All (Mayoral) Politics is Local?

Sanmay Das∗ Betsy Sinclair† Steven W. Webster‡ Hao Yan§¶

November 22, 2019

Abstract

American politics in the contemporary era has become highly “nationalized” – focused onnational issues instead of local issues. Previous research has estimated nationalization via elec-toral vote shares and found significant state-level nationalization. We challenge the existingnationalization narrative by examining an alternative source of data – the political rhetoricused by mayors, state governors, and Members of Congress on Twitter. We analyze rhetoric interms of partisanship and topic similarity. We find that gubernatorial rhetoric – but not mayoralrhetoric – closely matches that of Members of Congress. There are substantial differences in thetopics and partisan content of mayoral speech. Our core finding is that the emphasis of mayoralspeech is local; the office of the mayor has been more immune to the trends of nationalizedpolitics.

∗Department of Computer Science and Engineering, Washington University in St. Louis.†Department of Political Science, Washington University in St. Louis.‡Weidenbaum Center on the Economy, Government, and Public Policy, Washington University in St. Louis.§Department of Computer Science and Engineering, Washington University in St. Louis. Now at Facebook.¶This project was heavily influenced by the work conducted by the Green Team as part of the BUILD conference

at WUSTL in May 2018.

1 Introduction

At the heart of the republican ideal in the United States is the notion that elected officials properly

represent those to whom they are electorally accountable. That is, elected officials in a well-

functioning democratic setting are expected to articulate the preferences and desires of the governed.

Numerous studies of democratic representation compare some measure of aggregate constituency

preferences with some measure of elite behavior or outcomes (see, e.g., Esaiasson and Wlezien,

2016). Given the nature of American federalism, this implies that local officials should be concerned

with representing and articulating the preferences of citizens at the local level while federal officials

should be comparatively more concerned with national-level policies.

However, recent scholarship suggests that all levels of American politics have become increas-

ingly nationalized. Thus, rather than being contested over local issues that are of importance to

each respective constituency, all political campaigns are largely focused on national issues (Hop-

kins, 2018). What are the stakes of nationalized local politics? This focus on national politics

often comes at the expense of local issues that are typically of greater importance to citizens; the

nationalization of American elections has the potential to weaken the degree to which citizens are

effectively represented by their governing officials. Evidence for such a nationalization effect has

been found in U.S. House, Senate, and gubernatorial elections (Carson and Sievert, 2018; Sievert

and McKee, 2018; Aleman and Kellam, 2008).

In this study, we examine the extent to which the nationalization of American politics has

affected mayoral and gubernatorial representation. Presiding over the most local forms of govern-

ment within the federal system and managing the government services with which citizens interact

most frequently, both mayors and governors have substantial influence over the daily lives of their

citizens. Accordingly, understanding whether these elected officials have shifted their focus toward

national issues and away from the needs of their local constituencies is of tremendous importance.

To address this question, we depart from typical studies of nationalization by analyzing patterns of

elite speech with respect to partisanship and topic similarity via the Twitter social media platform

rather than election results. Adopting this measurement strategy allows us to both validate existing

findings from the literature on the nationalization of American politics as well as to introduce new

1

metrics that capture differences in the behavior of political elites themselves.

We collect a unique dataset of 404,049 tweets from U.S. mayors, 102,670 tweets from governors,

and approximately 964,810 tweets from Members of Congress. To establish a baseline from which

to compare the rhetoric of political actors, we also collect a corpus of tweets sent from an apolitical

set of Twitter users: soccer players. We analyze this data both in terms of partisanship as well as

topic similarity. In order to do so, we need to carefully construct measures of nationalization and

partisanship, and also consider the problem of how to measure similarity across groups. Using these

measures, we show that mayors are not completely immune to the trend of using more nationalized

rhetoric, doing so more than apolitical Twitter users. However, we further show that the degree of

nationalization in their communication is relatively small; mayors – regardless of party – talk about

qualitatively different subjects than their co-partisans in Congress. Governors, by contrast, are

more rhetorically in line with Members of Congress. Mayoral offices focus on predominantly local

issues. Such a finding complements work suggesting that the polarization and partisan antipathy

that is pervasive within the context of national politics is largely absent at the most local levels of

government (Jensen et al., 2019). We also present results that indicate that the degree to which

mayors discuss national issues rather than local issues is dependent upon the size of the city over

which they govern (consistent with their exposure to national trends) and the number of years they

have been in office (consistent with their ability to ignore national pressures). As the population

size of a city increases, mayors are more likely to engage in nationalized political rhetoric. More

time spent in office, on the other hand, is associated with lower amounts of nationalized rhetoric.

These results are robust to the inclusion of individual mayoral traits and political preferences.

This paper proceeds as follows. First, we characterize recent work on the nationalization of

American elections and develop a theory linking governors and mayors to nationalized behavior

manifested through elite speech. Next, we present a novel series of results based on tweet analyses

suggesting that while gubernatorial rhetoric resembles the type of speech typically found among

national politicians, American mayors are less partisan and less nationalized than their Congres-

sional counterparts. We proceed to show how mayors’ comparatively lower levels of nationalized

rhetoric vary as a function of individual mayor- and city-level covariates. Finally, we conclude with

2

a discussion about the implications of our results for democratic governance.

2 Elections, Accountability, and Nationalization

While canonical models suggest that voters make rational decisions about candidates within each

electoral context (see, e.g., Downs, 1957), recent evidence suggests that American political behavior

has become increasingly nationalized (Hopkins, 2018; Abramowitz and Webster, 2016). National-

ization, as typically conceptualized by the literature, conventionally refers to one of two things. In

one case, nationalization refers to the process whereby voters judge politicians – regardless of the

electoral level – by their evaluations of the national parties. In the second case, nationalization

occurs when state and local elections are largely fought over national (largely symbolic) issues.

Among other causes, these two different forms of nationalization are thought to occur when parties

offer similar candidates across electoral levels or the media market changes in a way that priori-

tizes national over local news (Hopkins, 2018; Martin and McCrain, 2019). To the extent that this

nationalization phenomenon causes state and local officials to prioritize national interests and the

concerns of ideologically-motivated donor bases to the exclusion of the local citizenry’s needs, as

Hopkins (2018) suggests, the nationalization of American politics has the potential to drastically

alter the relationship between citizens and their elected officials at the local level.

The existing literature on the growing nationalization of American politics has largely focused

on U.S. House and Senate elections (see, e.g., Carson and Sievert, 2018; Aleman and Kellam, 2008).1

Jacobson (2015), for instance, notes that the incumbency advantage in American politics, long seen

as the source of high re-election rates across the country, has been declining over time, and the

explanatory power of partisanship in predicting election outcomes for House elections has increased

tremendously. Such a shift indicates that Americans today care less about the specific person who

represents them and more about the partisan balance of power in Congress. This implies that,

unlike in earlier eras, it is increasingly difficult for politicians to court the “personal vote” in their

districts (Mayhew, 1974; Fenno, 1978).

1Two notable examples are Hopkins and Schickler (2016) and Hopkins (2008). The former analysis focuses on the

nationalization of state party platforms, while the latter examines the nature of urban policy agendas.

3

Outside of federal elections, scholars have focused almost entirely on how the nationalization of

American politics has affected gubernatorial elections.2 Hopkins (2018) shows that the state-level

correlation between voting for the Democratic presidential candidate and the Democratic guber-

natorial candidate has increased considerably over time. Nearly thirty years ago, the correlation

between voting patterns at these two electoral levels was a moderately strong .61. By 2010, the

correlation had strengthened to just under .9. A similar relationship exists between the percentage

of the two-party vote accruing to the Democratic presidential candidate and the percentage of the

votes received by the Democratic gubernatorial candidate (see also, Sievert and McKee, 2018).3

Mayors may be somewhat immune from these national trends: voters cast ballots in mayoral

elections based upon retrospective local economic conditions, such as local unemployment, and this

effect typically dwarfs the effect of national economic conditions (Hopkins and Pettingill, 2017).

That voters are casting ballots for mayors in a way that does not necessarily channel national

political trends suggests that mayoral representation may be based on local, not national, politics.

Relevant for this project, moreover, they find that “[i]n cities with their own TV stations and

newspapers, there is a robust relationship between city-level unemployment and the [electoral]

performance of the incumbent mayor” (Hopkins and Pettingill, 2017). That information about

local economic conditions changes political behavior, particularly relative to national economic

conditions, suggests that mayoral politics remains more focused on local issues.

That the existing literature measures nationalization by examining the correlation between

partisan vote shares at various levels of the federal electoral system is sensible as elections allow

citizens to hold elected officials accountable for their actions in office (Ferejohn, 1986; Fearon, 1999;

Fiorina, 1981), ensure proper representation (Verba, Schlozmann and Brady, 1995), and provide an

avenue for expressive political participation (Hamlin and Jennings, 2011). Yet while this approach

has been useful in establishing the rise of nationalized political competition, it necessarily focuses

2One notable exception is Fouirnaies and Hall (2018), who study the electoral incentives of state legislators. They

find that state legislators are responsive to their constituencies – a finding which runs counter to the phenomenon of

the nationalization of American elections.3This relationship holds whether one examines gubernatorial elections that occur during midterm elections or

concurrently with presidential elections.

4

on voter behavior just as much – if not more – than elite behavior. Because we are interested here

in the nationalized behavior of political elites, we need a measure that is less dependent upon the

actions of the mass public. For this reason, we measure nationalization by examining the similarity

in mayoral and gubernatorial speech with Members of Congress, captured via Twitter.

Why might politicians use social media? Using social media may provide policy benefits to

constituents (through sharing of information) as well as personal benefits to politicians (through

self-promotion). We anticipate that governors and mayors will use social media when the benefits

outweigh the costs – that social media usage is associated with a governor or mayor’s electoral

incentives and that governors and mayors will be strategic with their social media usage. A politi-

cian’s social media content is a component of their strategic communication plan to engage with

constituents as well as a broader audience. Importantly for our purposes, social media provides us

with an opportunity to capture and record a component of this strategic communication plan.

That communication is a key component of strategic elite-level communication in modern poli-

tics has been shown by numerous scholars. One study, for instance, found that Republican officials

tend to use veiled religious language that appeals to white evangelical Protestants. This “GOP

Code” is designed to both signify a politician’s in-group partisan membership and solidify the

support of sympathetic voters (Calfano and Djupe, 2009). Additionally, research has shown that

politicians are increasingly and deliberately engaging in speech that seeks to outrage and belittle the

opposing political party and its supporters. This verbal strategy, which Grimmer (2013) refers to

as “partisan taunting,” is yet another way that political elites seek to signal their in-group member-

ship to their supporters in the electorate. Speech, then, has been shown to be an important aspect

of politicians’ presentational style. We extend this logic by arguing that, if they are experiencing

the forces of nationalization, mayors and governors should be using rhetoric and patterns of speech

that are similar to their congressional counterparts. Moreover, if these more locally-focused offices

are nationalized, mayors and governors should be discussing the same topics as their co-partisans

in Congress.

While analyzing nationalization via elite speech on Twitter seems appealing, one potential

problem with this approach is that citizens are not uniformly on Twitter. In fact, most citizens are

5

not active on the social networking platform. Accordingly, it is possible that mayors and governors

are engaging in nationalized rhetoric but no one observes this. However, such a concern appears to

be unfounded. Studies have shown that those individuals who are not on Twitter are nevertheless

exposed to politicians’ tweets via journalists’ coverage of political affairs. Indeed, because journalists

view tweets as newsworthy they oftentimes quote from politicians’ tweets – or link directly to the

tweet itself – when writing their articles (McGregor and Molyneux, 2018; Lawrence et al., 2014).

Accordingly, it is possible for citizens who are not themselves users of Twitter to still come across

tweets from their political representatives.

We expect governors to be engaging in rhetoric – and discussing topics – similar to their co-

partisans in Congress. Such an expectation comports with the existing literature, which argues that

governors have become nationalized along both of the dimensions described above. Indeed, gover-

nors are often evaluated by the electorate through the lens of national politics and gubernatorial

elections are increasingly fought over national issues. In addition to changes in the media landscape,

one potential reason to expect governors to engage in nationalized rhetoric is that governors often

have ambitions for higher office. Governors frequently are candidates for President of the United

States, and 17 presidents have previously held a state’s governorship (Shapiro and Lawrence, 2015).

Engaging in nationalized rhetoric, then, is a potential way for governors to engage with issues of

broad importance and build their national brand in preparation for a presidential bid.

In addition to their tendency to harbor ambitions for higher office, we expect governors to

use nationalized rhetoric due to the nature of intraparty federalism. As Hopkins (2018) notes,

the demise of the patronage era made state political parties increasingly dependent upon national

party organizations for money and support. Such a top-down organizational structure changes

the incentive structure for both state parties and state-level politicians, forcing a more national

focus over more a regional approach. Such a shift has been exacerbated by the end of the solidly

Democratic south and the polarization of state legislatures, two trends which have made state

politics more similar to national political competition (Shor and McCarty, 2011; Hopkins, 2018).

There is less of an ex ante reason to assume that mayors engage in nationalized rhetoric. In fact,

mayors may be qualitatively different than statewide and, in particular, national politicians. May-

6

oral partisanship may be associated with different issues than partisanship at a national level, and

mayoral offices may be sufficiently local such that national partisanship does not affect voting nor

decision-making (Adrian, 1952). Mayoral governments are focused on evaluating what can be built

where, on public goods such as streets and sanitation, and on public safety (Oliver, 2012). Further,

according to Gerber and Hopkins (2011), “the presence of Republican mayors in overwhelmingly

Democratic cities – consider Rudolph Giuliani in New York or Richard Riordan in Los Angeles –

provides a hint that partisanship may function differently at the local level.” Additionally, mayoral

elections do not necessarily occur at the same time as national campaigns, and thus may not benefit

from the media market trends that align gubernatorial campaigns with national issues.

Nevertheless, there are compelling pieces of evidence to indicate that mayoral representation

can and does align with national political agendas. Indeed, mayoral fiscal preferences are known to

align with their partisan labels (Einstein and Glick, 2018), and partisan control of local government

is predictive of local government spending at both the county level (Percival, Johnson and Neiman,

2009; Choi et al., 2010; de Benedictis-Kessner and Warshaw, 2018; Ybarra and Krebs, 2010) and the

mayoral level (Tausanovitch and Warshaw, 2014; Einstein and Kogan, 2015; de Benedictis-Kessner

and Warshaw, 2016, 2018).4

Accordingly, understanding whether mayors engage in nationalized rhetoric and the situations

under which they do so are important – and unanswered – questions. We theorize that mayors’

representational style is distinct from congressional representation, and moreover that we will be

able to evaluate these differences in representational style via elite speech on Twitter.5 However,

we also expect that a mayor’s likelihood of engaging in nationalized rhetoric will vary based off

of certain mayoral and city-level covariates. It is possible that mayors who represent cities with a

strong partisan tilt will be more likely to use nationalized political rhetoric, which is in accordance

4Earlier work did not necessarily support this finding (Ferreira and Gyourko, 2009; Gerber and Hopkins, 2011).5It is now well-established that text analytics can contribute to quantifying and measuring some important char-

acteristics of political speech, for example partisanship (Lin, Xing and Hauptmann, 2008; Grimmer and Stewart,

2013; Gentzkow and Shapiro, 2010; Ahmed and Xing, 2010; Iyyer et al., 2014; Yan et al., 2019). These methods

are often used to generate substantive conclusions. For example, Gentzkow, Shapiro and Taddy (2019) report that

partisanship in Congress, measured as the ease of telling which party a speaker is from based on a fragment of text

they generated, has been increasing in recent years.

