Analysis of Donald Trump's Election Cycle and Sentiment on Twitter

Abstract

In this thesis, the author examines the last 131 days of the 2016 election cycle. This analysis focuses on how sentiment is present on Twitter when people engage in political communication on social media.  With the increasing online political discussions, created on social media such as Twitter, an analysis of sentiment is critical.  The data could be obtainable for candidates to estimate the electorate’s opinion of each candidate.  A shift of sentiment offers a deeper insight for tracking changing attitudes toward candidates.  Because Twitter only allows each tweet to be 140 characters there is a simplicity that offers statements to be concise.  Trends for each candidate throughout the final days of the election cycle are correlated with national polls to assess if there is a relationship present.  This study applies sentiment to recognize trends that may estimate a candidate’s chance of winning the election and offers indications as to how the intended electorate may vote when a relationship is established between sentiment and national polls.

Chapter 1: Introduction

President Barack Obama and other political candidates running for public service offices began using social media during the 2008 presidential election cycle to broaden the means through which they could spread political messages.  In fact, President Obama sent a tweet after he won the election in 2008 stating, “We just made history.  All of this happened because you gave your time, talent, and passion.  All of this happened because of you.  Thanks” on Twitter from his personal Twitter account.  The use of social media altered the means by which the game of political communication occurred while also changing political campaigning forever.  As of 2012, there were 500 million Facebook users and accounts, with 100 million Twitter users, which increased to 170 million Twitter users in 2013 (Stieglitz & Dang-Xuan, 2012; Wladarsch, & Neuberger, 2013).  These numbers show tremendous growth within a span of one calendar year and continue to increase yearly.  It is important for candidates to stay current with technology and with the presumed electorates regardless of the method of social media they choose to utilize.

1.1 2016 Election Campaign

Political communication occurs on Twitter from political players and candidates daily.  The election cycle of 2016 was no exception.  However, the usage of social media differed from candidates in prior election cycles, as President Donald Trump and Secretary Hillary Clinton used social media as an avenue for hurling insults and personal attacks against each other.  The electorate joined in and insults were traded between party loyalists and candidates.  Self-efficacy of politics was present as issues and agendas were discussed between social media users, highlighting evidence of electorate involvement never seen in such a public eye.

Presidential debates and other events of the past offer material for social media users to discuss on Twitter as well as other social media outlets.  The first presidential debate in 2012 recorded 10 million tweets shared by 170+ million users (Flynn, 2016).  The 2016 election cycle was similar, except it recorded more viewers than 2012.  Sixty-seven million viewers, an estimated number, watched the first debate in 2012 via television networks, 80 million viewers watched it in 2016 (Flynn, 2016).  The same phenomenon can be observed on social media sites as some 369,000 users watched the debate via live stream on Twitter during the 2016 election cycle (Kafka & Wagner, 2016).  This probably is because Twitter currently has more than 313 million users to date, which continues to grow daily and has increased from 185 million users in 2012, essentially increasing the amount of political chatter in 2016 (Flynn, 2016).

Neil Young’s Rockin’ in the Free World played stridently at the Trump Tower on June 16, 2015 as Donald Trump rode down the golden escalator to announce his bid for presidency.  This was the beginning of the most tumultuous American presidential election cycle to date.  No one living within our borders or abroad will soon forget the 2016 presidential cycle.  Controversy occurred daily extending from the primaries until Election Day on November 8, 2016.  Media frenzy occurred in every newspaper, social media site, and on every televised network locally, nationally, and, internationally.  The presidential hopefuls utilized social media sites as a platform to reach the voting public.  The primaries in the election cycle of 2016 were as chaotic and tumultuous as the general election, considering the bid for Republican and Democratic nominations were up for grabs.  Seventeen Republicans eventually dove into presidential race of 2016 beginning with Cruz on March 23, 2015 (Bialik, 2016).  Only three candidates remained in the primaries as of March 16, 2016 until the Republican National Convention when Trump was nominated (Bialik, 2016).

Hillary Clinton was the last woman standing as she eagerly awaited entry back into the White House, this time not as First Lady, but as President of the United States.  History was being made as Clinton won primaries in California, New Jersey, South Dakota, and New Mexico securing the Super Delegates she needed to move forward on June 8, 2015 (Collinson, 2016).  The Democrats had a total of six candidates enter the race with the first candidate being Clinton who entered on April 12, 2015.  After the Iowa caucus on February 1, 2016, only two candidates remained in the race until the Democratic National Convention when Clinton was nominated.  After the Republican and Democratic National Conventions, there were two remaining candidates vying for a seat at the desk in the oval office.

1.2 Social Media Usage

Twitter is a micro-blogging social media site that allows 140 characters per tweet.  Twitter is used in many ways and for many reasons.  Generally, users share thoughts on various subjects such as foods, movie stars, and other newsworthy topics (Ronsenstiel, Sonderman, Loker, Ivancin, & Kjarval, 2015).  Some use Twitter for recreation, others use Twitter to promote business through marketing.  Twitter experts, marketing firms, and business leaders, often use Twitter for brand management and brand awareness (Lonoff Schiff, 2013).  Twitter is also a social media tool that can be used to make others aware of businesses and services (Lonoff Schiff, 2013).  Many use Twitter to tweet feelings about sports or whatever may be occurring in their daily activities.  Twitter has become a viable avenue for the access and consumption of news as Ronsenstiel et al. (2015) conducted a study that concluded that 81 percent of the participants checked their news daily on Twitter.

