This item is only available as the following downloads:
A Comparative Case Study Analysis on the Twitter Attack Strategies of Candidates in the 2012 United States General Senatorial Election What conditions are most conducive for political candidates' use of attack strategies on Twitter? Kimberly Greenplate University of Florida email@example.com Prepared as a 2013 Honors Thesis in the Department of Political Science Thesis Advisor: Professor Beth Rosenson University of Florida Gainesville, Florida
Twitter Attacks of Senatorial Candidates 2 Table of Contents I. Introduction 3 4 II. Literature Review 5 12 a. Overview 5 b. Theories 5 6 c. Sources and Types of Data 6 8 d. Methods of Analysis 8 10 e. Findings 10 11 f. Expanding on Previous Literature 11 12 III. Theory 12 13 IV. Data Gathering 13 17 a. Variable Introduction 13 14 b. Case Selection 14 17 V. Data: Variables and Measurement 17 20 VI. Hypotheses 20 23 VII. Uses of Data 23 26 a. Data Interpretation 23 24 b. Case Study Advantages and Disadvantages 24 26 VII. Qualit ative Analysis for Each Race 26 33 IX. Logistic Regression Model and Interpretation 33 35 X. Considering Problems Associated with Data 35 36 XI. Unobservable Effects Model and Interpretation 37 43 a. Interpretation for Random Intercept 38 40 b. Interpretation for Dummy Variables 40 43 XII. Confirmation or Denial of Hypotheses 43 45 XIII. Conclusion 45 47 XIV. Beyond This Study and Further Research 47 49 XV. Bibliography 50 54
Twitter Attacks of Senatorial Candidates 3 I. Introduction Launched in early 2006, Twitter has gained popularity around the world with more than 500 million registered users in 2012. This popularity has spread into the political arena, making major headlines as a powerful political tool in President Barack Obama's 2008 campaign. Still, the question has been asked, is the adoption of Twitter worth i t? ( Gulati and Williams 2010). More and more, candidates for political office are saying "yes" by adopting these social media platforms, including Twitter, to communicate with constituents. This is evidenced in the amount of recent studies devoted to unde rstanding which candidates adopt new communication tools like social media and why While Twitter provide s constituent communication benefits at little to no cost for candidates, it also creates the need for new communication strategies and content in 1 40 characters or less. With few specific institutional rules surrounding social media, candidates have the freedom to u se Twitter in the manner they desire. Like traditional political advertising, these tweets have the potential to go negative and attack t he opposition or take on positive strategies that play up candidates' strengths. Determining which strategies work in candidates' favor, especially as new communication o utlets emerge, is still a subject of differing opinion among political scientists. Previous political scientific research on the efficacy of particular campaign strategies, including their effect on voter decisions, is abundant. The debate about negative versus positive campaign strategies in pursuit of elected office has been ongoing m ostly because candidates believe their strategies play a large part in voters' per ception of their electability. A positive perception of electability turns into constituent votes and, eventually, candidates' success in obtaining office. However, both nega tive and positive campaigning strategies pose risks: the risk of appearing u nwilling/ unable to defend oneself through use of purely positive strategies or the risk of being
Twitter Attacks of Senatorial Candidates 4 seen as petty and playing the blame game through the use of purely negative strateg ies. Candidates' decisions to go negative have external effects that impac t voting constituencies, which ultimately can affect can didates' chance of being elected or retained. How will voters remember candidates at the polls? Negative campaigning has repea tedly been identified as more memorable and potent in influencing the electorate, usually favorable to candidates gaining office. More and more, candidates appear to be taking on negative campaign strategies as either an attack on the opposition or as a re action based on an attack from the opposition. The researcher is interested in these candidate attack strategies via the emerging political communication platform, Twitter. Specifically, the researcher will seek to answer: What conditions are most conduc ive for political candidates' use of attack strategies on Twitter? The question will add to already compelling research done on the outcome of candidates' utilization of attack strategies and will explain the use of such strategies for a new commun ication outlet. Politically, the type of office holder who is more likely to adopt and use Twitter has already been established which will be discussed in the literature review However, there is little information on the content used to convey candidates' messages on this emerging social media platform. The proposed research question seeks to fill in this gap by using candidates' Twitter content from the most recent national election cycle. In order to stay current with the use of different campaign strate gies and add to previous attack strategy research, t he researcher will analyze attack strategies among 2012 United States Senate candidates in the general election This will update studies already done on negative campaigning wit h more recent information on the use of a communication outlet that has only become mainstream and emerged as a vital campaign tactic in the past few years.
Twitter Attacks of Senatorial Candidates 5 II. Literature Review: Overview Previous literature has made sound conclusions about the traditional political advertisement strategies that candidates use when seeking to be elected to public office. While the relevant literature has concluded on the use of campaign strategies for political advertisements printed in newspapers and aired on te levision, there is a significant lack of lite rature on the efficacy of various types of messaging by politicians on new social platforms like Twitter. With new social media constantly emerging, researchers have not yet studied how these traditional campaig n strategies are used on new media. Specifically, it is unknown if candidates utilize Twitter in the same way as traditional forms of communication. This is mostly due to the fact that Twitter has only recently come into mainstream use by cand idates for public office, having only been available for use in the C ongressional elections of 2006 (barely), 2008, 2010, and 2012. Here, the researcher seeks to showcase previou s findings about negative campaigning and its effects leading up to the age of attacks on Twitter Theories Previous literature has shed light on theories about why candidates chose certain political campaign strategies and the effects of such on voters. Specifically, r esearch on negative campaigning has utilized the mutually assu red destruction theory, which stems from international relations. For example, i n studying Senate campa igns from 1988 to 1998, Lau and P omper assumed that candidates employ one for one feedback according to their opponents' negativism ( 2001). Essentially, if a candidate were attacked, the opposition candidate would defend himself or herself by attacking back for a to taled one to one attack ratio. Also utilized to explain why
Twitter Attacks of Senatorial Candidates 6 candidates use attack strategies was the natural selection theory This theory explains that c andidates have an incentive to use arguments that evoke fear, anxiety, and anger in their opposition through the use of attack strategies (Jerit, 2004). Natural selection theory supports the idea that attack strategies are us ually more emotional and resonate with the electorate more, allowing for the fittest candidate with the most effective discredit of their opponent plan to gain office. T he theory of demobilization of the electorate as a n effect of negative campaigning is s till highly debated. While Freedman and Goldstein (1999) may have found evidence to the contrary (mobilization instead) other well established researchers have concluded the demobilization theory to be true (Ansolabehere, I yengar Simon, and Valentino, 19 94). There was also substantial support for formal theory or game playing theory offered by Skaperdas and Grofman in 1995, which explains that the campaign spending on negatives is done to prompt opponents' supporters to become undecided in their voting ch oice due to a discrediting statement This formal theory of negat ive advertising effects has greater support than the candidate att ribute explanation for a candidate's use of attack strategies. For example, Theilmann and Wilhite studied the political appli cability of these two different theoretical models and found that the decision makers reflect the behavior that theory predicts ( 1998). Thus, the decision by candidates to go negative in their advertising is done based on the behavior of the opponent, or a ttacking in this case, and not the attributes of the candidate. The current study seeks to build on theories that explain why candidates opt for attack strategies, but not so much on theories that measure voter behavior as a result of those utilized strate gies. Sources and Types of Data For data collection about negative campaign advertising, a handful of different tracking programs and websites have been used in previous literature The majority of these data sources
Twitter Attacks of Senatorial Candidates 7 consisted only of television and newspaper p olitical advertising, but website source data sets were also utilized. In television advertising, ad tracking systems for national broadcast news networks and local news networks have been developed over time For national networ ks, data sources used include the Polaris Ad Detector, which Freedman and Goldstein used to measure the impacts of negative advertising (1999). Other media studies for nationwide general elections have used the Julian Kanter Political Commercials Archive ( Damore, 2002). For more detailed information about political television advertisements from candidates such as their airtimes, outreach, station, and market, the Campaign Media Analysis Group (CMAG) served as one of the lar gest sources with detection in th e 100 largest national media markets ( Walter and Vliegenthart, 2010; Freedman and Goldstein, 1999) Additionally, the Wesleyan Media Project provided insight into the roots of political advertising and who funds them, including interest groups, the candida tes themselves, or the political parties. For example, Fowler and Ridout used the Wesleyan Media Project to examine the frequency of campaign advertisements from candidates (2010). Local broadcast outreach and television viewing habits were commonly determ ined through coding of local news stories' tone or individually distributed surveys in a regional area. (Franz and Ridout, 2008; Freedman and Goldstein, 1999). For newspaper data sources, most previous literature used the Lexis/Nexis database as a source. There were also a few instances of unoriginal data collection, where the researchers used data collected and coded by previous studies to supplement for unobtainable election data. For example, Lau and Pomper utilized coding of the tone in every election newspaper article from a previous 1992 study done on Senate elections from 1988 to 1990 to fill in early data they were not able to retrieve from Lexis/Nexis (2001). Others went directly to individual candidate s' websites in order to analyze strategy conte nt. Druckman, Kifer and Parkin analyzed the content
Twitter Attacks of Senatorial Candidates 8 of candidates' websites during three election cycles from document coding alone (2009). To gain insight into specific decision making for campaign strategies, Theilmann and Wilhite studied political consultants' choices in comparisons to theory by selecting cases from the American Association of Political Consultants (1998). While previous literature focused on sources from databases of traditional communication outlets, there are f ar fewer social media focus ed sources available that could provide such extensive data across decades of elections. These previous findings push the researcher to focus the current study on only sources from the most recent election in order to stay up to date with current elected officials' campaign strategies and to seek sources specific candidates' social media use. Methods of Analysis Much of the previous literature on negative campaign advertising relied on coders to perform content analysis, or to t urn qualitative data into quantitative measures for analysis. No matter the election type or year, content from qualitative advertisement content was turned into usually dichotomous or c ategorical measures of campaigns For example, Druckman, Kifer and Par kin's study of candidate s' websites utilized a new data source compared with most previous literature and focused on one of the most modern communication outlets studied regarding campaign strategies. These authors had coders analyze content across various pages on candidates' websites, and opted for a dichotomous measure of each website instead of attempting to count every single negative mention across the websites, which could easily result in measurement error (2009). In order to avoid such error, the c urrent study will also rely on dichotomous measures of tweets instead of measuring each attack within one tweet although multiple attacks within one tweet seems unlikely based on the 140 character limit.