7

with Grimmer’s (2013) theory of polarized elite rhetoric. Moreover, in line with the approach

in de Benedictis-Kessner and Warshaw (2018), we also expect that mayors’ use of nationalized

rhetoric will depend upon the institutional and population features of their city. We expect mayors

to be more nationalized in their rhetoric when their city employs a city manager, as city managers

engage in more of the apolitical day-to-day activities of running a city. Freed from the requirements

of such activities, mayors whose city employs a city manager should have more time to engage in

nationalized political rhetoric. Finally, we expect the degree to which mayors engage in nationalized

rhetoric to be increasing in the size of their city’s population. As a city’s population increases each

individual becomes less likely to have any form of personal engagement with the mayor’s office. As

a result, mayors will have less of an incentive to discuss local and personal issues and will have a

greater opportunity to focus on national issues.

3 Twitter Data

We collected Twitter data for mayors, governors, members of the House of Representatives, and

English Premier League Soccer Players to examine the extent to which the nationalization of

American politics has affected gubernatorial and mayoral behavior.

We collected the official Twitter handle for every mayor who is a member in the United States

Conference of Mayors as of May 2018.6 Though mayors are not uniformly on Twitter, we obtained

Twitter handles for 587 mayors (42% of the mayors eligible for inclusion in our study).7 We

excluded all mayors who posted less than 100 tweets from our data, which left us with a sample

of 372 mayors, 27% of all mayors in the National Council of Mayors. In late December 2018, we

scraped the previous 3,200 tweets for each of these Twitter handles.8 Our focus is on tweets that

6The United States Conference of Mayors is the official non-partisan organization of cities with populations of at

least 30,000. We chose this sample because we anticipated that mayors of medium to large-sized cities have access to

Facebook and Twitter and possibly even staff to help monitor these accounts.7We first collected official accounts, and if an official account was not available, we then used a campaign account.8This is the maximum number of tweets we were allowed to get due to Twitter’s API limit. We were able to get

200 tweets in each request. For each of our targeted politicians, we continued sending requests until we reached the

3,200 limit or their total tweet limit.

8

were posted by the user, so we excluded all re-tweets. This generated a total of 404,049 tweets

from mayors. Full summary statistics of our mayoral data are shown in Table 1. Our aim is to

compare mayoral rhetoric with more nationalized speech, so using an identical collection protocol

we collected 102,670 tweets from 49 governors and 964,810 tweets from 475 Members of Congress.9

Finally, to ensure we can compare this data to an apolitical group, we collect Twitter data from

the English Premier League soccer players.

For mayors, we also collected individual-level attributes. This includes mayors’ gender, racial

identification, length of term in office, the percentage of the vote share they received in their most

recent election, and each mayor’s partisan affiliation.10 In instances where mayoral elections are

conducted without party labels, mayors are classified as non-partisan. We also collected information

on the institutional features of each city in our dataset.11 To these individual- and institutional-

level variables we added information on the demographic composition of each city.12 A summary

9The Governor of Alaska, Mike Dunleavy, has a private Twitter account which we were unable to scrape – we are

otherwise able to collect all gubernatorial Twitter data.10This data was obtained from a combination of mayors’ campaign websites, newspaper biographies of the mayors

or of the mayoral election, or it was inferred from the mayor’s collection of endorsements or source of campaign

funds. For example, mayors who received endorsements from prominent national Democrats were classified as a

Democrat. On the other hand, mayors who received funding from conservative political action groups were classified

as a Republican.11This includes indicators for whether the city government operates with a mayor-council system or a council-

manager system, as well as whether the city employs a city manager. In mayor-council systems, a city’s mayor is

elected separately from the city council and oftentimes acts as the city’s chief executive officer. Council-manager

systems, by contrast, select the city’s mayor from within the city council. Under the council-manager system of local

government, the mayor is comparatively weak and the day-to-day operations of the city are overseen by an appointed

or elected city manager. Though all city-manager forms of government contain a city manager, city managers are

also occasionally used in mayor-council systems.12Using data from the most recent American Community Survey (ACS) estimates, we collected the mean age of

residents of each city, as well as the percentage of male and female residents. We also collected information on the

percentage of Black and Hispanic residents, as well as the percentage of individuals who own or rent their homes. We

added to this data proxy measures of each city’s election results for the 2016 presidential election. These estimates

are constructed by using county-level presidential election returns that are weighted by the percentage of the city that

is in each county. For example, because 93.3% of Atlanta, GA, is in Fulton County and 6.7% is in DeKalb County,

the presidential vote share for Atlanta is a weighted average where the Fulton County vote shares receive a weight

9

of the mayoral data can be seen in Table 1.

Statistic N Mean St. Dev. Min Max

White 372 0.798 0.402 0 1Black 372 0.118 0.323 0 1Hispanic 372 0.059 0.236 0 1Male 372 0.769 0.422 0 1Female 372 0.231 0.422 0 1Republican 372 0.390 0.488 0 1Democrat 372 0.608 0.489 0 1Nonpartisan 372 0.003 0.052 0 1Council-Manager 372 0.470 0.500 0 1Mayor-Council 372 0.546 0.499 0 1City Manager 372 0.546 0.499 0 1Avg. Trump Vote Share 372 43.041 14.551 8.706 82.514

Table 1: Summary Statistics of Mayoral Data. This table shows summary data for our mayoral data, conditionalupon having an active Twitter handle.

4 Quantifying text data

We consider two distinct strategies for text representation in this paper – first, to focus on words/phrases

and second, to focus on topics. We rely on topic differences between the different sets of actors

in our data – mayors, governors, Members of Congress, soccer players – to not only illuminate

the different foci of these groups but also to illustrate the differences within and between them.

Our findings show that mayors are less similar to congressional representatives, for example, than

governors.

Word/phrase vector representation Research over the last two decades in machine learning

has established that simple representations of documents in terms of either presence or absence, or

adjusted frequency, of terms in a given vocabulary can be coupled with regularized linear models

to produce powerful out-of-sample predictions in both classification and regression settings. For

our models, we first establish the vocabulary by splitting text into bigrams after stopword removal,

of .933 and the DeKalb County vote shares receive a weight of .067. For cities that are contained within only one

county, the city’s estimated presidential election results are the county-level results.

10

tokenization, and lemmatization, and then using all bigrams13 that appear at least in three different

documents in the data. Once this is done, each document is represented by the TF-IDF (term

frequency inverse document frequency) reweighting of how many times each bigram appears (the

number of times that bigram appears in the document divided by the number of documents in

which that bigram appears at least once). We use the generic term document to refer to a single

instance depending on the unit of analysis (typically an individual’s entire set of tweets).

Topic representation Topic models are typically generative models, where the text generating

process is assumed to be that each document has a particular distribution over topics, and each

topic has a distribution over words (note that the language of documents and words is again

generic; in our case, documents are typically aggregated Twitter feeds and words are single words

as opposed to bigrams, as in the word vector representation above). A document is generated by

repeatedly first sampling a topic from its topic distribution, and then sampling a word from that

topic’s word distribution (note that this loses sequence information, but is consistent with the “bag

of words” style of document modeling). We use latent Dirichlet allocation (LDA)14 to learn topic

models.15 The LDA algorithm itself specifies and uses sparse Dirichlet priors over the topic and

13Bag-of-words feature only utilizes the frequency of words in documents. But it ignores the relative position

information between adjacent words, which will be preserved in bag-of-bigrams features. Actually, Wang and Man-

ning (2012) demonstrated that bag-of-bigrams feature improves the performance of classification models in multiple

sentiment analysis tasks.14Structural topic models (Roberts et al., 2013) have been widely applied in social science research problems. They

incorporate covariates (e.g. time or author) of the source text data in topic models. However, we are using the

representation in the topic space to create a useful measure of distance. In this case we do not want to allow the

model to condition on these covariates since they may contribute to the distances. Therefore we use a standard LDA

model, which only uses text data for training.15There has been some debate in the literature about the appropriateness of training LDA for short texts, such

as those found on Twitter. Some researchers have proposed different models that could be more appropriate in

certain settings; for example, Zhao et al. (2011) propose a Twitter-LDA model based on the assumption that each

tweet only belongs to exactly one topic. However, models like that require many topics in order to infer useful topic

distributions. Since our main task is to use topic models to infer distances between individuals, we require a smaller

number of topics, and preliminary analysis of tweets shows that they often cover multiple topics when the topic space

is constrained to be smaller (e.g. veterans affairs and healthcare). Further, LDA has successfully been applied in

11

word distributions, and then learns the distributions, typically using either variational methods or

Markov chain Monte Carlo sampling techniques.

For both types of representations of our text, when using them for prediction, we use regu-

larized logistic regression as our learning algorithm; while simple, this method (sometimes called

“maximum entropy”) has had tremendous success in text analytics. So, for example, we can view

the machine learning problem of partisanship or ideological lean prediction as follows. Given a

set of n instances of the form (xi, yi), where xi is the word vector or topic vector representation

of the document, and yi is either the partisan label (0 for Democrat and 1 for Republican) or a

real-valued variable (the first dimension of the DW-Nominate score), return a function h(x). For

the classification problem, we require h to return a real number between 0 and 1 that can be inter-

preted as the probability of being a Republican. We use L2 regularized logistic regression in this

case (thus h is the maximizer of the sum of the log likelihood and an L2 penalty on the weights).

Letting w be the weight vector representation of h, the objective function that is maximized is

then ln Πi Pr(yi|xi,w)− λ ‖w‖22.

5 Empirical Analyses

Our main empirical strategy is to use quantifications of the text produced by different types of

political actors in order to analyze how similar they are to reach other in terms of their Twitter

rhetoric. Our first set of analyses focuses on measuring differences between these actors in topic

space, while the second set leverages the ease or difficulty of predicting individual membership in

a particular group as a measure of how similar two groups are.

5.1 Topic differences: Mayors, Governors, and Members of Congress

To understand the degree to which mayors engage in nationalized rhetoric, we analyze content dif-

ferences between three different groups of politicians: mayors, governors, and Members of Congress.

If mayors and Members of Congress post similar content, this suggests that mayoral Twitter speech

numerous Twitter content and topic analysis tasks (Kang and Lerman, 2012; Vosecky, Leung and Ng, 2014; Kling

and Pozdnoukhov, 2012). Therefore, we employ a standard LDA approach.

12

is more nationalized – that is, mayoral content is more addressed to issues of the United States as

a whole and less specific to their constituents.

We start with a simple validity check on the plausibility of our hypotheses. We separately infer

topic models using latent Dirichlet allocation (Blei, Ng and Jordan, 2003) on Members of Congress’

tweets, governors’ tweets, and mayors’ tweets and visually inspect the words associated with all

topics. We list some of these for illustration in Table 2.16 We can see that, at least to a quick

human inspection, the congressional twitter topics are more “national” than mayors’. Mayoral

topics include terms such as “community, family, city, fire” and congressional topics include terms

such as “bipartisan, american, economy, health.” Gubernatorial topics more closely resemble those

of Members of Congress.17

We now turn to a more systematic investigation of topic differences. In order to establish

distances in topic space, we use the following methodology. We train 100 different topic models

(seeded with different random numbers) to account for the instability (Koltcov, Koltsova and

Nikolenko, 2014) of the LDA training process. Since some users post tweets with a higher frequency

than others, training the LDA model on our dataset directly will bias the model towards active

users.18 Therefore, in each round, we first randomly sub-sample 100 tweets from each of the

individual Members of Congress, governors, and mayors. We concatenate these 100 tweets into a

single document for each individual and then train an LDA model and use it to calculate a topic

distribution vector for each mayor, governor, and Member of Congress.19 Next, we calculate the

Jensen-Shannon divergence (Fuglede and Topsoe, 2004) between every pair of topic distribution

vectors in our sample, representing the differences in topics that show up in the content generated

by two individuals in our sample. The Jensen-Shannon divergence is a symmetrized and smoothed

version of the Kullback-Leibler divergence, and is commonly used to measure distances between

probability distributions.

16Using 25 topics.17Full word lists by topic are available in Table 7.3, Table 7.3, Table 7.3, Table 7.3, Table 7.3, and Table 7.3.18Directly comparing documents of different lengths is still an evolving area of research (Hongyu Gong and Xiong,

2018).19We use the gensim package for our LDA implementation (Rehurek and Sojka, 2010). We use 25 topics (based

on optimizing the cross-validated coherence score – see the appendix for more details). Each model is trained for 45

epochs. All other settings are the package defaults.

13

Table 2: Congress vs. Governor vs. Mayor tweet topics

Congress Governor Mayor

Rep

right, vote, health,senate, care, judge,

democrat, court, voted, ...

thank, service, woman,veteran, men, honor,

military, honored, grateful, ...

make, step, mile,fitbit, sure, better,

community, job, help, ...spending, trade, federal,

government, regulation, debt,role, rule, farmer, ...

tax, vote, cut,budget, property, get,

income, illinois, just, ...

meeting, town, hall,fire, enjoyed, tonight,

sterling, chamber, city, ...family, prayer, thought,victim, stand, attack,praying, lost, one, ...

state, flag, emergency,local, law, county,

stay, storm, weather, ...

county, school, mayor,ribbon, student, cutting,center, today, state, ...

business, small, honored,award, impact, receive,

support, job, economic, ...

business, great, good,meeting, small, leader,

news, economic, trade, ...

love, business, campaign,small, great, just,

people, big, new, ...glad, opioid, drug,fight, combat, see,

group, crisis, child, ...

texas, health, care,work, help, state,

effort, role, human, ...

thanks, work, great,public, well, community,

road, volunteer, congratulation, ...

