Pride, Love, and Twitter Rants: Combining Machine Learning and Qualitative Techniques to Understand What Our Tweets Reveal about Race in the US

Objective: Describe variation in sentiment of tweets using race-related terms and identify themes characterizing the social climate related to race. Methods: We applied a Stochastic Gradient Descent Classifier to conduct sentiment analysis of 1,249,653 US tweets using race-related terms from 2015–2016. To evaluate accuracy, manual labels were compared against computer labels for a random subset of 6600 tweets. We conducted qualitative content analysis on a random sample of 2100 tweets. Results: Agreement between computer labels and manual labels was 74%. Tweets referencing Middle Eastern groups (12.5%) or Blacks (13.8%) had the lowest positive sentiment compared to tweets referencing Asians (17.7%) and Hispanics (17.5%). Qualitative content analysis revealed most tweets were represented by the categories: negative sentiment (45%), positive sentiment such as pride in culture (25%), and navigating relationships (15%). While all tweets use one or more race-related terms, negative sentiment tweets which were not derogatory or whose central topic was not about race were common. Conclusion: This study harnesses relatively untapped social media data to develop a novel area-level measure of social context (sentiment scores) and highlights some of the challenges in doing this work. New approaches to measuring the social environment may enhance research on social context and health.


Introduction
A social climate in which greater hostility towards minorities is manifested may cause psychological stress and increase risk of negative health outcomes.
Experiences with discrimination are commonly measured at the individual level by self-report [1,2].Self-reported racial attitudes and beliefs are subject to a number of limitations including social desirability bias and self-censorship [3,4], risking invalid exposure assessment [5,6].Self-reports of racial discrimination are also subject to reporting bias.This may include, among other things, coping (e.g., denial), trait-or state-based aspects of personality (e.g., stigma consciousness, race-based rejection sensitivity), and aspects of racial identity (e.g., internalized racism) [5].Cognitive measures have been developed to assess implicit racial bias [7,8].Audit studies have also documented discrimination in a variety of areas including housing [9] and employment [10].However, all these measures are aimed at characterizing individual level experiences and bias.
While individual level measures of racial bias and discrimination are valuable in documenting individual experiences, the social climate of a place can provide an ecological perspective for understanding one's experiences in relation to the broader social environment.An ecosocial approach to the study of discrimination views discrimination as operating across multiple levels over the life course and reflecting systemic prejudice, which has emergent properties of its own despite individual level experiences.For example, a landmark study examined the influence of contextual indicators of discrimination on birth outcomes among pregnant women of Arab descent.Comparing birth outcomes for women by race/ethnicity and nativity for the 6 months after 11 September 2001, to the same six-month period one year prior, only Arabic-named women experienced significantly increased risk of preterm birth and low birth weight post September 2001 [11].The study did not measure individual women's personal exposure to harassment, discrimination, violence, etc.However, after 11 September 2001, the social climate for Arabic-named women had changed, and was associated with increased risk of adverse birth outcomes for this population.Prejudice, antipathy towards a group based on poorly founded generalizations [12], has broad and important implications for health and development.This, and other studies, provides evidence relating racial prejudice to the health of communities [13].
Social media data offer an increasingly popular data resource for assessing social climate.Social media provides a space and opportunity for people to publicly express their ideas and viewpoints and represents what many believe to be a relatively untapped resource for assessing the contextual level social climate related to race.With 21% of US adults using Twitter and over 90% of Twitter users making their profile and communication public [14], social media provides researchers the opportunity to examine the public communications of a substantial proportion of the country.Social media therefore presents some advantages in illuminating national and potentially place-specific sentiments about race/ethnicity, providing a "temperature" of the social environment where the tweets are written.On Twitter, users are not required to report their age, sex, race, or geographic location, and as a result, grants people a level of anonymity.Studies have found that people feel less inhibited in expressing their views and beliefs online compared to in-person interactions [15][16][17].Twitter and other social media data have been used to describe national patterns in happiness, diet, and physical activity [18,19]; examine beliefs, attitudes, and sentiment towards various topics (e.g., vaccinations) [20]; track health behaviors and perform health surveillance [21], and investigate patient-perceived quality of care [22].
Few studies have used social media to examine more sensitive topics such as race and racism online.A previous study on internet-based racism found area-level racism, operationalized as the proportion of Google searches containing "n-word" was positively associated with all-cause Black mortality [23].Like this study, prior studies have primarily focused on the use of racial slurs [24,25].However, terms conventionally perceived as racial slurs can be used in non-derogatory ways, and such re-appropriation is common on Twitter; for instance, in popular culture the term "n*gga" is often used as an in-group term without valuation [24], making assessment of racial sentiment challenging.Furthermore, discussions conveying racial sentiment can occur without the use of racial slurs.A more comprehensive examination of tweets using race-related terms may include a sentiment analysis of tweets using racial slurs as well as neutral racial terms such as "Black", "African American", or "Asian".Fewer positive tweets using race-related terms in a state or county may indicate an environment that is less welcoming to racial and/or ethnic minorities, which may be a source of stress.A greater number of positive tweets using race-related terms may be indicative of an environment that embraces racial and ethnic diversity and is more inclusive.Therefore, examining positive and non-positive tweets may provide a fuller picture of the social context of a place.
To provide a measure of social climate in relation to race and ethnicity and address prior limitations of self-reported, individual-level measures, we employed mixed methods to (1) examine variation in sentiment regarding tweets using one or more race-related terms, and (2) identify emerging themes of the tweets.This paper describes the collection and sentiment analysis of Twitter data using one or more race-related terms and the qualitative content analysis of a subsample of tweets to provide a more contextual-level understanding of the social climate related to race.

