Abstract
Researchers have long investigated the conditions that promote the unambiguous communication of issues in political campaigns; however, previous research has been largely theoretic or has only tested ambiguous communication on a limited range of issues. We address these gaps, making three principal contributions. First, we provide an extensive empirical test of ambiguous communication, scoring every issue addressed on website “issues pages” from House candidates in 2014. Second, we accomplish this by using a supervised learning content analysis procedure that allows us to score a large volume of text based on a smaller subset of hand-coded text. This allows us to not only examine whether candidates comment on an issue but also the clarity of their message. Third, this article provides empirical support for how candidates’ communication strategies are shaped not only by the candidates’ personal characteristics but also the characteristics of the district. We find that district heterogeneity is an important predictor of ambiguous communication. Evidence also indicates that issue ownership and ideological extremity play a decisive role in the decision to “go vague.”
Clear communication of issue positions in campaigns is an essential component of healthy representative government. Promissory responsiveness, whereby citizens, as principles, exercise some control over their agents “through the explicit and implicit promises that the elected representative makes to the electorate,” is virtually impossible without unambiguous communication (Mansbridge, 2003). Campaigns are one way these promises are communicated: through debates, speeches, television advertising, websites, and so forth. This representative standard must be juxtaposed with an empirical reality, that is, campaign rhetoric is regularly steeped in ambiguity. As Page (1978) summarizes, “The most striking feature of candidates’ rhetoric about policy is its extreme vagueness” (p. 152).
With this normative issue in mind, scholars have long investigated the conditions that promote clear messages. However, several limitations constrain previous research. Much of the existing research is theoretic, deriving expectations about ambiguous communication without an empirical test (i.e., Alesina & Cukierman, 1990; Aragones & Postlewaite, 2002; Shepsle, 1972). Moreover, given the difficulty of measuring ambiguity, most empirical research has been limited to small samples of campaign communication on a limited range of issues. This article addresses these limitations, making three contributions. First, it provides an extensive empirical test of the theoretic propositions set out in the extant ambiguity research, scoring every issue addressed on website issues pages from every Republican and Democratic House candidate in 2014. Second, it accomplishes this by using a supervised learning content analysis procedure that scores a large volume of text based on a smaller subset of rigorously hand-coded text. Third, this approach offers an account of how candidate- and district-level factors interact to incentivize ambiguous communication. Findings generally support previous work indicating that clear issue statements are a function of features such as incumbency, the ideological distance between a candidate and the district’s median voter, issue salience, and issue ownership. This article extends these findings to argue that the district’s heterogeneity is equally important. Districts with a great deal of heterogeneity experience more candidate obfuscation, relative to homogenous districts. Thus, the decision to “go vague” is not simply a reflection of a candidate’s personal proclivities like incumbency and issue ownership. Rather, going vague should be seen as a strategic response to the underlying characteristics of the electorate to which the candidate is appealing.
Ambiguity in Campaign Communication
Rather than study the content of messages directly, research on political ambiguity has generally examined the incentive structure involved in advancing clear/unclear electoral signals. Early research followed Downs’s (1957) rational actor approach, concluding that parties have incentives to “becloud their policies in a fog of ambiguity,” a finding that has been extended by a great deal of formal theoretic research (i.e., Alesina & Cukierman, 1990; Aragones & Postlewaite, 2002; Shepsle, 1972).
The literature is mixed on the extent to which voters demand clarity in campaign communication. Tomz and Van Houweling (2009) find that, under the right circumstances, voters actually “embrace” ambiguity because they make optimistic inferences about candidates’ issue positions when campaigns send murky signals. Somer-Topcu’s (2015) work on European democracies concurs, finding that when parties have broad appeal (measured as perceptual disagreement on the part of voters), voters tend to think they are closely aligned with those parties. However, several criticisms warrant review. First, Krupnikov and Ryan (2017) note that preferences for ambiguous candidates are not equivalent to voting for those candidates. Voting is a costly act, and candidates may not “take action” on the part of ambiguous candidates, even if they think they are aligned with them. Ambiguity may affect the vote choice calculus in other ways. Alvarez (1997) shows that voters are less likely to vote for candidates about whom they are “uncertain,” and “certain” voters are more likely to engage in issue voting, a risky proposition for candidates who are unaligned with their district. As Alvarez (1997) summarizes, “The ambiguous candidate risks alienating risk-averse voters. However, the unambiguous candidate risks alienating voters with differing positions on the issues” (p. 156). Milita, Simas, Ryan, and Krupnikov (2017) add nuance to these findings, arguing that in the absence of issue statements, voters will rely on party heuristics to make inferences. The authors thus conclude that “There is little difference…between sending an ambiguous signal by speaking unclearly on an issue and the subject receiving no signal whatsoever about the candidate’s position” (p. 55). All this suggests that the strategy of ambiguity may be less valuable than Tomz and Van Houweling (2009) contend, with a trade-off between voters engaging in issue voting in high-information campaigns, and making inferences (or perhaps not voting at all) under conditions of uncertainty. Insofar as the strategy of ambiguity comes with a cost, then, the literature to date is somewhat question-begging: Why do so many candidates default to ambiguous stances?
