The overvaluing of expertise in discussion partner choice

1. An Information Shortcut

Since Downs (1957), scholars have noted that social communication is a potentially rational information shortcut that would allow individuals with high information costs to free ride off of the efforts of those with lower information costs (e.g., Mondak, 1995; Lupia and McCubbins, 1998). This process works most effectively when individuals choose discussion partners who share the individual’s preferences. Preference similarity is important because political discussion can be seen as a form of strategic communication in which the informant wants to persuade others to adopt his or her position. This desire to persuade others would be present even if discussion partners realize that changing one voter’s mind is not going to affect election outcomes. The discussion partner need only have the much smaller goal of having others agree with him or her for persuasion to be a goal of discussion.

As cheap talk models demonstrate, there is an incentive to misrepresent the positions of candidates when discussing politics with people with differing preferences (Crawford, 2003; Crawford and Sobel, 1982; Farrell and Rabin, 1996). ¹ This misrepresentation of candidate positions need not involve an absolute distortion of the candidates’ positions – e.g., saying a conservative is, in fact, a liberal. Rather, a potentially more persuasive message would involve slanting the facts. Slanting could be done by pulling the listener’s otherwise less preferred candidate closer to what the listener thinks is optimal: for example, ‘Obama’s position on health care is much more moderate than what the conservatives say’. Or the slanting could be pushing the speaker’s less preferred candidate away from the listener’s ideal: ‘the right-wing ideologues oppose reform because they are in the pockets of the insurance industry’.

This does not mean that signals from biased sources are never useful. For example, Calvert (1985) notes that an unbiased receiver can glean valuable information from biased sources when the source provides unexpected information (e.g., if in the USA a Republican endorsed a plan by President Obama). For receivers who tend to prefer one outcome or the other, however, signals from the other side are much less valuable. Imagine a Democrat who receives a signal from a Republican. A signal to vote Republican should be ignored as self-serving while a signal to vote Democratic will, at best, increase the Democrat’s confidence in a decision that he or she would have made without discussion. Information from a fellow partisan can increase confidence in the correctness of a party vote or highlight the situations where the candidate from one’s party is poor.

Why then would an individual ever discuss politics with someone with conflicting preferences? Ideal informants – experts who share the individual’s preference – may not be readily available and the search for these informants may be costly (Huckfeldt, 1983; Ahn et al., 2013). Individuals in this case may be forced to make compromises between political expertise and ideological proximity. Several studies of discussion patterns suggest that individuals value expertise over proximity (Huckfeldt, 2001; Huckfeldt et al., 2005). Unfortunately, the survey evidence cannot distinguish between whether individuals are seeking out expert informants or experts are constantly talking about politics. Surveys can show that experts play an outsized role in the social communication process with perceived experts having greater influence of the vote choices of their discussion partners (Ryan, 2011b; Richey, 2008b). Further, citizens often defect from their party when their social networks have members of the opposition party (Beck, 2002). Hence, it seems that voters place greater emphasis on expertise than shared preferences when evaluating social communicated messages which may lead citizens to cast votes for candidates they would not support if fully informed.

Not all citizens will be led astray by messages from experts who have conflicting preferences. The strategic sophistication of the speaker and the listener will affect the probability the speaker will try to mislead and the probability the listener will be led astray (Crawford, 2003). Discounting of communicated messages may be due to the listener’s knowledge of the diverging underlying preferences between the listener and speaker, a fundamentally strategic consideration. Listeners may discount messages that differ from one’s own prior beliefs (Ahn et al., 2010). ² Therefore, depending on (1) the availability of ideal political discussants, (2) the intention of and communicated message from the chosen discussant, and (3) the belief of the listener about the usefulness of socially communicated information especially when it comes from someone with different political views, the consequences of social communication may differ at the individual and aggregate levels.

