Party Incumbency and Information on Candidate Policy Positions

Taking these mechanisms in turn, one advantage incumbents have relative to challengers is the certainty voters have about their ideological position. Voters in general (Shepsle Reference Shepsle1972; Enelow and Hinich Reference Enelow and Hinich1981; Bartels Reference Bartels1988), and risk averse voters in particular (Morgenstern and Zechmeister Reference Morgenstern and Zechmeister2001; Eckles et al. Reference Eckles, Kam, Maestas and Schaffner2014), are more likely to vote for candidates they know because of certainty about their ideological positions. Applied to single-party regimes, voters should be less willing to vote for non-party alternatives compared to the party-backed candidates if there is less information about the positions of the non-party candidates.

The institutional design of electoral campaigns in single-party regimes means this is likely the case. Campaign restrictions endemic to single-party elections should lead to higher levels of uncertainty for non-party candidates. Single-party regimes do not allow independent candidates to mount electoral campaigns on either a collective or individual level. In Vietnam, for example, the election law explicitly bars any candidates, whether party-backed or independent, from engaging in any kind of campaigning. Their ability to campaign is restricted to meetings with voters organized by the party-backed mass organizations or through party-run provincial newspapers.Footnote ⁵ In contrast, in even restrictive hybrid regimes, opposition parties are permitted to formulate a message. In Zimbabwe in 2000, for example, although Mugabe's ruling ZANU-PF party harassed the opposition MDC Party throughout the campaign, the opposition was nonetheless able to develop a campaign message: “The MDC focused its campaign on Zimbabwe's socio-economic collapse, highlighting Mugabe's role in mismanaging national affairs” (Makumbe Reference Makumbe2002, 90). By sharp contrast, in single-party regimes information on the policy positions of non-party candidates is minimal for most voters.

Party Incumbency and Access to Resources

A second component of incumbency advantage in democracies centers on access to resources. In democracies, seniority may give incumbents access to pork that is not available to junior members (McKelvey and Riezman Reference McKelvey and Riezman1992; Muthoo and Shepsle Reference Muthoo and Shepsle2014; Hall and Shepsle Reference Hall and Shepsle2013). As with ideological certainty, the effect of access to resources—what Slater calls infrastructural power (“the ability to get things done”)—should be stronger in single-party systems (Slater Reference Slater2003, 86). Single-party regimes, by definition, allow only one party to control government, and electoral rules are often structured in such a way that the party is guaranteed a legislative majority over the combined share of non-party candidates. Therefore, because of their outsider status and lack of organizational backing, non-party candidates might be perceived as less able to enact policies or deliver particularistic benefits, even if voters prefer their positions. Non-party members will not be able to attend party meetings that precede and parameterize legislative debates (Malesky, Schuler, and Tran Reference Malesky, Schuler and Tran2011). Furthermore, they lack viable connections to top leadership and have no connection to the party apparatus that penetrates all the way down to the local level. This mechanism aligns with a large literature on the importance of clientelism in elections in hybrid regimes (Lust-Okar Reference Lust-Okar2006; Blaydes Reference Blaydes2011; Magaloni Reference Magaloni2006; Greene Reference Greene2010). In a single-party setting, party membership may be an important sign of a candidate's ability to deliver rents and/or the locality from deleterious central policies, regardless of how citizens evaluate the past performance of the party overall.

In the context of Vietnam ahead of the most recent 2016 National Assembly election, focus group discussions suggested the plausibility of this mechanism at the national level (Nguyen Reference Nguyen2016). In those groups, participants rated party membership as one of the most important factors determining their vote choice. According to follow-up discussions, voters consistently echoed Slater (Reference Slater2003) in citing the perceived ability of party members to “get things done” as important. One respondent explained the value of party membership and having an official position in the state in her vote choice, which generally correlates with party membership, thusly: “Having a position means having an influential voice, being effective locally, and having a greater ability to get work done.”Footnote ⁶

The Case: Elections in Vietnam

We test our argument in Vietnam, a single-party regime where the only legal party is the CPV. While the Vietnamese regime bans opposition parties, it allows non-party members to compete in direct elections for seats in the National Assembly.Footnote ⁷ As Table 1 shows, most voters in Vietnam have experienced elections with non-party candidates. In the three elections since 2007, nearly 14 percent of the candidates nationwide were independent, with each region in the country fielding at least 10 percent non-party members on the ballot (see Appendix A for a map of the distribution by province).