Little research has examined the ambiguity of campaign communication itself, but there are exceptions. Page (1978) finds that in presidential campaigns, policy stances are a rarity, relative to statements meant to reveal candidates’ character. Lim (2002) argues that presidential rhetoric has been characterized by an increase in abstraction. While the “vacuousness” described by Lim is conceptually similar to ambiguity, Lim’s primary purpose is to document this trend, not build a model explaining why some politicians employ this rhetorical strategy.
Druckman, Kifer, and Parkin (2009) code a large sample of campaign websites for the presence or absence of issue discussion (see also Druckman, Hennessy, Kifer, & Parkin, 2010). Silence on a topic is motivated by the candidate’s assessment of the campaign environment and their judgment regarding whether communication will help or hinder their electoral chances. For example, incumbents are much less likely to hazard an issue statement because doing so would pose a risk in an electoral environment where they have already built a winning coalition. Druckman et al. (2009) advance the literature by directing attention to how candidates balance risk and reward in communication decisions and how the relative rewards of a strategy are constrained by factors like incumbency and party. However, while Druckman and colleagues examine an extensive set of campaign websites, the purpose of the analysis is to predict the decision to hazard an issue statement, not evaluate the clarity of that statement.
Milita, Ryan, and Simas (2014) offer the most rigorous test to date of the predictions set out in the formal theoretic literature. The authors examine candidate webpages from 2006 to 2008 for evidence of clear policy stances on the Iraq War and same-sex marriage, distinguishing between “silence” on an issue (not commenting at all) and ambiguity on the issue (whether the statement was clear). The authors find that salience compels candidates to address the issue and address it clearly. As with other work indicating that incumbents are less likely to advance clear policy stances (Druckman et al., 2009), Milita et al. found that incumbency incentivizes clear position taking. Consistent with the theoretic literature, Milita et al. also find that candidates are more likely to make clear statements when they are ideologically congruent with their district.
Much of the literature focuses on the candidate, and how candidate characteristics (like incumbency status) incentivize ambiguity. However, little is known about how the underlying composition of the district influences the decision to go vague. This omission is surprising. Even the notion of ideological congruency, for example, assumes that the district’s median preference is knowable, that the electorate is sending a clear signal to the candidate, and that the candidate is reacting accordingly. This assumption may be difficult to meet in every district, suggesting that the manner in which district interests are revealed to candidates ought to be more completely theorized. Rhetorical choices are ultimately best seen as being influenced by the interaction between candidate and district-level factors, not candidate characteristics alone.
One such district-level factor that will condition candidate rhetoric is heterogeneity. There are two schools of thought with respect to how candidates will respond to district heterogeneity, but both ultimately yield similar substantive predictions. One approach builds on the median voter theorem and suggests that when candidates have a difficult time pinpointing the median voter with any degree of certainty, candidates deploy ambiguous language (Glazer, 1990). While the consequences of heterogeneity around the median have not been tested in the context of electoral campaigns, we know that heterogeneity influences legislators in other ways. For example, Gerber and Lewis (2004) demonstrate that legislators are most closely aligned with the district median when the district is homogenous. A second approach recognizes that districts are made up of numerous cross-cutting constituencies (involved in candidate selection both in the primary process and the general election) and the concept of a “median voter” operating in one-dimensional policy space has limited utility. By this logic, Ensley, Tofias, and de Marchi (2009) argue that a key challenge for candidates is survival under conditions of uncertainty. While heterogeneity is a problem for all candidates, Ensley et al. (2009) argue that incumbents have an advantage in more complex environments because they have already devised a platform capable of threading the needle in a variable electoral landscape. While the notion of complex constituencies presents a different underlying logic than does the median voter theorem, candidates will likely address complex constituencies similarly with respect to message ambiguity: all else equal, candidates should obscure policy positions in complex districts, lest they wish to alienate part of their district. Following Ensley et al.’s (2009) logic, this incentive may be particularly strong for challengers who have yet to build a winning electoral coalition.
Why Do Candidates Go Vague?
As this discussion makes clear, candidates’ messaging decisions can be understood in a risk/reward framework (Druckman et al., 2009). Going vague is a strategy to manage the risk environment, and risk can be understood as the product of candidate- and district-level factors. Candidate factors might include previous public statements or (for incumbents) well-known votes that constrain their ability to use ambiguous language. Of course, a clear public position might be grounds for credit claiming if a candidate’s constituency/district is aligned with that position. For this reason, Milita et al. (2014) and others have advanced the literature by arguing that the ideological distance between the candidate’s position and the district’s median is critical to understanding the strategic deployment of ambiguity.
Candidates have two decisions to make: whether to speak on a particular issue and how much clarity/ambiguity to articulate. There are several reasons to treat “avoidance” and “ambiguity” as distinct strategies, responsive to different underlying electoral pressures. Candidates seek to prime their strengths and avoid their weaknesses (Druckman, Jacobs, & Ostermeier, 2004), making avoidance the default strategy for candidates confronting problematic issues. There is little incentive to address an issue ambiguously when avoidance will do. However, there are cases when candidates are compelled to address an issue because of its salience in the public at-large (Druckman, Kifer, & Parkin, 2017; Milita et al., 2014) or because a previous voting record requires defense, explanation, or credit claiming (Shepsle, 1972). Thus, there are circumstances when a candidate must address an issue, but constituent and candidate constraints make clarity a liability. With this basic framework in place, we identify four specific constraints that can alter the candidate’s perception of risk/reward and ultimately influence the decision to go vague: incumbency status, candidate ideology (and the “congruency” of this ideology with the district), district heterogeneity, and finally the specific issue at hand.