Currently survey and experimental research offer mixed evidence on the value of social communication as an information shortcut. ³ Both sides of this literature, however, are in agreement with the models that preceded it (e.g., Downs, 1957; Crawford and Sobel, 1982) in the importance of shared preferences when evaluating social messages. Our argument is that some individuals tend overvalue expertise in choosing discussion partners and undervalue the importance of shared preferences. This leads to negative consequences for those who follow the advice of these experts with conflicting opinions. Hence, individuals may understand the rational value of the information shortcut, but have less understanding regarding how to exercise that shortcut efficiently.

2. Experimental design and procedure

At the heart of this paper is a tension between what individuals should do according to formal models and what individuals actually do according to survey research. ⁴ Previous theories argue individuals should place greater emphasis on shared preferences when engaging in social communication; but citizens in surveys appear to place greater emphasis on expertise. We ask if individuals would choose based on shared preferences more frequently if we highlight the strategic considerations? Further, when subjects do choose based on shared preferences, do they make better decisions?

We answer these questions through an incentivized experiment. The experiment does not test if subjects behave rationally. Rather, it is in the style of what Vernon Smith (1982) calls ‘heuristic’ or ‘exploratory’ focusing on the consequences of experimental treatments that are previously unexplored. The experimental framework allows us to observe aspects of how interconnected voters make decisions in ways that researchers could not using traditional survey methods. While the experiment is too complicated to fully analyze as a non-cooperative game, it allows us to set some analytical benchmarks, observe the entire communications process and, set an objective measure of a correct decision. Hence, we can figure out which subjects make better choices and what decisions in the process lead to those choices.

Subjects in this experiment were recruited from political science classes at a public university in the United States. They participated via networked computers using zTree (Fischbacher, 2007). A total of 10 sessions were run using seven subjects in each session. In the experiment, subjects are choosing between two computer generated candidates with unknown positions on an abstract policy scale. We randomly assign: (1) the candidate’s positions on the policy scale; (2) the subject’s positions on the policy scale; (3) how much money the subjects will have to pay for information regarding the candidate’s position on the scale.

Below, we describe the main elements of the experiment and then summarize the experimental procedure. ⁵

3. Making the right discussion partner choice

The experimental setting is too complicated to analyze in full as a non-cooperative game, but it is not devoid of an underlying theoretical framework. Thus, partial analyses on some aspects of the strategic incentives are possible based on the explicit assumptions of the setup.

In these analyses, we assume that subjects are self-interested – this assumption is common to all the previous models. As such, we do not account for concerns regarding fairness (Fehr and Schmidt, 1999) or reciprocity (Bolton and Ockenfels, 2000), although, some subjects likely do have these concerns. Further, we do not account for potential reputation effects that may occur due to the repeated nature of the experiment. Given the uncertainty about the information that subjects receive, it would very difficult for a receiver to judge whether a sender intentionally lied in sending a misleading message or simply truthfully reported the information gleaned from poor private information. Our analysis focuses only on expertise and shared preferences because the rational, experimental and survey literature regarding the efficacy of social communication has primarily examined the effects of these two factors.

We begin the analysis by determining the value of information. Recall that that candidate A’s position is an integer between 1 and 6, candidate B’s position is an integer between 2 and 7 and the candidates’ positions are randomly and independently drawn from uniform distributions. Thus, there are 36 possible and equally likely configurations of the two candidates’ positions with the two candidates’ positions identical in 5 of the 36 configuration.

If the positions were known, voter 1(7) should vote for candidate A(B) in 21 cases, for B(A) in 10 cases, and the voter is indifferent in 5 cases. If voter 1(7) cast her vote with the public information only, she should vote for candidate A(B) and the probability that the vote is cast correctly is 26/36 or about 72%, counting the votes when indifferent as correct. Similar calculations show that the ex ante probability of voting correctly is 25/36 (about 69%) for voter 2 and 6, 24/36 (about 67%) for voter 3 and 5, and 23/36 (about 64%) for voter 4. ⁸ This analysis implies that if the cost of purchasing private information is the same, the middle voter (voter 4) has the largest incentive to purchase private information and the extreme voters (voter 1 and voter 7) have the smallest incentive.