Twitter Attacks of Senatorial Candidates 9 Defining the concepts analyzed is also an importa nt step in determining how candidates campaign For example, in both Franz and Ridout (2008) and Freedman and Goldstein (1999) advertisements were sorted into categories based on tone (positive, negative (attack), and comparative/contrast) which directly related to candidates' campaign strategies. While contrast ads were placed into a neutral third category and similarly defined as ads that mentioned both candidates, the definition of negative ads varied in the literature. Lau and Pomper defined negativity as just the mention of another candidate and not a specified, directed attack statement about the opposition (2001). Some of the literature defined a negative strategy as synonymous with an attack strategy, or a criticism of the opponent. This current stu dy will adopt this critical tone measurement to ensure reliability that truly negative statements will be coded as attack tweets, meaning that a discrediting statement must be included when an opponent is mentioned or the tweet will not be considered an at tack. I n addition to categorizing the coded strategies, previous literature typically categorized the communication outlet's content or type of negativity/positivity based on topic. For example, Walter and Vliegenthart also coded and showed importance to the content of the strategic appeals, which were based on if candidates mentioned a policy issue, moral values, or a character trait of the opposition (2010). T hey found 93 percent of negative appeals dealt with these issues. According to Damore, candidates also selected ads on issues that they appear more credible than their opponent s (2002). Previous research has also promoted salience as a method of issue selection in negative campaigning, meaning that issues the pubic deems important wil l likel y be the basis of a candidate attack For purposes of this analysis, it is not essential to categorize the content of tweets or the type of attack much like Walter and Vliegenthart did across multiple
Twitter Attacks of Senatorial Candidates 10 communication c hannels. Determining the issue in each t weet is beyond the scope of this analysis for the sake of time, but issue themes can be briefly described qualitatively. Findings Fowler and Ridout found that political advertising for the 2010 midterm elections was the most negative advertising in recent history and that growth in funding fu eled these attacks (2010) These authors also concluded that more than half of the ir studied advertisements after September 1 for the general election were negative The majority of the literature concurred that candid ates were, overall, more likely to go negative for their campaign strategies This study ultimately wants to see if candidates continued to go negative in 2012 using the newly popularize d social media outlet Twitter. "More often than not, candidates critic ize, discredit, or belittle their opponents rather than promoting their own ideas and programs" ( Ansolabehere, Iyengar, Simon and Valentino, 1994). Also, m ultiple researchers supported that the closer to Election Day, the more candidates were likely to use attack strategies. Due to th e heavy emphasis on negativity's proximity to Election Day in previous literature, the current study has adopted this finding as a basis for theory and hypotheses. Furthering this, Jerit conc luded that candidates mostly used arguments that were intended to evoke fear and anger in the opposition (2004). For repeated attacks, Damore conclude d that candidates' prior attack behavior did not stop them from continuing to attack at subsequent points in a campaign (2002). This researcher also concluded that candidates' negative campaign strategies were functions of their poll standings, o f opponent behavior at previous points in the campaign, of issue characteristics on the agenda, and of proximity to Election D ay Franz and Ridout also determined that no matter the measure of campaign tone, the
Twitter Attacks of Senatorial Candidates 11 measures all tap into the same general concept and did not yield significant differences in tone determination (2008) They gave reliability to the study of ca mpaign attack strategies as a whole in that all different measurements yielded similar findings However, there were distinctive partisan differences found in previous literature regarding the use of attack strategies. Theilmann and Wilhite found that 75 p ercent of Republicans believe negative advertising can be effective if candidate s lack funding and that about half of Democrats think candidates must have at least $50,000 before using an attack strategy (1998). G enerally, Republicans were found to be more likely to attack in previous literature. Lau and Pomper found that negative campaigning was also more likely to be utilized if the candidate is a challenger, if it was an open seat contest, or if the race was expected to h ave a tight margin of victory (2001). Druckman, Kifer and Parkin support this finding with their conclusion that challengers are "risk takers" and engage in such behavior to combat incumbency advantages (2009). Thus, e xperience in o ffice will be a major de terminant of attack tweet use and critical in determining the key differences between incumbents and challengers for this study. P revious literature has made sound conclusions on who uses negative campaign strategies and why they are utilized in traditiona l media outlets. T he researcher will utilize these conclusions in the formation of hypotheses. Expanding on Previous Literature Research on attack strategies has mostly focused on traditional communication outlets such as television and newspaper, sometim es studying several out lets at once or its effects on voter turnout. Unlike the previous l iterature, the current study focus es on just one communi cation outlet, Twitter, and focus es on certain selected cases within one election cycle. P articular Senate rac es in the 2012 general elections were selected updating previous literatures' findings from 2010 (or earlier) on campaigning This analysis is a comparative case study which
Twitter Attacks of Senatorial Candidates 12 is somewhat limited in that it is not all encompassing of every one of the 33 contested Senate seats in 2012. However, t he case study style can give depth to insights on attack strategies, especially for social media campaign tactics Like most quantitativ e analyses in previous lit erature, a content analysis was preformed to turn qualitative data from T witter into codified quantitative variables. The content and tone in candidates' tweets were studied for three months prior to Election Day, which is a long er study period than most previous literature. Also unlike most of the previous literatur e, the current study has a mixed method analysis with attention given to both qualitative a nd quantitative analysis on a hierarchic al basis: each individual tweet the candidates who sent the tweets and the races in which the candidates run. Because candidates' use of Twitter is so new, much of research has come to the conclusion that more time was needed to determine its manner of use. Thus, updating the findings on t he conditions under which attack strategies are utilized for a newly politically utilized communication outlet is the focus of this study. This study also intends to show the evolution of candidate Twitter use from its creation to 2012 and how it h as mai nstreamed into candidates' usual campaign practices. III. Theory: The researcher seeks to determine the conditions under which candidates for office use attack strategies via Twitter in campaigns leading up to a general election The answer to this question can provide circumstances under which candidates utilize attack strategies and can suggest the posit ive or negative effects of such used in the campaign process Conditions like the social media climate will continue to evolve, and it seems likely that conditions will continue to change in ways that may or may not impact strategies" ( Druckman, Kifer and Parkin, 2009). Even if they are negative, s uccessful campaign s are always sought after by American public
Twitter Attacks of Senatorial Candidates 13 officials despite ever changing media conditions. Since a ttack strategies have shown to emotionally re sonate with the public, the researcher proposes the theory that candidat es are utilizing attack strategies (tweets for this study) in an effort to raise negatives of the opposit ion and increase their own chances of being elected to public office. Essentially, attack strategies are alternative s and desperation when candidates do not think they can win elections based on merit alone they m ust discredit the opposition as well Thi s is a theory used to explain campaign rhetoric in several studies of negativ e attack strategies, but this study applies it to a new communication outlet Twitter. Twitter may not be used for the same purposes as traditional campaign communication outlets and take on a different role in negative campaigning or lack thereof. The researcher seeks to find the conditions under which these attacks strategies do occur through candidates' Twitter accounts. To provide this robust analysis about the variation in attack strategy use, t his study require s a data set with a hierarchy: While the actual un it of analysis is a single, independent tweet, this unit of analysis is clustered within each candidate. Further still candidates are clustered within each Senate rac e. IV. Data G athering Variable Introduction The following variables were recorded for analysis in this study's data gathering process: W hether a given tweet is an attack tweet or not is the dependent variable upon which the entire study relies The following characteristics, used as independent variables for this study, were also recorded in the data gathering process: Time: Three months total, recorded by date and categorized in months leading up
Twitter Attacks of Senatorial Candidates 14 to Election Day (August 6, 2012 through November 6, 2012) Party affiliation: Republican or Democrat (this is also indicative of minority/majority party status for the Senate) Experience in office: Incumbent for Senate seat or new candidate for Senate seat/open seat race Competitiveness: The polling spread (margin) between candidates in a race State population: 2012 state population in which the Senate race took place Case Selection T he researcher use d various politically oriented online sources to select case s from the 2012 election cy cle These sources were also used to record characteristics about each Senate race and candidate ( the variables mentioned above) Particular criteria were met in selecting cases that theory and previous literature have deemed important: 1. Candidates from 2012 contested Senate races Presidential ele ctions do not have enough possible cases using Twitter having had only two races since Twitter's initial launch. U.S. Senate races have higher profile s than races in the U.S. House of Representatives or r aces wi thin state legislatures, which permits for more readily available inf ormation about these races Previous f indings support that members of the Senate have also been more likely to adopt Twitter than other members of Congress due to a lack of constituent connection across such a large a rea an entire state To find the contested seats and relevant information about each candidate, t he researcher use d the United States Election Atl as website However, the Real Clear Politics website served as a source for the monthly spread of competitive polling data for each race between candidates le ading up to the