Dem

work, make, gun,together, people, congress,

get, can, need, ...

law, signed, bill,opioid, legislation, help,take, prevent, today, ...

new, park, opening,downtown, center, summer,

grand, bike, project, ...health, care, access,

million, affordable, american,insurance, coverage, healthcare, ...

new, stand, trump,president, fight, administration,woman, fighting, washington, ...

art, food, west,local, tree, festival,early, music, tell, ...

bill, house, bipartisan,act, congress, senate,

legislation, pas, floor, ...

health, care, access,education, insurance, get,

coverage, early, affordable, ...

white, health, volunteer,care, plain, house,

word, court, ticket, ...tax, cut, gop,

american, republican, family,bill, middle, plan, ...

gun, violence, safety,domestic, must, delaware,letter, drilling, hand, ...

meeting, love, live,council, city, watch,

power, talk, mayor, ...job, worker, american,

economy, drug, company,cost, create, price, ...

justice, criminal, opening,full, show, reform,

schedule, discussion, step, ...

police, family, fire,officer, friend, chief,

life, thought, department, ...

For each pair of individuals in our dataset, the mean value of these 100 different distances is

then taken as the distance between their tweets in topic space. Figure 1 shows a histogram of

these differences, within- and between- different groups (color-coded appropriately). We observe

smaller differences within mayors (in green) and within Members of Congress (in blue) than we

do between mayors and Members of Congress (in red).20 That the average distance is greatest

between Members of Congress and mayors suggests that these groups are systematically focused

20Paired t-tests across each of these three pairs indicates these mean values are statistically distinguishable from

each other using 95% confidence intervals. Two-sample Kolmogorov-Smirnov test results also indicate that these

distributions are statistically distinguishable from each other using 95% confidence intervals.

14

on different topics. Importantly, we find that the distance between governors and Members of

Congress is smaller than the distance between mayors and Members of Congress. Substantively,

this indicates that governors are likely using more nationalized rhetoric than mayors.

While these results suggest that mayoral rhetoric is less nationalized than the rhetoric employed

by governors and by Members of Congress, the distances in topic distributions between mayors,

governors, and Members of Congress presented in Figure 1 do not allow us to claim that mayoral

rhetoric is devoid of nationalized political rhetoric. It is useful to compare with a baseline group

of Twitter users who could be considered generally apolitical. In order to do so, we compare

the differences in topic distributions between English Premier League (EPL) soccer players and

Members of Congress. 21 In order to verify that this group is actually apolitical, we collected all

the hashtags in soccer players’ tweets and checked the hashtags that occur more than 100 times in

our dataset. None of them are directly related to any specific U.S. politicians or political events.

The list of hashtags can be found in Appendix 7.2.

The distance comparison between this apolitical group, mayors and Members of Congress is

also shown in Figure 1. The soccer players are represented in the same topic space that is inferred

without using them as part of the training data, allowing for an “apples to apples” comparison

on the presumably political topic space. The topic distribution distance between our apolitical

group and Members of Congress (in purple) is the largest distance in topic distributions across all

of our pairwise examinations. That the distance between soccer players and Members of Congress

is greater than the distance between mayors and Members of Congress suggests that mayoral

rhetoric does, in fact, exhibit a degree of nationalization – of course, as the mayoral group and

the congressional group are both producing political speech. Yet again, note the ordering – the

relative differences between the topic distances of our apolitical group and Members of Congress

and mayors and Members of Congress suggests that mayoral rhetoric exhibits at most a modest

amount of nationalization compared to governors and Members of Congress.

To lend more credence to the idea that mayors and Members of Congress use different rhetoric,

21A recent report from Deloitte found that the English Premier League is the most lucrative and most watched

soccer league in the world. The full report can be accessed here: https://www2.deloitte.com/uk/en/pages/

sports-business-group/articles/annual-review-of-football-finance.html.

15

0.1 0.2 0.3 0.4 0.5 0.6Tweets' distances in topic model space

0.00

0.05

0.10

0.15

0.20

0.25

Freq

uenc

y

Cong. vs. Cong. (median=0.093)Mayor vs. Mayor (median=0.125)Cong. vs. Mayor (median=0.176)Governor vs. Governor (median=0.107)Cong. vs. Governor (median=0.123)Soccer Players vs. Soccer Players (median=0.137)Cong. vs. Soccer Players (median=0.242)

Figure 1: Histogram of topic distances.

we employ a test that asks whether we can reliably distinguish mayors from Members of Congress

based on their choices of topics about which to tweet. For each of the 100 rounds of the above LDA

analyses, we ran a 5-fold cross-validation test using logistic regression trained on the (joint) topic

distribution vectors, with the classification task being to distinguish between mayors and Members

of Congress (labeled 0 and 1, respectively). Cross-validation operates by randomly permuting the

data, separating it into k (in our case k = 5) “folds” and then repeatedly training a classifier on

k− 1 of the folds and testing on the remaining one. We find that the average area under the ROC

curve (AUC) over the 100 rounds of analyses is 0.9848±0.0004.22 Thus, the results of this analysis

22AUC is generally considered a more reliable way to evaluate classifiers than accuracy because of potential calibra-

tion issues. Through this paper, we apply Delong’s method to compute the standard deviation of the AUC (DeLong,

DeLong and Clarke-Pearson, 1988). We use the package provided by the Yandex School of Data Analysis, which

16

show that mayors and Members of Congress are easily distinguishable from each other based solely

on the distribution of topics about which they choose to tweet.

5.2 Variation in Nationalized Mayoral Rhetoric

While the preceding analyses suggest that gubernatorial rhetoric closely matches that of Members

of Congress, mayors appear to be using a different rhetorical style. However, these results do not

indicate that all mayors are similarly focused on local issues. In fact, it is possible that some mayors

use language more similar to congressional and gubernatorial rhetoric than others. Examining the

sources of variation in nationalized mayoral rhetoric is the task to which we now turn.

To more thoroughly examine the sources of variation in mayoral rhetoric, we created two mea-

sures that compare mayors to Members of Congress. First we computed partisanship scores for each

mayor (which can itself be done in two different ways); and, second, we measured the similarity

of mayors to Members of Congress. These two measures, which we used to examine variation in

nationalized mayoral rhetoric, are described below. In each case, “individual” refers to a mayor

or a Member of Congress, and “document” refers to the entire collected sample of an individual’s

tweets.

1. Partisanship scores We computed partisanship scores on top of the word vector represen-

tations of each document, one representing each individual.

• Mayor-internal partisanship (MIP) score. We used leave-one-out-cross validation on the

dataset of partisan mayors and their tweets in order to predict the partisanship of an individual

mayor. That is, to compute the MIP score of mayor i, we trained a logistic regression on the

tweet documents of each mayor other than i, labeled by their partisanship. We then predicted

the partisanship (probability of being a Republican) of mayor i using the logistic regression

model trained on the tweets of all other mayors.

• Congress-to-Mayor partisanship (CMP) score. In this case we trained a single regression

model of partisanship solely on the tweet documents of each Member of Congress, labeled by

implements a fast version of Delong’s method proposed by Sun and Xu (2014). Documentation is available at

https://github.com/yandexdataschool/roc_comparison.

17

the first dimension of their DW-Nominate score. We directly applied this model to each mayor

i to derive their CMP score.

2. Congressional similarity (CS) score The CS score is built on top of the topic vector rep-

resentation. An LDA model is trained on the union of congressional and mayoral tweet documents,

and each document is then represented by its distribution over topics. Each document is labeled as

either “Member of Congress” or “Mayor” (rather than by partisanship). Then, for each mayor i,

we trained a logistic regression classifier based on the documents corresponding to all other mayors

(labeled as 0) and Members of Congress (labeled as 1). We then predicted the probability that

mayor i is a Member of Congress using this leave-one-out model, and consider that prediction the

CS score.23

5.3 Partisanship: Mayors vs Congress

To begin our examination of the variation of the degree to which mayors resemble Members of

Congress we analyzed partisanship expression in mayoral tweets. First, we investigated whether it is

possible to correctly predict (out-of-sample) the party of a mayor based on their tweets. Restricting

ourselves to mayors, we ran a 5-fold cross-validation model on the word vector representation with

party identification as the target. The AUC is 0.7330 ± 0.0256. To calibrate that number, we

replicated this process using congressional tweets and found an AUC of 0.9938 ± 0.0027. Thus,

it appears that mayors express substantially less partisanship in their tweets relative to Members

of Congress. This is not to suggest that the expression of partisanship is completely absent from

mayors. However, our results suggest that mayoral partisanship is significantly harder to distinguish

23The following mayoral twitter ids have topic distributions most similar to Members of Congress: patownhall,

joshfryday, mayorcorymason, mike spence, andrewgillum. Andrew Gillum was the mayor of Tallahassee, Florida, and

was the Democratic candidate for Governor of Florida in 2018. As a candidate in a race that received significant

national attention, Gillum’s Twitter account frequently mentioned national issues and politicians including President

Trump. Those who are most different include: clvhtsgov, mayorhoye, mayornoak, bethlehemmayor, votemike4mayor.

This includes Smith Joseph, mayor of North Miami, Florida, and Manuel Lozano, mayor of Baldwin Park, California.

These accounts largely focus on local issues and events, such as community back to school fairs and announcements

about the retirement of city fire officials.

18

based on language used on social media than congressional partisanship.

It is possible that it is harder to distinguish language that is informative of partisanship from

mayoral tweets in part because the identification and use of language is less disciplined and, hence,

noisier. There is evidence that partisan language moves from Members of Congress to the media

(Yan et al., 2019) and can be used to assess the bias of media sources (Gentzkow and Shapiro, 2010)

as well as to quantify changes in polarization over time (Gentzkow, Shapiro and Taddy, 2019).

Similar phenomena might occur in social media, so we investigated the possibility of classifying

mayoral partisanship using congressional language. We first trained a linear regression model on

the Twitter records of Members of Congress, with the target being the first dimension of their

DW-Nominate score. We then used this model to estimate mayors’ partisanship using AUC (which

relies only on relative ranking, another advantage over accuracy). The AUC on the mayoral data

is 0.7796± 0.0250, which is better than the above 5-fold cross-validation experiment with mayors’

tweets. We call these scores the Congress-to-Mayor partisanship (CMP) scores. This is suggestive

that the nationalization of parties may be driving the increase in partisan speech.

We regressed mayors’ partisanship scores based on the above cross domain analyses and leave-

one-out cross-validation scores; this regression is shown graphically in Figure 2. While the scores

are clearly associated with each other, we rely upon the CMP scores in our next analysis.

We theorize that the mayors who discuss more similar topics to Members of Congress are more

likely to be classified as partisan via the CMP score (that is, when using the congressional partisan

language). To test this expectation, we created a measure of each mayor’s partisanship level.

Suppose mayor i’s CMP partisanship score is mi. Their partisanship level pmi based upon the k

co-partisan mayors is defined as:

pmi = |mi −median(mk)|

After creating this partisanship level measure for each mayor in our dataset, we regressed pmi

against our congressional similarity (CS) scores. The plot of this relationship is shown in Figure 3.

For mayoral tweets, there is a significant correlation between the topic similarities with congressional

tweets and partisanship level. The results of this section, taken together, indicate that there is a

19

0.30 0.35 0.40 0.45 0.50 0.55 0.60 0.65Mayor-internal partisanship (MIP) score

0.3

0.2

0.1

0.0

0.1

0.2

0.3

Cong

ress

-to-m

ayor

par

tisan

ship

(CM

P) sc

ore

Jorge Elorza

Andrew D. Gillum

Karen K. Goh

Jose A. AlvarezSlope = 0.718 ****

Figure 2: Congress-to-mayor partisanship (CMP) score vs Mayoral-internal partisanship (MIP)score.

significant correlation between expressions of partisanship – measured by the ease of predicting,

based on tweets, whether a mayor belongs to the Republican or Democratic Party – and the choice

of topics or issues one focuses on. We next turn to an analysis of the factors that are associated

with higher or lower levels of nationalized mayoral rhetoric.

5.4 Association between congressional similarity and covariates

The preceding analyses suggest that the nationalization of American politics has yet to reach the

mayoral level. However, as we have just demonstrated, the absence of nationalized rhetoric at

the mayoral level is not uniform. Indeed, based on our results, some mayors are likely engaged in

nationalized rhetoric at the expense of a more locally-focused representational style. To begin to un-

20

0.1 0.2 0.3 0.4 0.5 0.6 0.7Similarity score with members of Congress

0.0

0.1

0.2

0.3

0.4

Parti

sans

hip

leve

l

Cory Mason

Andrew D. GillumSlope = 0.171 ****

Figure 3: Mayoral congressional similarity scores vs. partisanship level.

derstand the sources of variation in mayoral rhetoric, we looked for associations between covariates

collected about mayors, their institutions, and the constituents they represent and the classifica-

tion of mayors vs Members of Congress using topic vectors (the congressional similarity score) or

mayoral partisanship level (the difference from the mean congress-to-mayor partisanship score for

co-partisans). Though these relationships are not necessarily causal in nature, they nevertheless

will help us to better understand the sources of variation in nationalized mayoral rhetoric.

To reiterate our theory from Section 2, we expect more nationalized rhetoric from mayors who

have had larger margins of victory or served longer terms in office. We also expect more nationalized

rhetoric from mayors who represent larger cities. Finally, we expect more nationalized rhetoric from

mayors who are less-involved in the day-to-day operations of their municipality – those who serve

21

under a council mayor system or have city managers.

We use our similarity to Congress as a metric of nationalized rhetoric. The greater the similarity

to Congress, the greater the likelihood that the mayor has adopted nationalized political speech.

We regressed this similarity measure on indicator variables for institutions (city manager, mayor-

council system, etc.), the total number of years the mayor has served in office, the city population

(re-scaled in units of 100,000), and the vote share won by the mayor in the previous election.24

We present models with and without additional covariates – we can also control for mayoral race,

gender, party, and Democratic vote share in the 2016 election – in Table 3.

Contrary to our theoretical expectation, we find that the number of years a mayor has served in

office is negatively associated with nationalized rhetoric. This suggests that longer-serving mayors

are more focused on their citizens’ local needs than their counterparts who have less experience

in office. Additionally, we find a positive relationship between our city population variable and

nationalized rhetoric. This indicates that, as city population increases, the similarity of mayoral

speech more closely mirrors that of a congressional representative. This lends support to our

theoretical expectation that, because citizens of larger cities are less likely to experience any form

of personal contact with their mayor, mayors can adopt a more nationalized representational style.

We find no statistically significant relationships between employing a city manager or operating

under a council-manager system and nationalized rhetoric.

While the results in this table should be interpreted with caution, as we are limited to a small set

of mayors, it is encouraging that we observe similar results when including additional controls for

race, gender, party and Democratic vote share in the 2016 election in the second column of Table 3.

Within this set of variables, the only meaningful associations are between gender, partisanship, and

congressional similarity.

24In cases where a city’s mayor is appointed to the position by the city council, we treat the mayor’s vote share in

the previous election as the percentage of the vote they received in their most recent election to the city council.