More information increases a voter’s likelihood of voting correctly which in turn increases the probability of victory for the voter’s favored candidate but information is costly for the majority of the voters in our experimental setting. How much should one pay for information? We determine upper limits of the value of information assuming that one’s vote is pivotal instead of a full equilibrium analysis.

We ran one million computer simulations for voters to assess the value of information. ⁹ Recall that one gains 50 cents if the candidate closer to the subject wins, loses 50 cents if the candidate further from the subject wins and a subject does not gain or lose any money if the two candidates care equally distanced. The marginal expected value of the first piece of information is about 11 cents, the second about 6 cents, the third about 4 cents and the fourth about 2 cents. Note that these are the maximum values assuming that a voter uses the private information fully and correctly and the voter is pivotal in determining the electoral outcome. The actual values would, thus, be smaller than these when voters do not use the correct Bayesian updating method (in which subjects use the information to narrow down the possible candidate positions) and/or when one’s vote is not pivotal.

What does this suggest for how individuals should choose discussion partners? Subjects face a tradeoff that is a function of (1) the subject’s information level (i.e., the ex post calculated probability of correct voting with private signals) and (2) the configuration of potential informants in terms of the preference similarity and the information level. The necessary condition is that by getting information from an informant the probability of correct voting should increase. The sufficient condition is that the informant is the best among the worthy informants (who pass the necessary condition).

The analysis suggests that in almost all cases subjects should place greater emphasis on choosing someone with shared preferences rather than someone who purchased more information for two reasons. First, the marginal value of additional information decreases rapidly – subjects would vote correctly two thirds of times even without private signals due to the availability of the public information and the cases in which two candidates are equally distanced.

Second, as Figure 1 shows, the probability of conflict rapidly increases as the distance between a pair of voters gets longer. In the figure, the light bars represent the probabilities of conflict for a citizen with a position of 1 while the darker bars represent the probabilities of conflict for a citizen with a position of 3. Figure 1 shows that as the distance between a pair of voters’ ideal points increases the probability of conflict greatly increases when the two voters are on different sides of the midpoint. In almost all cases, therefore, a rational voter would choose one of the immediate neighbors as the source of information. ¹⁰

OPEN IN VIEWER

Figure 1.

Probability that two members of a dyad will be in conflict for different distributions of possible candidate positions.

Do the subjects behave as we would suggest? Table 1 shows the extent to which subjects use information levels and shared preferences in their discussion partner choice by comparing the characteristics of all subjects to the characteristics of the chosen subjects. As the table shows, egos (the subjects choosing an information source) put a premium on the information levels of their alters (the subjects chosen as an information source). About 29% of subjects purchase the maximum of four pieces of information, while subjects who purchase the maximum information levels make up about 55% of the chosen alters.

OPEN IN VIEWER

Table 1.

Choosing alters by information level and shared preferences.

A. Information level
	All subjects	Chosen alters
0	13.9%	1.6%
1	24.6%	14.9%
2	28.2%	21.1%
3	4.8%	7.8%
4	28.6%	54.7%
N (Egos)	1050 (70)
B. Distance between ego and alter
	All dyads	Chosen alters
1	28.6%	44.0%
2	23.8%	24.6%
3	19.1%	14.0%
4	14.3%	7.6%
5	9.5%	6.0%
6	4.8%	3.8%
N (Egos)	6300 (70)	1050 (70)

While a discussion partner with more information is better, if subjects are reaching for ideologically divergent alters in order to choose a better informed discussion partner, then they are likely making a mistake. To some extent subjects recognize this as the modal case is choosing the alter next to them on the ideological scale and the probability that an alter is chosen decreases with distance. Nevertheless, we argue that subjects should choose an alter with the closest matching preferences in almost all cases. Subjects did not do this. Thirty-six percent of chosen ego-alter dyads lay more than two spaces away on the scale which means that many subjects crossed the median to choose an alter. As we have already shown, that means many subjects chose discussion partners with an incentive to mislead them.