Twitter Attacks of Senatorial Candidates 15 general election This website summarizes many polls that predict election outcomes to produce a competitive spread between candidates at distinct points in time (usually each month), which will be used as the measure of competition for thi s analysis. The polling margin between candidate s given as close to May 6, 2012, as possible was utilized as a competitive outlook on the race six months before Election Day. 2. A equal number of Republican and Democrat candidates A total of 10 races and 20 candida tes (a Republican and a Democrat f r o m each race) were used for the current study. An Internet search for candidates' websites, which indicates if they had a 2012 campaign and if their seats were up for election was conducted. By selecting an eve n amount of party affiliated candidates, this assures there is minimal partisan based selection bias. This also allow ed the researcher to see if previous findings' determinations about the campaign strategies of each political party apply to candidates' us e of Twitter. The United States federal government's official Senate website was utilized to verify political party affiliation for Senate incumbent s, while personal websites were used to verify part y affiliation for candidates who were not Senate incumbents or just new candidates in 2012. Candidate s w ho served in another office such as the U.S. House are not considered incumbents in this study due to the varying nature of different levels of political office. 3. Active official account Twitter use In order to measure candidates' reliable use of Twit ter, a threshold to determine active use was set. The current study adopt ed a threshold that candidates must tweet at least 12 times, equivalent to about once per week, during the three months prior to the general election in order to be considered active. This threshold account s for gaps in tweeting by using a total
Twitter Attacks of Senatorial Candidates 16 rather than a weekly or monthly quota. An y races with candidates that do not meet this threshold were not considered as a potential case to use in this analysis. The reasoning behind selecting candidates who are active on Twitter is that more tweets offer more chances for attack tweets to occur a nd, thus, provide more in depth insights into th e attack strategies for Twitter The current study use d the official Twitter website and Topsy a social analytics website that allows users to go back as far in time as desir ed in the Twitter archive, as source s to determine how often elected officials tweeted Additionally, the researcher verified the Twitter URL before selecting it for the study. If candidates had several Twitter accounts, the one used to the sole p urpose of campaigning, usually called "T eam X (name of candidate)," was analyzed. Due to the variability in the number of tweets over time the current study analyze d tweets over a three month period of time Based on the criteria above, t he researcher was left wit h the following Senate races as cases: Arizona Jeff Flake (R) and Richard Carmona (D) California Elizabeth Emken (R) and Dianne Feinstein (D) Connecticut Linda McMahon (R) and Chris Murphy (D) Delaware Kevin Wade (R) and Tom Carper (D) Massachusetts Scott Brown (R) and Elizabeth Warren (D) Michigan Pete Hoekstra (R) and Debbie Stabenow (D) Minnesota Kurt Bills (R) and Amy Klobuchar (D) Nebraska Deb Fischer (R) and Bob Kerrey (D) Utah Orrin Hatch (R) and Scott Howell (D) West Virginia John Raese (R) and Joe Ma n chin (D)
Twitter Attacks of Senatorial Candidates 17 These cases represen t candidates with varying : location within the United States experien ce in office, state population, and competitiveness. This variation makes the 10 cases more representative of all 33 contested Senate races in 2012. V. Data: V ariables and Measurement This study cover s three clustered levels of analysis : individual tweets are at the most basic level, candidates at the second level and races at the third level. When looking at the independent variables, t ime is expected to vary around the fundamental level of a single tweet. Time is an important variable to include for theoretical reasons. There is a theoretic expectation that there will be an increasing sense of candidate despe ration as it becomes closer to Election Day. Both e xperience in office and party affiliation vary at the candidate level. Finally, competitiveness and state pop ulation vary at the race level. The researcher performed content analysis on each tweet for the dependent variable. Then, information was gathered about each candidate and race for the independent variables. The researcher used Topsy, official Twitter account pages, and the other previously described sources for case selection in the three months leading up to the general election on November 6, 2012. D escriptive, qualitative information about each Senate race and its candidates' Twit ter activity frame the statistical data. Prior to quantitative analysis the researcher focus ed solely on interpreting the qualities o f the races. These qualities include characteristics of the race such as the candidates' political background, tweet patte rns, incumbency term, percentage of votes received, and any scandal. This part of analysis is purely quali tative, asi de from election outcome percentages of votes received, and premises condition s at the three clustered levels prior to statistical analysis Quantitative interpretation took the form of logistical regressions to accou nt
Twitter Attacks of Senatorial Candidates 18 for the lack of observation (tweet) independence) and c andidate heterogeneity. Ultimately, t his is a time series cross sectional (across three months) design that looks at dif ferent levels of analysis ( individual tweets, each candidate and each race ). Each variable was measured and analyzed in the following ways: Dependent Variable: Whe ther a given tweet is an a ttack tweet This variable is expected to vary across time, between candidates, and between races, which explains why a multi level data set is necessary. An attack is defined as a negative critique of the opposition (not a positive or indifferent mention). Any twe eted links to articles or documents with negative m entions of the opponent are also considered attack tweet s If the candidat e had one negative criticism of the opposition within a tweet, then the researcher counted that particular tweet as an a ttack tweet. As a dichotomous variable, atta ck tweets were rec orded as 1 and non at tack tweets were recorded as 0. T he gathered data defies the assumpt ion of being continuous thus this variable was analyzed as an odds ratio. An odds ratio measures the odds of an event happening relative to the odds of an event not h appening. An odds ratio can never be negative, but one less than 0 indicates a decrease in the odds of an event happening. Independent Variables: Time: Each tweet was coded through time by date and was split into categories with numbers assigned upward for each month according t o the date of the tweet: Coded 0 for Augus t 6 through September 6, coded 1 for Sept ember 7 through October 6, and coded 2 for October 7 through November 6 at midnight. By analyzi ng e ach month separately, this allow ed the researcher to determine variation in attack tweet use through time and as it gets closer to Election Day. The specific date of each tweet was also recorded to look for any
Twitter Attacks of Senatorial Candidates 19 general tweet patterns in qualitative ana lysis. This variable allows the researcher to measure candidates' desperati on through time in via their use of attack tweets Party affiliation: This is a dichotomous variable coded as 1 for Republicans (2012 Senate minority) and 0 for Democrats (201 2 Se nate majority). Races with I ndependents and other third party candidates receiving a significant portion of the votes were already omitted in case selection. Includ ing these other candidates would make it difficult to determine the "opposition." The opposition for Republicans is Democrats and vice versa, but adding a third party candidates can complicate this definition. Also, Republicans and Democrats are the two dominant political parties that ha ve con trolled Congress in modern day. T hus, they are the most relevant to analyze. Testing the relationship between p arty affiliation and the odds of attack tweets allow s the researcher to analyze both the differences among political parties' use of attac k tweets and minority/majority parties' use of atta ck tweets. Experience in office: Incumbents were coded as 0 and challengers or open seat candidates were coded as 1. Some races had two challengers and no incumbents because it was an open seat rac e or b eca use both candidates were not the politicians who serve d in that Senate seat just prior to the 2012 election Experience in office is an important variable considered in nearly every previous study on candidates' negativity in campaigning. This data allow s th e researcher to determine the campaigning differences among the types of candidates who run for office and account for variation among the candidates' experience in office s pecific to the Senate. Competitiveness: The competit iveness between candidates was assessed from the polling marg ins between candidates on the RCP website for the predicted outcomes of 2012 Senate
Twitter Attacks of Senatorial Candidates 20 candidates. For example, common polls from RCP include SurveyUSA and Quinnipiac. The competitiveness of a race was based on the results of RCP posted surveys taken as close to six months prior to the general election as possible. RCP creates a margin from this survey data, shown as polling points in one candidates favor. For each polling spread closest to May 6, 2012, candidates r eceive d a positive or negative code based quantitatively on how much they are winning or losing compared to the other candidate in that race. Hence, this spread was coded with the same number for each candidate, but the winning candidates receive d a positive competitive spread ( expected election win at that po int in time) and the loser had a negative competitive spread ( expected election loss at that point in time). These spread s aid in explain ing the attack tweet use in publically perceived compet itive races compared to those that are not viewed as competitive races. State Population: The state population was gathered from Bureau of Economic Analysis' population estimates by state for 2012, which is derived from the U.S. Census. This data can sug gest whether attack tweets are more common for states with a larger population or states with a smaller population. The researcher was pleased that states with drastically varying populations were present after the vetting process. VI. Hypotheses After a content analysis to quantify tweets and the independen t variables, the following hypotheses were analyzed based on the tweets of candidates within U.S. Senate races during the 2012 election cycle. Similar hypotheses stem from previous negative camp aigning research but are applied to a new communication outlet, Twitter.
