1. Electoral systems and cognitive demands

The role of the information environment in shaping voter behavior is well established. The availability of information enhances political learning (Jerit et al., Reference Jerit, Barabas and Bolsen2006). Information-rich environments allow voters to correctly identify the positions of parties (Banducci et al., Reference Banducci, Giebler and Kritzinger2017) and engage in elections more actively (Hayes and Lawless, Reference Hayes and Lawless2015). Detailed information about individual candidates also enables voters to make their decisions without relying on partisan cues (Peterson, Reference Peterson2017). Further, a rich information environment improves voters' monitoring capacity and politicians' responsiveness (Snyder and Strömberg, Reference Snyder and Strömberg2010).

However, the role of the information environment is not always positive. Some studies from political psychology suggest that the provision of too much information about candidates and parties quickly leads to information overload (Lau and Redlawsk, Reference Lau and Redlawsk2006; Lau et al., Reference Lau, Andersen and Redlawsk2008; Rogers and Tyszler, Reference Rogers and Tyszler2018). Excessive information also complicates decision-making environments, inducing confusion and choice fatigue among voters (Augenblick and Nicholson, Reference Augenblick and Nicholson2015). Consistent with these concerns, Cunow (Reference Cunow2014) provides experimental evidence that increasing the number of candidates on the ballot substantially reduces voters' intention to vote in elections. Augenblick and Nicholson (Reference Augenblick and Nicholson2015) also show that a large number of choice sets reinforce voters' tendency to abstain or simply vote for the status quo. In short, some information environments may constrain voters' behavior.

Importantly, electoral systems play a critical role in shaping the information environment of vote choice. Here, particularly important are two aspects of electoral rules: (1) district magnitude and (2) the distinction between candidate- and party-centric systems. First, district magnitude determines the number of candidates and parties competing in the district. As district magnitude increases, the number of choices also increases. Since voters have a limited cognitive capacity to discern among a large set of alternatives, larger district magnitude makes information processing more costly and challenging (Carey and Hix, Reference Carey and Hix2011; Taagepera et al., Reference Taagepera, Selb and Grofman2014). Therefore, other things equal, district magnitude is positively associated with the cognitive demands of elections.

However, how much district magnitude intensifies cognitive demands depends on the second dimension of electoral rules: candidate- versus party-centric systems (Carey and Shugart, Reference Carey and Shugart1995; Shugart et al., Reference Shugart, Valdini and Suominen2005).Footnote ³ Some candidate-centric systems introduce intraparty competition. The presence of several candidates from the same party in the district minimizes the usefulness of partisan cues and forces voters to gather detailed candidate-level information (McCubbins and Rosenbluth, Reference McCubbins, Rosenbluth, Cowhey and McCubbins1995). By contrast, party-centric systems allow voters to make their choices based simply on party-level factors. To the extent that candidate-centric systems make information processing more cumbersome than party-centric systems, the former impose greater cognitive demands on voters for a given magnitude than the latter (Shugart et al., Reference Shugart, Valdini and Suominen2005).

These two features of electoral rules allow us to map different systems by their cognitive demand. First, a SMDP is least informationally demanding. Under SMDP, district magnitude is by definition one, which suppresses the number of options that voters need to consider. SMDP also enables voters to use party labels as a meaningful information shortcut. By contrast, candidate-centered systems with intraparty competition, such as the SNTV and open-list proportional representation (PR) systems, have higher cognitive demands. Under these systems, information environments become far more cognitively demanding as district magnitude increases. Finally, under party-centric closed-list PR systems, cognitive demands are still an increasing function of district magnitude. However, the positive effect of district magnitude on cognitive demands under these systems is less acute than under candidate-centric systems because decision-making tasks are less complicated.