4. How alters affect correct voting

In the experiment, subjects vote correctly 78% of the time. We say a subject votes correctly if the subject votes for the candidate that will offer them the larger payoff or if the subject was indifferent between the two candidates. ¹¹ How then did an ego’s choice of alter affect the likelihood he or she votes correctly? Table 2 presents the results of Heckman probit models which model alter choice in the first stage and then correct voting in the second stage. All possible ego-alter dyads are cases in the choice model while only dyads where an alter was asked and provided information are cases in the correct vote model. ¹²

OPEN IN VIEWER

Table 2.

How does a discussion partner affect a subject’s ability to vote correctly? Selection model with standard errors corrected for clustering on subjects.

Correct vote model	Coefficient	Z-Value
Distance between dyad members	0.066	1.30
Dyad members on same side of median	0.446	3.73
Logged error in ego’s estimates	–0.518	–4.51
Logged error in alter’s estimates	0.006	0.06
Logged bias in alter’s messages	–0.212	–2.51
Constant	1.594	7.99
Selection model	Coefficient	Z-Value
Alter information level	0.541	7.36
Distance between dyad members	0.051	0.69
Distance*information level	–0.090	–3.81
Constant	–1.944	–8.76
ρ (Standard Error)	–0.253 (0.147)
AIC	5330.108
N	6153
Uncensored observations	903
Number of subjects	70

The alter choice model includes two independent variables, the potential alter’s information level and the distance between ego and alter on the abstract scale, as well as, the interaction between these two variables. The results show that subjects placed most emphasis on alter information level, but were trying to follow the Downsian recommendation. If an alter purchased four pieces of information, then there is a 46% probability that the ego will choose that alter if he or she is one position away. That probability drops to 5% if the alter is six positions away. If a potential alter is uninformed, then the distance between ego and alter is irrelevant because there is very little chance that alter will be chosen.

While egos are unlikely to request information from alters with clearly competing preferences, they do not grasp the non-linearity in conflict either. There is a 24% probability that an ego chooses an alter who is three places away, even though, that discussion partner will prefer a different candidate at least half the time. Subjects would prefer not to cross the median, but they are willing to do so if it means finding an expert discussion partner. Theoretically, they would be better off sticking with a less informed partner with preferences that are more similar.

Empirically, this is the case as well, as the correct voting stage of the model in Table 2 shows. The distance between dyad members did not affect the probability of a correct vote, but being on the same side of the median did. For example, a subject at position 3 would have an 81.1% probability of voting correctly if he or she communicated with an alter at position 1. That probability drops to 73.3% if the alter is at position 5. In both cases, the alter is only two positions away, but in the second case the alter is on the other side of the median.

The other variables in the model measure the accuracy in the initial estimates of the candidate positions for egos and alters as well as the bias in the messages the alter sends. The error in estimates variables are measured by adding together the absolute differences between a subject’s estimates of both candidates’ positions and the candidates’ actual positions. The bias in messages is measured in a similar manner by calculating the absolute differences between the alter’s initial estimates and the signals the alter sent to the ego. These values are logged because effects of the mistakes should diminish as the errors are larger. ¹³

Unsurprisingly, egos who made larger mistakes in their initial estimates were less likely to vote correctly. Interestingly, the accuracy of an alter’s initial estimate did not affect correct voting. The bias in an alter’s messages, however, did. This suggests how much an alter knows is less important than whether or not the alter is willing to lie. Because the subjects act anonymously, egos cannot use personality traits to judge the honesty of a potential alter. They can use the alter’s information level, however. Better informed alters should be more certain of the candidates’ positions and correspondingly be more confident in which candidate they support and which their egos support. This will allow better informed alters to send more biased messages.

This is shown in the OLS model in Table 3 which has the (unlogged) bias in alter messages as the dependent variable. If an alter has no information, then, on average, the alter will send signals that results in a total bias of about 1.3. That value increases, to 2 if the alter has four pieces of information. The information level of the ego had an effect of a similar size. Alters sent more unbiased messages to better informed egos because the egos could verify the accuracy of the alters’ signals (Lupia and McCubbins, 1998).

