1. Introduction
To become a member of Congress, most candidates must win two elections, with distinct incentives, actors, and electorates in each. Though positional differences between parties primary and general electorates appear minimal (Abramowitz, Reference Abramowitz2008; Hirano and Snyder, Reference Hirano and Snyder2019; Sides et al., Reference Sides, Tausanovitch, Vavreck and Warshaw2020), policy demanders active in the party network play an important role during the nomination (Cohen et al., Reference Cohen, Karol, Noel and Zaller2008; Masket, Reference Masket2009; Bawn et al., Reference Bawn, Cohen, Karol, Masket, Noel and Zaller2012) and have distinct and “extreme”Footnote 1 preferences (Saunders and Abramowitz, Reference Saunders and Abramowitz2004; Hill and Huber, Reference Hill and Huber2017; Kujala, Reference Kujala2019). Candidates must therefore appeal to non-centrist groups in the party network to become the nominee (Fiorina et al., Reference Fiorina, Abrams and Pope2005) before attempting to garner wider support among a general electorate who prefer moderate candidates (Ansolabehere et al., Reference Ansolabehere, Snyder and Stewart2001) and punish extremism (Canes-Wrone et al., Reference Canes-Wrone, Brady and Cogan2002). Accordingly, candidates are presented with a strategic positioning dilemma (Brady et al., Reference Brady, Han and Pope2007) across the electoral cycle: which constituency should they appeal to?
Some research suggests that candidates move away from the center in primaries (Burden, Reference Burden, Galderisi, Ezra and Lyons2001; Brady et al., Reference Brady, Han and Pope2007), but a systematic study of candidate positions across a primary and general election cycle remains lacking, in part due to the limited availability of positional time series data of elected officials and losing candidates. Traditional ideal point estimates are only available for elected members of Congress (McCarty et al., Reference McCarty, Poole and Rosenthal2006) or aggregated across an entire election cycle (Bonica, Reference Bonica2014). To fill this gap, we measure changes in candidate positions both during and after the primary using an original dataset of dynamic social media-based positions. We use supervised machine learning (Goet, Reference Goet2019; Green et al., Reference Green, Edgerton, Naftel, Shoub and Cranmer2020) to identify the liberal–conservative axis of 2,500,000 tweets by 988 candidates running for the US House of Representatives in 2020. We validate our measure using NOMINATE scores of candidates in the sample who had ever served in Congress, our scores correlate at 0.93.
We use this measure to test candidate responses to the strategic positioning dilemma over the electoral cycle. Importantly for our design, our method enables us to continue positioning candidates after they lose a primary. Given that voters punish inconsistency (Canes-Wrone et al., Reference Canes-Wrone, Brady and Cogan2002), we expect that primary winners will maintain positions taken during the primary to prevent accusations of “flip-flopping.” We argue instead that positional adaptation will only be observed among primary losers after their defeats, and use this movement to identify whether candidates took artificial positions during the nomination, comparing their communication during the primary campaign with their positions after they lose. In doing so, we test the adaptative rather than the selective effect of the nomination process—our interest is in the change in candidate behavior rather than election outcomes—and hypothesize that losing candidates will moderate after a primary defeat. In this paper, we focus solely on the candidate side of the dilemma, we are explicitly not capturing voter responses to or reception of candidate positioning.
Among Democratic candidates, losing a primary was clearly associated with moderation following a defeat, suggesting the adoption of artificial or strategic positions during the nomination. This finding aligns with other scholarship about candidate behavior in two-stage elections (Burden, Reference Burden, Galderisi, Ezra and Lyons2001; Brady et al., Reference Brady, Han and Pope2007) and similar research on rhetorical position-shifting by presidential primary winners (Acree et al., Reference Acree, Gross, Smith, Sim and Boydstun2020). We find no equivalent shift in the position of losing Republican candidates, indicating limited strategic position-taking and continued support for “conservative” sentiment even when electoral incentives were absent. The party-level differences are likely explained by the asymmetric nature of the Republican and Democratic parties (Hacker and Pierson, Reference Hacker and Pierson2006; Theriault, Reference Theriault2013; Grossmann and Hopkins, Reference Grossmann and Hopkins2016). Our findings are significant at both the party and candidate levels, and when we restrict our analyses to tweets that explicitly contain policy content.
We proceed as follows: First, we review the literature on strategic positioning in campaign communication. Second, we consider the ability of existing measures to fully answer our question, introducing our scaling technique based on Twitter text. Next, we present our data and findings. Finally, we discuss explanations and implications of our results at both the party and candidate levels.
2. Candidate incentives in primaries
Before candidates can compete in a general election, they must first earn the party's nomination. To win the nomination, candidates must appease various party stakeholders or “policy demanders” (Bawn et al., Reference Bawn, Cohen, Karol, Masket, Noel and Zaller2012). Both theoretical expectations (May, Reference May1973) and empirical evidence (Converse, Reference Converse1964; Abramowitz, Reference Abramowitz2010) indicate that these groups—by virtue of being highly engaged and politically active—hold positions away from the center and prioritize candidates positional congruence in their selection criteria.
