2. Background: uneven district magnitudes and subsequent gaps in turnout

The Japanese National Diet consists of two chambers: the LH and the UH. The LH has priority in only a few – but essential – areas such as electing the prime minister, casting a vote of no-confidence, approving the budget plan/ratifying a treaty, and overriding a veto to legislation by the UH with a supermajority.

Since the LH electoral reform in 1994, both chambers have used the mixed-member majoritarian electoral system. The LH electoral system consists of a PR tier that currently elects 176 members from 11 regional districts and a majoritarian tier with SMDs that elect 289 members from their respective districts. The LH term is up to 4 years but could be dissolved by the prime minister before the term.

The UH, on the other hand, uses an electoral system with uneven district magnitudes for its majoritarian tier. The UH term is fixed for 6 years, and the UH elects half of its members every 3 years. In the 2016 election, the UH elected 48 members from the national PR tier with the open list system and the remaining 73 members from 45 districts, which basically used prefectures as electoral districts within the SNTV system.Footnote ¹ The district magnitude varies from one to six based on each district's population, although many of the district magnitudes are either one or two. This paper focuses on the majoritarian electoral tier with its uneven district magnitudes. To the best of our knowledge, MMD-plurality electoral systems with varying district magnitudes are rare, at least at the national level. The system has been quite common in Japan for many years, not only in the UH but also in the LH (pre-1994 electoral reform) and local assemblies. However, outside of Japan, currently only Vanuatu and Afghanistan use the MMD/SNTV electoral system with different district magnitudes at the national level.Footnote ²

In the following sections, this paper refers to UH districts that elect one representative as SMDs, those that elect two representatives as double-member districts (DMDs), and those that elect three or more than three representatives as MMDs.Footnote ³ Figure 1 shows the geographical distribution of each district magnitude in 2007 and 2010, and Table 1 shows the number of districts and the average number of candidates with respective district magnitudes.

Figure 1. District magnitude map in the Japanese Upper House during 2007 and 2010.

Table 1. Number of districts with each district magnitude (1998–2019)

^a District magnitude is 6 (Tokyo district).

Since the electoral reform in 1994, the Japanese party system has gradually evolved into a two-party system. The new party system consists of the conservative, rural-based Liberal Democratic Party (LDP) and the center-left Democratic Party of Japan (DPJ), whose support base is mainly in urban and industrial areas. This shift has been caused by the importance of the LH and its considerable weight on the majoritarian tier; however, smaller parties with long histories still survive under the system, such as the Buddhist centrist Clean Government Party (CGP, also known as Komeito) and the leftist Japanese Communist Party (JCP). These parties attain seats primarily at the PR tiers of both chambers, in addition to the UH MMDs.

The Japanese party system became a two-party (or at least two-camp) system in 2003 when the Liberal Party merged with the DPJ, whereas the LDP has been in a coalition government and electoral alliance with the CGP since 1999. Over time, the DPJ became the only serious challenger to LDP dominance – it had ruled Japan since its formation in 1955 except for a 1-year break from 1993 to 1994. The LH seat-share of the two major parties was 75% in 2000. However, it jumped to 86% in 2003 and stayed in the range of 80% until 2012. Similarly, its seat-share in the UH election was 74% in 2001, but it has remained above 80% since 2004.Footnote ⁴ In parallel with the evolution of the two-party system, the effective number of candidates (Laakso and Taagepera, Reference Laakso and Taagepera1979) in the LH SMDs decreased from election to election. It was 2.95 in 1996, the first election under the new system, but decreased to 2.26 in 2009 when the DPJ broke the LDP's predominance and came to power. While that number increased again to 2.98 in 2012 due to the DPJ split and new parties' rise, it also decreased to 2.60 in 2014 (Ariga et al., Reference Ariga, Horiuchi, Mansilla and Umeda2016: 23).

