2. Empirical background

The Japanese Parliament has two chambers: the lower house (LH) and the upper house (UH). The LH, which constitutes the focus of this study, holds more power regarding a few important issues. These include electing the prime minister and deliberating budget bills. The term of an LH representative may last up to 4 years, but the prime minister can dissolve the chamber anytime to call for a snap election. The LH uses a mixed-member majoritarian electoral system. The electoral system combines two electoral tiers: single-member majoritarian and proportional representation (PR). Currently, the LH electoral system elects 289 members from the majoritarian tier with SMDs and 176 members from the closed-list PR tier with 11 regional districts.Footnote ¹

Between 1998 and 2016, there were two major parties in the Japanese party system, the Liberal Democratic Party (LDP) and the Democratic Party of Japan (DPJ), as well as a few smaller parties. In most of the SMDs of the LH, only candidates from the two major parties had a serious chance of obtaining electoral victories. There were substantial electoral swings between the two parties, especially in the LH general elections during 2005, 2009, and 2012.Footnote ^2, Footnote ³

The LDP is a conservative party with its support base consisting mainly in rural areas and has been predominant in Japanese politics since its formation in 1955. During its history, the LDP has fallen out of power only twice, each time for a relatively short period (1993–1994 and 2009–2012).

The DPJ was formed in 1998 as a center-left party. The party became a serious contender for the LDP during the 2000s, primarily by gaining seats in urban and industrial areas. The DPJ took power in 2009 by achieving a sound victory over the LDP. However, the DPJ soon lost its popularity owing to problems with mismanagement (Kushida and Lipscy, Reference Kushida, Lipscy, Kushida and Lipscy2013). The LDP returned to power in 2012, with a sweeping victory. The DPJ barely maintained its status as the second-largest party in the LH due to this defeat. The DPJ broke up before and after the election. The LDP-led coalition has continued to win elections in the LH and UH until the present (December 2021).

Historically, candidates in Japan ran candidate-centered electoral campaigns. This was especially the case when the LH used a multi-member majoritarian electoral system with single non-transferable voting (MMD/SNTV) before 1994 (Curtis, Reference Curtis1971; Carey and Shugart, Reference Carey and Shugart1995; Hirano, Reference Hirano2006; Scheiner, Reference Scheiner2006).

Under the MMD/SNTV system, the governing LDP fielded more than one candidate in each district to achieve a majority. These candidates had to compete not only with the other party candidates but also with their co-partisans; therefore, they established koenkai, a private campaign organization, to mobilize voters in the districts beyond their party platform. Although the LH electoral reform in 1994 was designed to ensure a more party- and (national) policy-oriented campaign, the candidates kept using koenkai to mobilize support in their districts. The candidates usually start organizing their koenkai with their local friends and acquaintances as a core group of supporters. They then expand it based on neighborhood associations, local alumni associations, firms and business associations (especially for LDP candidates), or labor unions (especially for DPJ candidates). Their electoral campaign relying upon these organizations should be well-tailored to district-specific interests, not only to their national party platform, which would enhance the link between the policy preferences of the constituencies and their representatives. The members of koenkai are comparable to personal and primary constituencies, which Fenno (Reference Fenno1978) discusses the US House members' perception of their constituency.

There has been debate regarding the nature of electoral campaigns in Japan after the 1994 electoral reform. Initially, scholars argue that constituents in Japan continue to favor candidates who serve their local districts' interests, much as they did before the reform (McKean and Scheiner, Reference McKean and Scheiner2000; Bawn and Thies, Reference Bawn and Thies2003). In recent years, many argue that voters have switched to caring more about national issues such as national security, while parties and candidates run campaigns based on these programmatic appeals (Kabashima and Steel, Reference Kabashima and Steel2010; Reed et al., Reference Reed, Scheiner and Thies2012; Catalinac, Reference Catalinac2016a).