Primary voters do not appear to share the distinct preferences of these policy demanders, with empirical studies of both presidential (Norrander, Reference Norrander1989; Abramowitz, Reference Abramowitz2008) and congressional primary electorates (Hirano and Snyder, Reference Hirano and Snyder2019; Sides et al., Reference Sides, Tausanovitch, Vavreck and Warshaw2020) finding little or no positional differences between primary and general election party voters. Despite these findings, primary electorates are frequently characterized as extreme by scholars (Burden, Reference Burden, Galderisi, Ezra and Lyons2001; Fiorina et al., Reference Fiorina, Abrams and Pope2005; Kamarck, Reference Kamarck2014) and politicians (Keisling, Reference Keisling2010; Schumer, Reference Schumer2014) alike. Here, the perceptions of political actors are of particular importance given our focus on candidate behavior, where candidates might adopt artificial positions because they believe that primary voters hold non-centrist preferences with which they try to align. DeCrescenzo (Reference DeCrescenzo2020) finds that elites behave as if primary voters want ideological candidates, despite limited evidence that these voters express any such preference.
Yet, winning a primary is not only dependent on positional alignment with voters. In presidential contests, Cohen et al. (Reference Cohen, Karol, Noel and Zaller2008) document the influence of party elites during the nomination. At the congressional level, Hassell (Reference Hassell2018) similarly finds that actors in the party network play a key role in candidate selection. The UCLA school of parties (especially Bawn et al., Reference Bawn, Cohen, Karol, Masket, Noel and Zaller2012) highlights the importance of “policy demanders”—including donors, activists, interest groups, and even friendly partisan media—in determining candidate selection outcomes. In part because US nominations are comparatively inclusive and decentralized (Hazan and Rahat, Reference Hazan and Rahat2010; Cowburn and Kerr, Reference Cowburn and Kerr2023), formal party organizations have been “hollowed out” (Schlozman and Rosenfeld, Reference Schlozman and Rosenfeld2019), transferring power from electability-focused formal structures toward comparatively non-centrist and policy-oriented “informal party organizations” (Masket, Reference Masket2009). Alignment with these groups can help candidates secure the nomination in several ways.
Fundraising is a key indicator of a primary campaign's viability. Donors—and large donors in particular—hold more extreme and consistent positions than primary voters (Kujala, Reference Kujala2019), with distinct preferences and policy positions from non-donors (Gilens, Reference Gilens2009). In short, “Democratic contributors are more liberal than other Democrats and Republican contributors are more conservative than other Republicans” (Hill and Huber, Reference Hill and Huber2017, 10) and donate to proximate candidates (Bonica, Reference Bonica2014). Consequently, non-centrist position-taking aligns with an increased ability to raise funds in both primary and general elections (Ensley, Reference Ensley2009).
Activists form an integral part of a wider network (Bawn et al., Reference Bawn, Cohen, Karol, Masket, Noel and Zaller2012) and are a vital resource during the nomination process (Masket, Reference Masket2009) constituting primary campaigns on the ground. Like donors, these partisans are further from the political center than primary electorates (Saunders and Abramowitz, Reference Saunders and Abramowitz2004; Hill and Huber, Reference Hill and Huber2017). Interest groups can play a similar role, with evidence that candidates with interest group support have had increased success in congressional nominations in recent years (Manento, Reference Manento2019). Both activists and interest groups hold distinct positions on the issues they care about and seek assurances that candidates are positionally aligned during the nomination. Providing assurances to multiple groups can pull candidates away from the center in a process of “conflict extension” (Layman et al., Reference Layman, Carsey, Green, Herrera and Cooperman2010), with evidence that primary candidates who receive more interest group support take positions further from the center (La Raja and Schaffner, Reference La Raja and Schaffner2015; Manento, Reference Manento2019). The proliferation of partisan media may have further elevated ideological candidates through favorable coverage to an audience of party sympathizers (Heft et al., Reference Heft, Knüpfer, Reinhardt and Mayerhöffer2021).
Taken together, these factors help explain why candidates further from the center appear to be preferred even when primary electorates are moderate (Cooper and Munger, Reference Cooper and Munger2000; Chen and Yang, Reference Chen and Yang2002). Consequently, there may be considerable benefit to candidates who can communicate non-centrist positions during the nomination.
2.1. Communication and positional change
Legislators signal preferences through roll-call voting (Canes-Wrone et al., Reference Canes-Wrone, Brady and Cogan2002) and other candidates need to make alternative credible claims of positions, such as by differentiating themselves through their policies, behavior, or language. Intra-party positioning may include drawing support from aligned allies, attacking a primary opponent on ideological grounds, or associating with an ideological faction (Blum, Reference Blum2020). These types of differentiation are difficult to change during an election cycle. Perceptions of candidates’ positions may also be based on information obtained prior to the election, giving campaigns limited ability to shift over time. Candidates may also perceive strategic disadvantages of moving positions, such as being labeled as inconsistent or of “flip-flopping,” which voters are liable to punish (DeBacker, Reference DeBacker2008). Under the assumptions of the strategic positioning dilemma, we expect candidates to adopt non-median positions during the primary, with limited moderation of nominees in general election campaigns due to the electoral penalties attached to moving position. Because we do not expect primary winners to adapt their positions, we focus on losing candidates’ positional adaptation after primary defeats to empirically identify artificial positioning during the primary.