The evolution of two-party competition in the LH creates uneven pressure on the UH electoral contest. The UH SMDs became competitive as a two-party system developed. This is especially so as the DPJ grows as a challenger to the LDP not only in urban or industrial areas but also in rural areas where the LDP has been dominant. In contrast, DMDs will be uncompetitive unless the major parties field two candidates to achieve both seats.Footnote ⁵ Nevertheless, fielding more than one candidate may not be an efficient strategy for these parties under an SNTV system like the Japanese UH because such a strategy may cause their two candidates to fail together.Footnote ⁶ As a result, there are often only two serious contenders in DMDs despite the expectation from the M + 1 ruleFootnote ⁷ (Reed, Reference Reed1990; Cox, Reference Cox1997). Also, MMDs can be competitive because smaller parties can elect their candidates, mainly because the major parties have second thoughts about running second candidates under the SNTV rules.

Figure 2 shows the electoral competitiveness of respective district magnitudes (SMDs, DMDs, and MMD) for each election since 1998. The competitiveness is measured by the vote-share ratio of the (last) winner to the first loser. In 1998, there was not much difference in the competitiveness of the respective groups, probably because the LDP still had second candidates in many DMDs, whereas the DPJ strived to achieve single seats in the DMDs but not the SMDs.Footnote ⁸ In contrast, smaller parties like the CGP continued to nominate candidates in the MMDs during that period, which, in turn, increased the competitiveness in these districts.

Figure 2. District magnitude and electoral competitiveness* (1998–2019).

Note: * The (last-) winner-to-first-loser vote ratio.

In contrast, between 2004 and 2010, there was a clear gap in the competitiveness of SMDs and MMDs with DMDs. The vote-shares of the first loser to the (last) winner, on average, hovered around 80% in SMDs and MMDs, while they were 40–50% in DMDs. The gap between the SMDs and DMDs reversed in 2013 when the DPJ split and was defeated in the 2012 LH and 2013 UH elections. As the DPJ recovered from that defeat and gradually united anti-LDP political forces in 2016, the SMDs again became more competitive than the DMDs; however, the DPJ also split in 2017, before the LH election, between the centrist and the leftist factions,Footnote ⁹ which explains the higher competitiveness of DMDs in 2019.Footnote ¹⁰

The apparent gap in competitiveness should influence the strategic calculation of electorates and political parties. Political parties spend little effort mobilizing electorates in uncompetitive districts. Moreover, the electorates are less likely to vote because they are less interested and feel less political efficacy in uncompetitive races, and are less likely to be mobilized by political parties. In the following sections, this paper examines the influence of district magnitude upon voter turnout and then analyzes electoral mobilization by political parties, which also influences voter turnout.

3. Analysis I: turnout

This paper first examines whether district magnitude influences voter turnout across the districts. The hypothesis is as follows:

This paper tests the hypothesis by regressing the district-level turnout of seven UH elections between 2001 and 2019 to the district magnitude of respective districts, i.e., SMDs and MMDs as a group on the one hand and DMDs on the other hand, by using DMDs as a reference group.Footnote ¹¹ The model is describable as:

$$Turnout_{it}\sim DM_{it}\beta + Cov_{itk}\gamma + \alpha _i + u_{it}, \;$$

such that Turnout_it is the outcome variable of the model, the turnout of district i at election t, DM_it is the explanatory variable, the district magnitude of district i at election t (i.e., either SMDs/MMDs or DMDs), Cov_itk are various covariates (1, 2, …, k) also of district i at election t, α_i is the time-invariant district-specific factor for district i, and u_it is a random factor of district i at election t.

Each district may have time-invariant-specific characteristics α_i, which could boost or lower the district's turnout. Therefore, the turnout model uses random effect or fixed effect for each district to cope with these district-specific characteristics.Footnote ¹² As the covariates, the turnout models control the turnout in the previous LH elections in the corresponding area.Footnote ¹³ The models also include the election year as the fixed effect (2001 as a reference) to control the year-specific national surge and decline in the turnout, and the number of candidates (square-rooted) to control the time-variant factors in respective districts.