Cognitively demanding electoral rules change voter behavior in many different ways. For example, Taagepera et al. (Reference Taagepera, Selb and Grofman2014) argue that as district magnitude and the number of candidates increase, voters overwhelmed with too much information are more likely to retreat from the electoral process. Hence, there should be a negative association between district magnitude and turnout once district magnitude becomes sufficiently large. Alternatively, voters under demanding systems may try to make the best possible choices by looking for information shortcuts that allow them to simplify their decision rules. Consistent with this argument, Shugart et al. (Reference Shugart, Valdini and Suominen2005) show that under open-list PR systems with large district magnitude, voters utilize the localism of a candidate as a heuristic (Jankowski, Reference Jankowski2016). Marcinkiewicz and Stegmaier (Reference Marcinkiewicz and Stegmaier2015) also find that when electoral rules become more demanding, voters are more likely to rely on ballot position cues provided by parties.Footnote ⁴

However, these accounts may not necessarily explain the behavior of every voter under cognitively demanding systems. In fact, there are also voters who continue to participate but fail to acquire appropriate information shortcuts (Cunow, Reference Cunow2014). Given the presence of these voters, what is not explicitly considered in the prior studies is that cognitively demanding rules can also limit voters' abilities to make informed decisions. This situation may happen because highly demanding information environments make it more challenging for people to correctly set their priorities and recall prior information (Keller and Staelin, Reference Keller and Staelin1987; Eppler and Mengis, Reference Eppler and Mengis2004). By so doing, demanding environments restrict people's abilities to filter irrelevant information, which inevitably leads to suboptimal choices (O'Reilly, Reference O'Reilly1980).

The primary implication of this argument is that when electoral systems generate more cognitively challenging electoral environments, voters' decisions may become more susceptible to the influence of politically irrelevant cues. Confused with a complicated set of options, some voters may find it more difficult to make appropriate judgments about the policy and performance of candidates. Under this circumstance, some of the voters may, perhaps even inadvertently, rely on seemingly minor cues that they easily find at the polling station without regard for their relevance to candidates' quality.Footnote ⁵

2. Candidate name complexity as an irrelevant cue

Prior studies on voters' cue-taking behavior at the polling station identify a range of subtle cues that they mistakenly use, including order, color, and other minor information displayed on a ballot paper (Niemi and Herrnson, Reference Niemi and Herrnson2003; Reynolds and Steenbergen, Reference Reynolds and Steenbergen2006; Moehler and Conroy-Krutz, Reference Moehler and Conroy-Krutz2016). Although what kind of politically irrelevant information shapes voters' decisions may vary from context to context, the setting of Japanese elections makes the visual complexity of candidate names a particularly interesting example.

In Japan, voters are exposed to a relatively limited number of “last-minute” cues. Inside the polling booth, the voters are provided with the names and party affiliations of candidates but have no access to other information, such as candidates' faces or policy positions. Given the type of information available to voters right before casting their votes, it is reasonable to expect that some of them become highly reactive to name-related cues (Fukumoto and Miwa, Reference Fukumoto and Miwa2018). Importantly, while names may contain politically useful cues, such as class, gender, and ethnicity, they also carry politically meaningless information, such as their smoothness and complexity. Some studies show that the latter type of name cue can have an unintended effect on voters' perceptions about the attractiveness of candidates (O'Sullivan et al., Reference O'Sullivan, Chen, Mohapatra, Sigelman and Lewis1988).

Moreover, an interesting feature of Japanese elections is the use of a write-in ballot, which obliges voters to write the name of their favored candidate on a blank paper. Requiring an act beyond simply checking a box on the ballot, write-in ballots increase the costs of voting considerably (Fujiwara, Reference Fujiwara2015).Footnote ⁶ This fact is particularly true in Japan because of difficulty in remembering and writing kanji (Chinese characters). Unlike alphabets, there are more than several thousand different kanji, and people do not remember all of them. In addition, some kanji are very complex and difficult to write. It is common for Japanese people, even well-educated individuals, to make writing mistakes. Given these features, write-in ballots may increase the possibility that voters take into consideration the cues that are associated with complex names, even if they are politically irrelevant.