Twitter Attacks of Senatorial Candidates 21 Hypothesis 1: The odds of an attack tweet being issued will increase the closer in time it is to Election Day. This hypothesis states that the odds of attack tweets being issued a re dependent on time. Time i s defined by each date and categorized into months for analysis purposes. The researcher expects to see the odds of negative tweets to be the greatest in the month p ri or to Election Day (October 7 through November 6 at midni ght) or in the month categorized as 2 This is the most critical time when candidates attempt to mob ilize voters, hopefully in their favor. Following from the stated theory, they will use attack tweets in hopes of having a profound effect on voters that will d iscredit their opponents and remind voters of this discredit, as it gets closer to Election Day. Hypothesis 2: The odds of an attack tweet being issued by a challenger or open seat candidate are greater than the odds of an incumbent issuing an attack tweet. This hypot hesis suggests the odds of issuing attack tweets dependent on a candidates seat hol ding status in the Senate. The researcher expects that the challengers or open seat candida tes wil l have greater odds of issuing attack tweets than the odds of incumbents issuing attack tweets Similar to minority party candidates challengers and open seat candidates have a higher burden to overcome than those already dominating an office Incumbents also have widespread name recognition that challengers must overcome or use to their own advantage by attacking easily recognized opponents Chal lengers also have a fundraising burden, which is usually significant ly less for incumbents based on contributions during their previous, successful election cycle. Hypothesis 3 : The odds of candidates from the Senate minority party (Republicans) issuing an attack tweet are gr e ater than the odds of candidates from the Senate m ajority party
Twitter Attacks of Senatorial Candidates 22 (Democrats) issuing an attack tweet T his hypothesis use s the odds of issuing attack tweets dependent on minority party affiliation in the Senate. The minority party in the 2012 Senate was the Republicans, and the majority party was the Democrats. In order to change the status quo, Republicans will use attack tweets to discredit their Democrat opposition in hopes of winning elections and regaining a majo rity status in the U.S. Senate. T he minority party has more of a reason to discredit the opposition in hopes of taking back control of th e Senate, as the theory states. This means that its members should have more attack tweets than member s of the majority p arty. Hypot hesis 4 : The odds that an attack tweet is issued will be greater in races with a smaller spread of polling data (competitive margin) between candidates or greater competitiveness. T his hypothesis use s the odds of attack tweets dependent on the competitive spread of an election. Competitiveness as a variable is defined through real time polling data obtained via RCP which summarizes other common polls' election predictions at selected points in time throughout an election Polli ng data typically does not vary too much between six months prior to Election Day and Election Day itself, creating accurate polling data to base competitiveness on for this study If a candidate is worried about his or her ability to win an election he o r she will utilize attack tweets to discredit the competition. By discrediting the opposition, candidates hope that this will make the competitive spread larger in their favor or at least narrow a competitor's advantage in the race. Hypothesis 5 : The od ds that an attack tweet is used will increase as a state's population increases. This hypothesis uses the odds of attack tweets dependent on a state's population. The
Twitter Attacks of Senatorial Candidates 23 hypothesis predicts that states wit h a greater population will host candidates who have greater odds of utilizing attack tweets. This hypothesis is based on theories from previous literatures' that Senate candidates are not as closely connected to their constituents because those who they serve are spread out over an entire state. To make up for the lack of constituent connection, Senate candidate s, especially those from populous states, will have to employ strategies that resonate with voters wh ich are traditionally negatives ( or attack tweets in this case ) It will be particularly interesting to examine the variation in attack tweet use between the California (large state population) Senate race and the Delaware (small state population) Senate race. VII. Uses of Data Data Interpretations The current study is a comparative case study of political candidates' attack strategies via Twitter through time in the three months leading up to Election Day. Part one of the analysis is the qualitative interpretation for the race and characterist ics that frame each election The qualitative interpretation take s into account characteristics from each race and its candidates as a narrative premise. Part two is the quantitative analyses, based at the fundamental level of each tweet. For the quantitative analyses, the researcher sought variation in attack tweet use based on the independent variables to make sound conclusions on which conditions (out of the independent variables) are most conducive for attack strategy use. T his ca se study encompass ed about a third of the available population of contested Senate seats in 2012. Trying to assess the Twitter content for each candidate in each of the 2012 contested Senate seats could have result ed in document measurement error due to such an abundance of in forma tion. It was not feasible to analyze that many tweets within a reasonable time period, consider ing 10 races contained almost
Twitter Attacks of Senatorial Candidates 24 6 ,000 tweets alone. To account for differences in what is considered an atta ck strategy, the researcher look ed at three month s worth of tweets twice, once while determining if tweets are attacks and once while recording the date of each tweet T his yielded greater reliability for the case study 's data collection T he researcher was then able to analyze the attack tweets odds. In this case, the o dds reveal the probability of an attack tweet relative to the probability of a non attack tweet A robust statistical analysis of the tweets was then perfor med: Logistic regressions were run to determine the odds ratio of attack tweets occurring relative to the odds of attack tweets not occurring The first logistic regression run did not take into account the observation ( tweet ) dependence or candidate heterogeneity which affects the standard errors in the models The nature of data indicates a positive autocorrelation. For example, candidates who use negative tweets may be more likely to use them in the future. However, the researcher included a second logistic regression to take into account the lack of observati on independence The second regression also used a random intercept for the candidate level cluster an d dummy variable for each race level cluster. By running this next logistic regression, the researcher gave the findings from the study greater validity by taking into account differences between candidates that the researcher cannot control. T hese analyses were aimed at find ing the most conducive conditions under which attack strategies are used. Case Study Advantages and Disadvantages: Th e researcher does expect this case study to give in depth information about what conditions motivate the selected candidates to use attack strategies against their opponents within a new media outlet, Twitter. Specifically, the researcher hopes to find the conditions under which
Twitter Attacks of Senatorial Candidates 25 attack tweets occur via a case study method to analyze the characteristics of 10 races 20 candidates and individual tweet s An all inclusive method like population statistics would supply potential outlier candidates who do not us e Twitter or barely use Twitter, which could dramatically skew the findings. If all contested seats were studied, this would make attack tweets appear much less frequent because not every candidate would use Twitter, totali ng zero attack tweets across thre e months. The researcher did not want to hide potential insights in mounds of irrelevant ca se data. Although a cas e study cannot be generalized an experiment or other method would not replicate real world conditions. This case study sought to study real world campaign decisions and actual cand idates. A case study also allowed the researcher to study these cases over time and analyze not only how often attack strategies were used, but also where in time there is variation in attack strategy use. T he current study also has analyti cal leverage because it studied the conditions from each utilized race qua litatively i n addition to quantitative analysis based on the gathered tweets. However, there are many disadvantages to selecting specific ca ses. For case studies, i t is difficult to statistically find the average conditions under which attack strategies occur across each tweet, each candidate, and each race Also, the qualitative analysis of each race cannot be generalized due to its case specific narratives, which do not apply to every race. While the lack of research on candidates' attacks within a new ly mainstreamed media outlet gives merit to any research for this important negative campaigning question, it is impossible to make th ese findings branch across all candidates with certainty due to the specific case selection. Reliability in coding from qualitative based sources continue s to threaten this case study, despite being accounted for by the researcher. There are also validity flaws, as the researcher may not be including some relevant independent variable. However, the researcher tried to include all
Twitter Attacks of Senatorial Candidates 26 relevant data for independent variables within reason For external validity, it is im possibl e to assure that these 20 candidates selected within 10 races in just three month s time are representative of all candidates within all r aces at all times. However, as King, Keohane and Verba arg ue, the researcher select ed cases based on the previously specified criteria and on the differen t values of the independent variables (not the dependent variable) (1994). While the researcher cannot confirm external v al idity, the cases selected vary on the independent variables to ensure a representative a sample as possible. VIII. Qualitative Analysis for Each Race Note: All background information was taken from the candidates' official campaign websites via Real Clear Politics All race election results are from the New York Times official 2012 Senate election results 1. Arizona Senate Race : Candidate Backgrounds : Before winning the Senate seat, Jeff Flake (R) was a U.S. Representative in Arizona's 6 th Congressional District. Flake is also a member of the Church of Jesus Christ of Latter day Saints. Dr. Richard Carmona (D) was the U.S. Surgeon General. Both had assumed a position of considerable political infl uence before running for Senate. Carmona is best known for Surgeon General's report revealin g the harmful effects of second hand smoke. This Senate seat was left vacant after Senator Jon Kyl's retirement. Tweet Pattern: Carmona tweeted much more often than Flake. Instead of being the candidate from the minority party in the Senate (Republican) who tweeted more often, candidates of the minority party of just that particular state that tweeted more often