OPEN IN VIEWER

Table 3.

Total bias in messages sent to egos by alters. OLS model with standard errors corrected for clustering on egos.

	Coefficient	Z-Value
Distance between dyad members	0.237	2.27
Dyad members on same side of median	–0.612	–2.71
Ego information level	–0.199	–2.13
Alter information level	0.181	1.73
Constant	1.480	3.87
N (Subjects)	903 (70)
R 2	0.08

The largest effect on bias relates to the differences between ego and alter’s positions on the scale. The amount of bias increases as the dyad’s positions are further apart, but this increase is decidedly non-linear as shown in Figure 2. Figure 2 takes subjects with ideal points of 1 (the darker points) and 3 (the lighter points) and shows the predicted bias in messages sent by all other subjects. As the figure shows, the amount of bias jumps if ego and alter are on opposite sides of the median. For example, if a subject has an ideal point of 3, then they would expect to see an average bias of 2.1 points on the scale from an alter with an ideal point of 5 and an average bias of 1.4 from an alter with an ideal point of 1. Both alters are two points away on the scale, but only subject 1 is on the same side of the issues as subject 3.

OPEN IN VIEWER

Figure 2.

Predicted bias for different ego/alter pairs based on OLS model in Table 3.

It should be noted that overall, the biases in the messages sent are fairly small. About 71% sent had a total bias of two points or less – either one point for each candidate or two for one candidate and an unbiased message for the other – with no bias at all in more than a third of the signals. Still, about a quarter of social messages misrepresented the alter’s beliefs about which candidate the ego should support. Often, there was no need to send a largely biased message to tell an ego that he or she should support a different candidate than the alter believes. Further, a message that is only slightly biased might be more persuasive to an ego who has some idea about where the candidates are placed.

5. Changing the ego’s mind

We can use the ego’s initial estimates to determine which candidate the ego would have chosen prior to social messages by determining which candidate the ego believes is closer to him or her. An ego is persuaded by social messages when the ego changes his or her vote from the vote implied by their initial estimates of the candidates’ positions. In some cases, the change will be helpful – the ego was persuaded to vote for the candidate closer to him or her – but in other the change will be harmful – the ego was going to vote for the right candidate if the social messages had not changed his or her mind.

Most subjects stuck with the candidate they were supporting prior to social information. About 75% of subjects did not change their vote following social information and in 89% of those cases they voted correctly. When a subject did switch his or her mind, that decision was equally likely to be correct as incorrect. Who then is aided and who is hurt by social information? The most basic question to address concerns the effect of an ego’s information level. Table 4 presents cross tabulation between information level and a three category dependent variable: (1) helpful change, (2) harmful change and (3) no change.

OPEN IN VIEWER

Table 4.

Whether it benefitted subjects to change their vote following social communication by private information levels.

	No information	Low information (1 or 2 pieces)	High information (3 or 4 pieces)	Total
Helpful change	26.17%	13.19%	8.90%	13.18%
Harmful change	16.82%	14.47%	8.90%	12.74%
No change	57.01%	72.34%	82.21%	74.09%
N	107	470	326	903

It is important to notice three things. First, information level is directly tied to whether or not a subject changed their vote: subjects with less information were more likely to change their vote. Second, information level is not as strongly tied to whether or not that change was a good idea. When uninformed subjects switched their vote, this led to a better vote choice more often than not. For subjects with any information, changing their vote was equally likely to be a good or a bad decision. Finally, in the aggregate harmful switches were as likely as helpful switches since most subjects had some information. This suggests that even if egos are able to choose their alters purely on the basis of the expertise and preferences of the alter, social communication may not lead to more ‘enlightened’ voting in the aggregate.