Table 1 Non-Party Candidates in Vietnam from 2007–2016 (3 Elections)

* Hanoi is in the Red River Delta

** Ho Chi Minh City is in the Southeast

As discussed above, vetting and limits on campaigns are also central features to election campaigns. It is the case that high profile candidates are prevented from running for the National Assembly elections (Gainsborough Reference Gainsborough2005). In previous elections, several prominent non-party candidates were barred from competing, such as Lê Công Định, Cù Huy Hà Vũ, and Nguyễn Quang A. It is important to note that the restrictions on prominent personalities running are not limited to non-party members. In 2007, Đặng Hùng Võ, a vociferous party member and former deputy minister of environment and natural resources, was also barred (Malesky and Schuler Reference Malesky and Schuler2009). Given that the public knows that the independent and party-backed candidates are vetted in a similar fashion, this should reduce the perceived quality differences between party-backed and non-party backed candidates, aside from the critical policy and organizational mechanisms we discuss above.

Additionally, as discussed above, Vietnamese elections feature limits on campaigning. The electoral code explicitly bars individual candidates from raising money or organizing campaign rallies. This effectively means that the only candidates able to take advantage of any policy messaging are the party-backed candidates, who are implicitly associated with the ruling party's ubiquitous and constant public propaganda campaigns. Given the vetting and campaign restrictions on party and non-party candidates, this actually makes the hypothetical survey question more realistic than it might be in a context where campaigns were allowed.Footnote ⁸

2016 Election Results

Observational data is consistent with our general argument. As Malesky and Schuler (Reference Malesky and Schuler2011) show, party members earned an electoral boost in the 2007 election even after controlling for other factors such as candidate quality and district competitiveness. Figure 1 demonstrates that this pattern held in the May 2016 contest, where 96 (11 percent) of the 870 candidates for the 500-seat legislature were non-party members. In the first panel, we report the unadjusted share of candidates that voters ultimately elected. In the second panel, we report the unadjusted average vote share received by party and non-party candidates. While 62 percent of party members won seats in the districts they contested, only 20 percent of non-party members were similarly fortunate with an average vote share of 39.4 percent. Consequently, non-party members make up only four percent seats in the current VNA.

Figure 1 May 2016 Election Results by Party Membership

The poor performance of non-party candidates may be overstated in Figure 1, as it does not account for the fact that party members tend to be more highly educated, hold more prestigious occupations, and occupy more high profile and powerful positions in the Politburo, Central Committee, and local and central governments. To address this, in Table 2 we employ a variant of matching called entropy balancing (ebalance) (Hainmueller Reference Hainmueller2012). After re-weighting, party and non-party match directly in terms of average value, variation, and skew on candidate level qualifications,Footnote ⁹ allowing us to determine how much of the excess support for party members is due to the label itself and not to individual candidate features. Table 2 presents the results of our analysis in three models. First, we replicate the simple bivariate analysis displayed in Figure 1 without ebalance in Models 1 and 2 for share elected and vote share. Then, we employ ebalance weights in Models 3 and 4. Finally, we introduce provincial fixed effects to compare only candidates running within the same province, to account for local variation in voter preferences and party support in Models 5 and 6.

Table 2 Performance of Party Members in May 2016 Election

Note: OLS, standard errors in parentheses (*** p < 0.01, ** p < 0.05, * p < 0.1)

In the fully specified models, which account for candidate and provincial level differences, we find that CPV candidates received 15 percent higher vote shares, leading to 25 percent more CPV candidates elected than non-party candidates. The bottom line is that party members enjoy a substantial unexplained level of support. Unfortunately, all of the theories discussed above (manipulation, preference falsification, performance legitimacy, and extreme incumbency advantage) are consistent with this result.