Presidential candidates Donald Trump and Hillary Clinton have both used social media more than prior candidates during any other election cycles.  President Trump and Clinton also used social media in a different manner than President Obama and Mitt Romney in the 2012 election cycle.  The continuing growth of the population on Twitter and the more in depth usage by political candidates offered a platform for political discourse in addition to political communication.  The election cycle of 2016 was unique in the ways social media was leveraged.  Twitter ran a live feed of all three presidential debates in the general election for the first time.  This allowed a population that may not watch news networks to tune into the debate.  In addition, Twitter users could interactively tweet during the debates.  In addition, candidates displayed their pleasure or displeasure of responses to certain debate questions tweeting comments and rebuttals after debates ended which continued the political communication and fueled discourse.  Twitter created an official hashtag and emoji for all three of the live debates in the general election: #Debates, #debatenight, and #Debates2016 (Flynn, 2016).  The 2016 presidential candidates used social media differently and more frequently than past candidates, attempting to tweet their way into the public sphere while spreading political messages.

During political election cycles, campaign managers traditionally use polls to drive important aspects of the campaign while interested people use it to predict which candidate will gain the electoral votes from certain states or regions.  Due to the evolution of current technology, individuals can communicate sentiment toward candidates on social media during important events of an election cycle and interested parties have the ability through software to track these sentiments.  Certain events often occur during a political race that cause concern or satisfy individual voters.  The electorate then express their approval or contempt, which could be positive, negative, or neutral (absent of negative or positive emotion).  By tracking the sentiment of electorates, candidates can adjust their talking points and pivot to speak more on particular issues that the polls show as most important to the electorates.  Candidates can also begin to address issues raised by the electorates in order to gain approval and win the election.  Other interested people also use polls to predict which candidate will gain the electoral votes from certain states during the election cycle.  Polls are also used during the last hours of the Election Day to predict the winner of the election as eager electorates leave the precinct after casting their vote.

Researching political communication and political constructs of the recent candidates in addition to social media sites such as Twitter set a litmus test for upcoming political races.  This litmus offers a helpful guide for upcoming political boxing matches.  Sentiment analysis in addition to opinion mining of the voting electorate demonstrates the usage of social media in which an exchange of political communication occurs.  Political communication often serves the purpose of framing messages in a certain way that ensures a certain candidate or personal favorite is viewed more positively or negatively.  Messages via social media are framed differently depending on the opinion of the particular electorate.  Candidates may find that opinion mining or sentiment analysis produce a helpful guide for candidates to follow in future elections.

1.3 Polling

Polls are often difficult to conduct.  They require a significant amount of labor for activities like making telephone calls or going door-to-door.  Therefore, polling demands human resources, which contributes to the high financial costs often, related to the conduction a polls.  However, polling social media users, especially Twitter users, lowers the cost of polling and increases the sample population.  Pollsters can easily poll people from every area of the country.  The issue with narrowing a population to one certain social media site would be getting an ideal sample population that would accurately represent the electorate, including both parties and all demographics.  However, great advantages become present through social media and modern technology, as interested individuals can poll electorates who communicate sentiments for political candidates at any time throughout the election cycle.  Polls are also used during the last hours of the Election Day to predict which candidate will be victorious.  Exit polls attempt to gauge which candidate is leading the political race as electorates exit voting locations, which at times can be inaccurate for different reasons.  Electorates may not be truthful during exit polls, as they do not care to share their vote with others.  Most exit polls were inaccurate in 2016 and attributed to the “shy Trumper” hypothesis, which states that those who voted for Trump did not want to say in fear of backlash (Mercer, Deane, & McGeeney, 2016).  Furthermore, early voting is not accounted for in exit polls.  Mercer et al., 2016 stated, “Statisticians say that exit poll data, while well-intentioned, is inherently flawed as a way to predict final vote totals.  Due to the need to compile nearly instantaneous results, exit pollsters rely on statistical models that may be outdated by the time an election rolls around” (Mercer, Deane, & McGeeney, 2016, para 4).

Digital polling, on the other hand, especially with Twitter users, lowers the cost of polling and increases the sample population.  Social media polls, especially those conducted on Twitter users, allows pollsters to reach a larger geographical reach.  Pollsters can easily poll people from every area of the country and outside the United States.  Because of the ease of use or possibility of limited funding, interested individuals can poll electorates about their sentiments for political candidates continuously throughout the election cycle.  However, polling a Twitter population has its disadvantages.  Minus the question of validity of social media polls, pollsters have to put the age distribution of the sample into context.  Approximately 24% of adults who use social media use Twitter (Greenwood, Perrin, & Duggan, 2016).  Greenwood, Perrin, and Duggan (2016) assessed “Younger Americans are more likely than older Americans to be on Twitter.  Some 36% of online adults ages 18-29 are on the social network, more than triple the share among online adults ages 65 and older” (para, 10).  The authors also found through Pew Research Center surveys that the Twitter is used more by adults with college educations (29%) opposed to the (20 %) with a high school diploma or less (Greenwood, Perrin, & Duggan, 2016).  It may not yet be possible to represent the electorate fully.  While an overall assessment can be made, the question remains whether Twitter sentiment analyses is a productive way to poll the electorate and can we confidently ascertain that sentiment analysis is viable as a blaring signal?  We may be unable to actually predict the winner of an election, however, as story may be told by collecting sentiment over a specific time-period of the election.

Chapter 2:  Literature Review

2.1 Social Media and Political Communication

Stieglitz and Dang-Xuan (2012) suggest microblogging on Twitter use may increase political participation for some users.  Stieglitz and & Dang-Xuan (2012) assess Twitter as different from other social media sites.  Facebook usually consist of communities of friends that are familiar with each other in real life settings and much of the time has actually met in person whereas Twitter users connect by using hashtags and similar preferences by entering key words preceded by hashtags in search bars within Twitter.  Following individuals is different on Twitter opposed to other social media sites and often people follow a population that has similar views to their own.  Pennacchiotti and Popescu (2011) state, “Intuitively, Democrats are more likely to follow the accounts of Democratic politicians and Republicans those of Republican politicians” (p. 9).