22

Table 3: Coefficients for Congressional Similarity

Covariates Coefficients Coefficients

Years in office -.003* -.003*(.001) (.001)

City population .008* .007*(.002) (.002)

Previous vote share -.00 -.00(.00) (.00)

Council-Manager System -.036 -.029(.019) (.019)

City Manager .026 .027(.019) (.019)

Democratic Pres Share 2016 .00(.00)

White .006(.013)

Male .039*(.012)

Republican -.024*(.011)

Constant .261* .213*(.019) (.031)

N 361 360R2 .081 .126

23

6 Conclusion & Discussion

Among the deleterious consequences of the nationalization of American politics is the concern

that elites’ focus on national issues comes at the expense of attending to more pressing, locally-

based problems. The results we have presented suggest that, while there is variation, mayors

have largely remained focused on local issues and use more apolitical speech than governors and

Members of Congress. The degree to which citizens’ preferences and needs are represented by their

elected officials is an important barometer of the health of a democratic society. In the case of the

United States, which has experienced a considerable amount of political nationalization over the

past few decades (Hopkins, 2018), this metric has trended in a negative direction at the federal

and gubernatorial levels. However, our analyses suggest that there is room for optimism about the

vibrancy of representation provided by mayors. Unlike elected officials at other levels of government,

the average American mayor currently appears to be unaffected by the nationalization of American

politics. On the contrary, mayoral rhetoric tends to remain focused on the needs and concerns of

the local citizenry. This finding persists even when controlling for a multitude of mayoral- and

city-level covariates.

However, our results also suggest that mayors do not uniformly focus on local issues. Na-

tionalized mayoral rhetoric is more likely to occur among mayors who govern cities with a large

population. Such a finding likely arises because a large municipal population diminishes the likeli-

hood that any citizen has a personal experience or point of contact with her mayor. As a result,

mayors are freer to engage in national rhetoric and abjure local issues. Though plausible, such an

explanation is only one potential reason that we observe a relationship between city size and na-

tionalized mayoral rhetoric. Future work should more thoroughly examine the mechanism linking

a city’s population to mayoral representational style.

One final concern pertains to the durability of these results. The nationalization of American

elections did not suddenly develop and alter congressional and gubernatorial behavior. In reality,

the nationalization of American elections has been a secular process that has resulted from the

complex interplay of ideological realignment, partisan sorting, and a changing media environment.

Whether these trends in nationalization will eventually affect mayors, or whether Americans’ most

24

immediate chief executives will remain focused on their local mandate, is a question that can only

be answered in time.

References

Abramowitz, Alan I. and Steven W. Webster. 2016. “The Rise of Negative Partisanship and the

Nationalization of U.S. Elections in the 21st Century.” Electoral Studies 41:12–22.

Adrian, Charles R. 1952. “Some General Characteristics of Nonpartisan Elections.” American

Political Science Review 46(3):766–776.

Ahmed, Amr and Eric P Xing. 2010. Staying informed: Supervised and semi-supervised multi-view

topical analysis of ideological perspective. In Proceedings of the 2010 Conference on Empirical

Methods in Natural Language Processing. pp. 1140–1150.

Aleman, Eduardo and Marisa Kellam. 2008. “The Nationalization of Electoral Change in the

Americas.” Electoral Studies 27(2):193–212.

Blei, David M, Andrew Y Ng and Michael I Jordan. 2003. “Latent dirichlet allocation.” Journal of

machine Learning research 3(Jan):993–1022.

Bouma, Gerlof. 2009. “Normalized (pointwise) mutual information in collocation extraction.” Pro-

ceedings of GSCL pp. 31–40.

Calfano, Brian Robert and Paul A. Djupe. 2009. “God Talk: Religious Cues and Electoral Support.”

Political Research Quarterly 62(2):329–339.

Carson, Jamie L and Joel Sievert. 2018. Electoral Incentives in Congress. University of Michigan

Press.

Chang, Jonathan, Sean Gerrish, Chong Wang, Jordan L. Boyd-graber and David M. Blei. 2009.

Reading Tea Leaves: How Humans Interpret Topic Models. In Advances in Neural Information

Processing Systems 22. pp. 288–296.

25

Choi, Sang Ok, Bae Sang-Seok, Sung-Wook Kwon and Richard C. Feiock. 2010. “County Limits:

Policy Types and Expenditure Priorities.” American Review of Public Administration 40(1):29–

45.

de Benedictis-Kessner, Justin and Christopher Warshaw. 2016. “Mayoral Partisanship and Munic-

ipal Fiscal Policy.” The Journal of Politics 78(4):1124–1138.

de Benedictis-Kessner, Justin and Christopher Warshaw. 2018. “Politics in Forgotten Govern-

ments: The Partisan Composition of County Legislatures and County Fiscal Policies.” Forthcom-

ing at The Journal of Politics. Available at https://www.dropbox.com/s/9tcinbnllhikbit/

CountyCouncils181121_public.pdf?dl=0.

DeLong, Elizabeth R, David M DeLong and Daniel L Clarke-Pearson. 1988. “Comparing the

areas under two or more correlated receiver operating characteristic curves: a nonparametric

approach.” Biometrics 44(3):837–845.

Downs, Anthony. 1957. An Economic Theory of Democracy. Harper.

Einstein, Katherine Levine and David M. Glick. 2018. “Mayors, Partisanship, and Redistribution:

Evidence Directly from US Mayors.” Urban Affairs Review 54(1):74–106.

Einstein, Katherine Levine and Vladimir Kogan. 2015. “Pushing the City Limits: Policy Respon-

siveness in Municipal Government.” Urban Affairs Review 52(1):3–32.

Esaiasson, Peter and Christopher Wlezien. 2016. “Advances in the Study of Democratic Respon-

siveness: An Introduction.” Comparative Political Studies 50(6):699–710.

Fearon, James D. 1999. Electoral Accountability and the Control of Politicians: Selecting Good

Types Versus Sanctioning Poor Performance. In Democracy, Accountability, and Representation.

Cambridge University Press pp. 55–97.

Fenno, Richard F. 1978. Home Style: House Members in their Districts. Pearson College Division.

Ferejohn, John. 1986. “Incumbent Performance and Electoral Control.” Public Choice 50(1):5–25.

26

Ferreira, Fernando and Joseph Gyourko. 2009. “Do Political Parties Matter? Evidence from U.S.

Cities.” Quarterly Journal of Economics 124(1):349–397.

Fiorina, Morris P. 1981. Retrospective Voting in American National Elections. Yale University

Press.

Fouirnaies, Alexander and Andrew B. Hall. 2018. “How Do Electoral Incentives Affect Legislator

Behavior?” Working paper. Available at http://www.andrewbenjaminhall.com/Fouirnaies_

Hall_Electoral_Incentives.pdf.

Fuglede, Bent and Flemming Topsoe. 2004. Jensen-Shannon divergence and Hilbert space em-

bedding. In Proceedings of the International Symposium on Information Theory (ISIT 2004).

p. 31.

Gentzkow, Matthew and Jesse M Shapiro. 2010. “What drives media slant? Evidence from U.S.

daily newspapers.” Econometrica 78(1):35–71.

Gentzkow, Matthew, Jesse M Shapiro and Matt Taddy. 2019. “Measuring group differences in

high-dimensional choices: Method and application to Congressional speech.”. Forthcoming at

Econometrica.

Gerber, Elisabeth R. and Daniel J. Hopkins. 2011. “When Mayors Matter: Estimating the Impact

of Mayoral Partisanship on City Policy.” American Journal of Political Science 52(2):326–339.

Grimmer, J. and B. M. Stewart. 2013. “Text as data: The promise and pitfalls of automatic content

analysis methods for political texts.” Political Analysis pp. 1–31.

Grimmer, Justin. 2013. Representational Style in Congress: What Legislators Say and Why It

Matters. Cambridge University Press.

Hamlin, Alan and Colin Jennings. 2011. “Expressive Political Behaviour: Foundations, Scope and

Implications.” British Journal of Political Science 41(3):645–670.

27

Hongyu Gong, Tarek Sakakini, Suma Bhat and Jinjun Xiong. 2018. Document Similarity for

Texts of Varying Lengths via Hidden Topics. In Proceedings of the 56th Annual Meeting of the

Association for Computational Linguistics (Volume 1: Long Papers). pp. 2341–2351.

Hopkins, Daniel J. 2008. “The Origins of the Urban Political Agenda.” Working paper. Presented

at the 2008 Annual Meeting of the American Political Science Association.

Hopkins, Daniel J. 2018. The Increasingly United States: How and Why American Political Behav-

ior Nationalized. University of Chicago Press.

Hopkins, Daniel J. and Eric Schickler. 2016. “The Nationalization of State Party Platforms, 1918-

2014.” Working paper. Presented at the 2016 Annual Meeting of the American Political Science

Association.

Hopkins, Daniel J. and Lindsay M. Pettingill. 2017. “Retrospective Voting in Big-City US Mayoral

Elections.” Political Science Research & Methods 6(4):697–714.

Iyyer, Mohit, Peter Enns, Jordan Boyd-Graber and Philip Resnik. 2014. Political ideology detection

using recursive neural networks. In Proceedings of the 52nd Annual Meeting of the Association

for Computational Linguistics. Vol. 1 pp. 1113–1122.

Jacobson, Gary C. 2015. “It’s Nothing Personal: The Decline of the Incumbency Advantage in US

House Elections.” The Journal of Politics 77(3):861–873.

Jensen, Amalie, William Marble, Kenneth Scheve and Matthew J. Slaughter. 2019. “City

Limits to Partisan Polarization in the American Public.” Working paper. Available http:

//williammarble.co/docs/CityLimits-Mar2019.pdf.

Kang, Jeon-Hyung and Kristina Lerman. 2012. Using lists to measure homophily on Twitter. In

Workshops at the Twenty-Sixth AAAI Conference on Artificial Intelligence.

Kling, Felix and Alexei Pozdnoukhov. 2012. When a city tells a story: Urban topic analysis. In

Proceedings of the 20th international conference on advances in geographic information systems.

ACM pp. 482–485.

28

Koltcov, Sergei, Olessia Koltsova and Sergey Nikolenko. 2014. Latent dirichlet allocation: stability

and applications to studies of user-generated content. In Proceedings of the 2014 ACM conference

on Web science. pp. 161–165.

Lawrence, Regina G;, Logan Molyneux, Mark Coddington and Avery Holton. 2014. “Tweeting

Conventions: Political Journalists’ Use of Twitter to Cover the 2012 Presidential Campaign.”

Journalism Studies 15(6):789–806.

Lin, Wei-Hao, Eric Xing and Alexander Hauptmann. 2008. A joint topic and perspective model

for ideological discourse. In Joint European Conference on Machine Learning and Knowledge

Discovery in Databases. pp. 17–32.

Martin, Gregory J. and Joshua McCrain. 2019. “Local News and National Politics.” American

Political Science Review 113(2):372384.

Mayhew, David R. 1974. Congress: The Electoral Connection. Yale University Press.

McGregor, Shannon C. and Logan Molyneux. 2018. “Twitters Influence on News Judgment: An

Experiment Among Journalists.” Journalism .

Mimno, David, Hanna M Wallach, Edmund Talley, Miriam Leenders and Andrew McCallum. 2011.

Optimizing semantic coherence in topic models. In Proceedings of the conference on empirical

methods in natural language processing. pp. 262–272.

Oliver, J. Eric. 2012. Local Elections and the Politics of Small-Scale Democracy. Princeton Uni-

versity Press.

Percival, Garrick L., Martin Johnson and Max Neiman. 2009. “Representation and Local Pol-

icy: Relating County-Level Public Opinion to Policy Outputs.” Political Research Quarterly

62(1):164–177.

Rehurek, Radim and Petr Sojka. 2010. Software Framework for Topic Modelling with Large Cor-

pora. In Proceedings of the LREC 2010 Workshop on New Challenges for NLP Frameworks.

ELRA pp. 45–50.

29

Roberts, Margaret E, Brandon M Stewart, Dustin Tingley, Edoardo M Airoldi et al. 2013. The

structural topic model and applied social science. In Advances in neural information process-

ing systems workshop on topic models: computation, application, and evaluation. Harrahs and

Harveys, Lake Tahoe pp. 1–20.

Roder, Michael, Andreas Both and Alexander Hinneburg. 2015. Exploring the space of topic

coherence measures. In Proceedings of the eighth ACM international conference on Web search

and data mining. ACM pp. 399–408.

Shapiro, Walter and Jill Lawrence. 2015. “Governors and the Presidency: How They Campaign,

How They Govern.” Report at the Center on the American Governor. Available at http://

governors.rutgers.edu/testing/wp-content/uploads/2015/08/JLWS_Digital_Final.pdf.

Shor, Boris and Nolan McCarty. 2011. “The Ideological Mapping of American Legislatures.” Amer-

ican Political Science Review 105(3):530–551.

Sievert, Joel and Seth C. McKee. 2018. “Nationalization in US Senate and Gubernatorial Elections.”

American Politics Research .

Stevens, Keith, Philip Kegelmeyer, David Andrzejewski and David Buttler. 2012. Exploring topic

coherence over many models and many topics. In Proceedings of the 2012 Joint Conference

on Empirical Methods in Natural Language Processing and Computational Natural Language

Learning. pp. 952–961.

Sun, Xu and Weichao Xu. 2014. “Fast implementation of DeLongs algorithm for comparing the

areas under correlated receiver operating characteristic curves.” IEEE Signal Processing Letters

21(11):1389–1393.

Tausanovitch, Chris and Christopher Warshaw. 2014. “Representation in Municipal Government.”

American Political Science Review 108(3):605–641.

Verba, Sidney, Kay Lehman Schlozmann and Henry E Brady. 1995. Voice and Equality: Civic

Voluntarism in American Politics. Harvard University Press.

30

Vosecky, Jan, Kenneth Wai-Ting Leung and Wilfred Ng. 2014. Collaborative personalized twit-

ter search with topic-language models. In Proceedings of the 37th international ACM SIGIR

conference on Research & development in information retrieval. ACM pp. 53–62.

Wang, Sida and Christopher D Manning. 2012. Baselines and bigrams: Simple, good sentiment and

topic classification. In Proceedings of the 50th annual meeting of the association for computational

linguistics: Short papers-volume 2. Association for Computational Linguistics pp. 90–94.

Yan, Hao, Sanmay Das, Allen Lavoie, Sirui Li and Betsy Sinclair. 2019. The Congressional Classi-

fication Challenge: Domain Specificity and Partisan Intensity. In Proceedings of the 2019 ACM

Conference on Economics and Computation. pp. 71–89.

Ybarra, Vickie D. and Timothy B. Krebs. 2010. “Policy Responsiveness in Local Government:

Adoption of Smoke-free Policies in U.S. Counties.” State and Local Government Review 48(1):6–

20.

Zhao, Wayne Xin, Jing Jiang, Jianshu Weng, Jing He, Ee-Peng Lim, Hongfei Yan and Xiaoming

Li. 2011. Comparing Twitter and traditional media using topic models. In European conference

on information retrieval. Springer pp. 338–349.