There are at least two ways in which candidate name complexity could affect vote choice. The first mechanism is related to the psychological effect of visual stimuli on people's cognition (Donderi, Reference Donderi2006; Eng et al., Reference Eng, Chen and Jiang2005). A number of psychological studies suggest that simpler images require less processing capacity and are more easily stored in short-term memory than more complex images. Accordingly, visually complex objects, such as ads, brand logos, and even names, have a negative impact on people's attention and affection (Pieters et al., Reference Pieters, Wedel and Batra2010; Van Grinsven and Das, Reference Van Grinsven and Das2016; Silva et al., Reference Silva, Chrobot, Newman, Schwarz and Topolinski2017). This kind of psychological effect may well extend to the ways in which Japanese voters approach candidate names at the polling booth.

In particular, when they enter a voting booth, some voters' attention may immediately go to candidates with simpler names. Further, such attention may be translated into some sorts of positive feelings about candidates, as commonly observed in the ways in which people perceive others with simpler names in non-electoral contexts (Silva et al., Reference Silva, Chrobot, Newman, Schwarz and Topolinski2017; Zürn and Topolinski, Reference Zürn and Topolinski2017).Footnote ⁷ Not surprisingly, the cumbersome tasks of writing a candidate name should reinforce this tendency. The fact that some kanji are difficult to recognize and reproduce may provide an additional psychological reason to like simpler names that are easier to write. As a result, candidates with simpler names may take some votes away from those with more complex names.Footnote ⁸

The second mechanism is related to the error-prone nature of write-in ballots. Since they require voters to accurately write the name of a candidate, some votes may be invalidated (not counted) in a systematic manner (Fujiwara, Reference Fujiwara2015). In the context of Japanese elections, to the extent that more complicated kanji induce more writing errors, votes cast for candidates with more complex names may be more likely to be invalidated.Footnote ⁹ Hence, more complex names may be associated with a higher rate of wastage, especially when the information environment prevents overloaded voters from carefully reflecting on their choices.

The argument in the previous section suggests that in order to comprehend the impact of candidate name complexity, it is important to take into account how electoral systems moderate the cognitive demands of vote choice. Until 1993, the Lower House elections in Japan used SNTV, under which voters cast one vote in the multimember district.Footnote ¹⁰ District magnitude varied from 1 to 6, and the average number of candidates in the district was 7.2. Further, multiple candidates from the same party competed with each other. Intraparty competition necessitated that voters looked beyond party labels, which led to the extensive personalization of electoral campaigns (McCubbins and Rosenbluth, Reference McCubbins, Rosenbluth, Cowhey and McCubbins1995).

By contrast, under SMDP, which has been used since 1996, district magnitude is by definition always set equal to 1. The average number of candidates per district is 3.8, substantially lower than the number under SNTV. Moreover, since every district has only one seat, there is no intraparty competition. As a result, voters can generally rely on partisan cues to make their voting decisions, and electoral campaigns become more party-centric (Hirano, Reference Hirano2006; Catalinac, Reference Catalinac2016).

What is important here is that the cognitive demands of vote choice are far more extensive under SNTV than under SMDP. Because of larger district magnitude and the presence of intraparty competition, SNTV provides a more complicated set of options than SMDP, requiring voters not only to process more information but also to go over many different strategic scenarios (McCubbins and Rosenbluth, Reference McCubbins, Rosenbluth, Cowhey and McCubbins1995; Horiuchi, Reference Horiuchi2005). This cognitive difficulty may increase the chance that voters have difficulty in making fully optimal decisions. As a result, they may be more likely to make poorly informed choices, unintentionally resorting to the visual complexity of candidate names. A consequence is that candidates with simpler names may enjoy an electoral advantage over those with more complex names under SNTV, but not necessarily under SMDP.

