standard errors of 0.02–0.03 in any year.

The second pattern to note is the differing intercepts of the predicted probabilities for each party. The predicted value on the vertical axis evaluated at a policy utility of 0, shown with the horizontal lines in Figure 2, roughly corresponds to each party’s valence advantage on the probability scale. Consider the following examples. In each election, the probability that a voter with an average preference for the LDP’s policy bundle (⁠$(W~i,LDP=0)$⁠) intends to vote for the LDP is markedly high (35% in 2017, 43% in 2021). In contrast, the probability that a voter with the same utility for the JCP policy bundle (⁠$(W~i,LDP=0)$⁠) intends to vote for the JCP is only 3–8%. In other words, holding individuals’ preferences for parties’ policy bundles fixed at 0, individuals prefer the LDP by a wide margin when choosing between parties as a whole.³² The intercepts of the regression show the same pattern: the coefficient on the JCP intercept (where the baseline category is the LDP) has a point estimate of approximately −2 on the logit scale.

While policy preferences and valence advantage are both clearly important in respondent’s intended vote choices, valence alone is more predictive of the ultimate vote choice compared to policy alone. We quantified the goodness of fit by the correspondence between the model’s fitted probability using coefficients fit from a training set and the actual votes. An intercept-only (i.e., valence only) conditional logit model predicted vote choices correctly for 43% of 2021 respondents, while a model including policy utility and its intercepts forced to be 0 only gave correct predictions for 22% of the sample.

In the 2024 results, two things changed. First, the positive relationship between policy preference and vote choice weakened, as seen from the attenuated slopes in Figure 2. The coefficient on policy was 0.16 in 2024, which, although statistically distinguishable from zero, represents a threefold decrease from 2021 levels. Second, the LDP’s valence advantage over its opposition shrank, as seen from the drop in the LDP’s intercept and the twofold increase in the intercept for the center-left CDP.

Recall that in 2024, the LDP’s policy bundle became more popular, and the CDP’s policy bundle became less popular. If voters were maximizing policy utility, they should have voted even more for the LDP than they did in 2021. Yet, we see the opposite. Fewer survey respondents reported voting for the LDP, and the LDP lost more than a quarter of its seats. Consistent with the hypothesis that valence trumps policy, we find that reported trust in the LDP as a party, a measure of valence which we discuss in a subsequent section, decreased markedly in 2024.³³

Any disconnect between policy utility and vote choice could arise from a lack of understanding of parties’ policies rather than valence. To explore this possibility, the 2021 and 2024 surveys included a third set of conjoint exercises (without party labels) asking respondents to choose which policy bundle seemed closest to the LDP.³⁴ The discrepancy between policy preferences and party vote choices occurs even though respondents could reasonably guess which policy bundles were closer to the LDP, when forced to choose. While this also implies that part of our estimates of policy preferences (W ) may contain some valence considerations on the part of respondents (Dafoe et al., 2018), in our case this should downwardly bias the estimates for valence, which we estimate are still substantial.

The estimated policy preference structure in the previous analysis may suffer from omitted variables. We measure voters’ policy preferences based only on the main policy issues discussed in the actual party manifestos released at the start of the election campaigns, and may therefore miss some policy considerations behind party support. For example, right-wing voters might prefer the LDP because they prefer the LDP’s more hawkish position vis-à-vis China, which was not measured in our conjoint design because it was not a contested issue in the campaigns. If such omitted considerations exist, and are correlated with the error term in estimating Equation (5), what we attribute to valence will include unmeasured policy preferences.

We address this concern in three ways. First, in 2021 and 2024, we included a question asking respondents which issues were most important to them in the election. The answers to this question included the policy issues featured in the conjoint design, as well as an “Other” option with an open response field. If there were important issues in the campaigns that we failed to include, we might expect this latter option to be informative. However, only about 3% of respondents in both 2021 and 2024 selected this option.³⁵ This suggests that our design did not omit any issues of strong interest to many respondents.

Second, we re-estimated the model in Equation (5), but with the inclusion of each respondent’s self-reported ideology on the left-right scale as a control, reasoning that this summary measure may capture a major component of remaining policy preferences. After controlling for ideology, the party-level intercepts shrink by 15–30%, indicating that part of our valence effect could indeed be related to unmeasured policies. However, the intercepts remain substantially and statistically significant.³⁶ This indicates that our valence estimates in the previous section cannot be explained away by respondents’ ideology, which may proxy their residual policy considerations.