31

7 Appendix

7.1 Optimal Number of Topics

The number of topics k is an important parameter for LDA model performance. If k is too small,

words that belong to different subjects will be forced into the same topic. If k is too large, on the

other hand, there will be near-duplicate topics with overlapping words in the LDA model. In either

case, the model will not characterize the topic distribution well.

To measure the quality of a topic model, scholars have proposed multiple measurement methods.

In their original LDA paper, Blei, Ng and Jordan (2003) indicated that one could use a held-out test

set and compute its perplexity score as one way to evaluate a language model. However, Chang

et al. (2009) argue that likelihood based methods, such as perplexity, are not well aligned with

human judgment.

Topic coherence measurement (Bouma, 2009; Mimno et al., 2011; Stevens et al., 2012), which

calculates the coherence level of the top n words in each topic, is considered a reliable method that

improves the consistency between language model evaluation scores and human judgment (Roder,

Both and Hinneburg, 2015). To obtain the optimal number of topics k in our model, we use two

coherence measurement models: CV and CNPMI (Roder, Both and Hinneburg, 2015). As shown

in the benchmarking test in this paper, topic coherence ranked by these two models has strong

correlations with human ratings.

In order to find the optimal number of topics according to the coherence measure, we first

perform an incremental search to determine part of the range: we measure coherence when the

number of topics ki = 2i, for i ∈ {1, 2, . . . , 9}. For each ki, we first train an LDA model following

the same procedure as in Sec. 5.1. Then we evaluate the model with both CV and CNPMI as

measures of coherence. We repeat this process 20 times with different random seeds and take the

mean value as the final coherence evaluation scores for ki.

We use the Gensim package for our coherence test with both CV and CNPMI scores. We follow

the settings in the original paper (Roder, Both and Hinneburg, 2015) such that we only consider

the top-10 words in all topics for evaluation of coherence scores. We set the sliding window size as

32

50, and all other settings are the defaults in the Gensim package.

As we can see in Figure 4, the coherence score results based on both evaluation metrics CV

and CNPMI consistently indicate that the optimal number of topics is between 8 and 64. Then we

search within that space, testing ki = 2 ∗ i, for i = {4, 5, . . . , 24} (there is a clear tapering down of

performance at higher values, so we stop at 48 topics). The final grid search results are shown in

Figure 5. As we can see, [24, 26] gives the highest coherence scores. Therefore, we choose k = 25

as the number of topics throughout our experiments in this paper.

0 100 200 300 400 500Number of Topics

0.10

0.15

0.20

0.25

0.30

0.35

0.40

Cohe

renc

e Sc

ores

CV

(a) Based on CV scores

0 100 200 300 400 500Number of Topics

0.15

0.10

0.05

0.00

0.05

Cohe

renc

e Sc

ores

CNPMI

(b) Based on CNPMI scores

Figure 4: Number of Topics vs. Coherence Scores figures based on two coherence measurements –incremental search.

10 15 20 25 30 35 40 45 50Number of Topics

0.34

0.36

0.38

0.40

0.42

0.44

Cohe

renc

e Sc

ores

CV

(a) Based on CV scores

10 15 20 25 30 35 40 45 50Number of Topics

0.00

0.01

0.02

0.03

0.04

0.05

Cohe

renc

e Sc

ores

CNPMI

(b) Based on CNPMI scores

Figure 5: Number of Topics vs. Coherence Scores figures based on two coherence measurements –grid search.

33

7.2 List of Soccer Player Twitter Hashtags

After we exclude two garbled hashtags, the hashtags that occur more than 100 times in our soc-

cer players Twitter dataset are: #COYS, #CFC, #YNWA, #COYG, #COYI, #UCL, #MUFC,

#saintsfc, #mufc, #cfc, #ForzaJuve, #AFC, #CPFC, #LS19, #lcfc, #LFC, #PremierLeague,

#FACup, #YaGunnersYa, #MCFC, #WorldCup, #NoFuchsGiven, #teamkaching, #BeTheDif-

ference, #COYB, #preseason, #THFC, #LCFC, #cpfc, #premierleague, #coys, #W22, #legend,

#BHAFC, #ChampionsLeague, #mcfc, #Belgium, #WeMarchOn, #somosporto, #KTBFFH,

#boruc, #family, #EURO2016, #RJ9, #training, #aubameyang, #3points, #Repost, #mancity,

#FinoAllaFine, #WHUFC, #vamoss, #EFC, #GodIsGreat, #Arsenal, #nufc, #London, #Eng-

land, #comeonchelsea, #TBT, #SomosPorto, #SaintsFC, #afcb, #manchestercity, #bvb, #Three-

Lions, #UEL, #buzzing, #ManCity, #TeamOM, #blessed, #together, #CmonCity, #tbt, #Deusno-

controle, #matchday, #HereToCreate, #UmPassoDeCadaVez, #DieMannschaft, #behappy, #Ha-

laMadrid, #geezers, #nike, #Legend, #love, #, #bhafc, #FirstNeverFollows.

34

7.3 Complete Word Lists By Topic

Table 4: Full Word List: Congress – Republican

Congressional Word List (Rep.)

0.022*“local” + 0.020*“north” + 0.016*“community” + 0.014*“center” + 0.014*“county”

+ 0.013*“water” + 0.012*“south” + 0.011*“grant” + 0.011*“help” + 0.011*“project” +

0.010*“state” + 0.009*“rural” + 0.009*“official” + 0.008*“new” + 0.008*“disaster”

0.075*“house” + 0.026*“senate” + 0.024*“bill” + 0.023*“budget” + 0.021*“read”

+ 0.019*“committee” + 0.018*“congress” + 0.015*“colleague” + 0.015*“letter” +

0.012*“vote” + 0.011*“today” + 0.011*“passed” + 0.010*“white” + 0.010*“repeal” +

0.010*“republican”

0.077*“great” + 0.030*“meeting” + 0.025*“enjoyed” + 0.020*“today” + 0.016*“discus”

+ 0.014*“met” + 0.014*“visit” + 0.014*“wa” + 0.013*“issue” + 0.013*“morning” +

0.012*“meet” + 0.012*“county” + 0.012*“visiting” + 0.012*“time” + 0.010*“discussion”

0.037*“right” + 0.030*“vote” + 0.026*“health” + 0.020*“senate” + 0.018*“care” +

0.016*“judge” + 0.016*“democrat” + 0.015*“court” + 0.015*“voted” + 0.012*“vot-

ing” + 0.011*“conservative” + 0.010*“supreme” + 0.009*“ha” + 0.009*“amendment” +

0.009*“case”

0.056*“join” + 0.035*“take” + 0.032*“please” + 0.023*“getting” + 0.021*“share” +

0.020*“ready” + 0.018*“story” + 0.018*“washington” + 0.017*“win” + 0.017*“big” +

0.015*“day” + 0.015*“week” + 0.011*“early” + 0.010*“call” + 0.010*“moment”

0.042*“honor” + 0.029*“woman” + 0.028*“service” + 0.026*“veteran” + 0.021*“today”

+ 0.020*“thank” + 0.019*“men” + 0.016*“wa” + 0.016*“remember” + 0.015*“life” +

0.011*“year” + 0.010*“sacrifice” + 0.010*“nation” + 0.010*“american” + 0.010*“hero”

0.036*“student” + 0.031*“congrats” + 0.028*“school” + 0.027*“congressional” +

0.022*“high” + 0.019*“great” + 0.017*“dc” + 0.017*“young” + 0.016*“wa” +

0.013*“team” + 0.012*“congratulation” + 0.012*“art” + 0.011*“academy” + 0.010*“meet”

+ 0.010*“proud”

35

0.053*“new” + 0.039*“check” + 0.032*“show” + 0.029*“welcome” + 0.029*“sign” +

0.025*“wa” + 0.023*“latest” + 0.022*“update” + 0.018*“report” + 0.018*“weekly” +

0.017*“follow” + 0.013*“miss” + 0.011*“read” + 0.009*“last” + 0.009*“newsletter”

0.060*“tax” + 0.035*“job” + 0.028*“american” + 0.017*“cut” + 0.017*“reform” +

0.014*“plan” + 0.012*“million” + 0.012*“economy” + 0.011*“cost” + 0.010*“family” +

0.009*“new” + 0.008*“health” + 0.008*“ha” + 0.008*“rate” + 0.007*“obamacare”

0.036*“president” + 0.024*“obama” + 0.020*“ha” + 0.014*“american” + 0.012*“must” +

0.012*“trump” + 0.012*“administration” + 0.010*“iran” + 0.008*“action” + 0.008*“nu-

clear” + 0.008*“wa” + 0.008*“deal” + 0.008*“say” + 0.007*“threat” + 0.007*“u”

’0.024*“spending” + 0.020*“trade” + 0.020*“federal” + 0.018*“government” + 0.017*“reg-

ulation” + 0.015*“debt” + 0.015*“role” + 0.014*“rule” + 0.014*“farmer” + 0.011*“epa”

+ 0.010*“play” + 0.009*“without” + 0.009*“fed” + 0.008*“govt” + 0.008*“heard”

’0.103*“good” + 0.026*“make” + 0.022*“news” + 0.016*“luck” + 0.016*“sure” +

0.015*“best” + 0.013*“stay” + 0.011*“little” + 0.011*“see” + 0.010*“got” + 0.010*“word”

+ 0.009*“thing” + 0.009*“just” + 0.008*“even” + 0.008*“lot”

’0.058*“family” + 0.039*“prayer” + 0.034*“thought” + 0.019*“victim” + 0.019*“stand”

+ 0.017*“attack” + 0.016*“praying” + 0.012*“lost” + 0.012*“one” + 0.011*“wa” +

0.010*“loved” + 0.010*“hear” + 0.010*“heart” + 0.010*“friend” + 0.009*“entire”

’0.054*“watch” + 0.053*“hearing” + 0.052*“tune” + 0.047*“live” + 0.035*“discus” +

0.024*“talk” + 0.023*“morning” + 0.021*“joining” + 0.020*“listen” + 0.020*“floor” +

0.020*“today” + 0.018*“talking” + 0.015*“now” + 0.014*“interview” + 0.012*“speaking”

’0.093*“thank” + 0.076*“first” + 0.049*“rt” + 0.035*“state” + 0.021*“united” +

0.017*“wa” + 0.017*“former” + 0.017*“john” + 0.016*“president” + 0.013*“capitol” +

0.012*“senator” + 0.009*“leadership” + 0.009*“u” + 0.009*“responder” + 0.009*“george”

’0.044*“work” + 0.041*“can” + 0.035*“get” + 0.027*“need” + 0.023*“know” +

0.021*“want” + 0.020*“people” + 0.019*“help” + 0.015*“make” + 0.013*“hard” +

0.013*“working” + 0.013*“american” + 0.012*“let” + 0.011*“find” + 0.011*“together”

36

’0.092*“happy” + 0.042*“year” + 0.040*“day” + 0.027*“birthday” + 0.021*“congrat-

ulation” + 0.015*“wishing” + 0.015*“family” + 0.015*“celebrate” + 0.014*“wish” +

0.013*“friend” + 0.013*“today” + 0.013*“great” + 0.013*“celebrating” + 0.011*“one” +

0.010*“new”

’0.024*“border” + 0.021*“immigration” + 0.016*“photo” + 0.015*“u” + 0.014*“illegal” +

0.011*“facebook” + 0.010*“citizen” + 0.010*“security” + 0.010*“page” + 0.009*“clinton”

+ 0.009*“criminal” + 0.008*“secure” + 0.008*“gun” + 0.007*“system” + 0.007*“park”

’0.053*“bill” + 0.028*“act” + 0.023*“passed” + 0.021*“legislation” + 0.021*“support” +

0.019*“proud” + 0.017*“bipartisan” + 0.017*“house” + 0.017*“help” + 0.013*“care” +

0.012*“access” + 0.011*“va” + 0.011*“introduced” + 0.010*“health” + 0.010*“today”

’0.072*“business” + 0.050*“small” + 0.032*“honored” + 0.020*“award” + 0.020*“impact”

+ 0.017*“receive” + 0.014*“support” + 0.014*“job” + 0.012*“economic” + 0.012*“new” +

0.011*“proud” + 0.011*“owner” + 0.010*“local” + 0.010*“manufacturing” + 0.009*“great”

’0.058*“office” + 0.034*“town” + 0.034*“staff” + 0.032*“statement” + 0.031*“hall” +

0.021*“hour” + 0.016*“full” + 0.015*“tomorrow” + 0.015*“question” + 0.014*“stop” +

0.014*“hosting” + 0.012*“county” + 0.012*“holding” + 0.012*“mobile” + 0.012*“congress-

man”

’0.069*“keep” + 0.048*“every” + 0.041*“go” + 0.034*“law” + 0.023*“enforcement” +

0.022*“officer” + 0.022*“america” + 0.022*“life” + 0.017*“line” + 0.017*“protect” +

0.016*“safe” + 0.016*“put” + 0.016*“police” + 0.014*“keeping” + 0.014*“u”

’0.169*“thanks” + 0.078*“forward” + 0.057*“look” + 0.049*“looking” + 0.033*“great”

+ 0.029*“coming” + 0.027*“working” + 0.023*“everyone” + 0.017*“stopping” +

0.016*“came” + 0.011*“seeing” + 0.010*“sharing” + 0.009*“standing” + 0.009*“appre-

ciate” + 0.008*“see”

’0.038*“energy” + 0.015*“general” + 0.011*“create” + 0.011*“new” + 0.010*“ag” +

0.009*“job” + 0.009*“gas” + 0.008*“oil” + 0.008*“collins” + 0.008*“army” + 0.007*“tech”

+ 0.007*“technology” + 0.007*“via” + 0.007*“research” + 0.007*“innovation”

37

’0.061*“glad” + 0.026*“opioid” + 0.025*“drug” + 0.023*“fight” + 0.019*“combat” +

0.019*“see” + 0.017*“group” + 0.017*“crisis” + 0.015*“child” + 0.014*“abuse” +

0.012*“help” + 0.010*“end” + 0.010*“effort” + 0.009*“epidemic” + 0.008*“prescription”

Table 5: Full Word List: Congress – Democratic

Congressional Word List (Dem.)