3. Empirical analysis

To measure candidate name complexity, I use the total number of strokes in the candidate name. It is an appropriate measure because the visual complexity of kanji increases as a function of the number of strokes (Liversedge et al., Reference Liversedge, Zang, Zhang, Bai, Yan and Drieghe2014). At a glance, letters with a larger number of strokes look more difficult to write than those with a smaller number of strokes. Further, kanji with a larger number of strokes are less commonly used in daily life. As a result, some voters may subconsciously avoid candidates whose names consist of complex kanji or make writing mistakes when they try to reproduce these names.

For illustration, I show how to calculate the name complexity of two prominent politicians, Shinzo Abe (prime minister as of 2019) and Yukio Hatoyama (prime minister in 2009–2010). In Japanese, Shinzo Abe is written as 安倍晋三. Since the number of strokes in each letter is 安 = 6, 倍 = 10, 晋 = 10, and 三 = 3, his name complexity score is 6+10+10+3=29.Footnote ¹¹ In the case of Yukio Hatoyama (鳩山由紀夫), the number of strokes in each letter is 鳩 = 13, 山 = 3, 由 = 5, 紀 = 9, and 夫 = 4. Therefore, his name complexity score is 13+3+5+9+4=34. Yukio Hatoyama receives a higher complexity score than Shinzo Abe.

I rely on the Reed–Smith Japanese House of Representatives Elections Data Set (JHRED; Reed and Smith, Reference Reed and Smith2017) to measure candidate name complexity. Observations include all candidates in the general elections of the Lower House between 1947 and 2014. Among unique candidates, the median name complexity score is 30, the minimum is 7, and the maximum is 72. Politicians with very low name complexity include Tsutomu Kawara (瓦力), Hisa Yoneyama (米山久), or Koichi Yamaguchi (山口好一). By contrast, legislators with very high name complexity are Kanjyu Sato (佐藤観樹), Fukushiro Nukaga (額賀福志郎), or Naomi Tokashiki (渡嘉敷奈緒美). Recognize that the names in the latter group are visually denser and more complex than those in the former.

The key assumption is that the number of strokes contains no politically useful information. Since many different combinations of letters result in the same number of strokes by chance, the number of strokes does not inform the past performance of politicians nor predict the type and competence of candidates. Similarly, it would be surprising if an increase or decrease in the number of strokes led to systematic changes in the policy stances or personal quality of candidates.Footnote ¹² Therefore, it is hard to believe that voters use name complexity as a policy- or performance-related cue when making voting decisions.

One caveat of the current approach is that candidate names in the JHRED are based on their actual names, but they may not always be the same names that voters saw at the polling station. This discrepancy can happen because there are several ways to simplify written names in Japanese. For example, one can write the name of Shinzo Abe as あべしんぞう instead of 安倍晋三, and the former looks simpler (has a lower number of strokes) than the latter. It is known that when running for office, some candidates deliberately use the simplified versions of their names to help voters remember their names (Smith, Reference Smith2018). The strategic manipulations of registered names can lead to changes in name complexity scores and measurement errors.

Although this issue should be seen as a limitation of the study, three points deserve emphasis. First, if I use names that voters saw at the polling station, estimating the unbiased effects of name complexity becomes more difficult. We do not know what explains candidates' decisions to change their registered names, and candidates who simplified their names may be systematically different from those who did not in both observed and unobserved ways.Footnote ¹³ This increases a concern about omitted variable bias. By contrast, using candidates' actual names, I may alleviate this problem. Since their actual names were “exogenously” determined at birth in most cases, the assumption about no confounder may become more plausible (section A of the appendix). Given the reliance on actual names, the estimates of name complexity in this study can be interpreted as intention-to-treat effects (e.g., Angrist, Reference Angrist2006).Footnote ¹⁴ This approach only produces conservative estimates of name complexity by biasing them toward zero.