Twitter Attacks of Senatorial Candidates 27 and tweeted more attacks (Democrat). The twe ets were mostly spread o ut evenly over the three months. Attacks tweets focused on similarities between the Senate candidates and their party's presidential candidates. Obamacare" and endorsements from members of the opposing party were specifically mentioned in several attack tweets. Election Outcome : This election was moderately competitive compared to the other selected races Flake received 49.7 percent of the total votes, and Carmona received 45.8 percent of the total v otes. This was one of the c losest election outcomes of all the cases studied 2. California Senate Race: Candidate Backgrounds : Dianne Feinstein (D) is the long time incumbent in this race having served in the Senate for more than 20 years She is currently the Chairwoman of the Senate's Select Commit tee on Intelligence and Chairwoman of the International Narcotics Control Caucus. Elizabeth Emken (R) also previously ran for and lost a U.S. House seat in California. She is currently an offici al lobbyist for Autism Speaks. Tweet Patterns: Emken, as a challenger to a senior senator, probably had one of the biggest hurd les to overcome in order to win the Senate seat. Emken tweeted hundred s of times more than Feinstein. Most of the attack tweets mentioned Feinstein not agreeing to debate Emken or legislative failings in the state under Feinstein's terms in office. Feinstein never sent an attack tweet leading up to Election Day. Election Outcome: This was one of the leas t competitive races selected for the case study. T he outcome: 61.6 percent of th e votes went to Feinstein and 38.4 percent of the votes went to Emken. However, this was the closest a candidate has come to
Twitter Attacks of Senatorial Candidates 28 defeating Feinstein since she was elected to the Se nate. 3. Connecticut Senate Race : Candidate Backgrounds : This race was for an open se at, left open by retiring Senator Joe Lieberman. Linda McMahon (R) was the former CEO of the World Wrestling Enterta inment a position that was scrutinized when she ran for the Connecticut Senate in 2010. Chris Murphy (D) served in the U.S. House t he Connecticut Senate and the Connecticut House prior to running for the U.S. Senate. Tweet Patterns: McMahon ran a negative campaign in general including her tweets Murphy played largely defensive in his tweets. More tweets were sent in the third month, closest to Election Day, than any other month studied in this race. This race also had a large number of tweets from both candidates compared to the other races. McMah on's attack focused on Murphy's lack of a plan, especially for job creation, in America. Murphy's tweets criticized McMahon's plans as a senator that some tweets mentioned a petition against a Social Security initiative that McMahon endorsed Election Out come: This was one of the most competitive races chosen and a race where the predicted e lection winner fluctuated in the months leading up to Election Day Murphy ended up with 55 .2 percent of the votes and Mc Mahon ended up with 43.2 percent of the votes in this traditionally Democrat dominated Northeastern state 4. Delaware Senate Race : Candidate Backgrounds: Tom Carper (D) is the long time incumbent, having served in this Senate seat since 2001. Carper is also the Chairman of the Senate Committee on Homeland Security and Governmental Affairs. Kevin Wade (R) is the
Twitter Attacks of Senatorial Candidates 29 founder of Philadelphia Control Systems, Inc. a l arge engineering firm, and this election was his first run i n politics. Tweet Patterns: Wade tweeted more overall and atta cked more as a challenger. Wade's attacks focused on what negative e ffects electing Carper would have on Delaware and its voters. Wade took on the "shiny new penny" theme for America during his campaign. Carper did not send an attack tweet at all. Delaware despite being the state with the smallest population in this case study did not have fewer tweets from its candidates compared to ca ndidates from states with large populations Election Outcome: Expected for an incumbent in a Democrat dominated state, Carper took 66. 4 percent of the votes. Wade took just 29 percent of the votes. 5. Massachusetts Senate Race : Candidate Backgrounds : Scott Brown (R ) was the incumbent in this race, as he was specially elected to the Senate seat in 2010 af ter the death of Senator Ted Kennedy. Elizabeth Warren (D ) held appointed positions as the Special Advisor for the Consumer Financial Protection bureau and the Chairperson of the Congressional Oversight Pane l. She is also an author and a professor of law. Tweet Patterns: Both candidates had similar tweeting habits and attacked almost an equal amount. Both candidates barely ever mentione d the other candidate in any of the tweets studied. Despite polling as the most competitive race six months prior to Election Day, the attack tweets were not more common than in other races. Attacks centered around the debates between the two candidates. E lection Outcome: This was the most competitive race st udied w ith a polling margin of 0 at six months before Election Day. Warren received 53.7 percent of the votes,
Twitter Attacks of Senatorial Candidates 30 whil e Brown received 46.3 percent. This was the only race studied where the challenger defeated an incumbent. 6. Michigan Senate Race : Candidates Background: Debbie Stabenow (D) is the incumbent for this Senate seat, having assumed the position in 2001. She is the Chairwoman of the Senate Agriculture Committee and was formerly a U.S. House member. Pete Hoekstra (R) was a member in the U.S. House before running for the Senate. He has also run for g overnor, but lost that election as well. Hoekstra aired a controversial negative television advertisement against Stabeno w during the Super Bowl in 2012 which called his opponent Debbie "Spend It Now" and focused on immigrants taking jobs away from Americans. Tweet Patterns: Stabenow and Hoekstra both ran heavily negative Twitter campaigns in this race. They tended to tweet negatively around the same times throughout the three months studied, supporting the mutually assured de struction theory. These two candidates ha d more negative tweets than all of the races studied. Attacks focused on debate dodging and national budget issues. Election Outcome: Stabenow won with 58 percent o f the vote and Pete Hoekstra won 38 percent of the total vote. Despite frequent attacks from both candidates, the incumbent overwhelmingly prevailed in this race. 7. Minnesota Senate Race : Candidate Backgrounds: Kurt Bills (R) was a member of the Minnesota House of Representatives and a high school economics teacher. Amy Klobuchar (D) has served in this Senate seat since 2007. She was the first woman elected as a Minn esota
Twitter Attacks of Senatorial Candidates 31 senator and formerly served a county attorney. She has even been mentioned as a potential presidential candidate. Tweet Patterns: Klobuchar did not have a nega tive tweet during the entire three months studied. Bills tweets much more often than Klobuchar, but did not attack nearly as much as the other challengers Tweets were evenly spread out during the three months. These candidates had a fairly positive campaigns compared to the other candidates ones studied. Both Twitter campaigns focused on constituent support. Election Outcome: Klobuchar won with 65.3 percent of the total vote, and Bills had just 30.6 percent of the total votes. With such a large victory margin, this race supported incumbency as a strength and asset to candidates. 8. Nebra ska Senate Race : Candidate Backgrounds: Deb Fischer (R) served in the Nebraska state legislature. She is currently the Chairwoman of the Transportation and Telecommunications Com mittee, the Executive Board, and the Revenue Committee. Bob Kerrey (D) was Nebraska's s enator from 1989 to 2001 and Nebraska's governor from 1983 to 1987. This was a unique candidate set up as b oth of these candidates came in with extensive political experience, but neither were current incumbents. Tweet Patterns: Fischer barely sent any attack tweets and tweeted much less often than Kerrey. Kerrey tweeted a lot during the primary election in early August 2012. Most of his tweets were concentrated in the first two months of the time studied with very few tweets in the last month. His negative, attack tweets focused on refuting Fischer's claims made at debates between the two candidates. Election Outcome: Fischer took 58.2 percent of the votes, while Kerrey took 41.8
Twitter Attacks of Senatorial Candidates 32 percent. Like Fischer, most senators from Nebraska have been Republican candidates. Despite having served as a senator in the 1980s, Kerrey was not re elected. 9. Utah Senate Race : Candidate Backgrounds: Orrin Hatch (R) is the incumbent for this seat, having served in it since 1977. He served on the Senate Judiciary Committee and on the Senate Finance Committee. Scott Howell (D) was formerly a member of the Utah state Senate fro m 1989 to 2000 and an IBM executive. Both ca ndidates are Mormon like many Utah constituents Tweet Patterns: Hatch did not tweet at all during the third month studied and barely tweeted compared to Howell. He also did not tweet any attacks. Most of Howell's tweets centered on the se cond month studied and included some attacks. Howell frequently tweeted about how to finance his campaign (in order to defeat his opposition) and mentioned individual contributors to his campaign on Twitter. Election Outcome: Hatch won a seven th term in t he Senate with 65.2 percent of the v otes. Howell received just 30.2 percent of the votes. 10. West Virginia Senate Race: Candidate Background: Joe Ma n chin (D) is the incumbent, having assumed this Senate seat in 2010 after a special election after Robert Byrd died in office. John Raese (R) has run and lost this Senate seat in 1984, 2006, 2010 and 2012. He has also lost a primary election of West Virginia governor. Tweet Patterns: This race has the least tweets of any race studied with about 100 tweets pe r candidate for the three months studied. Ma n chin did not attack in any of