0.042*“work” + 0.038*“make” + 0.025*“gun” + 0.020*“together” + 0.020*“people”

+ 0.018*“congress” + 0.017*“get” + 0.017*“can” + 0.017*“need” + 0.015*“hard” +

0.014*“sure” + 0.014*“keep” + 0.011*“come” + 0.011*“working” + 0.010*“must”

0.069*“great” + 0.027*“thanks” + 0.021*“school” + 0.020*“join” + 0.019*“see” +

0.016*“student” + 0.015*“town” + 0.015*“meeting” + 0.013*“hall” + 0.011*“morning”

+ 0.011*“high” + 0.010*“today” + 0.009*“event” + 0.009*“annual” + 0.007*“hosting”

0.098*“wa” + 0.033*“forward” + 0.025*“look” + 0.021*“looking” + 0.016*“year” +

0.016*“first” + 0.010*“last” + 0.009*“welcome” + 0.009*“one” + 0.009*“new” +

0.009*“great” + 0.008*“story” + 0.008*“honor” + 0.008*“today” + 0.008*“photo”

0.035*“know” + 0.027*“want” + 0.024*“go” + 0.024*“let” + 0.020*“day” + 0.019*“open”

+ 0.016*“get” + 0.016*“dont” + 0.014*“sign” + 0.014*“every” + 0.014*“can” +

0.013*“just” + 0.013*“time” + 0.012*“u” + 0.011*“american”

0.059*“rt” + 0.046*“change” + 0.032*“can” + 0.029*“climate” + 0.026*“white” +

0.020*“need” + 0.020*“follow” + 0.019*“key” + 0.017*“safety” + 0.012*“speech” +

0.012*“house” + 0.012*“count” + 0.010*“stand” + 0.009*“net” + 0.009*“program”

0.106*“health” + 0.068*“care” + 0.033*“access” + 0.031*“million” + 0.019*“affordable”

+ 0.018*“american” + 0.018*“insurance” + 0.016*“coverage” + 0.014*“healthcare” +

0.010*“people” + 0.010*“family” + 0.008*“quality” + 0.008*“lose” + 0.008*“woman” +

0.008*“condition”

38

0.053*“woman” + 0.026*“american” + 0.022*“better” + 0.021*“future” + 0.014*“must”

+ 0.013*“every” + 0.012*“work” + 0.011*“pay” + 0.010*“equal” + 0.010*“america” +

0.009*“make” + 0.009*“generation” + 0.009*“u” + 0.009*“time” + 0.009*“economic”

0.031*“child” + 0.029*“policy” + 0.025*“trump” + 0.025*“family” + 0.021*“adminis-

tration” + 0.013*“must” + 0.011*“immigrant” + 0.011*“immigration” + 0.011*“parent”

+ 0.011*“border” + 0.010*“ha” + 0.007*“end” + 0.007*“president” + 0.006*“crisis” +

0.006*“cruel”

0.064*“proud” + 0.042*“support” + 0.034*“veteran” + 0.034*“good” + 0.028*“honored”

+ 0.014*“va” + 0.013*“join” + 0.012*“service” + 0.012*“congrats” + 0.011*“receive” +

0.011*“work” + 0.011*“award” + 0.010*“vet” + 0.010*“serve” + 0.009*“community”

0.021*“discus” + 0.019*“local” + 0.019*“business” + 0.016*“met” + 0.016*“new” +

0.016*“help” + 0.013*“community” + 0.013*“small” + 0.012*“today” + 0.010*“federal”

+ 0.009*“grant” + 0.009*“support” + 0.008*“program” + 0.008*“issue” + 0.007*“infras-

tructure”

’0.040*“big” + 0.033*“give” + 0.026*“question” + 0.016*“money” + 0.015*“think” +

0.012*“ask” + 0.012*“answer” + 0.012*“moment” + 0.011*“video” + 0.010*“using” +

0.010*“interest” + 0.009*“silence” + 0.009*“special” + 0.009*“back” + 0.009*“get”

’0.049*“bill” + 0.036*“house” + 0.028*“bipartisan” + 0.023*“act” + 0.019*“congress” +

0.017*“senate” + 0.016*“legislation” + 0.016*“pas” + 0.016*“floor” + 0.013*“passed” +

0.012*“need” + 0.010*“address” + 0.010*“support” + 0.010*“today” + 0.010*“funding”

’0.069*“state” + 0.021*“united” + 0.019*“congratulation” + 0.018*“press” + 0.017*“love”

+ 0.013*“conference” + 0.012*“win” + 0.010*“capitol” + 0.010*“former” + 0.010*“county”

+ 0.009*“city” + 0.008*“rep” + 0.008*“governor” + 0.007*“mayor” + 0.007*“public”

’0.052*“read” + 0.044*“statement” + 0.033*“attack” + 0.029*“security” + 0.029*“full” +

0.024*“national” + 0.020*“election” + 0.018*“joe” + 0.017*“congressman” + 0.017*“al”

+ 0.016*“u” + 0.014*“water” + 0.013*“green” + 0.012*“russia” + 0.011*“democracy”

39

’0.030*“family” + 0.025*“thought” + 0.022*“safe” + 0.020*“victim” + 0.017*“prayer” +

0.016*“violence” + 0.014*“heart” + 0.013*“sexual” + 0.012*“loved” + 0.012*“community”

+ 0.011*“first” + 0.011*“law” + 0.010*“one” + 0.010*“lost” + 0.010*“must”

’0.019*“hearing” + 0.014*“trump” + 0.014*“investigation” + 0.012*“letter” + 0.012*“com-

mittee” + 0.010*“ha” + 0.009*“special” + 0.008*“must” + 0.008*“house” + 0.007*“wa”

+ 0.007*“senate” + 0.006*“fbi” + 0.006*“republican” + 0.006*“need” + 0.006*“secretary”

’0.100*“happy” + 0.049*“friend” + 0.034*“birthday” + 0.034*“leadership” + 0.030*“col-

league” + 0.029*“thank” + 0.028*“wishing” + 0.021*“celebrating” + 0.020*“member” +

0.017*“everyone” + 0.016*“join” + 0.016*“wish” + 0.015*“family” + 0.014*“commitment”

+ 0.014*“new”

’0.060*“tax” + 0.033*“cut” + 0.023*“gop” + 0.021*“american” + 0.020*“republican” +

0.017*“family” + 0.016*“bill” + 0.015*“middle” + 0.014*“plan” + 0.014*“million” +

0.013*“budget” + 0.013*“class” + 0.011*“pay” + 0.011*“working” + 0.010*“benefit”

’0.079*“thank” + 0.033*“honor” + 0.031*“live” + 0.030*“life” + 0.027*“watch” +

0.023*“today” + 0.023*“service” + 0.020*“tune” + 0.020*“joining” + 0.018*“remember”

+ 0.017*“u” + 0.012*“discus” + 0.010*“woman” + 0.010*“military” + 0.009*“men”

’0.050*“vote” + 0.023*“clean” + 0.022*“voting” + 0.022*“energy” + 0.022*“find” +

0.021*“voter” + 0.018*“got” + 0.017*“new” + 0.017*“early” + 0.015*“air” + 0.014*“to-

day” + 0.013*“talking” + 0.010*“check” + 0.009*“register” + 0.008*“ballot”

’0.057*“job” + 0.024*“worker” + 0.017*“american” + 0.016*“economy” + 0.016*“drug”

+ 0.014*“company” + 0.014*“cost” + 0.013*“create” + 0.010*“price” + 0.010*“new” +

0.010*“ha” + 0.010*“fair” + 0.009*“consumer” + 0.008*“rate” + 0.008*“street”

’0.035*“student” + 0.022*“office” + 0.020*“learn” + 0.017*“help” + 0.016*“please”

+ 0.015*“stay” + 0.013*“can” + 0.013*“college” + 0.012*“call” + 0.009*“loan” +

0.009*“puerto” + 0.009*“disaster” + 0.008*“visit” + 0.008*“need” + 0.008*“assistance”

40

’0.040*“agree” + 0.034*“la” + 0.031*“de” + 0.023*“someone” + 0.022*“el” +

0.022*“leave” + 0.021*“social” + 0.015*“en” + 0.015*“paid” + 0.014*“georgia” +

0.012*“y” + 0.012*“los” + 0.011*“nh” + 0.011*“flag” + 0.010*“medium”

’0.054*“president” + 0.048*“trump” + 0.020*“ha” + 0.010*“said” + 0.009*“administra-

tion” + 0.009*“say” + 0.009*“oil” + 0.008*“internet” + 0.008*“rule” + 0.008*“free” +

0.007*“just” + 0.007*“must” + 0.007*“obama” + 0.006*“protect” + 0.006*“hold”

’0.066*“right” + 0.035*“fight” + 0.026*“stand” + 0.023*“must” + 0.016*“continue” +

0.016*“court” + 0.016*“justice” + 0.013*“american” + 0.012*“year” + 0.011*“human” +

0.011*“civil” + 0.011*“protect” + 0.011*“ha” + 0.010*“today” + 0.010*“supreme”

Table 6: Full Word List: Governors – Republican

Governor Word List (Rep.)

0.032*“great” + 0.014*“speak” + 0.013*“capitol” + 0.013*“night” + 0.012*“way”

+ 0.011*“wa” + 0.010*“see” + 0.009*“christmas” + 0.009*“last” + 0.009*“tree” +

0.008*“ribbon” + 0.008*“annual” + 0.008*“cutting” + 0.008*“thank” + 0.007*“morning”

0.082*“thank” + 0.032*“service” + 0.028*“woman” + 0.025*“veteran” + 0.018*“men” +

0.017*“honor” + 0.013*“military” + 0.013*“honored” + 0.012*“grateful” + 0.011*“today”

+ 0.011*“sacrifice” + 0.010*“life” + 0.009*“serve” + 0.009*“protect” + 0.009*“brave”

0.029*“west” + 0.022*“great” + 0.019*“state” + 0.014*“park” + 0.014*“beautiful” +

0.011*“virginia” + 0.011*“company” + 0.010*“see” + 0.009*“welcome” + 0.009*“one” +

0.009*“road” + 0.008*“day” + 0.008*“enjoy” + 0.008*“home” + 0.007*“another”

0.065*“happy” + 0.035*“hard” + 0.025*“work” + 0.021*“thank” + 0.018*“everyone” +

0.018*“good” + 0.017*“safe” + 0.017*“luck” + 0.016*“de” + 0.016*“best” + 0.015*“wish”

+ 0.015*“state” + 0.014*“birthday” + 0.012*“working” + 0.011*“wishing”

41

0.076*“tax” + 0.015*“vote” + 0.013*“cut” + 0.013*“budget” + 0.013*“property” +

0.011*“get” + 0.011*“income” + 0.011*“illinois” + 0.011*“just” + 0.010*“reform” +

0.008*“money” + 0.008*“want” + 0.008*“back” + 0.008*“state” + 0.007*“can”

0.034*“wa” + 0.027*“day” + 0.027*“proud” + 0.022*“today” + 0.022*“honor”

+ 0.018*“american” + 0.016*“year” + 0.011*“celebrate” + 0.010*“celebrating” +

0.009*“great” + 0.009*“state” + 0.009*“anniversary” + 0.008*“world” + 0.007*“life” +

0.006*“one”

0.043*“first” + 0.042*“congratulation” + 0.020*“lady” + 0.013*“win” + 0.012*“great”

+ 0.012*“proud” + 0.011*“award” + 0.011*“rt” + 0.008*“congrats” + 0.008*“today” +

0.008*“ground” + 0.008*“ha” + 0.008*“winning” + 0.007*“going” + 0.007*“team”

0.036*“student” + 0.025*“workforce” + 0.014*“education” + 0.014*“college” +

0.013*“program” + 0.012*“career” + 0.011*“future” + 0.010*“opportunity” +

0.010*“training” + 0.010*“skill” + 0.010*“higher” + 0.009*“statement” + 0.008*“energy”

+ 0.007*“need” + 0.007*“development”

0.041*“state” + 0.034*“new” + 0.032*“governor” + 0.017*“budget” + 0.013*“read” +

0.012*“first” + 0.010*“hampshire” + 0.009*“year” + 0.009*“today” + 0.008*“ha” +

0.008*“wa” + 0.007*“star” + 0.007*“case” + 0.006*“day” + 0.006*“last”

0.036*“state” + 0.021*“flag” + 0.018*“emergency” + 0.014*“local” + 0.014*“law” +

0.013*“county” + 0.012*“stay” + 0.012*“storm” + 0.011*“weather” + 0.010*“honor” +

0.009*“enforcement” + 0.009*“official” + 0.009*“response” + 0.008*“order” + 0.007*“hur-

ricane”

’0.033*“south” + 0.026*“congrats” + 0.025*“city” + 0.013*“great” + 0.012*“carolina”

+ 0.012*“north” + 0.011*“today” + 0.011*“fire” + 0.010*“county” + 0.010*“thanks” +

0.009*“dakota” + 0.008*“take” + 0.007*“team” + 0.007*“list” + 0.007*“best”

’0.017*“new” + 0.016*“opioid” + 0.016*“signed” + 0.012*“legislation” + 0.011*“today” +

0.011*“force” + 0.010*“bill” + 0.009*“sign” + 0.009*“law” + 0.009*“ha” + 0.008*“task”

+ 0.008*“work” + 0.008*“baker” + 0.007*“public” + 0.007*“assistance”

42

’0.026*“great” + 0.026*“wa” + 0.026*“family” + 0.024*“thanks” + 0.019*“friend” +

0.017*“prayer” + 0.011*“president” + 0.011*“thought” + 0.010*“first” + 0.009*“time”

+ 0.008*“passing” + 0.007*“governor” + 0.007*“loved” + 0.007*“life” + 0.007*“one”

’0.057*“school” + 0.035*“state” + 0.019*“student” + 0.018*“high” + 0.011*“hb” +

0.010*“need” + 0.009*“help” + 0.009*“great” + 0.009*“health” + 0.009*“together” +

0.008*“safety” + 0.008*“can” + 0.008*“government” + 0.008*“work” + 0.008*“people”

’0.055*“live” + 0.051*“forward” + 0.040*“watch” + 0.036*“looking” + 0.029*“tune”

+ 0.026*“look” + 0.019*“now” + 0.017*“joining” + 0.017*“talk” + 0.017*“discus” +

0.013*“justice” + 0.012*“listen” + 0.012*“join” + 0.011*“morning” + 0.011*“tomorrow”

’0.016*“video” + 0.015*“indiana” + 0.013*“follow” + 0.012*“check” + 0.012*“general” +

0.012*“water” + 0.010*“manufacturing” + 0.010*“link” + 0.009*“view” + 0.008*“click”

+ 0.008*“close” + 0.007*“right” + 0.007*“u” + 0.007*“new” + 0.006*“thrilled”

’0.027*“texas” + 0.020*“health” + 0.012*“care” + 0.012*“work” + 0.010*“help” +

0.009*“state” + 0.008*“effort” + 0.008*“role” + 0.007*“human” + 0.007*“recovery” +

0.007*“ha” + 0.007*“people” + 0.007*“transportation” + 0.007*“play” + 0.005*“can”

’0.025*“great” + 0.025*“tour” + 0.017*“enjoyed” + 0.015*“speaking” + 0.015*“excited” +

0.013*“today” + 0.012*“see” + 0.012*“big” + 0.011*“stop” + 0.010*“time” + 0.009*“fight-

ing” + 0.008*“chamber” + 0.007*“term” + 0.007*“visiting” + 0.007*“taking”