Dang-Xuan et al. (2013) believe through research that social media is a legitimate avenue for shared political information.  Social media has broadened the scope of gate keeping and agenda setting.  In previous presidential elections, network television performed the roles of gate keeping and agenda setting.  Media and political communication amongst electorates can occur every minute of every day, as opposed to waiting on the nightly news to share political information.  This allows electorates to participate in their own agenda setting.  Presently, micro-blogging through Twitter allows several million authors to frame messages in ways that apply to personal agendas.  Dang-Xuan et al. (2013) ascertain that contemporary democracies develop avenues through social media to engage with constituents before, during, and after political campaigns.  Citizens are able to spread political messages or information through retweets on Twitter (Dang-Xuan et al., 2013).

Twitter studies have produced empirical results for several paradigms that link political communication to candidates and electorates.  Journalists also find Twitter to be useful for political communication as a tool for spreading messages or setting agendas in many other instances.  Broersma and Graham (2012) conducted a content analysis of Twitter messages concerning the 2010 British and Dutch elections.  The goal of the study was to understand who was using Twitter, and if Twitter contributed to the print and online news headlines (Broersma & Graham, 2012).  Broersma and Graham (2012) found evidence that tweets were used in headline news stories for journalists in newspapers and for politicians in tweets when scandals appeared to develop on Twitter.  Evidence of tweets triggering newsworthy stories and headlines occurred more in the British election than the Dutch election (Broersma & Graham, 2012).  The tweets were classified by those who authored them and by their purpose in tweeting.  The category of authors was segmented into politician, expert, or cultural producer.  The functions of tweets were coded as either triggers or sources of news’ headlines generation (Broersma & Graham, 2012).  The researchers looked at mainstream media such as respected newspapers in the U.K. and Netherlands.

2.2 Prediction based on Twitter Sentiment

Researchers have explored whether Twitter sentiment emotions have predictive power.  These relationships are not means causal.  Research completed in these areas help to ascertain the usefulness of Twitter sentiment for game observations, which in essence, may work in political paradigms (Diakopoulos & Shamma, 2010).  Bollen, Mao, and Zeng (2010) conducted studies for predicting stock market value rise/fall.  The authors found predictive values relating to emotions on Twitter more than measurements of negative and positive sentiment.  As previously stated the prediction can only truly represent a relationship, as emotions cannot be a causation of market rise and fall.

Intensive work has been done in the past in measuring predictive factors on social media in the political arena.  Tumasjan, Sprenger, Sandner, and Welpe (2011) analyzed 100,000 tweets during a German election in 2009.  An aggregated sentiment analysis found that the population in fact used Twitter to deliberate opinions of candidates.  Tumasjan et al. (2011) stated,

Our results provide evidence supporting our theory that microblogging forums provide a mechanism for weighing information and that, despite individual biases, errors can cancel each other out.  The predictive accuracy is even more impressive when compared to the track record of the IEM, a prediction market set up with the explicit purpose to predict election results (pg. 414).

Prediction values tended to coincide with traditional polling as researchers suspected.  Nonetheless, Tumasjan et al. (2011) suggest that predictive values of Twitter are not a stand- alone method and should only compliment traditional polling as opposed to replacing polling entirely.

O’Connor, Balasubramanyan, Routledge, and Smith (2010) composed a time series sentiment analysis of public opinion while comparing the sentiment of candidates on Twitter to the sentiment in traditional polling.  The authors implemented a forecasting value that would indicate what polls would offer in future predictions.  In essence, text sentiment proved to be a superior predictor at a certain milestone in the empirical work cited.  Certain issues and periods offered predictive values through textual analysis as an increase of electoral confidence occurred, which also is evidence of a linear relationship (O’Connor et al., 2010).

On the other hand, some feel that predicting election results with Twitter may be an impossible feat.  Burch (2015) conducted a primary study using Sysomos machine learning to mine opinions in a sentiment analysis.  The question arose repeatedly throughout this literature review about whether social media sentiment analysis was a more effective prediction tool than traditional polls in recent elections.  In this study, volumes of mentions as well as sentiment were analyzed, along with traditional polling, which had been the focus in many previous studies mentioned.  Several periods were relevant in collecting data, as this was a two-fold experiment, which continued with the goal of predicting the primary winners in several states.  The evidence in this study to date displayed a larger following for some candidates such as Bernie Sanders on Twitter, with Sanders continuously ahead of Clinton on social media when measuring volume (Burch, 2015).  According to Burch (2015), Clinton was ahead in traditional polling by an overwhelming 54% to Sanders’ 33%.  However, looking at the Twitter conversation volume around each candidate, Sanders led the way and did consistently over the time analyzed.

2.3 Sentiment Studies

Researchers have also compared the outcome of sentiment studies that run congruently to traditional surveys.  Mitchell and Hitlin (2013) conducted a yearlong research study comparing results of sentiment from Twitter and traditional surveys, using eight political events that occurred throughout the 2012 election cycle as a measuring tool.  Mitchell and Hitlin (2013) ascertained that depending on whether the topic was considered more liberal or conservative, the sentiment rated higher or lower for social issues attached to certain party lines.  In some instances, such as gay marriage rulings, sentiment is altered according to social settings or exposure.  People often pretend to be more liberal on Twitter but more conservative when actually answering a traditional survey (Mitchell & Hitlin, 2013).  These findings are hypothesized differently because of the population tweeting about the specific topic changes according to the topic as mentioned above.  Topics that leaned more conservative were tweeted more by conservatives and vice versa with topics that leaned more liberal.  They disproved the belief that Twitter polls result in more liberal results than surveys.  Twitter conversations occurring about the presidential candidates Obama and Romney in 2012 were overwhelmingly more negative than positive.  However, Romney had a larger negative sentiment percentage in national polling and in Twitter sentiment in most instances except by the first debate (Mitchell & Hitlin, 2013).  Obama and Romney sparred in the first debate bringing back hope to conservatives as the Obama stumbled several times (MacAskill, 2012).  CNN conducted a poll that evening with 67 percent saying Romney was a clear winner (MacAskill, 2012).  Twitter sentiment varied after President Obama was re-elected.  Twitter represented a more positive sentiment than did polling by Pew Research polls (2013).  Mitchell and Hitlin (2013) explained that limitations were present in their study because, “those who get the news on Twitter and those who tweet news are very different demographically from the public” (p. 1).