Incumbency
We expect incumbency to be related to fewer issue statements but also higher degree of clarity on the set of issues candidates comment on. Milita et al. (2014) found challengers were the most likely to advance issue positions on websites and also tended to stake out these positions with greater clarity. Similarly, Druckman et al. (2009) have found that incumbents tend to comment on fewer issues than their challengers. In contrast, Shepsle (1972) argues that incumbents have less ability to use ambiguous rhetoric, given that they are constrained by past voting behavior, greater media scrutiny, and previous position statements. Similarly, Tetlock (1981) argues that presidents become increasingly complex in their public statements following an election because “Occupants of key governmental roles are confronted daily with painful choices in which conflicting objectives must be acknowledged and integrated” (Tetlock, 1981, p. 208). While there is tension in the literature, it is possible that these arguments are not mutually exclusive. As sketched above, avoidance and ambiguity are different strategies, and while we expect incumbents to avoid issue discussion, they may speak with clarity when their hand is forced, for the reasons articulated by Shepsle (1972). Moreover, Milita et al. (2017) acknowledge that ambiguous communication may be grounded in genuinely mixed feelings about an issue. As Tetlock (1981) suggests, these mixed feelings may come into sharper focus as incumbents have to make actionable decisions on the issues. Thus, we expect that incumbents will opt for a strategy of silence on the issues (Hypothesis 1a). However, when incumbents do hazard issue statements, these statements will be marked by a higher degree of clarity (Hypothesis 1b). 1
Candidate Ideology and Ideological Congruency
Most research predicts that candidates will be more ambiguous as their own preferences deviate from the district median. Chappell (1994) argues that candidates will be most likely to advertise when they are ideologically aligned with the district. Milita et al.’s (2014) study of campaign websites also finds greater ambiguity (and avoidance) as ideological distance increases. This might seem to imply that ideological distance will increase as candidates stake out extreme positions, but this is not necessarily the case. In some districts, centrist candidates may find themselves mismatched with an extreme electorate. This is important, because moderate and extreme candidates face different constraints on their ability to deploy extreme language. Candidates who take extreme positions are typically associated with a clearer issue agenda, as they offer a distinct alternative to the opposing party (Abramowitz, 2010; see also Levendusky, 2009). When a candidate caters to the extreme wing of their party, they leave themselves little room for obfuscation. For this reason, an ideological mismatch should be associated with greater ambiguity and ideological extremity with greater clarity. Candidates have the strongest incentive to go vague when they are ideological moderates mismatched with an extreme electorate. Candidates running extreme campaigns in moderate districts also have incentive to go vague, though this incentive is tempered by the fact that far-right conservatives and far-left liberals did not earn their branding by operating in an ideological gray area.
This discussion informs two specific expectations. First, all else equal, ideological distance will be associated with greater ambiguity and a higher probability of silence (Hypothesis 2a). Second, controlling for ideological distance, ideological extremity should correspond to less ambiguity and a lower probability of silence (Hypothesis 2b).
District Heterogeneity
Evidence indicates that candidates are also sensitive to the heterogeneity of their district (i.e., Bailey & Brady, 1998; Campbell, 1983; Fenno, 1978); however, as noted, there are competing mechanisms (grounded in the median voter theorem and the notion of complex constituencies) as to why this is the case. While observing any sort of relationship between heterogeneity and the decision to go vague represents an advance in the literature, we propose two different approaches to studying heterogeneity to help adjudicate between competing mechanisms. While this is not a definitive test of the aforementioned causal mechanisms, each of these heterogeneity approaches approximates different ways candidates see their districts and is suggestive of the causal mechanisms involved in the decision to go vague.
First, in a heterogeneous district, the “signal” provided by the median voter is less precise, and taking a stand on an issue involves more risk. Assuming district preferences are arrayed on a single dimension, then greater uncertainty about the median preference of voters will be associated with incentive to adopt ambiguous positions (Glazer, 1990). To operationalize the concept of variability around the mean, we use the standard deviation of district ideology estimates (described more completely in the Method section). This approach, which uses variance to operationalize heterogeneity, is comparable to other approaches in the literature (i.e., Campbell, 1983; Gerber & Lewis, 2004; Tausanovitch & Warshaw, 2013). The prediction is that more variance in attitudes will correspond to a higher degree of ambiguity in rhetoric (Hypothesis 3a).
Second, heterogeneity may present itself to candidates as multiple cross-cutting constituencies (rather than a one-dimensional policy space), such that the candidate will have a difficult time formulating a platform that can appeal to enough voters to win a nomination and a general election. To approximate this dynamic, we borrow an approach used by Sullivan (1973), who uses census data to quantify the diversity of subgroups across multiple demographic categories (described more completely in Method section). Adopting Sullivan’s (1973) approach, Bailey and Brady (1998) find that more homogenous states tend to produce constituency-oriented legislative behavior, which suggests that legislators are responsive to the contours of demographic subgroups in their district. The prediction is that a more heterogeneous mix of subgroups will also correspond to a higher degree of ambiguity in rhetoric (Hypothesis 3b).