Social Media Data Collection and Processing
From March 2015-April 2016, we utilized Twitter's Streaming Application Programming Interface (API) to continuously collect a random sample of publicly available tweets.Twitter's Streaming API gives users access to a random 1% sample of tweets.The Twitter API is freely available to everyone and this API allows users free access to subsets of public tweets.Users may request tweets for a certain geographic area that contain a particular set of keywords, or just random subsets of tweets.Depending on the search criteria used by the researcher, the number of tweets returned may comprise less than 1% of all available tweets.In our case, we restricted the data collection to tweets with latitude and longitude coordinates that were sent from the contiguous United States (including District of Columbia).We dropped duplicate tweets according to their "tweet_id" (each tweet has a unique identification).We removed job postings according to the hashtags "#job" and "#hiring."We manually examined outliers in our datasets (the top 99th percentile of tweeters) and eliminated automated accounts and accounts for which the majority of tweets were advertisements.In total, we collected 79,848,992 million general topic tweets from 603,363 unique Twitter users.
To identify potentially race-related tweets, a keyword list of 398 race-related terms was compiled (online supplementary materials Table S1 with the most commonly occurring terms bolded) from racial and ethnic categories used by the US census and an online database of racial slurs [26].Tweets using at least one or more of the race-related terms were identified resulting in a final analytic dataset of 1.25 million tweets.Location information from the tweets was used to map the tweets to their respective county and state using Python and R-tree to build the spatial index [27,28].Count of tweets using race-related terms by state can be found in the online supplementary materials (Table S2).This study was determined exempt by the University of California, San Francisco Institutional Review Board (Ref: 18-24255).