Suppose the candidates were to run an individual candidate-centered campaign. Could they influence the policy process to fulfill their promises, such as policies tailored to their districts? Japan has a parliamentary system with solid legislative discipline; however, conventional wisdom (and many scholarly publications) indicates that backbenchers still have an apparent influence on the policymaking process (Krauss and Pekkanen, Reference Krauss and Pekkanen2010; Oyama, Reference Oyama2011). For example, the Members of Parliament (MPs) for the LDP participate in a few divisions of the Policy Affairs Research Council (PARC). This intra-party institution discusses government policy proposals before officially submitting them to the legislature as bills. The divisions correspond to each policy area, such as foreign affairs, education, and agriculture. The MPs can affiliate with any of the divisions based on their own or their constituencies' interests. At these divisions, they exert their influence on legislation and appropriation.Footnote ⁴ The DPJ has a similar intra-party institution to the PARC, although the policy influence of their backbenchers could be smaller than that of the LDP (Hamamoto, Reference Hamamoto, Maeda and Tsutsumi2015). The constituents in Japan can expect to influence policy by electing candidates who have policy preferences close to their own, not solely by choosing the party with the most preferable national platform. Even if MPs cannot single-handedly alter the policy of their party, voters should be expected to want a representative that will advocate for them, which provides an incentive for candidates to accommodate (or a least to pay lip service) to their district preference.

This study examines the candidates' policy positions regarding rural–urban economic issues, such as public spending on employment, highway construction, or trade protection of agricultural products. These economic issues are convenient tools that can be used to examine dyadic representation in Japan. First, these issues have been visible and salient in Japanese politics, contributing to a more robust link between constituency and candidate preferences (Miller and Stokes, Reference Miller and Stokes1963; Kuklinski and Elling, Reference Kuklinski and Elling1977). The urban–rural divide is one of the most salient cleavages in Japanese politics (Shimizu, Reference Shimizu, Pekkanen, Reed and Scheiner2013). These economic issues have probably been the second most frequently raised issue for public discussion, only after foreign affairs/defense issues (Taniguchi, Reference Taniguchi2006). In addition to supporting fiscal transfers to rural areas or trade protection for agricultural products, LDP politicians advocate expansionary government spending, especially in public construction projects to create jobs in rural areas, while the opposition criticizes these strategies. These policies are arguably a compensation for rural sectors who would lose by trade liberalization, which the government has been promoting for collective goods (Naoi, Reference Naoi2015).

Although it seems paradoxical, the conservative LDP prefers more public expenditure, while in contrast, the liberal and urban-oriented DPJ supports less spending in this field. These issues should be relatively easy for regular voters to understand to evaluate their candidates in Japan, strengthening the policy link between representatives and those represented (Hurley and Hill, Reference Hurley and Hill2003).

These economic issues constitute lines of political disputes along with geographical boundaries, while most other political conflicts in Japan (such as those concerning foreign policy or inter-class redistribution) go across these boundaries. Because the majoritarian electoral system draws district boundaries geographically, some districts contain metropolitan areas, while others cover predominantly rural areas, which yields considerable variation in policy preferences across the districts regarding the urban–rural issue.

Given the tradition of a candidate-centered campaign, in addition to the saliency of the urban–rural economic issue in Japan, it is very likely to observe within-party variation regarding this issue, at least to some extent.Footnote ⁵ Furthermore, the development of the two-party system with a national campaign platform and their partisan discipline in the legislative process implies a clear partisan gap in policy preferences regarding the issue.

3. Data and methodology

To measure the positions of the candidates, this study used a candidate survey from the University of Tokyo and the Asahi Newspaper Survey (UTAS).Footnote ⁶ The UTAS has been conducted in every national election since 2003, covering seven LH general elections and six UH regular elections, both for candidates and voters. This study examined candidate surveys at the LH general elections between 2003 and 2012 (2003, 2005, 2009, and 2012) and excluded those taken in 2014 and 2017. Before these elections, there was significant redistricting,Footnote ⁷ which made it difficult to discuss the links between the districts and their candidates' preferences. The survey data in 2021 are not open to the public yet.