Conjoint Survey Experiment Design

Because the observational results are consistent with not only the advantages of incumbency bias, but also the preference falsification and performance legitimacy arguments, we designed a conjoint survey experiment that tests extreme incumbency advantage while reducing the confounding effects of manipulation, preference falsification, performance legitimacy, and candidate quality. Conjoint analysis allows researchers to design multidimensional treatments in survey designs and to evaluate which dimension has the most weight in determining the outcome (Hainmueller, Hopkins, and Yamamoto Reference Hainmueller, Hopkins and Yamamoto2014). This design is especially useful in determining the factors that influence voter decisions, because the object of the choice—the candidate—varies on a number of dimensions including policy position, gender, age, education, and party membership.

The conjoint design helps our analysis in three ways. First, and most obviously, the hypothetical nature of the question ensures that vote rigging is not a concern. While the abstract nature of the question may hinder generalizability, we take several measures to ensure that the hypothetical candidates match as nearly as possible the actual decisions voters face in the ballot box. Second, because a conjoint survey randomizes the features of the candidate, it can ensure that the party affiliation of a candidate is perfectly orthogonal to the features of candidate quality that are available to voters, something that was not fully achievable with the observational data above.

A third benefit is that the survey design provides shielding for respondents, such that it should reduce the role of preference falsification. Similar to a list experiment, respondents can select the non-party candidate without having to admit that that was the reason they did so (Coutts and Jann Reference Coutts and Jann2011). They can always attribute their choice to other innocuous features of the treatment. In this way, preference falsification should be limited by design. Indeed, as Hainmueller and his colleagues note: “conjoint analysis has the potential to reduce bias as compared to traditional survey experiments, because conjoint respondents are presented with various attributes and thus can often find multiple justifications for a given choice” (Hainmueller, Hopkins, and Yamamoto Reference Hainmueller, Hopkins and Yamamoto2014, 27). For this reason, conjoint analysis has been used to research sensitive topics such as HIV stigma (Schulte Reference Schulte2002). While we think the conjoint survey approach should reduce preference falsification in our tests, we take several additional steps to ensure that our results are not contaminated by preference falsification.

Data

Testing the conjoint survey responses on these different subgroups requires a large number of respondents, which we are able to obtain by including our instrument in a massive annual governance survey completed by roughly 14,000 respondents annually. The PAPI survey, conducted each year since 2011 by the United Nations Development Programme (UNDP), measures the perceived governance quality in Vietnam's 63 provinces. The survey is conducted on tablets through face-to-face interviews. Below the provincial level, sampling is clustered at the district, commune, and village levels using probability proportion to size selection procedures to select the units. In each province, 12 villages in six communes within three districts are chosen, with the goal of surveying 20 respondents per village.Footnote ¹⁰ Importantly for the purposes of this study, where interviewer effects could be a concern, PAPI interviewers are trained and managed by the Center for Community Support Development Studies, UNDP's local NGO partner, and are conducted without an official present.

Our baseline survey question, which we asked each year from 2016, approximated the information available to Vietnamese voters as nearly as possible. The design is based on information published via candidate lists in public newspapers, along with a rudimentary policy position they may have been able to glean from mandated candidate–voter meetings in electoral districts (see Appendix C for an example of an electoral district and the information available to voters). Thus, in the survey experiment, PAPI asked respondents to select between two hypothetical candidates for the Vietnam National Assembly. It then randomly assigned the two candidates to possess the characteristics listed in Table 3.

Table 3 Features of Candidates Varied in Conjoint Analysis in Multiple Years

Italicized items added in 2017 survey. In 2018, half of the respondents received the Employment information and half received the Experience information. Also in 2018, 20 percent of respondents saw candidate pairs with no policy positions.