Political communication—including press statements, interviews, and social media activity—allows more flexibility, enabling candidates not only to alter their policy positions but also to change emphasis (Meyer and Wagner, Reference Meyer and Wagner2019). Candidates can reposition not only by changing their stances on issues but also by changing the issues that they talk about (Budge and Farlie, Reference Budge and Farlie1983). Candidates who present themselves away from the center in their policy positions are also non-centrist in their communication, demonstrated here by the alignment of positions derived from voting behavior and social media communication for candidates in our data who ever served in Congress.
Most losing candidates in our sample did not run for alternative public office following their defeat. Though most—not all—remained active partisans, relatively few faced continued deliberation or public votes on their positions. Some candidates ran for or continued to hold local public office, but the vast majority did not. We consider losers’ social media communication after the primary as the best available approximation of “sincere” preferences. We recognize that even this communication does not take place in a vacuum, as unsuccessful candidates likely wish to remain in good standing with their party, either to run for public office again or to hold an appointed position. Yet, social media posts likely play a minimal role in fulfilling these goals, and, though we acknowledge that candidates will not want to communicate anything that causes reputational damage, they are likely less strategic than contributions in party meetings or other formal venues. We also recognize that the dominant linguistic frames used by party leaders and other elites likely influence candidate communication but minimize the extent of such effects by comparing candidates’ positions against themselves across a relatively short period. Empirically, we also expect that these strategic considerations likely decrease rather than accentuate positional movement compared to (unobservable) communication absent any external incentives.
3. Measuring elite positions
To determine whether candidates communicate artificial positions in primaries, we require positions over time. Common measures of positional estimation based on roll-call votes (Poole and Rosenthal, Reference Poole and Rosenthal1985) or campaign donations (Bonica, Reference Bonica2014) are either not available for all candidates or fail to provide the required temporal granularity. We therefore use an alternative measure placing candidates and officeholders on the same dimension by scaling social media communication. Social media allow political elites to communicate directly with potential voters in public. Twitter in particular has developed into an important campaign tool for parties and politicians that has gained substantial scholarly attention (Russell, Reference Russell2018; Barbera et al., Reference Barberä, Casas, Nagler, Egan, Bonneau, Jost and Tucker2019; Cowburn and Oswald, Reference Cowburn and Oswald2020; Cowburn and Knüpfer, Reference Cowburn and Knüpfer2023). Tweets have become part of the news cycle and Twitter is now a rich source of information about the thematic emphases of politicians and their positions. In line with established literature on the subject (see e.g., Boireau, Reference Boireau2014; Ceron, Reference Ceron2016; Sältzer, Reference Sältzer2020), we analyze Twitter text to position candidates over time. Unsupervised text classification methods include Wordfish, which enables comparisons of election manifestos (Slapin and Proksch, Reference Slapin and Proksch2008) and political speeches (Lauderdale and Herzog, Reference Lauderdale and Herzog2016). One challenge of these approaches is a lack of agreement that the extracted dimensions relate to political ideology. Supervised text analysis ensures a correct understanding of the underlying dimension but requires “training data” to teach algorithms which text aligns with different positions. Since ideology is continuous rather than categorical, methods such as Wordscores (Laver et al., Reference Laver, Benoit and Garry2003) use scaling, but set fixed endpoints using anchor documents. Similar approaches have also been applied to newspapers (Gentzkow and Shapiro, Reference Gentzkow and Shapiro2010) and television channels (Martin and McCrain, Reference Martin and McCrain2019). To identify the dimension of partisan conflict, Goet (Reference Goet2019) and Green et al. (Reference Green, Edgerton, Naftel, Shoub and Cranmer2020) use supervised learning on party labels to identify positions. We follow this approach here.
3.1. Data
We collected the timelines of social media accounts of candidates running as a Republican or Democrat in a contested primary for the US House of Representatives in 2020. In line with the established literature (Boatright, Reference Boatright2013, Reference Boatright2014), we consider primaries as contested when two same-party candidates feature on a ballot. Twitter accounts were collected based on a search list created by sourcing ballotpedia.com. We restricted our sample to candidates in contested primaries with identifiable Twitter accounts who tweeted regularly enough for us to position them both before and after their primary election date. We include positional data from 988 of the total of 1772 candidates that stood in a contested primary as a Democrat or Republican for the US House of Representatives in the 2020 election cycle. Our sample is heavily skewed toward candidates with a realistic chance of winning the nomination, where a large proportion of excluded candidates did not raise money or actively campaign and received single-digit vote shares. Unsurprisingly, higher-performing candidates were more likely to have an active social media presence.Footnote 2 Our data include candidates from 49 states, as Louisiana does not hold congressional primaries.Footnote 3
Accounts were cross-referenced with manually collected candidate data (Cowburn, Reference Cowburn2022), compiled throughout the 2020 primary cycle using certified data from state's websites. Tweets were collected using the Twitter API implementation rtweet (Kearney, Reference Kearney2018) for all candidates with Twitter accounts in June 2020. Having gathered the list of accounts in June, we constructed our dataset between June 2020 and March 2021. To prepare the data, we removed all URLs, lower-cased, and cleaned for HTML code (such as emojis). We removed names, punctuation, numbers, and Quanteda's (Benoit et al., Reference Benoit, Nulty, Müller, Obeng, Watanabe and Matsuo2018) default English stopword lists to reduce computational requirements. We remove all hashtags and mentions in our main analysis after comparing validity across specifications (see supplementary materials).