’0.027*“senate” + 0.020*“conference” + 0.018*“pleased” + 0.016*“legislative” +

0.016*“bill” + 0.015*“state” + 0.014*“house” + 0.013*“today” + 0.013*“m” +

0.011*“press” + 0.011*“session” + 0.010*“court” + 0.010*“judge” + 0.009*“district” +

0.009*“board”

’0.050*“job” + 0.024*“ha” + 0.017*“state” + 0.017*“rate” + 0.016*“since” + 0.016*“wis-

consin” + 0.014*“new” + 0.014*“unemployment” + 0.011*“growth” + 0.010*“ohio” +

0.009*“year” + 0.009*“drug” + 0.009*“record” + 0.008*“nation” + 0.008*“number”

43

’0.048*“business” + 0.045*“great” + 0.027*“good” + 0.024*“meeting” + 0.018*“small” +

0.017*“leader” + 0.015*“news” + 0.013*“economic” + 0.013*“trade” + 0.012*“industry” +

0.010*“morning” + 0.009*“discussion” + 0.008*“discus” + 0.008*“economy” + 0.007*“to-

day”

’0.023*“may” + 0.022*“let” + 0.015*“day” + 0.015*“u” + 0.014*“never” + 0.013*“get” +

0.011*“great” + 0.011*“god” + 0.010*“awareness” + 0.009*“month” + 0.008*“remember”

+ 0.008*“time” + 0.008*“fall” + 0.008*“forget” + 0.008*“know”

’0.019*“teacher” + 0.018*“make” + 0.017*“can” + 0.013*“need” + 0.013*“education”

+ 0.013*“support” + 0.011*“opportunity” + 0.010*“like” + 0.010*“community” +

0.010*“child” + 0.010*“raise” + 0.009*“every” + 0.008*“continue” + 0.008*“help” +

0.008*“go”

’0.046*“join” + 0.021*“u” + 0.021*“national” + 0.017*“kentucky” + 0.017*“please”

+ 0.014*“food” + 0.014*“help” + 0.014*“find” + 0.012*“can” + 0.011*“state” +

0.009*“house” + 0.009*“support” + 0.008*“guard” + 0.008*“sb” + 0.007*“drive”

’0.048*“new” + 0.037*“million” + 0.022*“job” + 0.019*“create” + 0.016*“investment” +

0.013*“facility” + 0.012*“project” + 0.012*“help” + 0.012*“grant” + 0.012*“center” +

0.010*“expansion” + 0.009*“florida” + 0.008*“invest” + 0.008*“hurricane” + 0.008*“an-

nounce”

Table 7: Full Word List: Governors – Democratic

Governor Word List (Dem.)

0.038*“mt” + 0.036*“thanks” + 0.033*“great” + 0.017*“wa” + 0.016*“rt” + 0.010*“morn-

ing” + 0.008*“talk” + 0.008*“can” + 0.008*“got” + 0.008*“glad” + 0.007*“u” +

0.007*“hear” + 0.007*“first” + 0.007*“meet” + 0.006*“youth”

44

0.054*“great” + 0.050*“happy” + 0.024*“looking” + 0.024*“forward” + 0.015*“day” +

0.014*“town” + 0.014*“congratulation” + 0.013*“join” + 0.013*“meeting” + 0.012*“wish-

ing” + 0.012*“hall” + 0.012*“celebrating” + 0.011*“look” + 0.010*“city” + 0.010*“today”

0.043*“student” + 0.026*“every” + 0.021*“school” + 0.017*“can” + 0.015*“need”

+ 0.013*“child” + 0.013*“make” + 0.012*“pa” + 0.012*“better” + 0.011*“kid” +

0.011*“north” + 0.011*“get” + 0.010*“way” + 0.009*“help” + 0.009*“college”

0.047*“time” + 0.040*“bill” + 0.031*“today” + 0.029*“signing” + 0.020*“mahalo” +

0.019*“ban” + 0.016*“welcome” + 0.015*“noon” + 0.014*“watch” + 0.014*“aloha” +

0.010*“voter” + 0.008*“new” + 0.007*“sign” + 0.007*“now” + 0.007*“california”

0.032*“work” + 0.026*“make” + 0.022*“working” + 0.018*“right” + 0.015*“can” +

0.014*“keep” + 0.014*“together” + 0.013*“state” + 0.011*“need” + 0.010*“hard” +

0.009*“sure” + 0.009*“pay” + 0.009*“people” + 0.009*“family” + 0.008*“protect”

0.025*“life” + 0.020*“year” + 0.018*“family” + 0.018*“honor” + 0.014*“one” + 0.013*“to-

day” + 0.013*“wa” + 0.013*“lost” + 0.012*“ha” + 0.011*“remember” + 0.010*“thought”

+ 0.010*“day” + 0.009*“officer” + 0.009*“american” + 0.009*“prayer”

0.029*“law” + 0.019*“signed” + 0.015*“bill” + 0.015*“opioid” + 0.015*“legislation” +

0.013*“help” + 0.012*“take” + 0.011*“prevent” + 0.009*“today” + 0.009*“safety” +

0.009*“medical” + 0.008*“important” + 0.008*“sign” + 0.007*“step” + 0.007*“crisis”

0.149*“new” + 0.051*“york” + 0.023*“state” + 0.022*“jersey” + 0.012*“yorkers” +

0.007*“day” + 0.007*“just” + 0.006*“airport” + 0.006*“first” + 0.005*“ha” + 0.005*“wa”

+ 0.004*“mean” + 0.004*“stronger” + 0.004*“year” + 0.004*“city”

0.043*“live” + 0.043*“budget” + 0.035*“watch” + 0.028*“malloy” + 0.026*“governor” +

0.022*“edward” + 0.020*“discus” + 0.019*“state” + 0.015*“conference” + 0.015*“tune”

+ 0.014*“today” + 0.011*“special” + 0.010*“vote” + 0.010*“proposal” + 0.009*“press”

45

0.040*“school” + 0.023*“court” + 0.020*“high” + 0.013*“wa” + 0.012*“supreme” +

0.011*“force” + 0.010*“state” + 0.009*“task” + 0.008*“justice” + 0.008*“judge” +

0.008*“ha” + 0.007*“university” + 0.007*“class” + 0.006*“congratulation” + 0.005*“stu-

dent”

’0.033*“new” + 0.019*“stand” + 0.017*“trump” + 0.012*“president” + 0.012*“fight” +

0.010*“administration” + 0.009*“woman” + 0.009*“fighting” + 0.009*“washington” +

0.008*“continue” + 0.008*“policy” + 0.008*“protect” + 0.008*“ha” + 0.008*“state” +

0.008*“york”

’0.026*“justice” + 0.021*“criminal” + 0.015*“opening” + 0.014*“full” + 0.012*“show” +

0.010*“reform” + 0.009*“schedule” + 0.009*“discussion” + 0.008*“step” + 0.008*“public”

+ 0.008*“happening” + 0.007*“excited” + 0.007*“blue” + 0.006*“grand” + 0.006*“amer-

ica”

’0.024*“emergency” + 0.022*“state” + 0.019*“stay” + 0.019*“storm” + 0.017*“update”

+ 0.016*“local” + 0.014*“road” + 0.012*“weather” + 0.012*“county” + 0.011*“winter” +

0.010*“follow” + 0.009*“condition” + 0.009*“response” + 0.009*“please” + 0.008*“official”

’0.044*“business” + 0.024*“small” + 0.012*“state” + 0.011*“see” + 0.009*“project” +

0.009*“one” + 0.009*“improve” + 0.008*“part” + 0.008*“investment” + 0.008*“support”

+ 0.008*“across” + 0.007*“biz” + 0.007*“program” + 0.007*“local” + 0.007*“help”

’0.053*“job” + 0.021*“new” + 0.015*“create” + 0.015*“economy” + 0.013*“mn” +

0.013*“investment” + 0.012*“economic” + 0.011*“workforce” + 0.011*“opportunity” +

0.010*“expand” + 0.009*“company” + 0.009*“help” + 0.009*“expansion” + 0.009*“busi-

ness” + 0.009*“ha”

’0.047*“gov” + 0.014*“holiday” + 0.013*“annual” + 0.012*“state” + 0.012*“congrats” +

0.010*“capitol” + 0.010*“beautiful” + 0.009*“tree” + 0.009*“little” + 0.009*“tour” +

0.009*“season” + 0.008*“join” + 0.007*“va” + 0.006*“today” + 0.006*“open”

46

’0.032*“energy” + 0.027*“clean” + 0.026*“climate” + 0.016*“change” + 0.012*“air” +

0.010*“goal” + 0.010*“u” + 0.009*“new” + 0.009*“plan” + 0.009*“ha” + 0.008*“state”

+ 0.008*“power” + 0.007*“wind” + 0.007*“renewable” + 0.007*“water”

’0.038*“good” + 0.027*“flag” + 0.021*“house” + 0.021*“honor” + 0.019*“victim” +

0.015*“go” + 0.014*“state” + 0.014*“luck” + 0.014*“senate” + 0.014*“bill” + 0.011*“wa”

+ 0.010*“ordered” + 0.009*“passed” + 0.009*“flown” + 0.008*“lowered”

’0.054*“help” + 0.017*“community” + 0.014*“can” + 0.014*“grant” + 0.012*“assistance”

+ 0.011*“today” + 0.011*“state” + 0.010*“recovery” + 0.010*“disaster” + 0.009*“re-

source” + 0.009*“federal” + 0.009*“call” + 0.009*“program” + 0.008*“ha” + 0.008*“new”

’0.083*“health” + 0.052*“care” + 0.020*“access” + 0.020*“education” + 0.014*“insurance”

+ 0.014*“get” + 0.011*“coverage” + 0.010*“early” + 0.010*“affordable” + 0.009*“system”

+ 0.009*“child” + 0.008*“need” + 0.008*“quality” + 0.008*“mental” + 0.007*“can”

’0.043*“state” + 0.019*“government” + 0.016*“federal” + 0.011*“new” + 0.010*“pub-

lic” + 0.009*“fund” + 0.008*“save” + 0.008*“land” + 0.008*“million” + 0.008*“park”

+ 0.008*“taxpayer” + 0.007*“billion” + 0.007*“ha” + 0.007*“highway” + 0.007*“drive”

’0.070*“thank” + 0.024*“service” + 0.020*“best” + 0.018*“veteran” + 0.017*“woman” +

0.016*“state” + 0.014*“member” + 0.013*“celebrate” + 0.012*“proud” + 0.012*“today” +

0.011*“national” + 0.010*“great” + 0.010*“military” + 0.010*“congrats” + 0.009*“men”

’0.109*“governor” + 0.055*“dayton” + 0.034*“minnesota” + 0.028*“cooper” +

0.024*“read” + 0.023*“order” + 0.022*“statement” + 0.021*“executive” + 0.012*“school”

+ 0.012*“today” + 0.011*“mark” + 0.010*“ha” + 0.009*“state” + 0.009*“roy” +

0.008*“learn”

’0.054*“gun” + 0.027*“violence” + 0.015*“safety” + 0.012*“domestic” + 0.010*“must”

+ 0.008*“delaware” + 0.008*“letter” + 0.008*“drilling” + 0.007*“hand” + 0.007*“oil” +

0.007*“today” + 0.007*“survivor” + 0.007*“threat” + 0.007*“keep” + 0.006*“can”

47

’0.028*“tax” + 0.018*“rhode” + 0.016*“free” + 0.015*“check” + 0.011*“wage” +

0.011*“get” + 0.010*“day” + 0.009*“time” + 0.009*“today” + 0.008*“island” +

0.008*“minimum” + 0.008*“state” + 0.008*“middle” + 0.008*“ha” + 0.008*“list”

Table 8: Full Word List: Mayors – Republican

Mayor Word List (Rep.)

0.073*“good” + 0.030*“morning” + 0.017*“via” + 0.013*“weekly” + 0.013*“meeting” +

0.012*“message” + 0.011*“budget” + 0.010*“wa” + 0.009*“today” + 0.007*“health” +

0.007*“board” + 0.006*“already” + 0.006*“real” + 0.006*“transportation” + 0.006*“talk-

ing”

0.053*“make” + 0.026*“step” + 0.025*“mile” + 0.017*“fitbit” + 0.016*“sure” +

0.014*“better” + 0.010*“community” + 0.009*“job” + 0.008*“help” + 0.008*“voting” +

0.008*“take” + 0.007*“election” + 0.007*“get” + 0.007*“know” + 0.007*“can”

0.052*“day” + 0.025*“one” + 0.022*“christmas” + 0.021*“happy” + 0.020*“great” +

0.014*“birthday” + 0.011*“parade” + 0.011*“old” + 0.011*“wa” + 0.010*“favorite” +

0.010*“yes” + 0.010*“beautiful” + 0.010*“memorial” + 0.008*“wonderful” + 0.008*“an-

niversary”

0.019*“win” + 0.017*“street” + 0.014*“wa” + 0.013*“candidate” + 0.009*“view” +

0.009*“way” + 0.008*“main” + 0.007*“located” + 0.007*“just” + 0.007*“spring” +

0.007*“cookeville” + 0.006*“back” + 0.006*“today” + 0.006*“mayor” + 0.006*“square”

0.032*“meeting” + 0.028*“town” + 0.026*“hall” + 0.016*“fire” + 0.015*“enjoyed” +

0.011*“tonight” + 0.010*“sterling” + 0.010*“chamber” + 0.010*“city” + 0.009*“join” +

0.009*“height” + 0.009*“pm” + 0.008*“today” + 0.007*“annual” + 0.007*“station”

0.074*“go” + 0.038*“u” + 0.036*“join” + 0.030*“live” + 0.026*“let” + 0.015*“watch” +

0.012*“west” + 0.012*“way” + 0.009*“development” + 0.009*“god” + 0.008*“economic”

+ 0.008*“know” + 0.008*“time” + 0.008*“covina” + 0.007*“everybody”

48

0.143*“thank” + 0.031*“de” + 0.022*“la” + 0.016*“veteran” + 0.015*“el” + 0.014*“en”

+ 0.012*“great” + 0.009*“enjoying” + 0.009*“y” + 0.007*“con” + 0.006*“ground” +

0.006*“los” + 0.006*“ha” + 0.006*“bridge” + 0.005*“para”

0.123*“great” + 0.015*“opening” + 0.015*“week” + 0.014*“time” + 0.013*“romeoville”

+ 0.013*“job” + 0.012*“always” + 0.011*“grand” + 0.011*“thing” + 0.010*“new” +

0.010*“glad” + 0.009*“another” + 0.009*“enjoy” + 0.007*“see” + 0.007*“mayor”

0.025*“honor” + 0.021*“congrats” + 0.021*“san” + 0.021*“proud” + 0.017*“today” +

0.016*“thank” + 0.013*“honored” + 0.011*“wa” + 0.010*“great” + 0.009*“serve” +

0.008*“support” + 0.008*“award” + 0.008*“full” + 0.007*“diego” + 0.007*“congratula-

tion”