Second, for a random sample of candidates, I check how often they simplified their registered names. Comparing names presented in the JHRED and campaign manifestos that they made, I find that a fairly large number of candidates did not simplify their names. Moreover, even when they did, the correlation in name complexity scores between actual and registered names remains fairly high (section B of the appendix). Finally, I also reanalyze the main results below using an alternative measure—which I call name difficulty scores—that only moderately correlates with name complexity scores. In this way, I demonstrate that the main findings are robust to the presence of some measurement errors (section E of the appendix).

Turning to the model specification, I rely on a multilevel linear model. Formally:

(1)$$ \eqalign{ {\rm Vote} \ {\rm Share}_{\rm ct} =& \alpha + \beta_{1} {\rm Name} \ {\rm Complexity}_{{\rm c}} + \beta_{2} {\rm Name} \ {\rm Length}_{{\rm c}} \cr & + \gamma_{1} {\rm Number} \ {\rm of} \ {\rm Candidates}_{{\rm d}} + \gamma_{2} {\rm District} \ {\rm Magnitude}_{{\rm d}} \cr & + \nu_{{\rm c}} + \nu_{{\rm p}} + \delta_{{\rm t}} + \epsilon_{{\rm ct}} } $$

where, the outcome is the vote share (0–100 scale) of candidate c in election-year t (belonging to party p and running in election-district d). α is a common intercept, and β₁ is a coefficient on name complexity. I also control for name length to account for the fact that the additive index of name complexity makes longer names more likely to have higher complexity scores.Footnote ¹⁵ Further, since the vote share of the candidate is mechanically affected by the number of candidates and district magnitude, I control for these two variables.Footnote ¹⁶ The number of candidates is log-transformed because it is right skewed. Then, ν_c and ν_p are random effects by candidate and party, respectively. I assume $\nu _{{\rm c}} \sim {\cal N}(0, \sigma _{{\rm c}}^2)$ and $\nu _{{\rm p}} \sim {\cal N}(0, \sigma _{{\rm p}}^2)$. Finally, δ_t is election-year fixed effects that account for time difference, and ε_ct denotes an idiosyncratic error term. I fit the parameters of the above model with Hamiltonian Monte Carlo implemented in Stan with weakly informative Gaussian priors for the regression coefficients.Footnote ¹⁷

I estimate models for the SNTV and SMDP observations separately. The rationale for this is that since the observations of this study come from a relatively large time period, empirical models need to control for the differences in the electorate or electoral environments that existed across time. For example, during the period under study, there were gradual changes in education levels among voters. The characteristics of voters who cast their votes also varied across years due to changes in the levels of turnout. Further, media environments changed over time. If I pool the SNTV and SMDP observations and include an indicator of the electoral rules, I cannot control for these time-varying factors that may affect voters' cue-taking behavior because the electoral reform in Japan happened at one point in time. By contrast, analyzing the SNTV and SMDP observations separately, I can include election-year fixed effects δ_t. By eliminating any time difference that existed under the same rule, this approach yields conservative estimates of name complexity.Footnote ¹⁸

3.1. Main findings

Table 1 shows the results of models with the SNTV observations. Model 1 is estimated without candidate random effects. The posterior mean of the coefficient on name complexity is − 0.027 with a 95 percent credible interval of [− 0.037, − 0.017]. In model 2, I include candidate random effects. The posterior mean of name complexity is − 0.031, which is again statistically reliable with a 95 percent credible interval of [− 0.046, − 0.015]. Therefore, under SNTV, candidates with more complex names tend to receive lower vote shares than those with simpler names.

Table 1. The effect of name complexity on vote share under SNTV

Note: The outcome is the vote share of the candidate. Models 1–4 use all observations under SNTV, while models 5 and 6 only use first-time runners. Average letter complexity = name complexity/name length. Standard deviations of the parameter posteriors are in parentheses.

The effect size of name complexity is substantial. The median difference in name complexity scores among candidates in the same district is 23. According to model 2, an increase of 23 in stroke count accounts for a 0.71 percentage point decrease in vote share. Since the margin of victory under SNTV is very small (e.g., median 1.36 percentage points), the median increase/decrease in name complexity can change the winners of the election in 30 percent of the districts. Further, the maximum difference in name complexity scores among candidates in the same district is 58. This difference leads to a 1.79 percentage point decrease in vote share, which is large enough to change the winners in 58 percent of the districts.