The response rate to the UTAS was very high among candidates – more than 93% of the candidates who participated in the surveys used for this paper, and this percentage was even higher among major party candidates. This could be partly because the Asahi newspaper writes articles based on responses seen during the campaign, both at the national and the local/candidate level. The UTAS is widely used by many scholars studying Japanese politics to decipher candidate policy preferences (Ito, Reference Ito2015; Miwa, Reference Miwa2015; Kubo, Reference Kubo2016; Smith and Tsutsumi, Reference Smith and Tsutsumi2016; Hamzawi, Reference Hamzawi2021).Footnote ⁸

From various questions asked in the UTAS, this study utilizes the survey response regarding rural–urban-oriented economic policies from candidates running in the majoritarian tier of the mixed-member electoral system, from either the LDP or the DPJ during this period.Footnote ⁹ Respondents are usually provided with a statement and asked to select one of five responses: agree, somewhat agree, neutral, somewhat disagree, and disagree. The question set on the questionnaire was not necessarily the same in each survey. Some questions, such as promoting public works for increasing employment, were asked every time across the surveys, but others have been asked only once. The wording of each question and the election year in which they were asked are as follows.Footnote ¹⁰

1. 1) Public work for employment: Public construction projects are necessary to secure employment in rural areas (2003), or public construction projects are necessary to secure employment (2005, 2009, and 2012).

2. 2) Balance budget vs fiscal stimulus: Government should spend fiscal stimulus for a while rather than cutting spending for fiscal reconstruction (2003, 2005, 2009, 2012).

3. 3) Postal service privatization: Three postal servicesFootnote ¹¹ should be privatized (2003).

4. 4) Free vs building new highway: The government should make existing highways charge-free, rather than building new ones (2003).

5. 5) Fiscal transfer from urban to rural areas: Fiscal transfers to rural areas should be abolished in principle (2003).

6. 6) Keep spending for road construction: The government should maintain the current level of spending on road construction (2009).

7. 7) Should join TPP: Japan should participate in the Trans-Pacific Partnership (TPP) (2012).

To utilize all responses for each questionnaire, this study used the graded response model (GRM) of the item response theory (IRT) with the ‘ltm’ package of R (Rizopoulos, Reference Rizopoulos2006). IRT is a latent variable model. Latent variable models assume that a few latent variables explain the interrelationships in a set of observed response variables. The models assume that unobserved variables such as political attitudes, which cannot be measured directly by conventional means, can be quantified by assuming latent variables. The IRT considers a class of latent variable models that link mainly dichotomous and polytomous manifest variables to a single latent variable (Rizopoulos, Reference Rizopoulos2006). The IRT is applied in political science to measure political knowledge (Delli Carpini and Keeter, Reference Delli Carpini and Keeter1993), legislators' policy positions (Clinton, Reference Clinton2007; Bafumi and Herron, Reference Bafumi and Herron2010; Hirano et al., Reference Hirano, Imai, Shiraito and Taniguchi2011; Shor and McCarty, Reference Shor and McCarty2011), Supreme Court judges (Martin and Quinn, Reference Martin and Quinn2002), and the voters' policy positions in legislative districts and municipalities (Tausanovitch and Warshaw, Reference Tausanovitch and Warshaw2013, Reference Tausanovitch and Warshaw2014). In particular, Hirano et al. used the model with the UTAS data to measure the distribution of policy positions of Japanese candidates, although they did not directly examine the association between the districts' and their candidates' policy preferences (Hirano et al., Reference Hirano, Imai, Shiraito and Taniguchi2011).

The two-parameter (binominal) IRT models use binominal responses to exam question items (such as correct and incorrect). The responses depend on θ_i: the latent trait level for person i, β_j: the difficulty level for item j, and α_j: the discrimination power for item j, such that:

$$P( X_{i, \;j} = 1{\rm \vert }\theta _i, \;\;\beta _j, \;\;\alpha _j) = {\rm \;}\displaystyle{{\exp [ {\alpha_j( {\theta_i-\beta_j} ) } ] } \over {1 + \exp [ {\alpha_j( {\theta_i-\beta_j} ) } ] }}$$

$$s.t.\;\;i = 1, \;\;2, \;\;3, \;\ldots , \;\;n; \;\;j = 1, \;\;2, \;\;3, \;\ldots , \;\;m\;\;\;$$

The GRM extends the two-parameter IRT model with a binomial response to the ordinal multilevel response. Suppose that the response is provided with five categories: agree (1), somewhat agree (2), neutral (3), somewhat disagree (4), and disagree (5) for a statement as the UTAS. P _ij1(θ) represents the probability of person i choosing the first category for item j, P _ij2(θ) represents those for the second category, and so on.