3.2. Positions from Twitter text
Following Goet (Reference Goet2019) and Green et al. (Reference Green, Edgerton, Naftel, Shoub and Cranmer2020) we use a supervised machine learning model to estimate candidates’ positions in Euclidean space (Laver et al., Reference Laver, Benoit and Garry2003; Slapin and Proksch, Reference Slapin and Proksch2008). We classify each candidate based on their party identification using a Naïve Bayes classifier. Our model uses a bag-of-words approach to predict the party membership of each candidate. Each word in the dataset is assigned a partisan value which can then be applied to any document to score how “partisan” it is. Traditional classifiers use binary classification to estimate the outcome, but, because we want a continuous measure, we use the (normalized) relative log-likelihood, giving a score that a document has a certain partisan “identity.” In the case of individual positions (as in the validation) this “document” is all tweets by a candidate in a given period.
Uncertainty
One disadvantage of this approach is the absence of confidence intervals. As the model estimates the likelihood of a text's partisanship, there is no natural interpretation of uncertainty. We can quantify how dependent the results are on specific cases and features, for example, if a candidate uses specific terminology in a manner distinct from their colleagues and changes the meaning. To account for this possibility, we compute bootstrapped positions. Instead of computing a single Naïve Bayes model, we resample all data by drawing 90 percent of them 400 times, rerunning the model, and storing the term weights. When predicting the positions of documents, we again predict 400 positions, computing the standard deviation to get an approximation of error. The results are normally distributed positions around a mean, allowing us to quantify potential uncertainty.
To apply our data to our research question we compute candidate positions at different time points, before and after their respective primaries. We use a three-step process: training the Naïve Bayes model, computing positions of members of congress, validating these positions, and aggregating the data at different levels. We predict the party membership of a validation set of 30 percent of candidates using the other 70 percent as training data. We achieve an accuracy of 0.946, precision of 0.955, recall of 0.926, and F1 score of 0.940, indicating that the model is very good at predicting candidates’ partisan affiliation.Footnote 4 Having trained the model at the individual level, we then apply the weights of these terms to tweets aggregated at the candidate level, the candidate level before and after the primary, and the party level over time (weeks). In other words, we train the model on partisan difference and then estimate the degree of partisanship.
Challenges of this approach include variation in the quantity of candidate-level data, with some candidates rarely tweeting and others so active that their tweets are capped by the API rate limitations Twitter imposes (3200 tweets). Perhaps most importantly, our dataset includes a combination of political tweets mixed with apolitical tweets that do not indicate position. This mix of content has the potential to produce problems when scaling positions, where higher rates of non-political tweets could result in candidates being interpreted as moving toward the center (Grimmer and Stewart, Reference Grimmer and Stewart2013). We deal with this problem explicitly by also applying our model to policy-related tweets only.
Our approach has several advantages. We use the simplest possible model, driven by our desire to avoid overfitting, as a model that was too tuned to classify partisanship might neglect intra-party differences. A second advantage is the computational requirements where, because of the speed of Naïve Bayes, large bootstraps can still run on a single computer. This type of model also does not require stop criteria or a loss metric as it is solved on the document feature matrix (DFM), meaning it does not need to converge in the way that a deep learning model would.
External validity
Introducing a new measurement for a latent construct requires external validation, we demonstrate our scores’ predictive validity against other known estimates of congressional candidates. Given that one motivation for this study is the absence of such measures for all candidates, we compare our results with a subset of our data. The most widely used measure is NOMINATE (Poole and Rosenthal, Reference Poole and Rosenthal1985), based on members’ roll-call voting in Congress. Of course, this measure is only available for members who have ever served in Congress. If these members are positioned in a meaningful way that captures the underlying dimension, other candidates placed on the same dimension should also align. In total, we validate our measure using over 2,000,000 Tweets by 518 members of Congress.
Figure 1 shows this validation, with NOMINATE scores on the x-axis. The y-axis shows the average positions predicted by Twitter communication over the entire electoral cycle. To increase the number of data points against which to validate, and to give our model a hard test, we also include US senators and incumbent representatives who retired in 2020 in this plot. Our model was not trained on these members’ tweets, providing an ideal independent corpus against which to validate.Footnote 5
Validation against NOMINATE for members of Congress.
The overall correlation is 0.93, with higher intra-party correlations than alternative recognized scaling measures such as follower network scores (Barbera, Reference Barberä2015) or CFscores (see Barber, Reference Barber2022). We also demonstrate semantic validity by labeling some notable representatives’ positions. In both parties, representatives who are commonly perceived as “moderates”—including Abigail Spanberger, Henry Cuellar, John Katko, and Fred Upton—are also moderate by our measure. Similarly, representatives such as Pramila Jayapal and Jim Jordan, viewed as highly liberal and conservative respectively, are away from the center on our scale. In addition, Democratic representatives such as Alexandria Ocasio-Cortez and Rashida Tlaib, who are incorrectly positioned as moderates by NOMINATE due to their opposition to some Democratic bills,Footnote 6 are positioned as more liberal under our measure. These correlations give confidence that our measure is aligned with the liberal–conservative dimension structuring roll-call voting behavior, and suggest that in some cases where they differ, our measure may even serve as a more accurate proxy for ideology than NOMINATE.