Issue-Specific Predictions
We expect that candidates will differ in the issues they choose to discuss, and the clarity with which they choose to discuss them. While we expect the incumbency pattern to replicate the pattern hypothesized above, partisan differences in clarity and silence should emerge with respect to specific issues. Two theoretic concepts guide expectations: salience of the issue and partisan issue ownership. Salience can goad candidates into making some sort of issue statement: When an issue is of high national importance, candidates will be compelled to say something. Candidates can be punished by voters for a failure to engage highly salient issues (Druckman et al., 2010; Milita et al., 2014). Issue ownership dynamics should also play a role. The theory of issue ownership contends that political parties are judged to have competence within certain domains, and that this competency is relatively stable over time for most issues (Egan, 2013; Petrocik, 1996). All else equal, candidates are more likely to comment on an issue when it is owned by their party because it is an issue upon which they are likely to be favorably judged (Druckman et al., 2010).
Druckman et al. (2010) posit that salience and ownership may operate in tandem to incentivize position-taking, noting that “in some cases, the saliency of an issue may force candidates to engage on it even if one party is seen to have a clear disadvantage” (p. 5). This suggests that under the circumstances described above (a candidate adopting an issue position in uncomfortable territory), the incentive to use ambiguity also increases. In sum, we expect less silence on an issue when the issue registers high levels of national importance (Hypothesis 4). However, while salience will promote discussion of an issue, it will not translate into clarity (Hypothesis 5). Traditional conceptions of issue ownership (i.e., Democrats and social spending, Republicans and guns) will translate into more discussion of an issue (Hypothesis 6). Likewise, lacking issue ownership will be associated with a higher degree of ambiguity when candidates are forced to venture onto “unowned” turf (Hypothesis 7). That is to say, when Democrats breach the topic of guns (or Republicans decide to talk about social security), they will do so with ambiguity.
Method and Data
Dependent Variables: Operationalizing Ambiguity and Silence
We assume that ambiguous statements have words and word pairs in common across a variety of policy and partisan contexts. The notion that words and word-pairs can serve as reliable indicators of ambiguity is supported by academic literature (Berry, Kamsties, & Krieger, 2003). For example, words like “because” and “else” tend to signal cause/effect reasoning and policy rationalization and will covary with message clarity. Other words might carry multiple meanings or indicate a form of pragmatic ambiguity within a sentence. For example, quantifier operators like “some” serve as a marker of ambiguous speech.
Our approach to scoring websites takes advantage of the assumption that certain words and word-pairs will occur with greater frequency in ambiguous passages. This approach uses hand-coded statements to identify patterns of word use in ambiguous/clear statements and then used the program Wordscores (Laver, Benoit, & Garry, 2003) to score text from 2014 congressional campaign websites, using hand-coded statements as a training set. While Wordscores was developed to assign ideology scores to political texts that could be meaningful placed on a left–right spectrum, it is also a sensible way to score texts with varying degrees of clarity/ambiguity. In this case, the process allows us to score the 2014 data relative to how well it mapped on to the set of hand-coded training set.
To perform the hand-coding, it is important that training texts that are as close as possible to the 2014 virgin texts, so that sample-specific variations do not introduce error. Accordingly, we drew a random sample of 5% of the major party congressional candidates from 2012 House races. Next, using an archive of campaign websites (Druckman, Kifer, & Parkin 2014), we collected text from every home page, biography page, and issues page in the sample. This yielded 357 unique webpages to be hand-coded. Coders then scored the website statements for their level of clarity and ambiguity. Each website was initially scored by two coders. When there was substantial intercoder disagreement, a panel of three additional coders rescored the statements, blind to the scores initially given. 2 In short, every hand-coded passage was scored by at least two (and in some cases five) trained coders. After hand-coding was complete, these scales were then normalized, summed, and averaged across all coders to create an overall ambiguity index.
In addition to hand-coding a sample of 2012 congressional websites to serve as the training set, we collected text from every major party candidate website from the 2014 cycle. Because some races did not field candidates from each party or the candidate did not have a website, we located a total of 761 candidates. Website data have a number of advantages. Because virtually every candidate produces a website, we are able to examine candidates who are unable to afford television advertising or who receive little media attention. Websites are also a good indicator of the “thrust” of a candidate’s overall campaign strategy (Druckman et al., 2009). In one recent study, congressional campaign insiders reported that campaign websites are the best medium to communicate a campaign’s overall message, leading the authors to conclude that “websites, relative to other media, can serve as digital hubs that encapsulate their entire campaign message” (Druckman et al., 2017). Thus, while the present analysis is limited to digital communication, there is reason to suspect that the dynamics we identify will hold for a campaign’s overall communication strategy.
From these websites, research assistants selected all nonupdating text from the main page, the biography page, and any statement the candidate made under the heading “on the issues” or “policy priorities” or the like. 3 In all, this yielded a sample of 6,283 statements for further analysis. Statements were sorted into categories based on issue type (i.e., “energy policy” or “jobs and economy”). 4 This approach provides for the flexibility to both examine when candidates chose to be silent on a particular issue or to aggregate across all issue statements to generate an overall ambiguity score.
Once sampling was complete, we used Wordscores (Laver et al., 2003) to score the 2014 website data using the 2012 hand-coded data as reference texts. As with other supervised approaches, this process uses a training set to “teach” the computer to learn what counts as “ambiguous” or “clear.” Wordscores uses the 2012 training set as input and outputs an ambiguity score for each 2014 statement based on similarities with the training data. We adopted a two-word “phrase frequency” approach, which uses all two-word phrases in the training set as reference. Higher scores indicate greater clarity in issue positions, and lower scores indicate more ambiguity. Compared to other content analysis procedures, supervised learning is a black box, such that the analyst never sees the “dictionary” used to score the virgin texts. To address this, we used the Wordscores data frame to generate a list of words that most frequently distinguishes ambiguous/clear texts. These results, which go above and beyond the usual procedures used in supervised content analysis, are reported in Online Appendix A.