Computer Modeling: Sentiment Analysis of Twitter Data
To prepare the dataset of tweets for analysis by the computer algorithm, each tweet was divided into tokens, which roughly correspond to words, using the Stanford Tokenizer [29], an open access software tool.Below, we briefly describe the algorithms we used to create variables for sentiment.
To conduct sentiment analysis on tweets with references to racial and ethnic minorities, we used the Stochastic Gradient Descent Classifier (SGD), an optimization method that minimizes a given loss function [30], in Python software version 2.7 (Python Software Foundation, Wilmington, DE, USA).In SGD, weights in the sentiment models are updated for each example in the training dataset [30], and several iterations are made over the training data until the algorithm converges.A strength of SGD is that it is quick to train and has been applied to large-scale and sparse-learning problems [31].Preprocessing of the tweets was undertaken to remove inconsequential variation and allow the sentiment model to focus on the relevant features of the tweets [32].These preprocessing steps included removing stop words (e.g., the, a, is), additional white space, punctuations, hashtag symbols, URLs, and Twitter usernames.All words were converted to lowercase, and the repetition of a character (symbol or alphabet letter) was replaced by one instance of the character.
Labeled tweets from Sanders Analytics (n = 5513 tweets) [33] and Kaggle (n = 7086 tweets) [34], and emoticons derived from Sentiment140 (e.g., smiley face to indicate happiness, n = 1.6 million tweets) [35] were used to train the computer algorithm to analyze the tweets.The computer algorithm then uses these labeled tweets to learn what a human considers a "positive" or "negative/neutral" tweet.Once trained, the computer algorithm then categorizes "positive" and "negative/neutral" tweets from our sample of 1.25 million tweets.
Preliminary analyses revealed that the SGD algorithm's accuracy against manual annotations was substantially greater for dichotomous sentiment compared to multiple sentiment categories.Specifically, the model was able to distinguish positive vs. non-positive tweets but had much lower accuracy when three (positive, neutral, and negative) categories were used.According to the race-related terms referenced in the tweet, we grouped tweets into four main racial/ethnic categories: Blacks, Hispanics, Asians, and Middle Eastern.The latter included tweets that were anti-Islamic or related to Muslims.
Tweets were assigned values of 1 (for positive) or 0 (for negative/neutral).State-level sentiment variables were created by averaging the dichotomous sentiment of tweets referencing various racial/ethnic groups.Next, state-level sentiment was mapped in order to examine geographic variation.Statistical analyses were implemented with Stata MP15 (StataCorp LP, College Station, TX, USA).
After training the computer algorithm and conducting the sentiment analysis on our novel Twitter dataset, we evaluated the accuracy of the computer algorithm using a randomly selected, manually labeled subset of tweets.Three coauthors (LP, TR, and SD) categorized the sentiment of the tweet as negative, neutral, or positive.In addition to labeling for sentiment, the coders indicated whether the tweet used discriminatory or stereotyping language about a racial or ethnic group (Yes/No).A sample of 150 tweets was labeled by all coders, and discrepancies were resolved.Coders then analyzed 150 more tweets, and an acceptable level of agreement was reached (kappa = 90%).Coders then independently coded 2000 tweets.Agreement on the coding of positive and non-positive tweets between the computer algorithm and the manual labels was computed.Please see Figure S1 in the online supplementary materials for a diagram presenting the analytic sample and analysis steps.

Content Analysis
A content analysis exploring themes emerging from a random sample of 2100 tweets using race-related terms was conducted.Two members of the research team developed a codebook (i.e., a list of codes and definitions representing the emerging themes) based on a literature review and coding and discussing 150 tweets from the above sample.This consensus building process enhanced the codebook by clarifying operational definitions.The final categories of tweets included an overall categorization of sentiment (negative, positive, or neutral) and the following sub-headings by sentiment conveying the emergent themes: casual use/slang, food, stereotypes, and related to sexuality/relationships. Complaint, insult, and hostility were additional sub-headings specific to negative sentiment.Using this coding scheme, two members of the research team coded a randomly selected sample of 2100 tweets.The researchers double-coded 300 tweets throughout the process (100 tweets in the beginning, 100 tweets in the middle, and 100 tweets near the end of the sample).Any disagreements in coding were discussed until consensus was met.Kappa agreement of 80% was met for each inter-coder session.In addition to the 300 double-coded tweets, each of the two team members independently coded 900 tweets.