Semantic validity
Though we obtain predictive validity by comparing the positions generated with roll-call votes, we need to qualify our analysis by understanding the language that identifies our dimension. To do so, we interpret influential words that produce scores further from the center. Our measure can be said to have semantic validity if these scores are associated with parties’ positions, campaign rhetoric, or policy issues.
Figure 2 shows the terms for each end of the dimension surrounding the positions estimated in Figure 1 that occur more than 1000 times in the entire corpus of tweets. Positions from Figure 1 are shown in the center of Figure 2. The lower (higher) the position of a word on the y-axis, the more indicative it is for the Democratic (Republican) Party and contributes to a score further to the left (right). Accordingly, representatives that tweet a lot about “illegals” and “rioters” receive scores further to the right than those who tweet about more moderate identifying terms such as “manufacturers” or “regulations.” The positions of words on the x-axis are for presentation purposes only and have no substantive meaning. Figure 2 demonstrates that the terms that score highly in either a liberal or conservative direction are in line with partisan expectations, where terms at the bottom would be words expected to be used by Democrats and terms at the top of the figure expected to be used by Republicans. In other words, Figure 2 indicates that our approach has semantic validity.
Validation with terms.
These terms can broadly be grouped into three categories: policy-related, own-party rhetoric, and negative terms. Policy-related terms to the right included “illegals,” “censorship,” and “unborn.” Republican own-party rhetorical terms included “patriots” and “conservatives.” The terms “rioters,” “communist,” and “leftist” were used by Republican candidates to talk negatively about the Democratic Party and their supporters and were similarly scored to the right. Liberal policy-related terms included “uninsured,” “ubi,” and “for-profit.” Democratic own-party rhetorical terms included “canvass” and “progressive,” and terms such as “lgbtq” and “trans” referred to demographic groups who favor the party. The terms “inhumane” and “cruelty” were negative liberal identifiers. Given that the terms at each end of our scale can be broadly understood as having a partisan valence, we can say that our approach has semantic validity.
4. Findings
Following validation, we trust the model to infer positions. In our first analysis, we produce a model at the party level and focus on dynamics over time. To test the effect of primaries, we are not interested in the date, but the relative time to or since candidates’ respective primaries. Because states hold nomination contests on different dates, we center the time around each intra-party election, using a time-to-primary variable for each tweet as weeks before or after the primary. We then aggregate at the following levels: party, whether the candidate won their primary, and time-to-primary (weeks). Each observation is the aggregate of terms used by members of a party who won or lost the nomination at the same relative time before or after their primary.Footnote 7
4.1. Shifting after the primary: the party perspective
Figure 3 shows the positions of winning and losing candidates in both parties as groups aggregated by week to or from their respective primary. As the figure indicates, Democratic candidates who do not become the nominee shift their position toward the center directly after their primary. Republican losers do not moderate following primary defeats.
Party level positions over time.
To test the statistical significance of this effect, we run a comparative interrupted time series analysis (ITS) with the below specification (see also Linden, Reference Linden2015). Our data are repeated observations of candidates’ communication positions and we expect positions to change following the “intervention”; the primary election date. We use a (comparative) ITS model given the obvious differences between many candidates who win and lose primary elections. Many candidates who win primary elections are either incumbent members of Congress or highly experienced and well-financed challengers. In contrast, many primary losers receive little to no support from party elites, have little financial support, and may be relatively unknown. Put simply, we conceive that there are too many differences between winning and losing primary candidates to control for, even using approaches such as matching, synthetic controls, or propensity score weighting. Instead, we use an ITS which allows us to compare groups and compare candidates’ positions to themselves prior to the intervention. We do not expect primary winners to moderate immediately after the primary in this design. Conversely, we expect that losing candidates will be more moderate after the primary than they were during the nomination. Using an ITS rather than a two-way fixed effects model also allows us to include group characteristics that change gradually during the election cycle. Given that our data-generating process is independent for each time period, we do not include lagged variables in our models (see also Warner, Reference Warner2019).Footnote 8 One drawback of this design is that the differences—both between winners and losers, and losers versus themselves in the previous period—mean our results are associational, and we cannot infer that the presence of the primary is what caused candidates to adopt artificial positions. We run separate models by party, with the following specification for our first models:
Where Y it is candidate position Y given membership of groupFootnote 9 i measured at week t. T t is the time in weeks to or since the primary. X t is a dummy variable representing the primary election, where pre-primary observations take the value zero and post-primary observations the value one. X tT t is the interaction term between post-primary and time, meaning β 2 is the immediate change following the primary and β 3 gives the ongoing movement among all observations. Z i is the group we expect to moderate, which takes the value one if a candidate lost and zero if a candidate won their primary. Coefficients β 4 to β 7 are the same as β 0 to β 3 interacted with losing (Z i), meaning β 6 gives the immediate change among losing candidates immediately after the primary and β 7 gives the ongoing movement following the primary. We expect moderation from losing candidates immediately after they lose their primary, meaning β 6 (Z iX t) is our main object of interest for the first models.Footnote 10
Given that our goal is not the causal identification of differences between winners and losers, we also include a second set of models that are restricted to losing candidates only. These models take the same form as the above specification with the removal of the loser variable Z i and subsequent interactions, meaning X t is the object of interest in these models. Our first models indicate how losing candidates were positioned relative to winners in the same week, whereas the second set of models identify how candidates moved relative to themselves in the previous period.