Starting with the main theoretical variable, party membership, 50 percent of the hypothetical candidates were party members, while another 50 percent were not. We randomized the other features of the hypothetical candidates in a similar manner. As Table 3 shows, in addition to the straightforward categories of gender and age, we also included career. To make the analysis tractable, we divided the candidates’ professional experience into two categories. The first group includes members of the Fatherland Front, Vietnam's umbrella group of mass organizations (i.e. the peasant and women's unions) charged with representing mass stakeholder views in Vietnam's Leninist system, and the second group includes business candidates. In 2017 and 2018, the survey expanded the careers to doctors and lawyers. In practice, each employment category is associated with large shares of party and non-party members, so career is not correlated with party status in voters’ minds. Another feature varied was nomination status. Vietnam requires most candidates to be nominated by a provincial- or central-level political organization. However, a subset of candidates is allowed to self-nominate (Malesky and Schuler Reference Malesky and Schuler2009). In 2017 and 2018, education was included to better account for candidate quality, as qualitative evidence suggests that voters tend to cue on education level as a measure of talent. It is important to note that all combinations of features in the hypothetical candidates are possible and occur frequently in Vietnamese elections. Because party members are allowed in business, there are members of the business community that run for election who are also party members. There are also non-party members who are nominated by the state to run and party members who self-nominate.

The experiment also allowed the candidates to declare a policy priority. Although candidates cannot campaign, they are allowed to express priorities in meetings with voters during the 30-day period ahead of the election. The experiment varied among four different policy priorities that candidates might have emphasized in these meetings, including support for increased foreign trade, support for environmental protection, support for economic growth, and support for poverty reduction. With this set of baseline variables, we can estimate the degree of support for candidates based on their party status independent of other candidate qualities that might co-vary with party membership.

The 2018 Factorial Survey Experiment Design

While the baseline survey allows us to measure support for the party, it is less able to directly analyze the specific mechanisms of incumbency bias. In particular, the baseline survey does not allow a test of how party support changes based on the informational environment or organizational backing. Therefore, in the 2018 iteration of the PAPI survey, we enhanced our research design to better test the informational (H1) and organization mechanisms (H2) for incumbency bias. To do this, we exposed randomly assigned respondents to small changes in the 2017 version of the conjoint instrument in order to better isolate their reactions to uncertainty in information about candidate policy and infrastructural power. These two experiments were administered in the form of the 2X2 factorial design shown in Figure 2.

Figure 2 Factorial Design to Test Incumbency Bias Mechanisms (2018)

In the first experiment, we randomly assigned 20 percent of respondents to “no policy” conditions, where neither candidate offered a policy priority. In line with H1, this experiment enhances uncertainty about candidates’ policy positions, so we expect that respondents with no information on policy positions will gravitate toward the party candidate. The delegates receiving the “no policy” treatment are compared to the counterfactual group that received the full set of four policy priorities in Table 3. In the second experiment we randomly assigned half the respondents to learn about the candidates’ experience in the VNA instead of the candidate's current career (business, lawyer, doctor, or VFF), varying whether they candidate had 0, 5, or 15 years of working experience in the government. Consistent with the literature on the US Congress, which equates seniority with the ability to access resources (Hall and Shepsle Reference Hall and Shepsle2013), our expectation is that respondents view experience in the government as similar to organizational power, and therefore would see non-party candidates with more experience as possessing infrastructural power. Consequently, if H2 is correct, respondents exposed to the experience treatments should be more likely to vote for non-party candidates.Footnote ¹¹

Figure 2 shows our research design. The pure control is visible in the second bar. These are respondents that have been exposed to neither the candidates’ policy nor their experience. According to our theory, under this condition of high uncertainty, incumbency bias should be the highest, and non-falsifying voters should favor the party candidate at the highest level. By contrast, the third bar indicates the treatment group where respondents were exposed to both policy priorities and experience. According to our theory, incumbency bias is lowest under these conditions, and party candidates will receive the lowest support from non-falsifying voters. In designing the question, we had no expectations about which of the two treatments (information on policy or experience) would have the substantively larger effect on voting for party members.