In models 3 and 4, I test the effect of an alternative measure of name complexity. I use average letter complexity, or the average number of strokes in the name letters (i.e., name complexity score divided by name length). Model 3 does not include candidate random effects, whereas model 4 does. In both models, the posterior distributions of the coefficient on average letter complexity point to a statistically reliable negative effect, with 95 percent credible intervals of [− 0.151, − 0.073] and [− 0.190, − 0.065], respectively. Therefore, regardless of whether I use the total or average number of strokes, the results do not change.

In models 5 and 6 of Table 1, I reestimate models 1 and 3 including only candidates running for office for the first time.Footnote ¹⁹ The rationale of this exercise is that because voters have relatively little information about first-time candidates, those who are confused with informational demands may be even more likely to rely on politically irrelevant cues. As a result, the negative impact of name complexity on vote share may be larger among first-time candidates. In models 5 and 6, the posterior means of the coefficients on name complexity and average letter complexity are − 0.039 and − 0.153, respectively (95 percent credible intervals [− 0.056, − 0.022] and [− 0.216, − 0.091]). Comparing the posterior means in models 1 and 5, the negative effect of name complexity is greater in the latter. Model 5 suggests that name complexity can explain up to a 2.26 percentage point decrease in vote share. Importantly, models 5 and 6 also suggest that the statistically reliable findings of models 1–4 are not necessarily the artifact of a relatively large sample size. Even when a subset of the observations is used, candidate name complexity continues to show a negative and statistically reliable effect on vote share.

In Table 2, I turn to the results of the models with the SMDP observations. Models 1–6 correspond to those in Table 1. Models 1 and 2 test the effect of name complexity without and with candidate random effects. Models 3 and 4 test the effect of average letter complexity without and with candidate random effects. Finally, models 5 and 6 are the same as models 1 and 3 except that they only include the first-time candidates. Unlike the results of the SNTV models, I fail to find the negative effect of name complexity on vote shares under SMDP. The posterior means of name complexity and average letter complexity never show a statistically reliable effect, some of which are even signed incorrectly.Footnote ²⁰ Further, comparing the results of Tables 1 and 2, the posterior estimates of name complexity under SNTV are smaller than those under SMDP at the 95 percent level.

Table 2. The effect of name complexity on vote share under SMDP

Note: The outcome is the vote share of the candidate. Models 1–4 use all observations under SMDP, while models 5 and 6 only use first-time runners. Average letter complexity = name complexity/name length. Standard deviations of the parameter posteriors are in parentheses.

In summary, the results suggest that candidates with more complex names tend to be disadvantaged under SNTV, but not necessarily so under SMDP.Footnote ²¹ These findings are consistent with the argument that cognitively demanding electoral systems may increase the chance that voters' decisions are inadvertently affected by politically meaningless information. Regardless, any observational study like this must show more than a correlation between the key variables to make a stronger case. In the following subsections, I take several steps to validate my argument.

3.2. Mechanism

I argued that candidate name complexity can shape voter behavior and election outcomes in two different ways: (1) inattentive voters are attracted to a candidate whose name is less visually complex and easier to write and (2) votes cast for candidates with complex names are more likely to be invalidated due to major writing errors. Although the first mechanism cannot be assessed with observational data, the second one is indirectly testable using data on invalid votes at the district level. If the second mechanism partly explains the above findings, district-level name complexity should be positively associated with the proportion of invalid (uncounted) votes under SNTV, but not under SMDP.