We also scored websites for issue silence. All issue statements were categorized according to topic by a trained coder (see Online Appendix B). We counted all separate issue statements that were demarcated by a separate title or subtitle. This strategy allowed us to create a dichotomous variable for every candidate, indicating whether they discussed the issue. We also summed issue statements into an overall issue count variable.
Independent Variables
We model ambiguity and silence as a function of district- and candidate-level characteristics. District characteristics include both measures of heterogeneity and the district’s ideological lean. Candidate characteristics include the candidate’s party and ideological position. Finally, we consider the candidate’s ideological congruence with their district.
District heterogeneity
While previous research points toward heterogeneity as a predictor, it leaves little guidance as to what type of heterogeneity will influence candidates’ rhetorical choices. For this reason, we test two measures of heterogeneity, each which could create an electoral risk if an unambiguous message were to appeal to one part of the district while simultaneously alienating another. First, to approximate the heterogeneity of diverse constituent subgroups, we adapt an approach used by Sullivan (1973). This approach measures the “proportion of characteristics upon which a randomly selected pair of individuals will differ” (Sullivan, 1973, p. 70). We use data from the 2013 American Community Survey, which reports categorical memberships for race (seven categories), education (seven categories), occupation (five categories), and poverty status (two categories—above or below the poverty line) for each U.S. congressional district.
5
Following Sullivan (1973), heterogeneity is computed as:
District lean
To control for the ideological lean of the district, we use Tausanovitch and Warshaw’s (2013) district ideology estimates. 7
Candidate ideology
To estimate a candidate’s left–right orientation, we used data from Parsneau and Chapp (2017), who estimate candidate ideology from campaign websites using Wordscores. This approach used sample of “.gov” websites as a training set, and then scored campaign websites accordingly. Higher scores on this variable indicate a more partisan-Republican campaign style, while lower (negative) scores indicate a more partisan-Democratic campaign style. Importantly, the training set used to estimate candidate ideology is entirely different from the set that was used to estimate ambiguity scores. To address concerns about overreliance on text-as-data for ideology scoring, the Online Appendix replicates this approach with Bonica’s (2016) campaign finance-based ideology estimates, reaching similar substantive results.
District mismatch
To calculate ideological congruency scores, we regressed the candidate’s ideology scores on the district ideology score and calculated predicted ideology for every district. Mismatch is the absolute value of the prediction error for every district, such that higher scores indicate candidates who are further away from the district median.
Party × Ideology interaction
Our expectation is that candidate rhetoric will become clearer as the candidate’s issue positions move to the ideological extremes. Accordingly, we interact the dichotomous party variable with candidate ideology, since we expect Democrats to be more clear with lower ideology scores (more liberal) but Republicans to be more clear with higher ideology scores (more conservative).
Results
Validation of Supervised Learning as a Strategy for Ambiguity Detection
We demonstrate the validity of using Wordscores to score texts for ambiguity in two ways. First, we examine how well Wordscores is able to reproduce hand-coded ambiguity scores. Second, we examine how well the ambiguity scores mesh with expectations about ambiguity on campaign websites.
One mark of a valid scoring procedure is whether the algorithm reproduces hand-coded results. We use 10-fold cross validation on the training and testing sets. That is, we use 90% of the hand-coded data to act as the training set, which in turn writes an algorithm to test the remaining 10% of hand-coded websites. After obtaining ambiguity scores from Wordscores, websites were classified as clear or ambiguous by splitting the set at .5. The 10% sample was then checked them against the remaining hand-coded data. This procedure was 10 times, each time checking against a different 10% testing set. Wordscores predicted 92.5% of passages correctly. 8
In addition, the Wordscores approach to ambiguity closely meshes with our a priori expectations about ambiguity on campaign websites. First, we expect different ambiguity levels depending on where on the website the text was located. Text from the home page, biography page, and issues pages dedicated to abstract principles (like “philosophy of government”) and ideology (like “true conservative” or “progressive values”) should be the most ambiguous, since they are not expressly dedicated to policy. In contrast, we expect “easy issues” to have the highest clarity scores in the sample because simple “policy ends” discourse discourages obfuscation (Carmines & Stimson, 1980). Likewise, following Geer’s (2006) observation that negativity is associated with factual information, we expect opponent-directed webpages (like “my opponent’s record”) to contain higher than average clear policy content. These expectations were largely confirmed. Figure 1 displays means for the most common 10 issues in the sample, along with biography pages, main pages, and other pages mentioned above, which allow us to test specific expectations. 9 Consistent with expectations, significantly lower clarity levels were observed on home pages (x¯ = .318; p < .05, two-tailed), biography pages (x¯ = .344; p < .05, two-tailed), and pages dedicated to ideology and other abstract principles (x¯ = .377; p < .05, two-tailed). Consistent with our easy issues expectation, the highest clarity scores in the sample were found on pages dedicated to equality issues like women’s rights, LGBT discrimination, and racial discrimination (x¯ = .569; p < .05, two-tailed), followed by “gun rights/restrictions” (x¯ = .514; p < .05, two-tailed). “Moral values issues” such as “traditional marriage” and “abortion” were also significantly higher than the mean (x¯ = .488; p < .05, two-tailed). Opponent-dedicated pages were clearer than the sample mean (x¯ = .467), although this difference was not significant.