Twitter Attacks of Senatorial Candidates 33 his tweets, but Raese attacked in many of his tweets. Raese attacked Manchin in tweets by comparing him to President Obama in the negative. These attack tweets included links to neg ative television advertisements. Election Outcome: Ma n chin won and received 60.5 percent o f the vote. Raese received 36.5 percent of the vote. IX Logistic Regression Model and Interpretation Logistic Regression Model: Twitter Attack Strategies from 2012 Senate Candidates Ind. Variables Odds ratio B S.E. Month of Tweet 1.222 ** .201 .043 Challenger 2.229 ** .801 .132 Minority Party 1.504 ** .408 .076 Competitiveness .995 .005 .002 State Population (per million) 1.033** .032 .003 Constant .062 ** 2.777 .126 B = logged odds ( odds ratios before exponentiation ) ; S.E. = standard error; = P value st atistical significance at the 1 level ; ** = P value statistical significance at the .05 level. Logistical Regression Table Analysis: Time : The odds ratio for time indicates a positive relationship. This means, p er month closer in time to Election Day the odds that an attack tweet was issued increased by a factor of 1.2 or were more likely by 22 percent all else equal
Twitter Attacks of Senatorial Candidates 34 Significance: This odds ratio is stat i stically significant at the 05 level. Thus, the null hypothesis that time and the odds of an attack tweet are independent of each other, can be rejected. Experience in O ffice : The odds ratio for experience in office (as a challenger) indicates a positive relationship. This means that t he odds of c hallengers or new candidates issuing an attack tweet were greater by a factor of 2.2 or 1 2 0 percent more likely relative to the odds that an incumbent issued an attack tweet all else equal Significance: This odds ratio is statist ically significant at the .05 level Thus, the null hypothesis that experience in office ( as a challenger) and the odds of an attack tweet are independent of each other can be rejected. Party Affiliation : The odds ratio for party affiliation (with the minority party ) indicates a positive relationship. This means that t he odds of a minority party candidate (Republican ) issuing an attack tweet was greater by a factor of 1.5 or 50 percent more likely relative to the odds that a majority party candidate (Democrat) i ssued an attack tweet, all else equal Significance: This odds ratio is sta tistically significant at the .05 level Thus, the null hypothesis that party affiliation (minority party) and the odds of an attack tweet are independent of each other can be rejected. Competitiveness : The odds ratio indicates a negative relationship for competitiveness. Per 1 point increase in the competitive margin t he odds that attack tweets were issued decreased by a factor
Twitter Attacks of Senatorial Candidates 35 of .99 5 or were .5 percent less likely, all else equal. Significance: This odds ratio is not stat istically significant at the .05 level but it is at the .1 level. Thus, the null hypothesis that competitiveness and the odds of an attack tweet are independent of each other can be rejected. Population ( per million): The odds ratio indicates a positive relationship fo r the state population variable This means that, p er million person incr ease in the state population the odds that attack tweets were issued increased by a factor of 1.03 or 3 percent all else equal Significance: This odds ratio is stat istically significant at the .05 level Thus, the null hypothesis that state population (millions) and the odds of an attack tweet are independent of each other can be rejected. X. Considering Problems Associated with Data: The data used in this study, as previously mentioned, was clustered. T he collected data (individual tweets) breaks the assumption that observations are independent from one another which the researcher must control for in order to have accurate results Similarly, candidates are no t independent of one another in the same race because they are competing for the same Senate seat. Simply, there is unit hete rogeneity, meaning that candidates are just different from each other and are just naturally more or less likely to tweet The potential ways to deal with data clusters: One can model the cluster, or one can control for it by adjusting the standard errors. This study seeks to model the cluster using a variance compon ent, which accounts for the hierarchy of tweets within candidates. The previous statistics in the logistic regression model above do not satisfy this independence condition and do nothing to acc ount for unit
Twitter Attacks of Senatorial Candidates 36 heterogeneity (also called lack of observation independence). Thus, t his next model an unobservable effects model, accounts for differences bet ween candidates, such as if a candidate is inherently more negative or more lik ely to use attack tweets based on previous negative Twitter usage The largest source of clustering in this study is within the candidate level cluster. It would be most efficient to show random intercepts for both the candidates and races, but the researc her thought that displaying dummy variables for the race level cluster would be useful in display ing differences between the states. The researcher is interested in state negativity, as well as candidates' negativity. It was most efficient to include dummy variables for the states rather than candidates, considering only 9 dummy variables were necessary compared to a necessary 19 dummy variables for candidates Also, note that the B (log ged odds) column was left out in this next model to make room for the statistical significance, which is more important in making conclusions for this study The logged odds are difficult to understand and not needed once they are converted into odds ratios using the exponent function.
Twitter Attacks of Senatorial Candidates 37 XI. Unobservable Effects Model and Interpretation Random Intercept for Candidates and Dummy Variable s for Races : Twitter Attack Strategies from 2012 Senate Candidates Ind. Variables /Races Odds Ratios S.E. Sig. Month of Tweet 1.172 ** .04431 .000 Challenger 1.350 .68437 .662 Minority Party 1.047 .38750 .905 Competitiveness 0.938 ** .01707 .000 California 1.522 .75939 .580 Connecticut 3.062 .64419 .082 Delaware 0.354 .91874 .258 Massachusetts 2.312 .72725 .250 Michigan 4.380 ** .73351 .044 Minnesota 0.090 ** .86478 .005 Nebraska 0.891 .68153 .866 Utah 0.010 ** .00306 .022 West Virginia 0.341 .91676 .241 Constant 0.069 ** .83508 .001 = P value statistical signi ficance at the .1 level; ** = P value statistical significance at the .05 level. Note: Arizona is the reference category for races. Population variable has been dropped because population and the dummy variables for the state are collinear.
Twitter Attacks of Senatorial Candidates 38 Variance Component for Candidate s 0 .36992 The variance component for candidates is an extra parameter included in this table to estimate how much (independent of) each independent variable, candidates typically vary in their use of attack tweets. Essentially, it i s how much candidates typically vary in their propensity to use ne gative tweets (controlling for) the independent variables The variance component also ensures the accuracy of the model's odds ratios. Interpretation for Random Intercept: Time: The odds ratio for time indicates a positive relationship. This means, per month closer in time to Election Day the odds that an attack tweet was issued increased by a factor of 1.2 or were more likely by 20 percent all else equal. This is equal to the odds ratio before the researcher took into account the differences between candidates with a random intercept and differences between races with dummy variables Significance: This odds ratio is st atistically significant at the .05 level Thus, the null hypothesis that time and the odds of a n attack tweet are independent of each other, can be rejected. Experience in Office: The odds of challengers or new candidates issuing an attack tweet were greater by a factor of about 1.4 or were about 40 percent more likely, relative to the odds that an incumbent issued an attack tweet, all else equal. This is much lower t han the original logistic regression's odds of an attack tweet issued by a challenger relative to the odds of
Twitter Attacks of Senatorial Candidates 39 an attack tweet issued by an incumbent. Significance: This odds ratio is not statistically s ignificant at the .05 or .1 levels. Thus, the null hypothesis that experience in office (challenger) and the odds of an attack tweet are independent of each other can not be rejected, like it could in the original logistic regression that did not take into consideration the data clustering. After this more thorough analysis, the challenger variable was not as significant in the role of attack tweet usage as presumed. Party Affiliation: T he odds of a minority party candidate (Republican) issuing an attack tweet were g reater by a factor of 1 .05 or 5 percent more likely, relative to the odds that a majority party candidate i ssued an attack tweet (Democrat), all else equal. Significance: This odds ratio is not statist ically significant at the .05 or .1 levels. Thus, the null hypothesis that party affiliation (minority party) and the odds of an attack tweet are independent of each other can not be rejected. Competitiveness: Per 1 point in crease in the competitive margin the odds that attack tweets were issued decreased by a factor of .94 or less likely by 6 percent, all else equ al. Significance: This odds ratio is statistically s ignificant at the .05 level Thus, the null hypothesis that competitiveness and the odds of an attack tweet are independent of each other can be rejected. This table with clustering accounted for varies greatly from the original logistic regression in that competitiveness is n ow sta tistically significant at the .05 level, wh ereas before it was only significant at the .1 level
Twitter Attacks of Senatorial Candidates 40 Interpretation for Dummy Variables: Compared to Arizona 's Senate race Note: Arizona had a moderate competitive margin and an average number of attack tweets. California : The odds of an attack tweet being issued in the California race were more likely by a factor of 1.5 or 50 percent more likely than the odds of an attack tweet being issued in the Arizona race. Significance: This odds ratio is not st a tistically significant at the .05 or .1 levels. Thus, the null hypothesis that the likelihood of an attack twe et is equal in Arizona (dummy variable reference state) and California cannot be rejected. The likelihood o f an attack tweet in California is not significantly greater than an attack tweet in Arizona. Connecticut: The odds of an attack tweet being issued in the Connecticut race were more likely by a factor of 3.1 or about 210 percent more likely than the odds of an attack tweet being issued in the Arizona race. Significance: This odds ratio is statistically significant at the .1 level, but not at the .05 level. Thus, the null hypothesis that the likelihood of an attack twe et is equal in Arizona (dummy variable reference state) and Connecticut can be rejected. The likelihood of an attack tweet in Connecticut is significantly greater than an attack tweet in Arizona. Delaware: The odds of an attack tweet being issued in the Delaware race were less likely by a factor of .35 or 65 percent less likely than the odds of an attack tweet being issued in the Arizona race.