Results
Of the 6600 tweets using a racial term which were manually coded, about 25% were coded as positive, 44% as neutral, and approximately 30% as negative.Only a small proportion (about 6%) of tweets with a race-related word used discriminatory or stereotyping language.Agreement on the dichotomous characterization between computer labels and manually generated labels was 74%, which is similar to the 76% accuracy found in a prior study labeling racist tweets [36].

Quantitative Analyses: Descriptive Characteristics of Tweets
Below, we describe results from our quantitative analyses, based on the SGD algorithmic classification of the sentiment (positive vs. neutral/negative) of 1,249,653 tweets containing at least one of the relevant keywords pertaining to a racial or ethnic minority group.Descriptive details of these tweets have been previously described [37].Briefly, approximately 620,000 tweets were about Blacks, 205,000 about Hispanics, 270,000 about Asians, and 60,000 about Middle Eastern groups.From a list of 398 terms, only 20 terms were necessary to characterize 84% of all tweets with references to a racial or ethnic minority group (top 20 terms bolded in list in Appendix A Table A1).
The top Twitter terms were "n*gga" (42.6%) (please note that when "*" is present, it was inserted by the study team and not part of the original text), "Mexican" (8.4%), "Thai" (4.2%), and "Asian" (4.0%).The automated sentiment classification characterized the tweet overall.Most tweets including a racial term were not specifically about race.We found that a "positive" tweet does not necessarily imply a positive racial sentiment, but merely a positive sentiment in which a racial term was used.For example, "@username I love you, n*gger."Overall, 15.2% of tweets using race-related terms expressed a positive sentiment; 13.1% of tweets containing the word "n*gga" were positive.The term "n*gger" was exceedingly rare, with only 507 mentions in our entire dataset, of which 12.7% were positive in sentiment.
States where tweets using a race-related term were least likely to express a positive sentiment were Nevada and Louisiana with 9-12% of tweets being positive.States with the most positive sentiment tweets were Utah, Oregon, and North Dakota with 17-20% of tweets being positive (Figure 1).Across the United States, tweets referencing Asians (17.7%) and Hispanics (17.5%) had the highest percentage of positive sentiment.Tweets that referenced Middle Eastern groups had the lowest positive sentiment (12.5% positive) followed by tweets that reference Blacks (13.8% positive).(See online supplementary materials Figures S2-S5 for maps of positive sentiment by race/ethnicity).When we weighted tweets by the number of followers (i.e., tweets sent by a user with more followers were weighted more heavily than tweets sent by a user with fewer followers), positive sentiment increased at least slightly for tweets using Black, Hispanic, and Asian related terms, but not for tweets using a term related to Middle Eastern groups.For Middle Eastern groups, the weighted percentage of positive tweets, 8.9%, was lower than the unweighted percentage of 12.5% (see Figure 2).We examined variation in sentiment by race across months and found relatively stable sentiment that maximally differed by 3-5% (Table 1).

Qualitative Content Analysis: Themes
In the random sample of 2100 tweets for the thematic analysis, 45% had themes expressing a negative sentiment, 25% had themes expressing a positive sentiment, and 15% had themes connected to intimate relationships (including negative and positive sentiment).Two percent of tweets were

Qualitative Content Analysis: Themes
In the random sample of 2100 tweets for the thematic analysis, 45% had themes expressing a negative sentiment, 25% had themes expressing a positive sentiment, and 15% had themes connected to intimate relationships (including negative and positive sentiment).Two percent of tweets were derogatory or expressed a racial/ethnic stereotype (Table 2).While all tweets use one or more race-related terms, the central topic for many of the tweets was not about race.

Spectrum of Negativity
Tweets ranged from innocuous statements about daily life, to complaints, insults using general derogatory language, and expressed hostility mentioning violence.Although uncommon, some tweets also mentioned racial or ethnic stereotypes or were racially derogatory (e.g., "Middle Eastern/Arabic accents piss me off more than most things," Table 2).The use of "n*gga" was pervasive in the negative sentiment tweets.However, most of the tweets using this term were not derogatory.Rather, Twitter users most frequently use this term casually as slang.The use of profanity was common in negative tweets.