Addressing Preference Falsification and Defining Non-Falsifiers

Two additional features of our research design are the ability of the conjoint instrument to reduce preference falsification and our ability to identify respondents who are not providing falsified support for the party. Starting with our evaluation of the ability of the experiment to shield the respondents, Table 4 shows the non-response rate for the conjoint survey compared to direct questions about voting behavior in the PAPI survey. The first question (D101d5a) asks, for those who remember there being a non-party member on the ballot in the 2016 election, whether they voted for the non-party member.Footnote ¹² Less than 20 percent of the sample in all three years remembered (or admitted to remembering) there being a non-party member. The skip pattern of the survey therefore ensured that the 80 percent who did not remember a non-party candidate were presented with a hypothetical question that is the closest direct analogue to the conjoint question in the survey (D101d5b): “Would you vote for a non-party candidate if one were on the ballot?”

Table 4 Average Support for Non-Party Candidate and Non-Response

Mean of D102 is calculated by running an OLS regression of vote choice on party and non-party while suppressing the constant.

The key thing to notice in Table 4 is the tremendous reduction in non-response from using the conjoint experiment (D102), which is shown in the third row.Footnote ¹³ Compared to its direct analogue (D101d5b), item non-response rates for the conjoint question were eleven percentage points lower in 2016 and eight percentage points lower in 2017 and 2018. We interpret this as additional evidence that respondents were more comfortable answering the conjoint question. Non-response leads to systematic measurement error, as respondents are not skipping questions randomly. When they refuse to answer, they clearly are trying to hide sensitive information or are frightened by the question.

Beyond the utility of the conjoint question in reducing measurement error, Table 4 offers another helpful remedy for skirting the preference falsification problem for those unconvinced that respondents answer the conjoint question honestly. Respondents who answered D101d5b affirmatively are clearly voters who are not afraid to admit their true views to authorities. We refer to these as the non-falsifiers. By limiting statistical analysis to their responses, we can capture the level of regime support from only the set of voters who are not afraid to admit their true preferences for non-party candidates. Therefore, while we might omit some non-falsifiers who are otherwise happy with the party, we can be confident that these non-falsifiers clearly are unafraid to say they would vote for a non-party candidate. Given their stated willingness to vote for a non-party candidate, any support for the party from these respondents should further bolster the argument that factors other than preference falsification underlie support for the party.

Regression Results

Figure 3 reports the results for all consistent covariates across the three survey waves among all the respondents exposed to the question.Footnote ¹⁶ We observe a consistently strong effect for party membership across the entire population and within the subpopulations. For the entire population, party membership increased the likelihood of selection above non-party membership by 14 percentage points in 2016, 11 percentage points in 2017, and 13 points in 2018. To put this figure in perspective, when all covariates are held at their mean level in 2016, respondents voted for party members 56.8 percent of the time and for non-party candidates 42.6 percent of the time (a 14.2-point advantage). Importantly, each of the iterations over the three years offer qualitatively similar results and are not statistically distinguishable from the 15-point party advantage we estimated from the actual vote data in Table 2 above, which is a strong indication of both the reliability and validity of our findings.

Figure 3 Marginal Effects of Candidate Characteristics on Vote Choice

In Table 5, we turn to the subpopulations analysis. For presentational purposes, we show the full conjoint results on all respondents in the first model for each wave. In the second model, we see a fascinating effect when include only those respondents that said in direct questions that they would support non-party members. As the results from 2016 and 2017 show, support for party members among this group drops but remains positive.Footnote ¹⁷ In other words, even those who say in direct questions they would be willing to vote for a non-party member still favor party members in the shielded, conjoint experiment.

Table 5 Effect of Party Membership in Conjoint Survey Experiment

Linear probability with marginal probability of voting for party over non-party shown. Standard errors, clustered at province, district, and commune in parentheses (*** p < 0.01, ** p < 0.05, * p < 0.1). Full regression results are available in Appendix J. 2018 results are divided into two panels based on whether respondents was exposed to career or experience treatment. “No falsification” includes only the respondents who said in a direct question that they would vote for a non-party candidate (see Appendix G for the question). “Econ neutral” includes only those who were neither positive nor negative on current economic situation (see Appendix E). The full specification includes the subset that was both neutral and directly expressed willingness to support non-party candidate.