To test this proposition, I estimate the following multilevel linear model:

(2)$$ \eqalign{ {\rm Invalid} \ {\rm Votes}_{\rm dt} =& \alpha + \beta_{1} {\rm Average} \ {\rm Name} \ {\rm Complexity}_{\rm dt} \cr & + \beta_{2} {\rm Number} \ {\rm of} \ {\rm Candidates}_{\rm dt} + \beta_{3} {\rm District} \ {\rm Magnitude}_{\rm dt} \cr & + \nu_{{\rm d}} + \delta_{{\rm t}} + \epsilon_{\rm dt} } $$

where the unit of analysis is the election-district (district d in year t), and the outcome variable is the logged proportion of invalid votes. I log-transform this variable because it is heavily right skewed. β₁ is a coefficient on average name complexity, or the mean name complexity score of the candidates running in the same district. The higher this value is, the larger the number of candidates with complex names in the district. I also control for the logged number of candidates and district magnitude as before. ν_d indicates random effects by district with $\nu _{{\rm d}} \sim {\cal N}(0, \sigma _{{\rm d}}^2)$. Further, δ_t denotes election-year fixed effects, which account for unobserved variation in voter behavior and other factors across years. Finally, ε_dt is an idiosyncratic error term. Due to data availability, the analysis is based on the Lower House elections between 1958 and 2014.Footnote ²² I again model the SNTV and SMDP observations separately.

Figure 1 summarizes the posterior means and 95 percent credible intervals of the coefficients on average name complexity.Footnote ²³ The first row shows the result of the SNTV model. The effect of average name complexity is positive and reliable with a 95 percent credible interval of [0.0004, 0.0104]. The coefficient estimate is 0.005, meaning that a one unit increase in average name complexity leads to a 0.5 percent increase in the proportion of invalid votes. Therefore, under SNTV, more votes are discarded in districts with a larger number of candidates with complex names. By contrast, the second row indicates that under SMDP, the effect of average name complexity is not statistically discernible from 0, with a 95 percent credible interval of [− 0.005, 0.002]. The null result is consistent with the findings of the previous exercises. In short, the analysis of district-level invalid votes provides some suggestive support for one mechanism by which candidate name complexity may distort election outcomes.

Figure 1. The effect of district-level name complexity on invalid votes.

Note: The figure shows the posterior means and 95 percent credible intervals of the effect of average name complexity on the proportion of invalid votes under SNTV (first row) and SMDP (second row). The outcome variable is log-transformed.

3.3. Within-SNTV heterogeneity

Finally, I test heterogeneous effects within SNTV to explore a more nuanced relationship between candidate name complexity and voter behavior. Recall that under SNTV in Japan, district magnitude varied from 1 to 6, and the number of candidates in the district varied from 2 to 23. According to Carey and Hix (Reference Carey and Hix2011), the information environment of low-magnitude multimember districts resembles that of single member districts. In fact, some of the SNTV districts in Japan are very similar to SMDP districts in terms of the number of candidates, and voters in these districts are no more likely to suffer from informational demands than those in SMDP districts. Consequently, if my argument is correct, the visual complexity of candidate names should not have any effect on election outcomes in these SNTV districts. To examine this possibility, I add the interaction term between name complexity and the number of candidates in the district to model 2 of Table 1.Footnote ²⁴

Figure 2 shows the posterior marginal effect of name complexity under SNTV conditional on the number of candidates in the district. The range of the x-axis is restricted to the 0th to 90th percentiles of the number of candidates in the data, which range from 2 to 14. Dashed lines represent 95 percent credible intervals. The figure shows that the negative effect of name complexity is statistically discernible from 0 only when there are a sufficiently large number of candidates in the district (≥ 6). By contrast, in districts with a relatively small number of candidates (2–5), which resemble SMDP districts, the effect of name complexity is not statistically reliable. Therefore, even within SNTV, voters' decisions are likely to be distorted by politically irrelevant information only when the information environment becomes sufficiently demanding.

Figure 2. Heterogeneous name complexity effect by the number of candidates under SNTV.

Note: The figure shows the marginal effect of name complexity on vote share conditional on the number of candidates under SNTV. Dashed lines indicate a 95 percent credible interval.