Pride
Most of the positive sentiment tweets related to pride.Some Twitter users indicated love for their culture and food.Some tweets had a tone of pride for being loyal to their friends throughout the years.In addition, several tweets were rejecting negative views of their race and offering alternative viewpoints.

Intimate Relationships
Many of the tweets focused on some aspect of an intimate relationship.Specifically, these tweets included comments about appearance, affinity to a particular race, cheating, frustration over behavior of men or women, sex, and introspection about one's own behavior."N*gga" was a common term used in the relationship tweets, though it was not generally used as a racial slur but rather as slang.Many tweets about appearance and affinity towards a particular race regarding dating and marriage were positive.The tweets about cheating, sex, and frustration over the behavior of men or women were mainly negative and many times included profanity, but were not generally about race although a race-related term was used.On the other hand, racial terms were, at times, used to aid in an insult, such as tweets about cheating and unfavorable views about men or women ("N*ggas only lie cus females can't handle the 100% truth", Table 2).Some users' tweets indicated introspection and reflection about past and future relationships (I'm f*cking up by pushing people away and acting like a n*gga," Table 2).

Discussion
In this study, we examined tweets using a race-related term that were collected over a year-long period.Sentiment analysis on the full sample of 1.25 million tweets using natural language processing revealed sentiment expressed in tweets differed across the race/ethnic group mentioned.Tweets mentioning Middle Eastern groups or Blacks were least likely to express positive sentiments.Sentiment varied moderately across calendar month, with the maximum variation being about 2-5%, indicating sentiment at the state level was relatively stable over the one-year period.The quantitative analysis revealed that tweets that expressed discriminatory sentiments were rare, highlighting the need for a more in-depth analysis of tweets utilizing race-related terms.Therefore, we conducted a qualitative content analysis of a random sample of 2100 tweets that elucidated three core themes: spectrum of negativity, pride, and various aspects of intimate relationships.Understanding the nuances of the sentiment can provide more insight about whether and why patterns in tweet sentiments could plausibly be linked to a particular health outcome.
Tweets may not represent beliefs.Tweets may reflect an image a person may be promoting or another motive.However, because tweets are expressed statements that people are willing to make publicly, they may reflect an area's social acceptability or willingness to showcase certain sentiments.As such, we leverage tweets to gauge the racial climate of a state.State-level differences in Twitter activity can reflect changes in willingness to vocalize negative sentiment.Increased vocalization could be an additional source of stress.Whereas before, negative sentiment was more hidden, changes in public acceptance or willingness to display negative sentiment can be an additional source of stress and impact health and well-being.Previous studies have found that social media can capture information about the social environment that has utility for predicting health outcomes.For example, Eichstaedt et al. found greater use of anger, negative emotion, and disengagement words on Twitter predicted county-level heart disease mortality [38].Notably, previous research found more negative area-level sentiment towards blacks and Middle Eastern groups was related to worse individual-level birth outcomes, and this was true for the full population and for racial minorities [37].Thus, a hostile social climate related to race may have implications for health and well-being, but developing valid and inexpensive measures remains an important research challenge.
In this paper, social media data were used to examine sentiment towards racial/ethnic minorities.Limitations of the computer algorithm include the inability to identify sarcasm or humor.In addition, the tweets using one or more race related terms may be a part of a conversation but only tweets using specific-race related keywords were analyzed.Thus, the context in which the tweet is embedded and the context in which the specific racial keyword is used may be lost.
The content analysis revealed the central topic of numerous tweets, despite using one or more race-related terms was not, in fact, about race.Our analysis was conducted at the state level, and there may be individuals within a smaller area that are exposed to a higher proportion of derogatory tweets than the state average.Overall, negative sentiment tweets that were not racially derogatory or making stereotypical statements were common.For example, one of the most frequently used race-related term was n*gga.While many tweets using this term had a negative sentiment in terms of the context of the tweet, the term itself was often used casually as slang.The content analysis revealed more clearly the limitations of the sentiment analysis to identify tweets that were not centrally related to race although using race-related terms.Nonetheless, the tweets using at least one of the 398 race-related terms covering a range of topics may represent a signal of the average level of racial attitudes, given our prior work showing an association between state-level sentiment scores and preterm, low birth weight, and very low birth weight [37].Future work can develop algorithms to categorize tweets by sentiment as well as by topic.
The present analysis was based solely on the text within the tweet; images and videos could not be analyzed.We used race-related keywords to identify tweets.However, tweets that do not use racial slurs or neutral racial terms may be race-related.Previous studies have found 1-2% of all tweets have latitude and longitude coordinates [39].The analysis includes only geotagged tweets, which may vary from tweets without geotagged information.This study assessed the social environment via online tweets, and these may differ from in-person interactions.However, when assessing expressed attitudes and beliefs online, people may feel less inhibited due to anonymity and invisibility of being online [15].
In the current study, tweets sent from that state were utilized to construct indicators of sentiment towards racial/ethnic minorities.Social media users tend to be younger than the general population with 36% of people 18-29 using Twitter in 2016 while 10% among those 65+ years reporting Twitter use [40].Thus, Twitter users may not be representative of the broader US population.However, social media has been increasing over time, and access to the internet using cellphones has resulted in people from all socioeconomic backgrounds using Twitter.Tweets also include information seldom found in conventional data sources with researchers being able to capture real-time conversations about daily life, activities, and beliefs.