Moreover, the random-effect model also includes the population ratio of those over 65 years old and the DID index (the population ratio living in an urban area)Footnote ¹⁴ in the 2010 census as covariates. These covariates reduce potential bias due to the correlation between the district magnitude and the random effect for each district. The UH election is widely regarded as less important than the LH election, which decides the next prime minister. Thus, it is reasonable to assume that the electorate originally with a stronger predisposition of voting to participate in the UH election (such as elderly voters and those in a rural area) even control the turnout in the previous LH elections of the area. These covariates are not used in the fixed-effect model because they are time-invariant.Footnote ¹⁵

Table 2 summarizes the analysis results; the random-effect model is in the third column, while the fixed-effect model is in the fourth column.Footnote ¹⁶

Table 2. The effect of the district magnitude on turnout (2001–2019)

Random-effect model is estimated using a linear mixed model with lmer() in the R ‘lme4’ package. The statistical tests are single-sided for the explanatory variables: SMDs and MMDs, because the turnouts in these districts are expected to be higher than those in DMDs. Statistical significance at ⁺10%, *5%, †1%, ‡0.1% levels. The coefficients for the electoral year fixed effect are abbreviated. Districts that were merged since 2016 (Shimane-Tottori and Tokushima-Kochi) are treated as new districts. The numbers in the parentheses are the standard errors of the coefficient estimates. The standard errors in the fixed-effect model are clustered by districts. The correction uses the ‘bias-reduced linearization’ adjustment (CR2) available in the R ‘clubSandwich’ package.

Both models show that the turnout is higher in SMDs and MMDs than in DMDs by 1.4% (both statistically significant). This result suggests that DMDs have lowered the turnout vs that in districts with different district magnitudes, even when controlling for the district-specific factors.

4. Analysis II: mobilization

Next, this paper examines how political parties made varying efforts to mobilize voters across the districts. The different levels of mobilization could contribute to the turnout gap discussed in the previous section and the other factors such as the electorate's political efficacy. Some previous studies in Japan suggest that the major parties exerted more effort in their campaigns, such as party leader visits to SMDs during the UH electoral campaign (e.g., Fujimura, Reference Fujimura2016). This paper instead focuses on personal contacts conducted by the parties through, for example, postcards or phone calls – arguably more effective than impersonal messaging such as TV campaign advertisements (e.g., Green and Gerber, Reference Green and Gerber2015).

This paper uses the Japanese Electoral Survey (JES) to measure the mobilization efforts of political parties. The JES is a series of pre- and post-electoral panel surveys conducted in Japan during national elections and nationwide local elections, mainly with a face-to-face design and a sample size of 2,000–3,000. The third series (JESIII) was conducted between 2001 and 2005, and the fourth series (JESIV) was completed between 2007 and 2011. This series covers three LH general elections in 2003, 2005, and 2009, and four UH regular elections in 2001, 2004, 2007, and 2010. Unfortunately, the more recent JESV does not cover party mobilization questionnaires, at least in a comparative way with JESIII and JESIV.Footnote ¹⁷

This paper uses the responses to the JESIII and JESIV questionnaire regarding (1) receiving postcards, (2) receiving party newspapers or campaign leaflets, and (3) being contacted via phone by the LDP, the DPJ, or other smaller parties as indicators of the parties' mobilization efforts. Each response is coded as one if the answer is Yes and zero if the answer is No.Footnote ¹⁸ This paper uses the sum of the three answers (zero to three for each) as the analysis' outcome variables.

This paper runs ordered probit models for the two major parties and small parties separately to measure the effects of district magnitude on mobilization efforts during the UH campaigns in 2001, 2004, 2007, and 2010. For this estimation, this paper uses a linear mixed model with the function clmm() in the R ordinal package (Christensen, 2019). The model enables the use of the random-effect model for the ordered probit model. The main hypothesis of the paper regarding the mobilization strategies of the major parties is as follows:

On the other hand, the hypothesis regarding smaller parties is as follows:

The models to test H2a and H2b can be described as follows:

$$Contact UH_{ijt}\sim DM_{\,jt}\beta + Cov_{ijtk}\gamma + d_j + r_i + u_{ijt}, \;$$

such that Contact UH_ijt is the outcome variable of the model (which takes the value of 0, 1, 2, or 3), the number of contacts from the party to respondent i in district j in UH election t, DM_jt is the explanatory variable, the district magnitude of district j at election t (i.e., either SMDs, DMDs or MMDs, by using DMDs as a reference group), Cov_ijtk are covariates (1, 2, …, k) of respondent i or district j at election t, d_j and r_i are the time-invariant factors for district j and respondent i, respectively, and u_ijt is a random factor of respondent i in district j at election t.