Obviously, everyone who is tweeting is not necessarily participating in political communication.  However, many were in the past three election cycles and sentiment analysis offered predictive nuances for debate winners.  Cody et al. (2015) states,

Twitter has also been used to examine human sentiment through analysis of variations in the specific words used by individuals.  Dodds et al. develop the “hedonometer” a tool for measuring expressed happiness—positive and negative sentiment—in large-scale text corpora (p. 2).

With these emotive expressions of happiness or disappointment through negative and positive sentiment researchers can hypothesize as to which candidates are in the lead in the debates and at different moments throughout the election cycle due to sentiment.

Pew researchers Rosenstiel and Jurkowitz (2011) conducted another detailed analysis of Twitter in the presidential election of 2012 that differed some from that done by Mitchell and Hitlin (2013).  Rosenstiel and Jurkowitz compared Twitter sentiment to the blogospheres, which was “more voluminous, more fluid, and even less neutral” (2011, p. 2).  After the first comparison of blogs to tweets occurred, a second comparison was analyzed by comparing blogs/tweets to mainstream news such as network televisions.  The sample contained 20 million tweets that fluctuated according to certain events throughout the 2012 presidential election cycle.  The authors used two methods for analyzing data and coding tweets and blogs for sentiment.  First, a content analysis ensued to ascertain the quantity of exposure on Twitter and blogs.  Secondly, Crimson Hexagon technology (computer coding) was used, which allows a computer to code a large data set containing millions of tweets while also coding a small number of tweets in the beginning manually to ensure categories are mutually exhaustive (Rosenstiel & Jurkowitz, 2011, p. 28).  Both political blogs and tweets were run through the Crimson Hexagon to gauge sentiment.  Because blogs often contain several assertions, only statements that contain the candidate’s names were utilized for sentiment (Rosenstiel & Jurkowitz, 2011).  Findings in this particular study displayed a greater negative sentiment on social media such as Twitter and blogs and less negative sentiment on television broadcasts concerning the candidates (Rosenstiel & Jurkowitz, 2011).  The presidential election of 2012 was similar to 2016 with several GOP candidates in the primaries.  Rosenstiel and Jurkowitz (2011) measured sentiment on the three outlets for all candidates and only found a positive sentiment being greater than a negative sentiment on a couple occasions between May 2 and November 27 for the two candidates.  Of course, negative and positive sentiment differed according to what milestones were occurring during each candidate’s campaign.

Bollen, Mao, and Pepe (2011) also conducted a sentiment analysis in the latter part of 2008 using tweets as data, along with a specific timeline from August 1 to December 1.  These particular authors chose to compare socio-economic events alongside mood patterns mined during a sentiment analysis.  Social and economic indicators could be events such as the presidential elections, Twitter mood, and stock market fluctuations, or the death of a favorite celebrity (Bollen, Mao & Pepe, 2011).  About 9,664,952 million tweets were compiled and compared to profile of mood (POMS-ex), a psychometric scale that originates from POMS (Bollen, Mao & Pepe, 2011).  “POMS measures six individual dimensions of mood, namely tension, depression, anger, vigor, fatigue, and confusion, not intended for a large scale textual analysis” (Bollen, Mao & Pepe, 2011, pg. 451).  Measuring POMS normally occurs through a questionnaire format given to live subjects.  POMS-ex differs in the way data is collected and received.  POMS-ex acquires a large amount of text virtually through social media or electronic media.  Questionnaires are not administered to human participants (Bollen, Mao & Pepe, 2011).  This analysis was more about proving that machine mining as well as machine learning produces accurate results from large data.  However, researchers did find that significant events occurring that are political in nature could be correlated with several mood dimensions that fluctuate throughout events (Bollen, Mao & Pepe, 2011).

Sentiment continues to be relevant when analyzing microblogging and political communication.  As candidates increase their following via Twitter, public opinion mining becomes more prevalent to understand how a candidate is performing on the campaign trail.  As with any campaign, there will be highlights and lowlights that alter positive and negative feelings from the vocal social media users.  Wang, Can, Kazemzadeh, Bar, and Narayanan (2012) conducted a real-time Twitter analysis during the 2012 presidential election cycle.  Many Twitter sentiment studies account for positive, negative, and neutral opinions toward candidates.  Wang et al. (2012) included a category, named unsure, that normally is not included in most sentiment analyses.  Therefore, instead of classifying tweets in the three categories normally used in twitter sentiment, four categories were utilized.  The stated goal and approach of this empirical endeavor was to combine real-time statistical sentiment through modeling, while gaining an understanding of social and political praxes through social media, especially on Twitter (Wang et al., 2012).  This study found that special events have the ability to increase tweet volume and the proposed sentiment model is sufficient to evaluate public sentiment during real-time events (Wang et al., 2012).

2.4 Studies Comparing Polls to Twitter

Anuta, Churchin, and Lou (2017) conducted an experiment to first gauge whether or not the polls of the 2016 election were biased toward one candidate over the other, and secondly, to research whether Twitter would be useful as a less biased predictor for the last presidential cycle than polls.  Data was gathered from several polls and several states throughout the election cycle to analyze if polls were biased regarding the popular vote.  The authors created a prediction model that would detect bias for the popular vote only.  Anuta, Churchin, and Lou (2017) chose nine states in total to analyze.  The authors assumed the states that leaned liberal, conservative, and finally that were battleground or swing states might produce superlative results.  For the popular and electoral data concerning Twitter sentiment, the authors used specific tweets from Twitter API generated from certain areas of the United States identical to the states used for polling information.  A sentiment analysis was completed on 750,000 tweets using a program named Python, which contains the sentiment tool VADER (p. 4).  The results of this study yielded biases from the eight named sources of media.  Anuta, Churchin, and Lou (2017) stated,

In the 2016 U.S. election, the media (as encapsulated by our 8 sources) was, quantifiably biased against Donald Trump by -2.0% in the popular vote and -1.6% in the state based votes over the entire election period.  Towards the end of the election (in the 3-month period before Election Day), the popular vote bias decreased slightly to a -1.0% bias against Donald Trump (p. 10).