Finally, we rely on the experimental nature of the party-label experiment in 2021 and 2024. We compute the Average Marginal Party-Label Effect (AMPLE), which is the difference between the marginal means for each attribute level with and without a party label.³⁷ In what follows, we first show the effects of the LDP party label, averaging across all respondents. In the next section, we decompose this total effect using the individual-level variation in AMPLE. We find evidence consistent with the idea that these effects are primarily capturing respondents’ attitudes toward valence attributes.

The left panel of Figure 3(a) shows the estimated marginal means of choosing profiles that include the corresponding policy position in 2021. For example, the bottom row shows the marginal means of choosing a profile that includes the LDP’s position on diversity and social policy (namely, same-sex marriage and the right for married couples to keep separate surnames). The marginal means without the LDP label are in black, whereas those with the label are in blue. The right panel of Figure 3(a) shows the differences in these marginal means in percentage points. Across all five issues in 2021, with six party positions each, the probability of choosing a given profile increases by approximately 8–10% points when the LDP label is attached, compared to the exact same profile without the LDP label. Specifically, the effect size shown in Figure 3(a) (in triangles) ranges from 7.6 to 12.8 percentage points, with a median of 9.1.³⁸

The AMPLE is uniformly positive, even when the policy attributed to the LDP comes from the far-left JCP. When it comes to security policy, for example, the JCP’s position is to abolish the US-Japan Security Treaty, while the LDP’s position is to strengthen the alliance. The two parties’ positions are directly in opposition. The JCP’s position is among the least popular, with a marginal mean of 0.41. But when the LDP label is added, the probability of choosing a profile with this unpopular position increases substantially: the marginal mean is 0.54. This kind of substantial change is also observed with the LDP’s unpopular position on energy policy (restarting nuclear power plants that were deactivated after the 2011 Fukushima nuclear meltdown).

In the 2024 experiment, as shown in Figure 3(b), the LDP’s label effect is still uniformly positive across the seven policy issues and six party positions, but the size of the effect shrank threefold to a median of 3.5 percentage points. In other words, the LDP party label was still a positive “brand,” but it lost some of its appeal among respondents.³⁹

Why does attaching the LDP label motivate so many respondents to change their minds in choosing between alternative policy bundles? Is the LDP’s partylabel effect capturing valence considerations, or is it merely capturing policy considerations unobserved by our design? Moreover, what exactly constitutes valence? As the existing literature on spatial voting acknowledges, valence is a catch-all term that includes everything from campaign-related advantages, like strong networks of organized party activists, to character -based advantages of parties or leaders, like perceived competence and trustworthiness (Adams et al., 2005; Ansolabehere and Snyder, 2000; Calvo and Murillo, 2019; Clark, 2013; Green and Jennings, 2017; Groseclose, 2001; Schofield, 2004; Schofield and Sened, 2006; Stokes, 1963; Stone and Simas, 2010).

To answer these questions, we conduct an exploratory analysis of which types of respondents display larger LDP label effects, using additional questions included in the 2021 and 2024 surveys. We measured, through direct questions following the conjoint exercises, several potential policy and valence factors in party support: respondents’ priorities in vote considerations (e.g., policy, candidate personality, local benefits, workplace benefits), evaluations of parties’ trustworthiness and capability, evaluations of party leaders’ trustworthiness and leadership, direct mobilization appeals from groups (a potential indication of the organized vote), and retrospective evaluations of cabinet performance in different policy issue domains.

In the context of the broader literature on spatial voting, this exercise can be understood as a “horse race” between competing sources of the party-label effect: policy preferences and several potential valence attributes. Many of these variables also speak to the existing explanations for LDP dominance in the Japanese politics literature — for example, a respondent’s consideration of local benefits would hint at clientelism (Catalinac, 2025; Horiuchi and Saito, 2010; Scheiner, 2006), while being asked to vote hints at the mobilization efforts of organized groups, including the religious supporters of Komeito (Ehrhardt et al., 2014; Liff and Maeda, 2019; Reed, 2022). We also consider district-level variables related to clientelism, including per capita intergovernmental fiscal transfers.