One potential issue with cross-sectional time series data is non-stationarity, where conditional means are dependent on the time period and where a variable has a unit root. To demonstrate that our models have I(0) balance (Pickup and Kellstedt, Reference Pickup and Kellstedt2022) and to understand the order of integration we perform (augmented) Dickey–Fuller (Dickey and Fuller, Reference Dickey and Fuller1979) tests on each of the four groups’ dependent variables, with results reported in the supplementary material. In each case, our tests return significant values, indicating no unit root on the left-hand side of our models. We also account for variation in the trend stationary dependent variable by including T t in our specification. Of our independent variables, both the primary (X t) and winning or losing (Z i) do not contain a stochastic component. The only term on the right-hand side of our equation that is stochastic is the error term; we demonstrate that the estimated errors (residuals) are indeed white noise in a further series of Dickey–Fuller tests, with the results reported in the supplementary material. These tests indicate that our equation is I(0) balanced.
In line with the visual trend depicted in Figure 3, our first model in Table 1 shows that Democratic losers became significantly more moderate than winners immediately after the primary (Z iX t). In contrast to the weak time trend, the effect is almost 5 percent of the total range of the variable, this is the strongest identifier of position other than partisanship. In other words, losers shift their position after their primary relative to winners, and this shift is more than 20 times greater than the average weekly positional change (T t). Losing Democratic candidates were more moderate than winners prior to the primary (Z i) yet moved much further rightward following the primary (Z iX t). All other Democratic coefficients in this first model are substantively close to zero.
ITS results: party level
Note: Newey–West standard errors shown in parentheses.
*p < 0.1; **p < 0.05; ***p < 0.01.
For Republican losers, Table 1 indicates no significant moderation following primary defeats relative to primary winners (Z iX t). It appears that Republican winners moderate slightly after the primary (X t) then quickly move back toward their pre-primary positions in subsequent weeks (X tT t), also seen in Figure 3. Across the whole period, losing Republican primary candidates are consistently further to the right than winners (Z i). All other coefficients in this first model are substantively close to zero.
In the second set of models, we consider the position of losers after the primary compared to their positions during the primary, indicated by the post-primary coefficient (X t). Among Democratic losers, our finding is virtually unchanged, with Democratic candidates again positioned significantly further to the right immediately after the primary compared to their previous positions. Among Republicans, we also see evidence of moderation of losers in the immediate post-primary period as compared to their position during the primary. As depicted visually in Figure 3, it appears that all Republicans moderated immediately after the primary and then returned to their original positions over time. This movement is substantively far smaller than among losing Democrats.
Unsurprisingly, partisanship—shown here in the form of the intercept—is the strongest predictor of position for candidates in both parties. At the party level, we find a clear moderating effect among losing Democratic candidates.
4.2. Robustness to the changing salience of non-political tweets
One identifiable problem of ideal point estimation over time is the changing salience of features that contribute to the dimension (Grimmer and Stewart, Reference Grimmer and Stewart2013). The appearance of moderation may stem from movement toward more centrist content—ideological moderation—or a reduction of political or policy-related content. Accordingly, it might be that candidates are merely tweeting less about politics and turning their account into a private platform after they lose a primary rather than continuing to discuss politics.
To ensure the robustness of our approach to this problem, we apply our method to a subset of explicitly policy-related tweets. To do so, we hand-coded a random set of 1200 tweets using three categories; political (y/n), policy-related (y/n), and policy area (using policy fields established in the Comparative Agendas Project). Though the sample was too small to analyze policy areas individually, roughly half of the tweets in the sample were policy-related. We then trained a classifier for these tweets, using an English-language Bidirectional Encoder Representations from Transformers (BERT) (Devlin et al., Reference Devlin, Chang, Lee and Toutanova2019) model, which achieves a satisfactory F1 score of 0.8. We use this model to predict whether all 2,500,000 tweets in our original sample were policy-related (again, roughly half were) and estimate positions.Footnote 11 We then re-ran our analyses on this subset.
The results are shown in Table 2 and align with our main finding, with substantively significant moderation among Democratic losers after the primary, either compared to Democratic winners in the same period or to themselves during the primary. Movement immediately after the primary is again more than 20 times the size of the average weekly movement and is the strongest indicator of position other than partisanship. Our finding that Republican losers were more moderate after than during the primary is no longer significant when restricted to policy tweets, suggesting that this finding was at least partly the result of a shift in focus. This additional analysis gives confidence that our main result for Democrats is not an artifact of the changing saliency of policy-related tweets after primary defeats and is instead evidence of positional adaptation by losing candidates.Footnote 12
ITS results: policy tweets only
Note: Newey–West standard errors shown in parentheses.
*p < 0.1; **p < 0.05; ***p < 0.01.