Twitter encapsulated results that were far more biased on the electoral and popular vote.  There was a filter bubble on tweets that were against Clinton and for Trump.

Stecanella (2016) conducted a sentiment analysis on MonkeyLearn from July 2016 until Election Day.  Millions of tweets were processed to gauge each day throughout the time span.  The author of this experiment was actually an engineer who created a social media tool that showed changing sentiment graphs for the timespan as mentioned.  Results returned more negative sentiment overall for each candidate than positive sentiment on a daily basis.  MonkeyLearn is said to have 70% accuracy for reading sentiment, whereas Sysomos is said to be 86% accurate when human accuracy normally falls between 70 and 85% (Stecanella, 2017; Bowers, 2017 para 2).

2.5 Problem with Polls

In the recent election, polls were not as accurate or reliable as they have been in the past.  Shirani-Mehr, Rothschild, Goel, and Gelman (2016) decomposed the margin of error in surveys that were given during statewide elections.  This study used polling data from more than 4,000 polls for 608 state presidential elections, political races for senators, and gubernatorial elections for nearly two decades (p. 3).  Shirani-Mehr et al. (2016) attempted to calculate biases while explaining the margin of error.  The findings of this study resulted in considerable election-level bias and superfluous variance.  Shirani-Mehr et al. (2016) estimated a standard absolute bias is “1.8 percentage points for senate races, 2.1 percentage points for gubernatorial races , and 1.0 percentage point for presidential races” (p. 22).  Polls in past presidential elections displayed small excess variance.  However, results yielded a larger standard error of .08 percent in senatorial and gubernatorial races (Shirani-Mehr, 2016).  Williams (2015) also agrees that national polling is in somewhat of a crisis.  Williams quoted a University of Michigan political science professor who specializes in political polling of elections who stated that “polling is a very important element of democracy and polls give the public an independent voice that’s not generally present” (2015).

2.6 Presentation of Study

Given the abovementioned literature, the present study seeks to assess the sentiment from Twitter messages regarding each candidate separately throughout the presidential election cycle ending on November 8, 2016.  This study also evaluates if there is a relationship present between Twitter sentiment and FiveThrityEight polls.

Chapter 3: Methodology

A descriptive analysis was done to examine trends in Twitter sentiment for Trump and Clinton between July 1, 2016 and November 8, 2016 using Sysomos.  A descriptive analysis was also done using the data collected from the election forecast from FiveThrityEight.  Sentiment results were compared to polls results using the Pearson product-moment correlation.  Tweet volume was also considered within the data set.

To conclude this study by answering the central questions we must examine polls and Twitter from numerous angles.  There were a few candidates running in the presidential election.  This study is only concerned with the candidates from the Republican Party and Democrat Party, and uses polls as a comparison that subtract for the third-party candidate.  The third party is excluded from the sentiment analysis statistics and adjusted for in the polls conducted by FiveThirtyEight.

3.1 Polls for Comparison

FiveThirtyEight uses several daily polls from all 50 states.  An average is then calculated for a final daily percentage displaying the popular vote, Electoral College, and chance of winning for each candidate.  FiveThirtyEight offers a poll-only forecast, which contains information from polls and does not factor in any other facets such as the economy or past elections.  They also have a now-cast, which gives the outcome if the election were to occur on that particular day.  Polls-only and now-cast are used for comparison in the current study.  FiveThirtyEight includes polls that were rated through an intense rating system for accuracy and integrity and belong to the National Council on Public Polls (NCPP) or the American Association for Public Opinion Research (AAPOR) (Silver, 2016).  Polls are excluded and placed on a banned list if the manager of FiveThirtyEight believes they have used fake data in the past or participated in unethical conduct such as robocalls to cellphones without the inclusion of live interviewers (Silver, 2016).  FiveThirtyEight takes several steps to assure the best accuracy possible.  Firstly, FiveThirtyEight adjusts its results by accounting for five major effects that could alter accuracy if not factored.  The five effects are likely voter adjustment, convention bounce adjustment, omitted third-party candidate adjustment, trend line adjustment, and house effects adjustment.  Secondly, poll outcomes are combined with other data that measure and account for third-party voting, undecided voters, projection of popular vote, national vs state polls, partisan voting index (PVI), demographic regression, and blending polls with regression, and state elasticity scores (Silver, 2016).  Lastly, they simulate the election, as uncertainty normally tends to decrease closer to Election Day.  FiveThirtyEight also accounted for national error, state-based error, and finally, regional or demographic error for the 2016 election forecast (Silver, 2016).