We use the individual-level estimate of the party-label effect on preferring a policy bundle (Individual Marginal Party-Label Effect, or IMPLE) as the dependent variable. The IMPLE is an individualized version of the averages shown in Figure 3. Recall that each respondent sees a particular conjoint profile twice, once with and once without a party label. We can modify our notation slightly and denote Y ^Ai=1 as whether respondent i chose the profile labeled as LDP (A = 1) in task k during the second set of tasks, so that it is 1 if the respondent chose the LDP-labeled profile (and 0 if they choose the paired profile that is labeled with another party). Then, denote Y ^Ai=0 as the respondent’s corresponding choice in the first set of tasks, i.e., whether they chose that same profile when there were no party labels. It is 1 if the respondent chose the non-labeled profile that would later be labeled with the LDP, and 0 if the respondent chose the alternative non-labeled profile that would later be labeled with another party. Then,

which takes on one of 13 values and holds the set of profiles considered fixed..⁴⁰

We regress the IMPLE on a host of respondent-level and district-level variables that could be antecedents of respondents’ inclination toward the LDP, as described earlier.⁴¹ To adjust for over 20 coefficients being tested for significance, we use a Bonferroni-adjustment for the standard errors. This analysis is not intended to capture all potential sources of the party-label effect, but identifies some of the respondent-level and district-level characteristics that are most plausibly associated with variation in the party-label effect. It estimates the amount of change in an individual’s party label effect when one of the independent variables changes by one standard deviation, holding the rest constant.

Figure 4 shows the results of this exploratory analysis, pooled across the two surveys (2021 and 2024). The results reveal that, apart from prior party support, the most relevant variable in explaining the cross-sectional variation in the effect of the LDP label is the degree to which respondents trusted the LDP as a party more than other parties. This variable is constructed as a difference in responses to a four-point Likert scale question, “To what extent do you trust each of these parties?”⁴² A one-standard-deviation increase in a respondent’s net trust for the LDP corresponds to a 5-percentage-point increase in the party-label effect, even controlling for prior party allegiance, issue considerations, and domain-specific cabinet approval.⁴³ To further test if trust is a component of valence, we revisited the conditional logit model in our main results. Respondent’s individual trust was a significant predictor of vote choice as well.⁴⁴

It is also worth noting which variables are not relevant predictors. First, although there is a small and significant effect for “perceived policy capability,” which is a valence attribute related to general policy competence, all other variables related to policy considerations are insignificant. If switching profiles in the presence of an LDP label were due to strong preferences for LDP policies that were not included in our conjoint design (such as China policy), we would have expected this policy motivation to be evident in direct questions as well. Second, in the literature on clientelism, the amount of fiscal transfers (pork-barrel spending) is viewed as critical to understanding voting behavior and outcomes (e.g., Catalinac, 2025; Horiuchi and Saito, 2010; Scheiner, 2006). However, the logged amount of per capita fiscal transfers at the SMD-level is not correlated with respondents’ party-label effects (bottom panel of Figure 4).⁴⁵

Related variables measured at the respondent level, such as whether they considered “local benefits” or “workplace benefits” when making voting choices, or whether they were asked to vote, are also insignificant. Finally, a variable measuring how much respondents considered candidates’ personalities (a candidate-level measure of valence) is also insignificant, supporting arguments that recent elections have become mostly party-centered (Hamzawi, 2022; Reed et al., 2012).

What sort of non-policy considerations does the large coefficient on trust in the LDP represent specifically? As noted, the concept of trust can mean any number of things to different voters, including competence, benevolence, and integrity (e.g., Devine et al., 2024; Hetherington, 1998; Levi and Stoker, 2000; Lukner and Sakaki, 2018; Mayer et al., 1995; Parker, 1989). In the 2024 survey, we included an open-ended question that directly asked respondents to explain the response they gave for trust in the LDP. Respondents who reported being distrustful of the LDP were much more likely to mention the money scandals in their open-ended responses.⁴⁶ Those who reported being trustful of the LDP were more likely to mention the cabinet’s ability to implement policy.

Ultimately, trust in the LDP may have multiple and overlapping components. But our analysis suggests that the electoral edge the party has maintained in recent elections, and its loss of support in 2024, might be largely attributed to two perceptions among voters. First, that the party is generally more competent and capable of governing the country compared to the opposition parties. This aspect of trust coincides with the argument that dominant parties like the LDP enjoy a reputation for being the natural “party of government” (Carty, 2022). We find additional evidence for this interpretation with a subgroup analysis focused on the JIP, which controls regional and local governments in the Kinki region, and whose local leaders were credited with competently handling the COVID-19 crisis (Pekkanen et al., 2023). The JIP enjoys larger party-label effects and higher levels of trust among respondents in the Kinki region compared to elsewhere in the country.⁴⁷

Second, the decline of trust in the LDP in 2024 is likely attributable to a perceived erosion in integrity. The 2024 open-ended text responses corroborated the salient news coverage of LDP politicians’ deep financial connections with the Unification Church, and dubious campaign fundraising practices leading up to the election. Neither of these scandals was exposed prior to the 2021 election. Once large swaths of the public started questioning the LDP’s financial integrity, its perceived trustworthiness and vote share declined, without significant changes in the coordination or policy appeal of the opposition parties.