The multinomial logit model in Table 5 examines the factors that determined helpful and harmful vote switching in greater detail. The dependent variable is the same as in Table 4 with no change as the baseline category. The independent variables are the same as in the correct voting model in Table 2: the distance between the ego and alter, a dummy variable coded 1 if the ego and alter are on the same side of the median, the logged error in the initial estimates for both the ego and the alter and the bias in the alter’s messages.

OPEN IN VIEWER

Table 5.

Did social information aid the subject? Multinomial logit model with standard errors adjusted for clustering on subjects. Baseline category is when the subject did not change his or her vote following social information.

	Harmful change		Helpful change
	Coefficient	Z-Value	Coefficient	Z-Value
Distance between dyad members	–0.115	–1.26	–0.265	–2.32
Dyad members on same side of median	–1.191	–3.87	–0.468	–1.58
Logged error in ego’s estimates	0.282	1.30	1.710	6.90
Logged error in alter’s estimates	0.035	0.17	–0.101	–0.40
Logged bias in alter’s messages	0.285	1.36	–0.208	–1.21
Constant	–1.721	–4.46	–2.810	–5.21
N (subjects)	903 (70)
AIC	1280.04

The best predictor of helpful change is the accuracy of the ego’s initial estimates. Egos who made larger mistakes in estimating the candidates’ positions were more likely to change their mind following social communication and make a correct vote. The initial estimates of the alter and the absolute bias in the alter’s messages did not have a statically discernible effect. Recall, that the bias in the alter’s messages were typically not very large. So, a subject who was very mistaken about the placements of the candidates would be aided by social messages that were slightly biased because they still may better reflect reality than what the subject believed.

None of the alter level variables affect probability of change in either direction, but the dyadic distance variables do. As the distance between ego and alter increases, the probability that the subject makes a helpful change decreases. Further, egos are also less likely to make a harmful change if they talk with an individual on the same side of the median. The sizes of these effects are displayed in Figure 3. The light bars in the top graph in Figure 3 display the probability that a particular alter will help a subject at position 1 while the dark bars display the same probabilities for a subject at position 3. The bottom graph in Figure 3 displays the probabilities that alters will lead to a harmful vote switch.

OPEN IN VIEWER

Figure 3.

The probability of helpful and harmful changes for different ego/alter pairs based on the multinomial logit model in Table 5.

The top graph shows egos who select alters on the opposite end of the policy scale have a decreased probability of a helpful change. There is an 11% probability that an alter at position 2 will help an ego at position 1 vote correctly. That probability drops to 4% if the alter is at position 7. For egos at position 3 − and hence, closer to the middle of the scale – there is no relationship between the position of the alter and the probability of helpful change because the distance between ego and alter is never wide enough for the probability of helpful change to decrease substantially.

While dyadic distance only affects the probability of helpful change for those with extreme positions, the likelihood of vote conflict affects the probability of harmful change for all subjects. As the bottom graph in Figure 3 shows, subjects are less likely to change their vote to the wrong candidate if they choose an alter on the same side of the median. There is less than a 7% chance that an alter on the same side of the median will lead a subject astray. The probability of a harmful change increases by as much as 12 percentage points if the alter is on the other side of the median.

What accounts for this difference? The vast majority of egos are going to vote for the correct candidate without social information: only 15% of subjects with any information are going to make a mistake based on their estimates of the candidates’ positions. If a subject chooses an alter on the same side of the median, the alter’s signals will typically suggest to the ego that they are making the right choice. Hence, they will rarely change their vote at all meaning there is little chance that they will make a harmful switch, but they are unlikely to make a helpful switch either. Egos who choose alters on the other side of the median will receive signals suggesting that they are making a mistake. Most egos, aware the signals may be biased, will ignore this information (Ahn et al., 2010; Lupia and McCubbins, 1998), but some will update their beliefs about the candidate positions and switch their vote to a candidate that benefits the alter, but not the ego.