Average ambiguity scores by website location.
Ambiguity and Silence in Congressional Campaigns
Because we are interested in both candidate- and district-level predictors and most districts have two candidates, we estimate a multilevel model, treating candidates as “nested” within congressional districts. 10 We fit a multilevel linear model to predict ambiguity, and a hurdle model with logit and Poisson components to predict the number of issues pages addressed by the candidate (a count variable). 11 Results, presented in Table 1, are consistent with predictions. First, the incumbency predictions are supported by the data. Consistent with previous research (Druckman et al., 2009 and Milita et al., 2014) and Hypothesis 1a, the negative coefficient in both page count model indicates that incumbents take on significantly fewer issues than their challengers. 12 However, the positively signed coefficient for incumbency in the ambiguity model indicates that incumbents are significantly clearer on the issues they choose to discuss. This fits with Hypothesis 1b, suggesting that while incumbents have greater incentive to avoid a discussion of issues, they are constrained by past legislative behavior and thus tend to offer clearer policy positions on the limited range of issues they chose to tackle. 13
Ambiguity on Congressional Webpages.
Note. Estimates are from a multilevel model, with candidates nested in congressional districts. Dependent variable in the silence models is the total page count. Dependent variable in ambiguity models is the ambiguity score. AIC = Akaike information criterion; OLS = ordinary least squares.
Our expectations for candidate congruence and candidate ideology are also supported by the data. Consistent with Hypothesis 2a, candidates who were a mismatch with their district were less likely to discuss an issue at all and more likely to use ambiguous language. Holding all other variables at their mean or mode, the coefficient of −.086 corresponds to an 80.6% lower clarity score when we move from a “perfect fit” candidate to the most mismatched candidate in our sample. 14 This fits with the expectation that avoidance and ambiguity are dual strategies employed by candidates to avoid taking unpopular positions with voters.
The significant, negative interaction between party and candidate ideology is illustrated in Figure 2. Candidates from both parties exhibit greater clarity when their ideologies tend toward the extremes of their party. This is reflected in a negative slope for Democrats, with greater clarity for liberal ideologies (negative scores) rather than centrist scores (near 0), but a positive slope for Republicans, with greater clarity for conservative ideologies (higher scores) rather than centrist scores. Centrist Democrats tend to take more clear positions than centrist Republicans, and the clarity of Democrats tends to grow faster as their ideologies move away from the middle than Republicans. The interaction term in the issue count model parallels the ambiguity model, indicating that for both parties ideological extremity is associated with addressing more issues (consistent with Hypothesis 2b).
Ambiguity by ideology and party.
While the interaction between party and ideological extremity was predicted, the partisan difference was not. We offer two explanations for this disparity. First, Republicans commented on more Democratic-owned issues. The decision to comment on topics like “jobs” and “health care” likely compelled more ambiguous rhetoric. In addition, Republicans were more likely to offer broad ideological or philosophical platitudes on their websites (under headings like philosophy of government), which tended to be associated with low levels of issue clarity.
The models offer evidence that ambiguity is a response district heterogeneity. Consistent with Hypothesis 3b, the heterogeneity of demographic subgroups is associated with both avoidance and ambiguity, suggesting these are strategies candidates employ in order to appeal to districts with cross-cutting constituencies. All else equal, moving from the most homogenous to most heterogeneous district is expected to lower the level of clarity by 51.0%. Hypothesis 3a, which predicted attitudinal heterogeneity would correspond to a lack of clarity, did not fare as well in the model. The logit component of the hurdle model is insignificant and substantively small. While the coefficients in the ambiguity model and the Poisson component of the hurdle model are significant and run in the predicted direction, the coefficients are smaller, suggesting that candidates are primarily responding to varied demographic subgroups. In general, the results are consistent with Hypothesis 3b but offer only limited support for Hypothesis 3a. This suggests that candidates are more responsive to the complexity of constituent subgroups than the signal they are receiving from the district’s median voter.
In summary, candidates who are nonincumbents, are Republicans, are more moderate, reside in districts with a greater heterogeneity, or have ideologies that deviate from the ideology of their districts all tend to engage in more ambiguous campaign communication. These variables also correspond to the decision to comment on an issue in the first place, with the notable exception that the incumbency variable is signed in the opposite direction. Incumbency, the results suggest, is related to less overall issue discussion, but more clarity.
Issue-Specific Ambiguity
Are candidates more ambiguous when their “hand is forced” to speak on an issue? To test this, we turn to an examination of avoidance and ambiguity on specific issues. We limit our analysis to the 10 most common issues: jobs/economy (n =551), health care (n = 432), budget/taxes (n = 349), education (n = 347), veterans (n = 291), foreign policy (n = 283), immigration (n = 275), seniors’ issues (n = 258), energy (n = 251), and guns (n = 191). We specify two models for each issue: a multilevel logistic regression predicting the probability a candidate will dedicate a specific page to an issue, and a multilevel linear model predicting the level of ambiguity/clarity among those candidates who chose to take on the issue. Each model contains the same variables from Table 1, excluding attitudinal heterogeneity and district ideology. 15 Because examining 20 models is cumbersome, regression outputs are available in Online Appendix D. The discussion below summarizes the main findings for each independent variable across all 10 of the issues.