If voters were purely policy-maximizing and parties were purely office-seeking, dominant parties like Japan’s LDP would have a hard time staying in power without routinely shifting policies to meet popular demands. The fact that they do exist in some democracies — and, in the case of the LDP, repeatedly win elections despite proposing unpopular policies — suggests that voters do not choose parties based entirely on policies, but instead are also motivated by non-policy (valence) considerations like trust.

In this study, we have investigated the role of policy and valence in voters’ support for the LDP in recent elections, proposing a novel use of conjoint analysis to measure voters’ utility from each party’s multidimensional policy bundle. In contrast to traditional left-right ideological placements, our conjoint-based measure better captures the policy utility represented by parties competing in a multidimensional issue space. Furthermore, unlike the standard use of conjoint data that focuses on disaggregating the effect of a single component of a profile averaged across voters, we provide a method to estimate an individual voter’s holistic preferences over the bundles of policy offered to them in election campaigns.

Our analysis documents a positive relationship between policy utility and vote choice among Japanese voters in three elections spanning eight years. However, this finding does not paint a complete picture of voting behavior. When the LDP’s policy proposals provided lower utility to the average voter than the policy proposals of the main opposition parties in 2017 and 2021, many voters nevertheless supported the LDP. In other words, policy voting, even with our improved measurement, was not large enough (and the policies of the LDP were not popular enough) to explain parties’ vote outcomes. Moreover, the LDP lost seats in 2024 even though the popularity of its policies improved.

We have attributed the gap between policy preferences and vote choice largely to the LDP’s valence advantage over other parties. Our analysis of individual-level variation in the preference for LDP’s party label suggests that trust in the LDP (potentially comprising competence and integrity) is an important component of the LDP’s valence. The LDP enjoyed the highest trust in 2017 and 2021 when its valence advantage was high, but it is precisely when its trust eroded in 2024 that its valence advantage halved (and it lost its seat majority). That said, our analysis cannot exhaustively explain the meaning of trust, which is also likely to change over time. Future research might examine how other components of valence, such as a history of fiscal transfers or strong organizational ties, may in turn generate trust.

Together, our findings help to explain an important aspect of the longstanding puzzle of LDP dominance in Japan. We argue that the LDP’s dominance in recent elections is partly because it holds a substantial valence advantage over the opposition. Challengers may try to appeal to voters with popular policies, and voters do seem to partly respond to policies they prefer, independent of party label. But our quantification of the policy and valence components of vote choice suggests that popular policy appeals alone are insufficient to overcome the LDP’s valence advantage. Nevertheless, no party’s valence advantage is invincible, and the LDP has lost voters’ trust in the past, as well as during the span of our study.

Our findings also suggest that dominant parties like the LDP can maintain support through a valence-based advantage over the opposition, rather than through proposing popular policies. However, the extent to which these findings travel beyond Japan to other democracies with dominant parties remains to be seen. Other dominant parties may be able to win support based on policy coalitions and populist platforms, for example, and the relationship between policy preferences and vote choice might be higher in the case of new parties for whom valence attributes are still unclear (e.g., Guntermann and Lachat, 2021). Future research should therefore investigate the range of conditions and issues that contribute to strengthening or weakening party valence over time.

There is also an open question as to whether a valence advantage inevitably leads to policy disconnect, or if it might sometimes provide the flexibility to adapt policies to appeal to the majority of voters (e.g., Adams and Merrill, 2009; Buisseret and Van Weelden, 2022; Calvo and Murillo, 2019; Clark, 2013; Greene, 2007; Groseclose, 2001; Zur, 2021). Our modeling approach can be a useful framework for future work that considers the implications of the relative importance of policy and valence in different settings.

We are grateful to Shusei Eshima for his valuable contributions to earlier versions of this study. We acknowledge financial support from Harvard University (Weatherhead Center for International Affairs, Program on US-Japan Relations), Dartmouth College (Faculty of Arts and Sciences), and the University of Pennsylvania (School of Arts and Sciences). For research assistance, we thank Minh Nam Pham. For feedback, we thank Michael Becher, Peter Buisseret, Erin Cikanek, Alexandra Cirone, Max Goplerud, Eric Guntermann, Nahomi Ichino, Kosuke Imai, Junko Kato, Lewis Luartz, Ingrid Mauerer, Cynthia McClintock, Yuki Shiraito, Milan Svolik, Chris Tausanovitch, Teppei Yamamoto, Soichiro Yamauchi, and Anna Yorozuya.

Online Appendix available from: http://dx.doi.org/10.1561/100.00024134_app

Supplementary Material available from: http://dx.doi.org/10.1561/100.00024134_supp