3.2 Sysomos MAP

Sysomos was the analytic tool used in the current study to acquire data from Twitter.  Sysomos is an analytic tool that performs machine analysis.  Sysomos Map contains an exclusive contextual sentiment engine where the entire text becomes classified mechanically through machine learning-based algorithms (Sysomos).  Sysomos (2017) states, “The sentiment engine has been trained on over 200,000+ human-tagged samples to understand and classify keywords as having negative, positive, neutral, or none” (para. 2).  A 4-step process is used to look for keywords, phrases, and language constructs associated with positive and negative meanings to determine sentiment.  Sysomos (2017) claims, “The MAP sentiment engine has been benchmarked at an accuracy rate of 85% (+/- 5%) however, it should be noted that assessing sentiment is a difficult task for a machine” (para. 3).  Firstly, words must go through a qualification phase that filters the several languages that Sysomos is able to read.  Secondly, an extraction of keywords that have passed through the qualification phase are extracted according to what Sysomos filters are set.  Thirdly, Sysomos sends all inquiries through the POV (point of view) verification, which is analyzing objectivity of the query requested, only sending subjective mentions to phase four.  Lastly, after the query passes the previous steps, the query is classified as negative, neutral, or positive.  Media Analysis Platform (MAP) was used to conduct a sentiment analysis and record volume of tweet mentions for Trump and Clinton.  Sysomos has the capability to track archived data for a total of one year or in real-time and has access to 100 percent of all tweets (Twitter Firehose) within a search criteria (Ampofo, Simon, O’Loughlin, Chadwick, Halfpenny & Proctor, 2015).  Sysomos is able to filter data in several different ways by demographics, country of origin, and state of origin if necessary.  This study only analyzes tweets that originate in the United States, filtering out every other location to understand how trends changed for candidates through possible electorates.

3.3 Data Collection

Collection of data began on July 1, 2016 and concluded on November 8, 2016 using Sysomos analytics.  A total of 655,500 tweets were collected.  Each calendar day was reported between these dates.  In total 131 days were reported in this study.  The time filter was set to run from 12:00 a.m. through 11:59 p.m. for every day included in the sample.  To acquire data for each candidate, altering the search queries according to opposing candidates was necessary.  This helped eliminate Twitter noise referring to family members or certain words that may be directly associated to either candidate.  Sysomos only allows for the collection of 5,000 tweets daily.

3.4 Volume of Mentions

To ensure the research obtained all tweets referring to Hillary Clinton, the keyword Hillary Clinton or Hillary or Clinton was used.  The Clinton name is associated with family members, foundations, and numerous other buzzwords (keywords) as found in searches.  Sysomos also allows researchers to filter out words that may not pertain to the candidate in question.  To eliminate Tweets that combine sentiment for both candidates the opposing candidates name was removed from each search query with the expectation that the sentiment rating was solely about one candidate.  For Clinton, the search query in Sysomos as follows (“Hillary”OR”Clinton”)AND NOT”Bill”OR”foundation”OR”Chelsea”OR”Donald”OR”Trump”).  Eliminating these words would allow for a distinct search for each candidate and offer a more concise sentiment analysis for the candidate in the data set.  It was also important to filter results in order to receive the Sysomos analytics solely from the United States.

To ensure the research obtained all tweets referring to Donald Trump, the keyword Donald Trump or Trump was used.  The Trump name is associated with family members, foundations, and numerous other buzzwords as found in searches.  Filters in Sysomos were also used to eliminate as much noise as possible that related to the Trump name that had nothing to do with the race for presidency.  To eliminate Tweets that combine sentiment for both candidates the opposing candidates name was removed from each search query with the expectation that the sentiment was just calculated for a one particular candidate.  For Trump, the search query in Sysomos was as follows (“Donald”OR”Trump”)AND NOT “Ivanka”OR”Melania”OR”Donald Jr.”OR”Barron”OR”Tiffany”OR”Eric”OR”Tower”OR”Hillary”OR”Clinton”).  Eliminating these words would offer a distinct search for Trump and offer a more concise sentiment analysis for the candidate in the data set.  Again, as stated above this data set also excluded any tweet not originating solely in the United States.

3.5 Sentiment Analysis

Sysomos analyzes everything in question such as word clouds, buzzwords, hashtags, volume of mentions, and sentiment analysis simultaneously.  However, because this study was comparing trends in sentiment throughout the campaign to FiveThirtyEight polling concurrently, a sentiment analysis was done separately for each day to acquire the maximum amount of tweets.  The daily maximum amount of tweets Sysomos allows for mining is 5000.  The same dates, days of the week, and search query was used for Trump and Clinton as named above.

The overall purpose of this study was to investigate existing relationships and trends that occurred during the presidential election cycle of 2016 by analyzing Twitter sentiment for the two remaining candidates (Trump, Clinton) of the major political parties in the United States along with polls that were conducted during this race.  This study seeks to display helpful information that candidates may consider in the future while on the campaign trail by recognizing various milestones and trends that occurred during this particular race that could possibly hinder or aid in the election process for candidates.  These trends in sentiment were recorded from Twitter and congruently from polling of FiveThrityEight (2016).  Limited research has been done regarding trends in Twitter sentiment for presidential candidates (Trump, Clinton) while comparing trends to polls.

Chapter 4: Analysis

4.1 Descriptive Statistics

This study used the 131 days leading to November 8, 2016, which was Election Day.  As seen in Table 1, each variable contained has been summarized.  For both candidates the highest mean and standard deviation was from the variable Twitter results which pertains to volume of tweets from possible electorates regarding each candidate with Trump having (M=645,713.27, SD=628,755.98) and Clinton having (M=379,903, SD=240,946.41).  Neutral sentiment as mentioned above is lacking negative or positive sentiment was also high for both candidates however, is virtually unimportant considering the experiment is looking at trends that would offer an indication as to which candidate would essentially win by looking at numerical parameters.  The chance of each candidate winning if the election were held on one of the 131 days of the sample (n=131) according to FiveThirtyEight (2016) shows a trend that is very different for each candidate.  Trump had substantially lower chance of winning statistically as his average was (M=26.442, SD=10.91), while Clinton had an average of (M=73.44, SD=11.052).  The intended popular vote by the electorate taken by FiveThirtyEight closed the gap some opposed to the chance of winning per day as Trump had (M=43.267, SD=1.143) and Clinton (M=49.9, SD=1.179).  Negative sentiment from Sysomos for both candidate also told a story through trends as in this race it seemed the candidate with the lowest negative sentiment eventually was victorious as Trump had a (M=11.769, SD=1.682) while Clinton had a (M=12.246, SD=2.119).  The lowest mean and standard deviation was calculated in the category of positive sentiment which was mined from Sysomos with Trump having (M=3.5, SD=.8025) and Clinton having (M=2.337, SD=.9681).