Incumbency
Of the 10 issues considered, incumbents were less likely to address the issue in seven cases, and this difference was statistically significant in four cases (jobs and the economy, education, guns, and immigration). When we change our dependent variable to ambiguity scores for particular issues (limiting our sample to only those candidates who took on a particular issue), the pattern reverses. Incumbents had higher clarity levels for each of the 10 issues, and this difference was statistically significant in seven cases. Considering the linear and logit models together, the pattern supports Hypothesis 1a and 1b, suggesting that while incumbents would prefer to avoid specific issue stances relative to challengers, when they do address an issue they face increased pressure to present their positions with clarity.
Demographic heterogeneity
Next, we examine the predictive power of demographic heterogeneity on silence and ambiguity. The heterogeneity coefficient is negatively signed for nine issues, and this difference is significant in six cases. Consistent with the results from Table 1, candidates are less likely to bring up any issue when the district is heterogeneous. Turning to ambiguity, the results are less impressive, albeit consistent with the results from the main model. The coefficient is negatively signed in six models, meaning that as the district grows more heterogeneous, candidates are more likely to obfuscate their specific issue stances. However, this result only reaches statistical significance in one case (seniors). In short, when confronting a heterogeneous district, candidates avoid the issues first and obfuscate second.
Candidate/district mismatch
The results for candidate mismatch should be interpreted with caution, given that the estimates for “candidate” and “district” used to build the mismatch estimate did not examine ideology on specific issues, but rather used an overall ideology measure. This said, given increased ideological sorting in the electorate, it is expected that a candidate’s general ideological agreement with a district will predict issue-specific avoidance and ambiguity. Results support expectations. For 9 of 10 issues, a larger mismatch is associated with a lower probability to address the issue (significant in eight cases). The effect of mismatch on issue-specific ambiguity is smaller. While a larger mismatch corresponds to less clarity in six of the models, mismatch is only significant in two cases. In other words, when candidates are ideologically distanced from their constituents, they are more likely to avoid the issue completely than use ambiguity, though ambiguity remains a viable strategy.
Puzzlingly, candidates are more likely to speak clearly on education when they are mismatched with their district. While this defies predictions, it is possible that education policy is amenable to a unique level of clarity even when candidates are not district congruent. Egan (2013) refers to education as a “consensus issue,” about which conservatives and liberals have similar policy goals. Conservatives in liberal districts can still prioritize education by trumpeting market-based reforms, while liberals in conservative districts can credibly claim to prioritize children. The domain of education policy provides opportunities for ideologically mismatched candidates to make clear policy appeals rather than shroud policy proposals in vague language.
Party and ideology
Finally, we turn to the role of party, candidate ideology, and the interaction between the two. Because these interaction terms can be cumbersome to interpret, Figure 3 displays predicted probabilities of dedicating web space to an issue page, by party and ideology. Figure 4 displays predicted ambiguity scores for specific issues, examining only those candidates who advanced an issue position. Examining the impact of party and ideology in each of the models, the decision to comment and speak clearly on an issue is highly contextualized, depending on the party’s positioning on the issue, the candidate’s ideology, and the salience of the issue. We organize our analysis into four sections: high-salience issues, Democratic-owned issues, Republican-owned issues, and two issues that defy a straightforward explanation.
Probability of a dedicated issue page, by party and ideology.
Predicted ambiguity scores on issue pages, by party and ideology.
High salience issues: The economy and health care
According to the 2014 National Exit Poll, the economy and health care were the most important issues facing the nation, registering perceived importance at 45% and 25%, respectively. Consistent with Hypothesis 4, congressional candidates reacted accordingly: 70% of candidates had an issue page dedicated to jobs and/or the economy, while 54% dedicated an issue page to health care. Despite Democratic issue ownership in these policy domains (Egan, 2013, p. 71), Republicans—and in particular ideologically conservative Republicans—were more likely to discuss these issues. This points toward the importance of salience because on high-salience issues Republicans were more likely to discuss an issue, even when doing so involved treading on Democratic turf. (It also indicates that many Democrats continued to face scrutiny surrounding the Affordable Care Act.) As Figure 3 (panel B) illustrates, more moderate Democrats were particularly likely to avoid discussion of health-care reform. However, while Republicans were compelled to venture onto Democratic turf, those who did so avoided clear policy stances. As Figure 4 (panels A and B) shows, on these issues, clarity became increasingly likely as candidates moved left of center (the interaction between ideology and party is significant for economic issues, while it misses the significance threshold for health care). This lends support to Hypothesis 5, that is, salience can compel a statement, though candidates may still be hesitant to advance specifics.
Democratic-“owned” issues: Seniors, education, and veterans
Consistent with Hypothesis 6, Democrats were more likely to comment on issues with a social service provision component: education, seniors’ issues like social security and medicare, and veterans’ issues. In these cases, party had a significant independent effect on the probability of bringing up the issue. The interaction between party and ideology is also significant and negative in the case of seniors and education, indicating that as Democrats move further to the left ideologically, the probability of a position statement on these issues increases (Figure 3, panels C–E). Turning to the ambiguity models, we observe a negative coefficient for candidate ideology on “seniors,” indicating that as candidates move to the left, their clarity on seniors’ issues increases. The interaction term is also significant in the veterans’ model.