Looking at the 2018 results, however, we see a marked shift among the non-falsifying respondents. For the respondents who were provided information about the candidates’ job and not the candidate's experience, we see that even the non-falsifiers prefer party members to non-party members by about 10 percentage points. This is nearly identical to the 2017 and 2016 results. However, for the non-falsifier respondents given information about the candidates’ experience, the difference in the support for the party member versus the non-party member candidate disappears. The marginal effect is only 1.8 percentage points and is not statistically significant. As our analysis below shows, this suggests that once organizational backing is controlled for as a determinant of vote choice, non-falsifiers withdraw their support for the party. Other respondents, however, are less affected. We return to this when we explore sub-group effects below.

Alternative Levels of Performance Legitimacy

It is possible that the subsample of respondents expressing neither satisfaction nor dissatisfaction with their current economic situation does not fully account for the possible confounding effects of performance legitimacy. If those with neutral opinions of economic performance have a positive evaluation of the CPV's performance, the positive effect of party membership could reflect this performance legitimacy rather than incumbency advantage. This is especially important because there is some theoretical overlap between performance (which is retrospective) and organization capacity (which is prospective) when performance history shapes future expectations of service delivery. Additionally, we must distinguish the ability to extract resources from the party from the evaluation of party performance. To account for this fuzziness, we repeat our analysis on all the different subgroups of respondents depending on the different economic evaluations shown in Appendix E. If we limit our analysis to only those with low evaluations of the economy, this should limit the possibility that retrospective evaluations of performance of the party as a whole are responsible for party support.

Figure 4 shows the results of a replication of Table 5 (Models 3, 7, 11, and 15), but at different levels of economic satisfaction. Not surprisingly, the results show some effect of performance legitimacy. In 2016, 2017, and 2018 the value of the party label declines based on the level of economic satisfaction. However, even at the lowest levels of economic satisfaction the party still receives a positive and statistically significant electoral boost. In 2016, among those who said their economic situation was bad, party members were 10 percentage points more likely to receive the vote than non-party members. Party support among this low group is eight percentage points in 2017 and 2018. This indicates that even at lower levels of economic satisfaction, party members have a sizable advantage, giving us reason to believe that factors additional to performance legitimacy and preference falsification must drive support for the party.

(QA11: As for your own family, how do you rate your economic situation today? Is it … ?)

Figure 4 Effects of Party Membership at Different Economic Satisfaction Levels

Conditional Marginal Means

Two methodological concerns influence our choice of specification for the 2018 analysis. First, recent work on conjoint experiments highlights the need to consider the conditional marginal means (CMM) rather than just the average marginal component effect (AMCE). This is because the AMCE relies on an arbitrary base category (Leeper, Hobolt, and Tilley Reference Leeper, Hobolt and Tilley2019). Therefore, when comparing across subgroups, it is not immediately clear from the AMCE whether the difference is a result of a change in the value for the base category or changes in the categories that are presented. For this article, of concern is whether the policy conditions and organizational backing are changing how respondents view party members or non-party members.

An additional concern is that our factorial design had two treatments, as shown in Figure 2. Our first subset includes respondents that are given policy traits versus those that are not. The second subset includes respondents that are given the experience (organizational backing) versus the job background of the candidates. Not interacting these two treatments could potentially bias findings, given that treatments may work differently depending on which other treatment the respondent viewed (Muralidharan, Romero, and Wuthrich Reference Muralidharan, Romero and Wuthrich2019). The solution to both these problems is to present the CMM of the triple interaction. This not only solves some of the methodological concerns above, but also provides a clear visualization of the effects of the two treatment conditions on all respondents as well as the non-falsifiers. Models 4 and 8 of Table 6 therefore present the following specification. We run the following the model shown in the equation below in Table 6. To calculate the CMM in Figure 5, we suppress the constant and include all potential covariates values in the matrix X, so that the results are not dependent on the selection of base category:

$$\eqalign{ vote_{i,j} &= \beta _0+\beta _1P_{i,j} + \beta _2Nopolicy_i + \beta _3Experience + \beta _4P_{i,j}\,\ast\, Nopolicy_i \cr & \quad + \beta _5P_{i,j}\,\ast\, Experience_i + \beta _6P_{i,j}\,\ast\, Nopolicy_i\,\ast\, Experience_i + X_{i\comma j} + \varepsilon _{i\comma j}} $$

We present the results of the analysis in Table 6, which is divided into two panels for all respondents and non-falsifiers. Each panel contains four models, indicating the direct effects of the three component terms (Models 1 and 5), the interaction between Party and No-Policy (Models 2 and 6), the interaction between Party and Experience (Models 3 and 7), and the triple interaction (Models 4 and 8).

Figure 5 Impact of Party Status Conditional on Policy Information and Organizational Backing

Table 6 Testing the Experience and Policy Uncertainty Mechanism (Randomized in 2018)

Linear probability model with marginal probabilities. Regressions include all conjoint control covariates. Standard errors, clustered at province, district, and commune in parentheses (*** p < 0.01, ** p < 0.05, * p < 0.1); No Policy indicates that the candidate pairs did not express a policy preference; Treatment Experience indicates respondents exposed to the experience covariates, as opposed to the job covariates. Non-falsifiers are those that answered in a direct question that they would vote for a non-party candidate if one was on the ballot.

Focusing on the fully specified Models 4 and 8, we see that when the policy is known and candidate experience is not stated, the preference for party members is 13 percentage points higher than non-party members among all respondents and 8 percentage points higher among non-falsifiers. However, when there is uncertainty about policy, all respondents (5.3 percentage points) and non-falsifiers (14.8 percentage points) shift their votes further toward party members. Informing voters about candidate experience also impacts support for the party, but only for the non-falsifiers. While the effect is insignificant for all respondents, it is a substantial 10.6 percentage points for non-falsifiers. Finally, the null effect on the triple interaction term illustrates that the “no policy” and “experience” treatments do not influence one another. Learning about policy and experience operate independently.

The net effect of the two experiments are illustrated in Figure 5, which shows the CMM of the triple interaction from Models 4 and 8 in Table 6 (Leeper, Hobolt, and Tilley Reference Leeper, Hobolt and Tilley2019). The range bars represent 95 percent confidence intervals around the probability of voting for the party under different conditions. In the first panel, which looks at all respondents, the probability of voting for the party is about 13 percent higher than for non-party members when policy differences are stated, The difference shifts to 18 percent when policies are not stated. There is little difference in the effect of party for the entire sample based on the experience treatment.

When we look at the at the second panel of the graph (non-falsifiers), we can see that the probability of voting for the CPV is sizable and positive (22 percentage points greater than the non-party candidate) when both candidates’ policy priorities and experience are unknown. In this scenario 60.5 percent of non-falsifiers select the party candidate versus 37.6 percent who vote for the non-party candidate. However, when non-falsifiers are provided with differentiating policy information and candidates’ experience, they are indifferent between the party and non-party member candidates. Consistent with our expectations in H1 and H2, the lowest preference for the party occurs when non-falsifiers are aware of both policy and experience (2.5 percentage points lower than non-party members). In this scenario the predicted vote for party members was 46.7 percent, but it was 51.25 percent for non-party members. Furthermore, unlike for all respondents, the organizational backing has an effect for the non-falsifiers. When no policy is stated, the effect of organizational backing drops the premium for party membership from 22 to 11 percentage points. When policies are stated, the effect of organizational backing drops support from an eight percentage points benefit to a two percentage point deficit vis-à-vis the non-party candidate. Thus, for the entire sample, information drives support for the party. However, for non-falsifiers, it is clear that party status signals clarity about both policy direction and organizational capacity.