4.2 Inferential Statistics

For this particular study in order to identify trends through relationships between sentiment variables and variables from polls conducted by FiveThirtyEight (2016) (positive, neutral, negative, intended voting by electorate, chance of winning), a series of analyses were conducted using a Pearson product-moment correlation.

A Pearson product-moment correlation coefficient was computed to assess the relationship between six variables during this analyzation individually for each candidate.  The six categories are as follows and can be seen in Table 2 and Table 3.  Twitter results measured the volume of tweets given on each day for each candidate (Trump, Clinton).  Positive (Trump, Clinton) is the measurement of positive sentiment reported by Sysomos during the sentiment analysis.  Neutral (Trump, Clinton) is the measurement of sentiment in which no negative or positive sentiment was detected in Sysomos.  Negative (Trump, Clinton) is the measurement of negative sentiment detected in the Sysomos sentiment analysis tool.  The next two categories were from the polling of FiveThirtyEight.  The first one is listed as Intended popular vote by electorate, meaning that the candidate would essentially win the popular vote from the electorate.  The final category is titled chance of winning.  The chance of winning variable accounts for the candidates (Trump, Clinton) actual chance of winning the election if it were to be held on that particular day.  All of the Pearson product-moment correlations were 2-tailed.

4.2.1 Chance of Winning Election

Beginning with Trump, a two-tailed Pearson correlation showed various significant relationships.  Strong significant relationships were reported with Trump chance of winning and Trump intended popular voting by electorate.  The positive correlation between these variables indicated that when intended voting by electorate increased so did his chance of winning, r = (131) = .788, p < .001.  Clinton also had a strong positive significant relationship as expected also indicating the chance of winning increased as the as the intended voting by the electorate increased, r (131) = .924, p < .001.  Trump also had a positive correlation relating to positive Twitter sentiment.  The more positive  Twitter sentiment from the electorates became, the more his chance of winning increased showing a moderate relationship between the two variables r(131) = .348, p < .001.  This trend continued with Clinton.  Clinton also had a moderately significant relationship between the variable chance of winning and positive Twitter sentiment.  As her positive Twitter sentiment increased so did her chance of winning the election.

4.2.2 Intended vote by electorate

The intended vote by electorates showed relationships for both candidates with Trump Clinton.  Trump had a relationship that was weak while Clinton’s correlations showed a moderate relationship.  The positive correlation between intended vote by electorates raised when his positive sentiment was calculated in Sysomos raised r (131) = .226, p < .001.  This trend was quite different for Clinton.  When the intended electorate showing in FiveThrityEight (2016) polls increased, the positive sentiment for Clinton decreased creating a negative correlation relating to positive Twitter sentiment r(131) = -.350, p < .001.

Relationships were also shown for both candidates between intended votes by electorates Twitter results, which were volumes of mentions.  Trump had a positive correlation with Twitter results the same trend occurred for Clinton.  When the intended votes by electorates increased during polls from FiveThrityEight, so did Twitter results.  This happened to be a weak relationship as r (131) = .244, p < .001.  This trend also repeated for Clinton.  The positive correlation between Clinton intended voters by electorates and Twitter results indicated  that the higher the volumes of mentions the better chance Clinton would win votes by the intended electorate r(131) = .246, p < .001.

4.2.3 Negative Sentiment

Trends for negative sentiment in Twitter were only weakly related for Trump to Twitter volume.  As the positive correlation between negative sentiment and volumes of tweets indicated the more negative sentiment in tweets occurred the more mention he was getting in volume r (131) =.134, p < .001.  Clinton had a negative non-significant relationship between negative sentiment and Twitter results (volume of tweets).

4.2.4 Positive Sentiment

Positive sentiment for Trump had a positive non-significant relationship with Twitter results.  The positive correlation for Clinton was much more significant than Trump, but only moderately.  The correlation between positive sentiment Clinton and Twitter results specified that the more positive the sentiment, the higher the Twitter Volume raised r (131) = .330, p < .001.

Chapter 5: Discussion and Conclusion

In this study, a general examination of sentiment was done regarding presidential candidates by investigating social media.  Tweets were analyzed to illicit a more in-depth understanding of associations between Twitter sentiment and national polls.  Correlations were made displaying relationships to sentiment and polls from FiveThrityEight (2016) for both candidates.  The intended electorate has demonstrated that Twitter is becoming more commonplace as a viable way to discuss candidates, which created trends in sentiment for each political candidate throughout the final 131 days of the 2016 election cycle.

5.1 Findings

Changing trends for each candidate offer insights about sentiment during each candidate’s presidential campaign and transpire in the results of this study when compared to national polls.  First, it was found that the most significant relationships or correlations between sentiment and polls appear when there is a positive or negative sentiment present for each candidate.  Secondly moderate relationships were established between positive/negative sentiment and intended vote by electorate in different ways for each candidate.  The more Trump’s positive sentiment grew the better his intended vote became from the electorate increasing his chance of winning the election.  However, Clinton’s results were different.  The relationship became negative according to positive sentiment when relating this to intended vote, which also decreased her chance of winning the election.

Table Four demonstrates the highest and lowest positive and negative sentiment, chance of winning, and intended vote by electorate for each candidate.  There is a possibility that this could be attributed to certain events or situations that occurred during the 2016 election cycle.  These events could reveal why trends of sentiment and polls rose or plummeted.  These events were not scientifically related to sentiment but ensued on dates when sentiment was at the maximum and minimum juncture for each participating candidate.  Scandal for Trump and Clinton continued during the entire campaign process.