Republican-owned issues: Guns, energy, and budget/taxes
This dynamic essentially reverses for traditionally Republican-owned issues. Republicans, and especially conservative Republicans, are more likely to bring up gun issues, though not any more likely to speak clearly on this topic (perhaps owing to a ceiling effect). Similar effects are present, though more muted, on the issues of budgeting and energy policy. This suggests that while issue ownership incentivizes candidates to speak on an issue, it does not appear to encourage policy clarity.
The puzzling cases of immigration and foreign policy
Immigration and foreign policy provide two interesting cases. Compared to the economy, neither immigration nor foreign policy is especially salient. Using the 2014 national exit poll as a benchmark, 14% of voters thought that “illegal immigration” was the most important problem in the nation, while 13% reported “foreign policy” was most important. Moreover, neither issue has clear party ownership. Foreign policy “ownership” regularly changes hands, often tied to incumbent administration performance (Petrocik, 1996). Likewise, immigration defies clear partisan ownership, and in fact strategic politicians on both sides of the aisle have used the issue for strategic coalition building (Monogan & Doctor, 2017). Despite low salience and unclear issue ownership, patterns of communication on these immigration and foreign policy looked similar. Republicans—especially conservative Republicans—were significantly more likely to bring these issues up, while Democrats—especially liberals—were more likely to discuss these issues with clarity.
Discussion
The decision to comment on an issue and the decision to communicate with clarity represent distinct strategic candidate behaviors. Candidates go vague when issue clarity comes with risks. Candidates who are ideologically mismatched with their district, for example, tend to use more ambiguous rhetoric, and incumbents are less likely to hazard public policy position-taking. However, the decision to speak on an issue is not the same as the decision to be ambiguous. When incumbents comment, they tend to offer more clarity than their challengers. More ideologically extreme candidates are also clearer and address more issues. The evidence also suggests that previous research has neglected how the “target audience” influences ambiguity. Greater heterogeneity is associated with both fewer issues discussed and less clarity. When running in heterogeneous districts, candidates do not engage in “opinion leadership” by trying to persuade the public. Instead, candidates in these districts strategically avoid concrete policy promises through silence or ambiguity.
These findings dovetail with research documenting new strategies politicians are adopting to respond to electoral cleavages and polarized issue publics. For example, it is well known that in an era of media fragmentation, candidates regularly “narrowcast” messages to distinctive subgroups (Overby & Barth, 2006). Likewise, campaigns increasingly invest in data-intensive tactics to “microtarget” specific voters with highly individualized messages (Hillygus & Shields, 2009). And, when messages are broadcast, candidates frequently employ “dog-whistles,” such that divisive (often racialized) messages are only heard by specific subgroups (Albertson, 2015; López, 2014). Each of these strategies suggests that candidates respond to heterogeneity by pitching arguments that are tailor-made for their target audience. Our research suggests an alternative strategy, whereby candidates, confronted with heterogeneous constituencies, choose to “opt out” of the type of promissory responsiveness democratic theorists laud. Instead, candidates send ambiguous signals to constituencies with divergent characteristics. As such, ambiguity should be understood as one of several strategies candidates have adopted to respond to heterogeneous groups of voters. These findings are concerning against the normative backdrop of promissory responsiveness. All else equal, voters who reside in the most heterogeneous districts will likely face the largest informational disadvantages.
This research is not without limitations. Congressional campaigns are complex and fluid, and operationalizing the risk environment across a large number of campaigns requires simplifying assumptions. In particular, we have identified a small number of factors—like issue ownership and district congruency—that might make concrete issue positions more attractive. It is likely that this simplification led to some misattributions about the incentives surrounding particular policy stances. This qualification in place, it is perhaps all the more impressive that we obtained broadly consistent effects for heterogeneity, ideological extremity, and district congruence across a variety of different congressional races. A second limitation recognizes that we are studying a single election year. Would these findings be generalizable to other campaigns? While we can only speculate, we think it is likely that risk-aversion guides clarity in campaign communication to a greater extent now than in the past. Campaigns’ messaging strategies are heavily dependent on networks of professional consultants paid to take stock of the campaign risk environment and craft a message accordingly. Candidates also have considerable access to data on voters in their districts. For these reasons, we think it is likely that the dynamics observed in the 2014 midterms are more foretelling of future campaign communication than campaigns of the past.
Supplemental Material
Online_Appendix - Going Vague: Ambiguity and Avoidance in Online Political Messaging
Online_Appendix for Going Vague: Ambiguity and Avoidance in Online Political Messaging by Christopher Chapp, Paul Roback, Kendra Johnson-Tesch, Adrian Rossing and Jack Werner in Social Science Computer Review
Footnotes
Authors’ Note
Replication data are available at 
. For website text files used to generate scoring algorithm and ambiguity scores, e-mail:
We are grateful to Kevin Coe, Paul Goren, and James Druckman for helpful comments on earlier versions of this article. We thank Gary Jacobson, Adam Bonica, Chris Tausanovitch, Christopher Warshaw, Jamie Druckman, Martin Kifer, Michael Parkin, and Kevin Parsneau for providing data for this project. We acknowledge support from the Center for Interdisciplinary Research at St. Olaf College.
Declaration of Conflicting Interests
The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This project was supported by NSF grant # 1045015.
Notes
References
Supplementary Material
Please find the following supplemental material available below.
For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.
For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.