6. The external validity of the experiment

There are two different ways to consider the external validity of this experiment: (1) the realism of the design; and (2) the comparability of the results to findings outside of the experimental laboratory. Starting with the design, the most glaring omission is party. The public information regarding the possible candidate positions provides similar information to party regarding which candidate the subjects should typically support and the probability that two subjects would support the same candidate. It does not take into account that individuals might dislike individuals once they discover they support a different party and thus will cease speaking with those on the other side of the issues (Brady and Sniderman, 1985; Buttice et al., 2009; Iyengar et al., 2012). We might, therefore, be overestimating how much individuals weigh expertise in discussion partner choice. On the other hand, the ‘affective’ polarization documented by Iyengar et al. (2012) is driven by exposure to political information and the experimental subjects most likely to be harmed by discussion are those exposed to the least information. Hence, our argument may be restricted to the group most in need of the discussion shortcut: those who do not regularly follow politics (Mondak, 1995).

How do these findings compare to results involving voters in actual elections? Like our experimental subjects, survey respondents discuss politics more frequently with those they believe are better informed about politics (Huckfeldt, 2001), but the best predictors of vote choice are the biases of the discussion partner (Beck, 2002; Ryan, 2010). Further, those survey respondents who report discussing politics with people with differing viewpoints vote correctly less frequently (Sokhey and McClurg, 2012). In the experiment, this result is driven primarily by small lies that the subjects tell regarding candidate positions. It is difficult to know how much lying takes place in actual political conversations because the researcher needs to know exactly what information the subject held in order to judge the truthfulness of a statement. Outside of the lab, individuals could send accurate messages about their beliefs regarding candidates, but those messages may be misleading due to partisan or ideological biases (Bartels, 2002). Hence, the mechanism we identify might not explain in all cases why individuals who speak with those on the other side vote frequently less often, but the larger conclusion would hold.

7. Conclusions

The experimental results presented here cast doubt on the ability of social communication to lead to more instances of correct voting in the aggregate. For every subject that voted correctly because of messages sent by their informant, there was a subject who switched to an incorrect vote following social information. Egos chose their alters based on two factors: the alters’s expertise and the distance between the alters’s position and their own position. Only the subject/informant distance, however, affected whether or not the subject voted correctly because informants sent biased messages to those on the other side of the policy scale. Hence, subjects who valued expertise over preference proximity in choosing informants increased the odds that they would make a poor voting decision.

This experiment is not meant to mimic the entire communication process that takes place when people discuss politics. Rather, it isolates two of the key variables in political communication and demonstrates the role those variables play in correct voting. Still, it can provide a mechanism to explain the findings of research that uses surveys to study similar questions. For example, Huckfeldt (2001) shows that people report speaking more frequently with better informed political informants and acknowledge expertise in people who have divergent political preferences. It is possible that it is the informants who want to talk about politics, but these results suggest that individuals would look to the experts for guidance if the experts did not offer unsolicited advice.

This study echoes previous research that questions the efficacy of discussion as an information shortcut for anyone except the truly uninformed (e.g., Mondak, 1995; Sokhey and McClurg, 2012). It advances previous studies by demonstrating that even when individuals are allowed to choose their discussion partners and the incentives to mislead are clear, social communication does not lead to an aggregate increase in correct voting. We argue that egos should choose alters on the same side of the median because they are less likely to be led astray by social signals. When subjects do this, however, they typically receive a signal telling them to vote for the candidate they were planning on voting for before social information. For this reason, there is little benefit from social communication in the context of our experiment and other experiments on social communication (e.g., Jackman and Sniderman, 2006; Ryan, 2011a).

That conclusion might change if our experiment included a costly vote choice. In that case, subjects may become more confident that they have chosen the better candidate because of the signals sent by likeminded alters (Levitan and Visser, 2009). Individuals who are more confident that their chosen candidate is the correct choice are more likely to pay the costs of voting (Krupnikov, 2011). Hence, the next step for experiments on correct voting is incorporating insights from the literature on social networks and participation (e.g., McClurg, 2006; Nickerson, 2008; Siegel, 2009) to consider all of the elements that go into casting a correct vote.