Twitter Attacks of Senatorial Candidates 41 Significance: This odds ratio is not statistically significant at the .05 or .1 levels. Thus, the null hypothesis that th e likelihood of an attack twe et is equal in Arizona (dummy variable reference state) and Delaware cannot be rejected. The likelihood o f an attack tweet in Delaware is not significantly greater than an attack tweet in Arizona. Massachusetts: The odds of an attack tweet being issued in the Massachusetts race were more likely by a factor of 2.3 or 130 percent more likely than the odds of an attack tweet being issued in the Arizona race. Significance: This odds ratio is not statist ically significant at the .0 5 or .1 levels. Thus, the null hypothesis that the likelihood of an attack twe et is equal in Arizona (dummy variable reference state) and Massachusetts cannot be rejected. The likelihood o f an attack tweet in Massachusetts is not significantly greater than an attack tweet in Arizona. Michigan: The odds of an attack tweet being issued in the Michigan race were more likely b y a factor of 4.4 or 34 0 percent more likely than the odds of an attack tweet being issued in the Arizona race. Significance: This odds ratio is statistically significant at the .05 level Thus, the null hypothesis that the likelihood of an attack twe et is equal in Arizona (dummy variable reference state) and Michigan can be rejected. The likelihood o f an attack tweet in Michigan is signi ficantly greater than an attack tweet in Arizona. Minnesota: The odds of an attack tweet being issued in the Minnesota race were less likely by a factor
Twitter Attacks of Senatorial Candidates 42 of .09 or 91 percent less likely than the odds of an attack tweet being issued in the Arizona race. Significance: This odds ratio is statistically significant at the .05 level Thus, the null hypothesis that the likelihood of an attack twe et is equal in Arizona (dummy variable reference state) and Minnesota can be rejected. The likelihood o f an attack tw eet in Minnesota is significantly greater than an attack tweet in Arizona. Nebraska: The odds of an attack tweet being issued in the Nebraska race were less likely by a factor of .89 or 11 percent less likely than the odds of an attack tweet being issued in the Arizona race. Significance: This odds ratio is not st atistically significant at the .05 or .1 levels. Thus, the null hypothesis that the likelihood of an attack twe et is equal in Arizona (dummy variabl e reference state) and Nebraska cannot be rejected. The likelihood o f an attack tweet in Nebraska is not significantly greater than an attack tweet in Arizona. Utah : The odds of an attack tweet being issued in the Utah race were less likely by a factor o f .01 or 99 percent less likely than the odds of an attack tweet being issued in the Arizona race. Significance: This odds ratio is statisti cally significant at the .05 level Thus, the null hypothesis that the likelihood of an attack twe et is equal in Arizona (dummy variable reference state) and Delaware can be rejected. The likelihood o f an attack tweet in Utah is significantly greater than an attack tweet in Arizona.
Twitter Attacks of Senatorial Candidates 43 West Virginia : The odds of an attack tweet being issued in the West Virginia race were less likely by a factor of .34 or 66 percent less likely than the odds of an attack tweet being issued in the Arizona race. Significance: This odds ratio is not statist ically significant at the .05 or .1 levels. Thus, the n ull hypothesis that the likelihood of an attack twe et is equal in Arizona (dummy variable reference state) and West Virginia cannot be rejected. The likelihood o f an attack tweet in West Virginia is not significantly greater than an attack tweet in Arizona. XII. Confirm ation or Denial of Hypotheses Hypothesis 1: The odds of an attack tweet being issued will increase the closer in time it is to Election Day. The researcher found the time variable to be statistically significant in both logistic regression models making it one of the strongest factors studied that impact attack tweet use. Thus, this hypothesis was confirmed by the models, which yielded the sa me results: F or each month in time closer to Election Day the odds that an attack tweet was issued increased by a factor of about 1.2 or more likely by about 20 percent, all else equal This shows not only how much of an effect time has on attack strategy use on Twitter, but also that these results were statistically significant after the lack of observation independence was considered. Hypothesis 2: The odds of an attack tweet being issued by a challenger or open seat candidate are greater than the odds of an incumbent issuing an attack tweet. While the researcher found experience in office to be statistically significan t before accounting for unit heterogeneity the second logistic regression model yielded results that were
Twitter Attacks of Senatorial Candidates 44 not statistically s ignificant. This means that being a challenger or open seat candidate does not h ave as much of an effect on issuing of attack tweets as the researcher had originally predicted. This makes this hypothesis false in a statistical sense meaning the experience in office also does not matter as the researcher had predicted Still, both models did show challengers and open seat candidates as having greater odds of issuing attack tweets relative to incumbents. Hypothesis 3 : The odds of candidates from the Senate minority party (Republicans) issuing an attack tweet are gr e ater than the odds of candidates from the Senate majority party (Democrats) issuing an attack tweet. The researcher found minority party status to only be statistically significant before accoun ting for the lack of observation independence While both logistical regression models had odds ratios that indicated a positive relationship, meaning that minority party status increased the odds of issuing attack tweets this variable does not matter as much as predicted The results of the model indicate that it is not the minority/majority party status of a candidate in Congress that determines the odds of using attack tweets, but the competitive nature of each race (despite party affiliation) that determines candidates' odds of issuing attack tweets. Hypothesis 4 : The odds that an attack tweet is issued will be greater in races with a smaller spread of polling data (competitive margin) between candidates or greater competitiveness. Along with time, competitiveness had the most significant effect on the odds of attack tweets being issued. Both logistic regression models at a statistically significant l evel supported this hypothesis in the sense that p er 1 point increase in the compet itive margin the odds that attack tweets were issued decr eased by a factor of about .94 or 6 percent less likely, all else equal. Projected winners h ad decreased odds of issuing attack tweets as the polling margin moved in a n
Twitter Attacks of Senatorial Candidates 45 increasingly positive direction above 0, which lessens the competitiveness of the race. Simply, the more a candidate was expected to win (less ening the competitiveness), the lower the odds that an attack tweet was issued. Hypothesis 5 : The odds that an attack tweet is used w ill increase as a state's population increases. T his hypothesis could only be tested in the first logistical regression model that did not take into account the data clustering, which could potentially skew the researcher's conclusion on this hypothesis. The statistically significant results from the first logistical regression model indicate d that per million person increase in the state population, the odds that attack tweets were issued increased by a factor of 1.03, or 3 percent, all else equal. Thus, it is difficult for the researcher to conclude that this hypothesis is supported by the data because the population variable presented a collinear problem in the second regression model. However, to assure the other data is valid an d takes into accou nt the lack of data independence t he dummy variables for the race level cluster control for state population. XIII. Conclusion The researcher sought to find the conditions under which candidates for office utilize attack strategies via Twitter. This research question was studied through a c omparative case analysis of 10 contested Senate races, each with one Republican candidate and one Democratic candidate, in the 2012 general election. It was theorized that candidates use d these resonate, negative attack strategies because they want to discredit their opposition in hopes of winning the general e lection. The current study use d tweets as the recording unit for campaign rhetoric and code d these tweets as attac ks or not attacks during the three months leading up to Election Day.