0.119*“city” + 0.040*“council” + 0.017*“best” + 0.016*“ha” + 0.016*“excited” +

0.015*“state” + 0.009*“year” + 0.008*“just” + 0.007*“mayor” + 0.006*“one” +

0.006*“new” + 0.006*“address” + 0.005*“member” + 0.005*“voted” + 0.005*“announce”

’0.039*“love” + 0.017*“business” + 0.012*“campaign” + 0.010*“small” + 0.010*“great” +

0.009*“just” + 0.008*“people” + 0.008*“big” + 0.008*“new” + 0.007*“wa” + 0.007*“ha”

+ 0.007*“walk” + 0.006*“one” + 0.006*“market” + 0.006*“day”

’0.043*“see” + 0.033*“back” + 0.027*“get” + 0.025*“everyone” + 0.010*“going” +

0.009*“hope” + 0.009*“please” + 0.009*“qo” + 0.009*“wait” + 0.008*“kid” + 0.008*“one”

+ 0.008*“come” + 0.008*“give” + 0.007*“believe” + 0.007*“stay”

’0.044*“night” + 0.029*“last” + 0.016*“tomorrow” + 0.015*“day” + 0.013*“art” +

0.012*“friday” + 0.011*“city” + 0.009*“see” + 0.009*“library” + 0.008*“worth” +

0.008*“lake” + 0.007*“park” + 0.007*“tonight” + 0.007*“center” + 0.007*“open”

’0.058*“happy” + 0.044*“check” + 0.033*“video” + 0.033*“new” + 0.018*“family”

+ 0.015*“miss” + 0.014*“wish” + 0.009*“wa” + 0.008*“mark” + 0.008*“year” +

0.008*“great” + 0.008*“news” + 0.008*“thanksgiving” + 0.007*“fun” + 0.007*“safe”

49

’0.024*“mayor” + 0.021*“city” + 0.016*“got” + 0.011*“great” + 0.010*“youth” +

0.010*“summer” + 0.009*“boy” + 0.008*“program” + 0.008*“girl” + 0.008*“neigh-

borhood” + 0.008*“league” + 0.006*“club” + 0.006*“conference” + 0.006*“now” +

0.005*“mesa”

’0.019*“chief” + 0.014*“like” + 0.011*“police” + 0.010*“weekend” + 0.009*“dog” +

0.009*“freedom” + 0.008*“wa” + 0.007*“personal” + 0.007*“day” + 0.006*“support” +

0.006*“fire” + 0.006*“update” + 0.006*“today” + 0.006*“george” + 0.006*“course”

’0.033*“county” + 0.027*“school” + 0.021*“mayor” + 0.016*“ribbon” + 0.013*“student”

+ 0.011*“cutting” + 0.011*“center” + 0.010*“today” + 0.010*“state” + 0.008*“new” +

0.008*“senior” + 0.007*“elementary” + 0.006*“home” + 0.006*“law” + 0.006*“high”

’0.032*“getting” + 0.029*“ready” + 0.025*“show” + 0.022*“park” + 0.015*“get” +

0.008*“awesome” + 0.008*“lord” + 0.008*“word” + 0.007*“thats” + 0.007*“want” +

0.007*“men” + 0.007*“woman” + 0.005*“day” + 0.005*“couple” + 0.005*“car”

’0.036*“can” + 0.023*“free” + 0.012*“feel” + 0.011*“please” + 0.011*“report” +

0.010*“just” + 0.010*“email” + 0.010*“park” + 0.009*“get” + 0.009*“power” +

0.008*“working” + 0.008*“crew” + 0.008*“via” + 0.008*“now” + 0.007*“help”

’0.084*“thanks” + 0.033*“work” + 0.027*“great” + 0.023*“public” + 0.020*“well” +

0.017*“community” + 0.013*“road” + 0.010*“volunteer” + 0.008*“congratulation” +

0.008*“many” + 0.008*“done” + 0.007*“safety” + 0.007*“traffic” + 0.007*“hard” +

0.006*“service”

’0.044*“looking” + 0.044*“look” + 0.040*“forward” + 0.021*“good” + 0.012*“north”

+ 0.012*“still” + 0.012*“luck” + 0.011*“now” + 0.010*“seeing” + 0.008*“like” +

0.008*“need” + 0.008*“great” + 0.007*“run” + 0.006*“next” + 0.006*“legislative”

’0.022*“tax” + 0.021*“nice” + 0.020*“time” + 0.011*“arcadia” + 0.010*“first” +

0.010*“white” + 0.008*“rate” + 0.007*“light” + 0.007*“home” + 0.007*“property” +

0.007*“ha” + 0.007*“red” + 0.007*“super” + 0.007*“house” + 0.006*“sale”

50

’0.128*“new” + 0.084*“facebook” + 0.080*“posted” + 0.078*“photo” + 0.028*“album”

+ 0.021*“welcome” + 0.008*“apex” + 0.007*“tn” + 0.006*“first” + 0.005*“firefighter” +

0.005*“just” + 0.005*“loving” + 0.005*“center” + 0.004*“coming” + 0.004*“california”

’0.035*“wa” + 0.023*“school” + 0.022*“police” + 0.016*“high” + 0.015*“officer” +

0.015*“prayer” + 0.014*“thought” + 0.010*“family” + 0.008*“just” + 0.008*“today” +

0.007*“say” + 0.007*“made” + 0.006*“department” + 0.006*“year” + 0.005*“lost”

’0.031*“u” + 0.028*“vote” + 0.018*“need” + 0.017*“please” + 0.017*“help” +

0.016*“think” + 0.013*“question” + 0.010*“join” + 0.010*“can” + 0.009*“keep” +

0.009*“today” + 0.008*“mayor” + 0.008*“like” + 0.008*“get” + 0.008*“american”

Table 9: Full Word List: Mayors – Democratic

Mayor Word List (Dem.)

0.042*“new” + 0.025*“park” + 0.018*“opening” + 0.016*“downtown” + 0.011*“center”

+ 0.011*“summer” + 0.011*“grand” + 0.010*“bike” + 0.010*“project” + 0.009*“open” +

0.009*“today” + 0.009*“city” + 0.009*“ribbon” + 0.007*“station” + 0.006*“cutting”

0.049*“can” + 0.031*“get” + 0.022*“know” + 0.019*“need” + 0.019*“vote” +

0.016*“please” + 0.015*“help” + 0.015*“just” + 0.014*“want” + 0.014*“let” + 0.013*“one”

+ 0.011*“call” + 0.010*“election” + 0.009*“find” + 0.009*“people”

0.026*“via” + 0.018*“tax” + 0.016*“budget” + 0.016*“city” + 0.015*“state” +

0.012*“bill” + 0.010*“plan” + 0.008*“yes” + 0.008*“council” + 0.008*“public” +

0.007*“update” + 0.007*“property” + 0.007*“fund” + 0.006*“proposed” + 0.006*“fund-

ing”

0.028*“art” + 0.024*“food” + 0.014*“west” + 0.012*“local” + 0.011*“tree” + 0.011*“festi-

val” + 0.011*“early” + 0.010*“music” + 0.008*“tell” + 0.007*“challenge” + 0.007*“south”

+ 0.007*“city” + 0.007*“shop” + 0.007*“number” + 0.007*“artist”

51

0.051*“join” + 0.043*“city” + 0.023*“hall” + 0.018*“u” + 0.017*“come” + 0.016*“tomor-

row” + 0.013*“town” + 0.013*“please” + 0.013*“day” + 0.012*“open” + 0.011*“event” +

0.010*“tonight” + 0.010*“today” + 0.009*“free” + 0.009*“center”

0.057*“make” + 0.046*“job” + 0.022*“sure” + 0.016*“win” + 0.016*“best” + 0.016*“take”

+ 0.015*“well” + 0.013*“thing” + 0.013*“top” + 0.012*“got” + 0.009*“making” +

0.009*“city” + 0.008*“great” + 0.008*“game” + 0.008*“team”

0.126*“thank” + 0.052*“forward” + 0.047*“looking” + 0.029*“mayor” + 0.027*“look” +

0.014*“smith” + 0.011*“joseph” + 0.011*“seeing” + 0.010*“joining” + 0.009*“future” +

0.008*“discussion” + 0.008*“talking” + 0.008*“north” + 0.008*“sharing” + 0.008*“wife”

0.227*“new” + 0.065*“check” + 0.061*“facebook” + 0.061*“posted” + 0.054*“photo” +

0.040*“video” + 0.014*“change” + 0.013*“climate” + 0.008*“exciting” + 0.007*“just” +

0.007*“page” + 0.007*“first” + 0.006*“york” + 0.005*“album” + 0.005*“time”

0.028*“white” + 0.027*“health” + 0.026*“volunteer” + 0.014*“care” + 0.013*“plain”

+ 0.011*“house” + 0.011*“word” + 0.010*“court” + 0.010*“ticket” + 0.008*“bay” +

0.008*“help” + 0.007*“spread” + 0.006*“call” + 0.006*“clear” + 0.006*“phone”

0.141*“great” + 0.045*“day” + 0.027*“time” + 0.017*“see” + 0.017*“thank” +

0.014*“night” + 0.013*“everyone” + 0.012*“happy” + 0.012*“last” + 0.011*“beautiful”

+ 0.011*“event” + 0.011*“wonderful” + 0.010*“fun” + 0.009*“wa” + 0.009*“community”

’0.097*“wa” + 0.046*“go” + 0.023*“right” + 0.013*“law” + 0.008*“recycling” +

0.007*“way” + 0.007*“enforcement” + 0.006*“missed” + 0.006*“just” + 0.006*“one” +

0.006*“thx” + 0.006*“civil” + 0.005*“sorry” + 0.005*“waste” + 0.005*“dog”

’0.053*“meeting” + 0.044*“love” + 0.042*“live” + 0.038*“council” + 0.034*“city” +

0.018*“watch” + 0.017*“power” + 0.014*“talk” + 0.014*“mayor” + 0.012*“tune” +

0.011*“discus” + 0.011*“san” + 0.010*“tonight” + 0.010*“medium” + 0.009*“committee”

’0.023*“run” + 0.013*“running” + 0.013*“club” + 0.013*“governor” + 0.013*“girl” +

0.012*“now” + 0.012*“candidate” + 0.011*“god” + 0.010*“race” + 0.009*“boy” +

0.009*“state” + 0.008*“democratic” + 0.008*“florida” + 0.007*“joe” + 0.007*“course”

52

’0.037*“police” + 0.034*“family” + 0.020*“fire” + 0.020*“officer” + 0.018*“friend” +

0.016*“chief” + 0.015*“life” + 0.013*“thought” + 0.012*“department” + 0.011*“honor”

+ 0.011*“prayer” + 0.010*“today” + 0.010*“heart” + 0.009*“lost” + 0.009*“remember”

’0.119*“mayor” + 0.040*“de” + 0.030*“la” + 0.021*“conference” + 0.020*“en” +

0.017*“press” + 0.016*“el” + 0.012*“y” + 0.011*“miramar” + 0.011*“twitter” +

0.010*“los” + 0.009*“week” + 0.008*“que” + 0.007*“con” + 0.007*“m”

’0.042*“proud” + 0.028*“congratulation” + 0.024*“city” + 0.017*“honored” + 0.016*“con-

grats” + 0.015*“thank” + 0.015*“business” + 0.012*“youth” + 0.012*“award” +

0.011*“wa” + 0.010*“year” + 0.010*“young” + 0.009*“great” + 0.009*“work” + 0.009*“to-

day”

’0.030*“street” + 0.019*“glad” + 0.017*“veteran” + 0.012*“king” + 0.010*“martin” +

0.009*“legacy” + 0.009*“bridge” + 0.009*“chamber” + 0.008*“annual” + 0.008*“traffic”

+ 0.007*“life” + 0.007*“lima” + 0.007*“luther” + 0.007*“part” + 0.007*“denver”

’0.062*“ha” + 0.035*“going” + 0.019*“im” + 0.017*“welcome” + 0.017*“gun” +

0.013*“since” + 0.011*“violence” + 0.011*“year” + 0.010*“city” + 0.010*“back” +

0.009*“stand” + 0.009*“made” + 0.008*“crime” + 0.008*“long” + 0.007*“progress”

’0.083*“school” + 0.044*“happy” + 0.036*“high” + 0.027*“student” + 0.026*“housing” +

0.014*“affordable” + 0.011*“birthday” + 0.011*“middle” + 0.010*“wait” + 0.009*“costa”

+ 0.009*“force” + 0.008*“city” + 0.008*“see” + 0.008*“mesa” + 0.007*“today”

’0.043*“like” + 0.029*“president” + 0.021*“just” + 0.016*“nice” + 0.013*“trump” +

0.010*“year” + 0.009*“look” + 0.009*“kid” + 0.009*“red” + 0.009*“turn” + 0.009*“even”

+ 0.009*“day” + 0.008*“happy” + 0.008*“fall” + 0.008*“obama”

’0.023*“water” + 0.020*“street” + 0.020*“road” + 0.015*“snow” + 0.015*“parking” +

0.015*“please” + 0.014*“crew” + 0.012*“stay” + 0.011*“emergency” + 0.011*“car” +

0.011*“drive” + 0.010*“main” + 0.010*“work” + 0.010*“due” + 0.009*“safe”

53

’0.182*“thanks” + 0.099*“u” + 0.093*“good” + 0.038*“follow” + 0.033*“following” +

0.031*“ready” + 0.030*“getting” + 0.025*“facebook” + 0.013*“luck” + 0.012*“morning”

+ 0.008*“cafe” + 0.008*“wednesday” + 0.008*“bagel” + 0.007*“join” + 0.007*“rock”

’0.040*“rt” + 0.025*“fairfield” + 0.023*“say” + 0.013*“via” + 0.011*“rain” + 0.010*“door”

+ 0.009*“nothing” + 0.009*“watching” + 0.009*“statement” + 0.008*“news” +

0.008*“sound” + 0.007*“really” + 0.007*“beat” + 0.006*“interested” + 0.006*“just”

’0.019*“work” + 0.017*“city” + 0.013*“people” + 0.011*“need” + 0.011*“must”

+ 0.010*“working” + 0.010*“community” + 0.009*“continue” + 0.009*“child” +

0.008*“help” + 0.008*“better” + 0.008*“together” + 0.007*“year” + 0.007*“can” +

0.006*“issue”

’0.029*“woman” + 0.028*“community” + 0.018*“public” + 0.016*“neighborhood” +

0.014*“support” + 0.014*“service” + 0.013*“city” + 0.012*“men” + 0.011*“leader” +

0.010*“help” + 0.010*“safety” + 0.010*“program” + 0.009*“resident” + 0.008*“strong” +

0.008*“thank”

54