$$P_{ijk}^\ast ( \theta ) = {\rm \;}\displaystyle{{\exp [ {\alpha_j( {\theta_i-\beta_{\,jk}} ) } ] } \over {1 + \exp [ {\alpha_j( {\theta_i-\beta_{\,jk}} ) } ] }}\;$$

$$s.t.\;\;i = 1, \;\;2, \;\;3, \;\ldots , \;\;n; \;\;j = 1, \;\;2, \;\;3, \;\ldots , \;\;m; \;\;k = 1, \;\;2, \;\;3, \;\;4\;$$

$$P_{ij1}( \theta ) = 1-\;P_{ij2}^\ast ( \theta ) $$

$$P_{ij2}( \theta ) = P_{ij2}^\ast ( \theta ) -\;P_{ij3}^\ast ( \theta ) $$

$$P_{ij3}( \theta ) = P_{ij3}^\ast ( \theta ) -\;P_{ij4}^\ast ( \theta ) $$

$$P_{ij4}( \theta ) = P_{ij4}^\ast ( \theta ) -\;P_{ij5}^\ast ( \theta ) $$

$$P_{ij5}( \theta ) = P_{ij5}^\ast ( \theta ) $$

The IRT model has two specific characteristics that are convenient for this study. First, the ‘test’ parameters do not depend on the particular group of ‘examinees’ of the population. Moreover, the examinees' trait assessment does not depend on a particular set of question items (Fan and Sun, Reference Fan and Sun2013: 45). In other words, the candidates responding to each survey do not have to be the same set of individuals. Certainly, those running in the 2003 LH election do not belong to the same group as those running in 2012. The IRT model works if there are some overlaps between the candidates: for example, candidate group A ran between 2003 and 2005, candidate group B ran between 2005 and 2009, and candidate group C ran between 2009 and 2012. Their parameters – latent policy preferences – can be measured on the same scale if some of them are answering the same survey items.

By utilizing the model, this study measures the latent policy position of each candidate, which is termed the Urban–Rural (U–R) Policy Score. For convenience, it assumes that larger numbers on the scale indicate support for a more rural-oriented position (e.g., more public spending for employment, road construction, rural area subsidy, and trade protection). In comparison, smaller numbers indicate support for an urban-oriented policy (e.g., less public spending, free trade, and economic efficiency).Footnote ¹²

While using the model, it was assumed that candidates had a latent and fixed policy preference regarding economic policy during the period in which their responses to the questionnaires could be measured.Footnote ¹³ Hence, each candidate first chose his/her policy position regarding economic policy based on what aligned with his/her district (or political parties nominated such candidates) and did not change it after that. As discussed above, the candidates in Japan organize their personal campaign organization, koenkai, which should be well-tailored to district-specific interest. They might change their responses to a survey question with the same wording across elections due to the campaign's context, while they do not (or cannot) change their (latent) policy preferences, which is firmly structured with the organization.

Figure 1 shows the distribution of the U–R Policy Score of the LDP and DPJ candidates measured using this approach. The white histograms on the upside represent the LDP, while the gray ones on the downside represent the DPJ. These figures demonstrate substantial partisan differences in policy positions; however, considerable within-party heterogeneity and some overlap between the parties remain.

Fig. 1. LDP/DPJ candidate urban–rural policy score.

It is much more difficult to measure the policy position of the constituencies. The UTAS asks for the same set of policy questions for both candidates and voters. Nevertheless, it is problematic to use a voter survey to estimate the policy position of the constituencies in the respective districts. First, the voter survey of the UTAS does not cover all of the LH districts, and the numbers of respondents in the respective districts are too small to estimate the policy position of the entire constituency with sufficient reliability.Footnote ¹⁴ Both low political interest and limited knowledge regarding individual respondents introduce a considerable measurement error to the estimate of the constituency preference, which is already limited by the small sample survey, even if the constituencies in the (district-) aggregate would have a reasonable opinion (Page and Shapiro, Reference Page and Shapiro1992).Footnote ¹⁵