4.3. Individual-level robustness
To avoid the ecological fallacy, we also analyze the individual level. We do not have enough tweets at the individual level to reliably compute positions in the same density as at the party levelFootnote 13 meaning we instead aggregate candidates’ positions before and after their primary to enable the direct comparison of candidate-level movement. In this model, we control for incumbency given that incumbents may face additional pressures and incentives to maintain their positions because they have political records to uphold which can be held accountable by voters and opposition candidates. Given that district partisanship influences positional incentives in both primary and general elections, we control using The Cook Political Report's (2017) partisan voting index (PVI), rescaled to a +/– Republican lean.Footnote 14
Figure 4 shows the individual-level results. These models use the difference (movement) in candidates’ positions before and after their primary as the dependent variable, where positive coefficients indicate rightward movement and negative coefficients indicate leftward movement. We test using two dependent variables: absolute movement, and a variable of significant movement. This variable takes the value 1 if a candidate moves rightward three standard error confidence intervals and the value –1 if a candidate moves left to the same degree.
Individual-level movement.
In line with our party-level findings, Democratic losers took more moderate positions after the primary in both individual-level models, giving further confidence in our party-level findings. Republican losers also move slightly to the right, but the effect is not statistically significant. As in the party-level model, partisanship—the intercept—indicates moderation among all candidates at the individual level following the primary. Democratic incumbents moved slightly to the left at the individual level, with no significant effect among Republicans. District partisanship had no relationship to Democratic positioning and a small but significant association for Republicans, who took less conservative positions in districts that were less favored for the party.
5. Discussion
Our results indicate that primaries are associated with artificial position-taking among Democratic candidates only. We interpret these findings as support for the strategic positioning dilemma among Democratic candidates, who adopted artificial positions during the primary which they did not continue to hold once absent the (perceived) incentives to do so. Among Republican candidates, we find minimal evidence of artificial positioning, suggesting that communication during the primary was done out of conviction rather than for perceived advantage. Absent electoral incentives, losing Republican primary candidates continued to communicate highly conservative positions.
The moderation of losing Democratic candidates after the primary indicates our theorized effect that intra-party competition is associated with artificial extremism during the nomination. Grossmann and Hopkins (Reference Grossmann and Hopkins2016) suggest that the Democratic Party is a diverse coalition of group-oriented actors. Rather than being defined by ideological conflict, candidates advocate for different groups which are understood primarily in terms of demographics and identity. Consequently, Democratic candidates are less frequently ideological purists and so may be more comfortable adapting their positions. Because ideology is not a central binding force in the party, candidates are able to be more flexible and change positions than their Republican counterparts. If candidates perceive that important policy demanders are to their left, they may have additional incentives to adopt artificial positions during the nomination. The Democratic Party might therefore recruit more strategic candidates or be more selective in recruitment by actively seeking out candidates who can adapt positions. The ability to be flexible and strategically appeal to many of the diverse interest groups that make up the Democratic Party appears one important characteristic sought out by party elites and policy demanders in the party network who play a central role in candidate recruitment (Cohen et al., Reference Cohen, Karol, Noel and Zaller2008; Hassell, Reference Hassell2018). These groups prefer candidates with a broad appeal during the nomination process (Masket, Reference Masket2020), in part out of necessity because the party needs to carry some swing or even marginally Republican-favored districts in general elections to control the House. In short, recruitment strategies matter and are likely asymmetric (Maestas and Stewart, Reference Maestas, Stewart and Carson2012). Intra-party power struggles likely provide further incentives for Democrats to moderate after a primary. Though progressives have made recent gains, the Democratic Party remains dominated by “establishment” center-left moderates, meaning losing candidates who want to continue a career in the party are wise to moderate to appeal to like-minded individuals.
For Republicans, our results align with scholarship that positions candidates for Congress as more extreme, or at least more ideologically consistent, than other groups and voters in their party (Bafumi and Herron, Reference Bafumi and Herron2010; Barber, Reference Barber2016). These results run counter to the expectations of the strategic positioning dilemma. Candidates in the Republican Party take non-centrist positions out of conviction both during and after the primary, where losing a primary was not associated with moderation. That losing Republicans largely continue to communicate non-centrist positions likely reflects a reality where the only candidates running are located so firmly on the right of the political spectrum that they perceive little concern over strategic positioning during the nomination. This explanation aligns with scholarship indicating that the Republican Party has moved sharply rightward in recent years (Hacker and Pierson, Reference Hacker and Pierson2006; Mann and Ornstein, Reference Mann and Ornstein2012; Theriault, Reference Theriault2013), meaning losing primary candidates have less incentive to moderate to help their future career in the party. Republican partisans are also less tolerant of elite positional heterogeneity (Dunn, Reference Dunn2021), meaning party elites and other actors in the formal party organization may be more disposed to recruit loyal (or sincere) believers who hold consistent positions away from the political center. Given the (perceived) position of primary voters and policy demanders in the party, moderate Republicans may simply decide that running for Congress is not worthwhile (Thomsen, Reference Thomsen2017). Institutional biases in general elections—including aggressive Republican gerrymandering in the previous redistricting cycle and the electorally inefficient clustering of Democratic voters in urban districts—may also have furthered a perception among Republican policy demanders and primary voters that candidates on the right of the political spectrum are electorally viable.