Discussing a few dates from Table Four should allow for some clarity.  Positive sentiment for Trump was lowest at the end of October 2016, and highest in mid July 2016.  At the end of October Jessica Drake, Trump’s eleventh accuser of sexual misconduct, came forward which may demonstrate why his positive sentiment was at an all-time low (Kenny, 2016).  Trump’s highest positive sentiment occurred in mid-July when he announced Mike Pence ad his Vice President during the RNC (Brander, Bush, & Lee, 2016).  Trump’s negative sentiment was lowest at the beginning of July 2016 even though accusations of sexual assault continued surfacing from an individual that said Trump assaulted her when she was thirteen.  His highest negative sentiment occurred in mid-September 2016 as he called inner cities crime-ridden and jobless.  In addition, Donald Trump Jr. referenced that they would be warming up the gas chamber if Clinton were a Republican in regards to her email scandal, which were noted in the media to anti-Semitic (REPUBLICINSANITY, 2016).

Clinton suffered similar setbacks and highlights for positive and negative sentiment that also offers insights as to why her trends continued changing.  Positive sentiment for Clinton was lowest in Mid-August of 2016 and highest at the end of September of 2016.  Clinton’s scandals differed from Trump’s however, were just as damaging.  The DNC email was hacked.  A group called Anonymous published several emails daily.  Early August when Clinton’s positive sentiment was lowest emails were released exhibiting ties with the State Department workers and the Clinton Foundation suggesting that” loyal supporters of Clinton should be found a position in Washington” (Fain, 2016).  Emails continued to be released during the duration of her campaign with some being more damaging than others were.  Clinton’s highest positive sentiment came directly after the first presidential debate as she was deemed the winner by several news outlets (Fain, 2016).  The lowest negative Sentiment for Clinton came in mid-July when the director of the FBI announced they found no wrongdoing by Clinton in the investigation into her having confidential emails on her personal email account (Hartig, Lapinski & Psyllos, 2016).  The highest negative sentiment occurred for Clinton in mid-September as she stumbled on the campaign trail raising questions of her health and transparency (Collinson, 2016).

Also by looking at Table Four, it is easily seen that the minimum and maximum dates for negative and positive sentiment occurred within a two-week range of each candidates highest and lowest chance of winning and intended voting by electorate.  When comparing Trump and Clinton’s positive and negative sentiment at a glance, it is easily seen that the candidates had similar lows and highs.  However, when observing variables from national polls (intended vote by electorate, chance of winning) these figures offer a different outcome.

5.2 Linking Related Work to Current Study

During the last decade, the concentration on sentiment from Twitter has rapidly grown.  This may be attributed to an increase of interest in personal opinions on various topics that users turn to Twitter to divulge.  Aforementioned studies used Twitter sentiment for predicting political outcomes, stock market, and feelings about climate change.  In the German election of 2009, Tumasjan (2011) associated Twitter volume served somewhat as a predictor to the winner of the election.  This study was not looking for predictive factors however, found relationships between Twitter sentiment and polls just as this study revealed.  Even though the awareness that these relationships are not casual, they are able to shed some light on the possibilities of studying Twitter sentiment for upcoming events.  In the political realm, these studies continue to grow as Wang (2012) established through his work the development of a system that performs real-time sentiment of the entire presidential election, which was recorded on an interface that tracked domineering keywords that were deemed positive or negative via a nave Bayes modeling system.  This study was similar in findings when relationships were established between Twitter sentiment and tweet volume in the final days of the election cycle.  Correlating polls to Twitter sentiment just as this study has done has benefited politicians in the area of measuring public opinion (O’Connor 2010).

5.3 Research Contributions and Practical Implications

This particular study makes several contributions to research.  First, it is demonstrated that Twitter is being used by intended electorates for political communication during election cycles as shown in other previous studies by way of tweeting opinions on policy and political candidates.  Secondly, this study offers an extension of current literature by probing how positive and negative sentiment attributes to a candidates chance of winning an election and the possibility as to how the intended electorate plan to vote.  Furthermore, this study addresses the relationships that Twitter sentiment has on existing polls.  Given the fact that social media has become relevant as an arena for political communication, the importance for politicians to use social media is growing more in each election cycle.

5.4 Limitations and Future Research

Many researchers are gauging for sentiment in different paradigms and are attempting to formulate a theory for a strong theoretical foundation.  This study and many other similarly lack a strong theoretical foundation to test research questions and hypotheses.  Limitation of this study also derives from the analysis of data restricted to Twitter.  Twitter is not fully representative of the electorate.  Firstly, not all electorate or Twitter users tweet about politics.  Secondly, the population on Twitter tends to be a younger generation, which would not represent an older voting population.  This raises issues of generalizability.  Future research may extend sentiment studies to other social media outlets and include Twitter to represent a broader electorate.  Another limitation derives from using tweets because of the short length of text offered.  This is another reason future research should include other social media sites or blogs when gauging public sentiment for political candidates.  Lastly, limitations of this study come from using Sysomos MAP as a method for collecting.  Sysomos cannot filter for sarcasm and restricts the number of tweets a researcher can obtain daily from Sysomos.  Future research might include gathering tweets with and without the use of electronic mining and use human coders to code for sentiment along with using Sysomos even though Sysomos claims accuracy in upwards to 85%.

5.5 Conclusion

The predominate goal of this study was to increase a thorough conception of trends in sentiment on Twitter for the two remaining candidates (Trump, Clinton) during the 2016 presidential election cycle and then compare these trends to national polls.  Twitter sentiment is a viable avenue to use for researchers when gauging public opinion.  Twitter sentiment has the unique capability of indicating public opinion and sentiment for political candidates when used congruently with polling and is practical for estimating what candidate may win upcoming elections.  The current study aides in helping both researchers and politicians to better understand the political discourse and the function of sentiment in information dispersion on Twitter.  Correlating polls with sentiment can provide important practical uses for politicians while also furthering goals in research.