Twitter Attacks of Senatorial Candidates 46 Prior to coding and quantitativ e analysis, the researcher interpret ed qualitative characteristics of each selected race as a premise for quantitative analysis. The researcher then observe d from the previously described s ources or official online Twitter accounts of the selected candidates to codify the independent variables for use in quantitative logistical regression analysis. The researcher addressed problems related to candidate heterogeneity and tweet dependence in a second logistical regression. Thus, the odds of attack tweets were studied, subject to the effects of time minority party st atus, experience in office, competitiveness and state population to determine the conditions under which the candidates use attac k tweets and any variation during the three months analyzed. This brought about sound conclusions for strategies on a newly politically mainstreamed social media outlet, Twitter, and update d the social media research conducted for the 2010 general electi on or prior. The research found evidence of Twitter's political mainstreaming in t he creation of specific accounts for campaign purposes alone, which is indicative of the increase in public officials' use of Twitter. The data analysis yielded results that demonstrated: in what ways candidates are different in their attack strategy use, how much greater or how much less likely candidates are to issue attack tweets, and if that increase or decrease in attack tweets is statistically significant. This case stu dy ended up telling a story of 10 Senate races in 2012 and how the attack tweets within those races were affected by various characteristics of the candidates/races. Time and competitiveness ended up being the m ost significant variables, statistically tha t ha d the greatest effect on attack tweet usage. Thus, th e closer in time to Election Day the more likely the candidates were to issue an attack tweet. In fact, per month closer in time to Election Day the odds that an attack tweet was issued increased b y 20 percent. While the other variables may have had significant effects, this significance did not hold true after the researcher included the lack of
Twitter Attacks of Senatorial Candidates 47 observation (tweet) independence. According to this study's findings, being a candidate of the Senate's minority part y or a challenger was not significant Instead, the individual races conditions of competitiveness determined the odds of an attack tweet being issued. In fact, per 1 point increase in the competitive margin (less competitiveness) the odds of attack tweets were less likely by 6 percent. Massachusetts, the most competitive race used in this study, had high odds of issuing attack tweets (compared to the other races studied) an d fits the researcher's conclusions on the conditions that det ermine attack tweet usage Ho wever, one race that defied the finding s was Michigan, the race with the greatest odds of issuing and attack tweet. Michigan's race was only moderately competitive. This case study was feasible because the source gathering pr ocess is free and available at any time on the Interne t. With the limitation of only so many races (10 used) fitting the vetting criteria ample time was left for quantitativ e analyses of candidates' tweets for a sample size of about 6,000 tweets at the fu ndamental level and 20 cases (candidates) at the next hierarchical level The qua litative narrative of the race also aid ed in the interpretation of the quantitative results by describing the real world settings in which the tweets took place. Candidates' attacks via Twitter have seldom been studied on the in depth l evel of a case study, and this study provide d new insigh ts into when candidates use such strategies Candidates who ch o ose to adopt Twitter in the future can build on the kno w ledge of better defined, attack tweet time cycles, especially those utilized by race winners, to craft their own campaigns on social media. XIV. Beyond This Study and Further Research As mentioned in the literature review, the researcher theorized the mutual destruction theory of candidates' Twitter use. The researcher hoped to include a variable to test this theory,
Twitter Attacks of Senatorial Candidates 48 but it would hav e required a more complex model An original hypothesis was constructed based on this theory, although the research was un able to test it. Perhaps, further research on Twitter use during political campaigns could test this hypothesis and look at candidate responses to their oppositions' attack tweets. The odds that a candidate issues an attack twee t will increase as the odds of his or her opponent issuing an attack tweet increases. In coordination with previous literature's mutually assured destruction theory, candidates will take on a one for one feedback approach when the opposition sends an at tack tweet by responding with their own attack tweet. It also follows from the theory that candidates want to discredit opponents' statements by responding to those attack tweets with equal force. Additionally, if candidates are directly @replied to by the opposition in an attack tweet, the odds of a responsive attack tweet may be even higher because the candidates have records of every time their accounts are mentioned by other users' accounts. Candidates will defend themselves by discrediting the attack t weets' accusations or negative c ampaigning from the opposition. Beyond this study, the researcher suggests a purely random sample to provide greater validity and a more representative sample. This case study only looks at one, historical moment in time August 6, 2012 through November 6, 2012. Although this study was limited to just U.S. Senate races, perhaps further studies could examine the Twitter attack strategy uses of U.S. House races and state legislature races. Further research could also examine the different characteristics of candidates not included in this study such as gender age, or ethnicity Also, the researcher did not consider that the competitiveness variable could reach a threshold in either direction, a point where candidates choose to stop tweeting because they realize they have an
Twitter Attacks of Senatorial Candidates 49 almost definite win or loss. While this study looks at 2012, the most recent election year, how candidat es use Twitter and their attacks on this social m edia outlet may change. One day candidates may atta ck nearly 100 percent of the time or attack strategies could become obsolete on social media. This study does not examine or predict what could happen in the future regarding negative campaigning on Twitter which is an important question for research on pol itical social media strategies. However, this study generated conclusions on the most conducive conditions for candidate attacks based on original Twitter data at one historical moment in time the 2012 Senatorial general election
Twitter Attacks of Senatorial Candidates 50 Bibliography Ansolabehere, S., Iyengar, S., Simon A., & Valentino, N. (1994, December). Does Attack Advertising Demobilize the Electorate? The American Political Science Review 88(4), 829 838 Retrieved from http://www.jstor.org/discover/10.2307/2082710?u id=3739600&uid=2&uid=4&uid=3739 256&sid=21102900525661 Battle for the Senate. (2012). Real Clear Politic s Retrieved from http://www.realclearpolitics.com/epolls/2012/senate/2012_elections_senate_map.html Bills, K. (2013). Tweets [@KurtBills]. Retrieved from https://twitter.com/KurtBills Brown, S. (2013). Tweets [@ScottBrownMA]. Retrieved from https://twitter.com/ScottBrownMA Carmona, R (2013). Tweets [@CarmonaforAZ] Retrieved from https://twitter.com/CarmonaForAZ Carper, T. (2013). Tweets [@SenatorCa rper]. Retrieved from https://twitter.com/SenatorCarper Damore, D. (2002, September). Candidate Strategy and the Decision to Go Negative. Political Research Quarterly 55(3), 669 685. Retrieved from http://prq.sagepub.com/content/55/3/669.full.pdf+html
Twitter Attacks of Senatorial Candidates 51 Druckman, J. N., Kifer, M. J., & Parkin, M. (2009). Campaign Communications in U.S. Congressional Elections. The American Political Science Review 103, 343 366. Retrieved from http://journals.cambridge.org/action/displayAbstract?fromPage=online&aid=6101636 Emken, E. (2013). Tweets [@Emken2012]. Retrieved from https://twitter.com/Emke n2012 Feinstein, D. (2013). Tweets [@SenFeinstein]. Retrieved from https://twitter.com/SenFeinstein Flake, J (2013). Tweets [@FlakeforSenate]. Retrieved from https://twitter.com/FlakeforSenate Fischer, D. (2013). Tweets [@DebFischer2012]. Retrieved fro m https://twitter.com/DebFischer2012 Fowler, E. F. & Ridout, T. N. (2010). Advertising Trends in 2010. The Forum 8(4), 4.! Retrieved from http://www.bepress.com/forum/vol8/iss4/art4 Franz, M & Ridout, T. N. (2008) Evaluating Measures of Campaign Tone. Political Communication 25(2), 158 179. Retrieved from http://www.tandfonline.com/doi/abs/10.1080/10584600801985409 Freedman, P. & Goldstein, K. (1999, October). Measuring Media Exposure and the Effects of Negative Campaign Ads. American Journal of Political Science 43(4), 1189 1208. Retrieved from http://www.jstor.org/stable/2991823
Twitter Attacks of Senatorial Candidates 52 Gulati, G. J. & Williams, C. B. (2010). Communicating with Constituents in 140 Characters or Less. Working Papers Paper 43.! Retrieved from http://opensiuc.lib.siu.edu/cgi/viewcontent.cgi?article=1042&context=pn_wp Hatch, O. (2013). Tweets [@OrrinHatch]. Retrieved from https://twitter.com/OrrinHatch Hoekstra, P. (2013). Tweets [@petehoekstra]. Retrieved from https: //twitter.com/petehoekstra Howell, S. (2013). Tweet [@VoteScottHowell]. Retrieved from https://twitter.com/VoteScottHowell Jerit! J. (2004, August). Survival of the Fittest: Rhetoric during the Course of an Election Campaign! Political Psychology 25(4), 563 575. Retrieved from Symposium on Campaigns and Elections: http://www.jstor.org/stable/3792409 Kerrey, B. (2013). Tweets [@kerreybob]. Retrieved from https://twitter.com/kerreybob King, G., Keohane, R. O. & Verba, S. Designing Social I nquiry. Scientific Inference in Qualitative Research Princeton University Press, 1994. Klobuchar, A. (2013). Tweets [@amyklobuchar]. Retrieved from https://twitter.com/amyklobuchar Lau, R. R. & Pomper, G. M. (2001). Negative Campaigning by the U.S. Senate Candidates. Party Politics 7, 69 87. Retrieved from http://ppq.sagepub.com/content/7/1/69.refs Manchin, J. (2013). Tweets [@JoeManchinWV]. Retrieved from https://twitter.com/JoeManchinWV
Twitter Attacks of Senatorial Candidates 53 McMa hon, L. (2013). Tweets [@Linda_McMahon]. Retrieved from https://twitter.com/Linda_McMahon Murphy, C. (2013). Tweets [@ChrisMurphyCT]. Retrieved from https://twitter.com/ChrisMurphyCT Raese, J. (2013). Tweets [@RaeseForSenate]. Retrieved from https://twitter.com/RaeseForSenate Senate Map Election 2012. (2012). New York Times Retrieved from http://elections.nytimes.com/2012/results/senate Stabenow, D. (2013). Tweets [@TeamStabenow]. Retrieved from https://twitter.com/TeamStabenow Theilmann J. & Wilhite, A. (1998, November). Campaign Tactics and the Decision to Attack. The Journal of Politics 60(4),1050 1062. Retrieved from http://www.jstor.org/stable/2647730?origin=JSTOR pdf Topsy. (2013). Twitter Search, Monitoring Retrieved from http: //topsy.com/ United State Election Atlas. (2013). 2012 Senatorial General Election Results Retrieved from http://www.uselectionatlas.org United State s Senate. (2013). Senator s of the 113 th Congress Retrieved from http://www.senate.gov/general/contact_information/senators_cfm.cfm
Twitter Attacks of Senatorial Candidates 54 Wade, K. (2013). Tweets [@kevinwade2012]. Retrieved from https://twitter.com/kevinwade2012 Walter, A. S. & Vliegenthart, R. (2010, August 13). Negative Campaigning across Different Communication Channels: Different Ball Games? The International Journal of Press/Politics 15, 441 461. Retrieved from http://hij.sagepub.com/content/15/4/441 Warren, E. (2013). Tweets [@ elizabethforma]. Retrieved from https://twitter.com/elizabethforma