Instead, this study uses the employment structure of each district – more specifically, the ratio of the labor force in agriculture, forestry, fishery, and construction – as a proxy for the rural–urban level, and hence for the policy preference of respective constituencies. The higher the ratio, the more electorates prefer a candidate with a rural-oriented policy position. Figure 2 shows the distribution of the rural employment structure in the LH districts based on the 2010 census.Footnote ¹⁶ These rural sectors are often the focus of LDP mobilization through a patron-clientelistic exchange (Scheiner, Reference Scheiner2006: 70–73) and indicate a high level of support for the LDP (Miyano, Reference Miyano and Katase1998; Mulgan, Reference Mulgan, Pekkanen, Reed and Scheiner2013).Footnote ¹⁷

Fig. 2. Employment structure in Lower House districts.

Note: 2003–2012 district boundaries, based on 2010 Census.

4. Hypothesis and model

The main claim of this study was that the within-party variation in the candidates' policy positions would correspond to the policy preferences of their district constituencies. More specifically, it was expected that the district characteristics measured with the employment structure would be associated with the policy position of the candidates regarding economic issues. To verify this claim, the following hypothesis was tested:

This hypothesis was tested with the ordinary least squares model, the U–R Policy Score – the policy positions of candidates in the urban–rural-oriented economic policy – as the outcome variable, and the district (rural–) employment structure (RES) and the candidate party affiliation (PARTY) as explanatory variables.

$${\rm Model\ }1\colon U-R Policy Score_i\sim \beta _0 + \beta _1RES_i + \beta _2PARTY_i + e_i$$

It was expected for voters to reward candidates representing the constituencies' policy preferences well in the election. Otherwise, the candidates would have little incentive to be faithful to their constituencies. The operational hypothesis to test this claim is as follows:

This study ran the linear mixed-effect model by the districts in which the two major-party candidates compete to examine the effect. The outcome variable of the model was the two-party vote-share margin (Margin): (LDP candidate votes – DPJ candidate votes)/(LDP candidate votes + DPJ candidate votes). The explanatory variables were the difference in the policy scores between the LDP and the DPJ candidates (PSD), the district employment structure (RES), and the interaction term between them. The interaction term was included in the model because the same difference between the LDP and DPJ candidates in the economic policy should affect differently to their electoral outcome between rural and urban districts. The specification examines whether more significant differences in the rural-oriented policy position between LDP and DPJ candidates in the same district are associated with larger LDP margins of victory over their DPJ counterparts in districts with a larger proportion of the workforce in the rural sector than in districts with smaller proportions of the workforce in the rural sector.

As covariates, the model used the effective number of candidates (ENC) and the election year as the fixed effect (2003 as a reference) and dummy variables for each district and candidate as the random effect.

$${\rm Model\ }2\colon \;Margin_{ij}\sim\beta_0+\beta_1 PSD_{ij}+\beta_2 RES_i+\beta_3 PSD_{ij}\times RES_i+\beta_4 ENC_{ij}+\beta_{5} 2005 \;Dummy_j+\beta_6 2009 \;Dummy_j+\beta_7 2012 \;Dummy_j+\gamma_i+\delta_k+\varepsilon_l+\delta_{ij}$$

where i denotes district, j denotes election year, k denotes specific LDP candidates, and l denotes specific DPJ candidates.

Due to the operationalization, the positive number in Margin indicates electoral results favorable to the LDP. The analysis focused on the coefficients for the difference in the respective policy positions (PSD), that is, β ₁, and its interaction effect with the district's employment structure (RES), that is, β ₃.

β ₁ was expected to be negative. In highly urbanized districts where almost no one is involved in the agriculture, forestry, fishery, or construction sectors, an LDP candidate with a relatively rural-oriented policy position should attract fewer votes than their DPJ opponent (with high PSD). It was predicted that β ₃ should be positive. In rural districts (with high RES), the constituencies are more likely to vote for the LDP if the LDP candidate is more favorable to rural interest than their opponent.

Finally, the marginal effects (and their standard errors) of the candidates' policy positions on electoral outcomes were examined, given the estimate of these coefficients and the employment structure of the districts (Brambor et al., Reference Brambor, Clark and Golder2006).