Given that our analysis is conducted over a single electoral cycle, we must also consider the relative effect of 2020 electoral conditions on the two parties. Boatright and Moscardelli (Reference Boatright and Moscardelli2018) demonstrate that congressional primaries have a “presidential pulse.” In 2020, the Democratic Party was favored to win the presidency and expected a strong down-ballot performance, with higher numbers of primary candidates as a result. Higher levels of primary competition may have served as a further incentive to induce Democratic candidates to adopt artificial positions.
The party-level differences may also relate to demographic and ideological differences between Twitter and non-Twitter users. Twitter users are Democratic-leaning and disproportionately come from demographic groups which favor the party, such as young college-educated Whites with higher incomes (Wojcik and Hughes, Reference Wojcik and Hughes2019). Even among Democratic partisans, those on Twitter tend to hold more progressive or left-leaning positions (Cohn and Quealy, Reference Cohn and Quealy2019), with fewer moderates active on social media (Hawkins et al., Reference Hawkins, Yudkin, Juan-Torres and Dixon2018). Democratic primary candidates may therefore have communicated positions on Twitter to appeal to a section of the electorate that they—correctly—perceived as non-centrist. In contrast, Republican candidates may perceive that fewer of their primary voters are on Twitter and so use the platform to communicate to journalists and media outlets, other candidates, or party figures.
Asymmetries in the parties’ financial structures may further explain our findings. Basedau and Kollner show that “centripetal tendencies are better avoided when the channels of party finance are controlled by the party leadership” (Reference Basedau and Köllner2005, 19), and recent literature highlights clear partisan differences in this regard. Boatright and Albert (Reference Boatright and Albert2021) show that independent expenditures were not particularly prevalent in financing primary challengers to Democratic incumbents in 2018. Assuming a similar pattern in 2020, the tighter financial control of the formal institutions of the Democratic Party may have incentivized losing candidates to moderate to retain favor with party leadership and advance their political careers. The asymmetric structure of media ecosystems, with greater pressure from the right and far-right of the ideological spectrum (Heft et al., Reference Heft, Knüpfer, Reinhardt and Mayerhöffer2021), may also have induced Republican candidates to maintain conservative positions. Pierson and Schickler (Reference Pierson and Schickler2020) find that meso-institutional structures pull Republicans away from the center more than Democrats. One interpretation of our findings is that these structures continue to affect Republicans’ positions following primary defeats.
For general election voters, these results are not encouraging when considered in terms of spatial models of voting. Given that we find limited evidence of moderation among primary winners in either party,Footnote 15 voters in November appear to have been presented with polarized choices—as theorized by Fiorina et al. (Reference Fiorina, Abrams and Pope2005)—albeit for contrasting partisan reasons, with Democratic candidates having strategically adopted artificial positioning during the nomination and Republicans sincerely holding non-centrist positions out of conviction. Non-moderation of Democratic primary winners may indicate a perception among candidates that they must continue to hew to the preferences of policy demanders beyond the primary or reflect candidates’ beliefs about the electoral risks associated with moving positions between a primary and general election. Among Republicans, our data suggest limited adaptation, and positions appear more deeply ingrained in the preferences of candidates.
6. Conclusion
We find that losing Democratic candidates moderate after the primary. We argue that this is evidence that candidates communicated artificial positions during the nomination to try and align with key policy demanders and the perceived positions of their primary voters during the nomination. Losing Republican candidates did not moderate following their primary defeats. These results align with scholarship indicating asymmetry in the ideological positions (Hacker and Pierson, Reference Hacker and Pierson2006; Theriault, Reference Theriault2013) and identities (Grossmann and Hopkins, Reference Grossmann and Hopkins2016) of the two major parties and the policy demanders active during the nomination process within each. These differences provide distinct partisan constraints and incentives to candidates both during and after primary elections.
The debate over whether primaries contribute to polarization in Congress is ongoing (Fiorina et al., Reference Fiorina, Abrams and Pope2005; Abramowitz, Reference Abramowitz2010; Sides et al., Reference Sides, Tausanovitch, Vavreck and Warshaw2020), yet, many studies only consider this question in terms of a selective effect from primary voters. We demonstrate a further way in which contested nominations may exacerbate partisan conflict in Congress: the adaptation of candidate positions during the nomination phase of the election cycle. If many candidates perceive that communicating artificial positions is beneficial during the primary and then feel compelled to maintain those positions during the general election, voters in November will be presented with more polarized choices as a result of the nomination process.
We find little movement among nominees in either party once they are selected, a potentially positive normative finding in terms of representation. Regardless of whether candidates adopt sincere or strategic positions, primary winners communicate positions in general election campaigns that are consonant with their positions during the nomination. How candidates communicate in a primary is at least consistent with what they advocate when they become the nominee—and, potentially, indicative of the policies they will support in Congress. This finding contrasts with the image of politicians as pandering to different groups for their own benefit (Lippmann, Reference Lippmann1955; Jacobs and Shapiro, Reference Jacobs and Shapiro2000).
Supplementary material
The supplementary material for this article can be found at https://doi.org/10.1017/psrm.2023.62.
To obtain replication material for this article, please visit https://doi.org/10.7910/DVN/.
Open access