As covariates, the model includes the respondent characteristics such as gender, age, education, urban level of residence, and years in current residence, which could be associated with the chance of party mobilization. Parties and candidates are expected to mobilize people they already know (those in a local social network), who are powerful, and who are more likely to respond (i.e., people with resources; Rosenstone and Hansen, Reference Rosenstone and Hansen1993). As voters age and live longer in their current residences, they become more likely to be contacted by political parties via local social networks. These voters might be affiliated with a candidate's personal support group (koenkai), quite common in Japan, which often contacts voters based on membership lists before the polling day (Tomasu, Reference Tomasu2012). Moreover, a higher education level (and subsequently higher income) provides people with more social and economic resources for political participation, which gives political parties an incentive to make contact. Next, the major party models use dummy variables for the second candidate of these parties in the DMD (one if either the LDP or DPJ nominates a second candidate, zero otherwise), which provides parties with an additional incentive to mobilize. Furthermore, the models control the election year as the fixed effect because parties could have varying mobilization efforts at the national level across years.

Finally, the variation among districts and individuals may not be attributable to the easily-observed covariates discussed above. To avoid this bias, our approach uses the following techniques: First, the model includes random effects for each UH district to control the district-specific factors.Footnote ¹⁹ Moreover, by utilizing the JES panel design, the model controls the level of party mobilization that respondents experienced during the LH electoral campaign in neighboring years as a covariate, measured using the same method for the UH election. Similarly, random effects for respondent ID are included in the model because the same respondents answered the survey in 2001 and 2003 or 2007 and 2010 (JESIII or JESIV, respectively). They will eliminate much of the bias caused by unobserved differences at either the district or individual level.

Figure 3 summarizes the results of the analysis in the coefficient plots.Footnote ²⁰ The dots show the point estimate of the effect: The black circles are for LDP, the black triangles are for DPJ, and the white squares are for the smaller parties. The bar indicates the confidence interval. The coefficient plots have dots, not at the middle but the upper end of the bars. This is because both H2a and H2b assume single-sided tests: Major parties are more likely to mobilize in SMDs than in DMDs (and MMDs), whereas smaller parties are more likely to mobilize in MMDs than in DMDs (and SMDs). Hence, if the point estimate were below zero, the null hypothesis could not be rejected. Moreover, because the statistical test is single-sided, the bar length is 1.64 times the standard error at the lower side of the point estimate for statistical significance at the 5% level – not 1.96 times at both sides of the estimate. The coefficients are statistically significant if the dots are over zero and the bars do not cross the line.

Figure 3. The effect of the district magnitude on party contacts (ref. DMD).

The figure shows that major parties are more likely to contact the constituents in SMDs, as H2a predicts. The figure also shows that smaller parties are more likely to reach out to the electorate in MMDs than in DMDs, as suggested by H2b. The effect size is non-negligible. Suppose that a voter in DMDs has a 40% chance of receiving at least one contact from respective parties during the UH campaign given his/her attributes, district, and year. Then the LDP and the DPJ contact the voters in SMDs sharing the same characteristics with a chance of around 53 and 48%, respectively. Similarly, at least one small party reaches the voters in the MMDs with a 50% chance.Footnote ²¹ In contrast, major parties seem to exert similar mobilization efforts between DMDs and MMDs, while small parties mobilize at the equivalent level in SMDs and DMDs. The point estimates for these coefficients are very close to zero and statistically insignificant – either the estimates are below zero or above zero, but the bars for the standard error cross the zero line.

In conclusion, the analysis indicates the expected result: The two major parties in Japan – the LDP and the DPJ – expended more substantial mobilization efforts in SMDs than they did in DMDs and MMDs during UH electoral campaigns. Moreover, smaller parties focused their mobilization efforts on MMDs where they had a better chance of electing their district candidates or at least nominated candidates. These estimates should be robust to the bias caused by differences across districts (other than the district magnitude) because the analysis controls the level of contact that the same respondents experienced during the LH election, which used SMDs for the majoritarian tier. These outcomes imply that the lower turnouts in DMDs presented in Table 2 might be attributable to political parties' lack of campaign mobilization. The following final section elucidates the implications of this result.