Conclusions
This paper is among the first to characterize the sentiment of tweets using race-related terms.In this paper, we address the limitations of self-reported measures of discrimination and racial bias to create an environmental indicator of the social context (state-level sentiment scores) from expansive and relatively untapped social media data.Self-reported measures provide valuable information about individual-level experiences, but these measures may be subject to reporting bias.Online discussions using race-related terms can be used to better understand the social context for areas across the United States.However, this area of research is in its infancy.Importantly, our findings revealed some of the challenges and complexities of characterizing tweets using race-related terms and the importance of better understanding what, exactly, social media data are capturing.The study findings highlight the difference between sentiment of the tweets, the topic of the tweet, and the ways in which the racial terms are being used (e.g., in derogatory or neutral ways).Despite these challenges, there remains much to be learned from these data.Future work is needed to further identify and categorize race-specific tweets, and to examine whether the social context influences the health, economic, social, and educational outcomes of minorities and immigrants to the United States and thus identify levers for alleviating disparities.S1: Race terms used in Twitter data Collection, Table S2: Count of tweets using race-related terms by state.

Figure 1 .
Figure 1.Geographic distribution of percent tweets using race-related terms that are positive, collected April 2015-March 2016.

Figure 2 .
Figure 2. Descriptive characteristics for sentiment of 2015-2016 tweets using race-related terms.Data sources: 1,249,653 geolocated tweets from the 48 contiguous United States plus the District of Columbia collected between April 2015-March 2016.Number of followers was used to calculate weighted percentages.

Figure 2 .
Figure 2. Descriptive characteristics for sentiment of 2015-2016 tweets using race-related terms.Data sources: 1,249,653 geolocated tweets from the 48 contiguous United States plus the District of Columbia collected between April 2015-March 2016.Number of followers was used to calculate weighted percentages.

Supplementary Materials:
The following are available online at http://www.mdpi.com/1660-4601/16/10/1766/s1,FigureS1: Analytic Sample and Analysis Flow Chart, Figure S2: Geographic distribution of percent tweets using Black-related terms that are positive, collected April 2015-March 2016, Figure S3: Geographic distribution of percent tweets using Middle Eastern-related terms that are positive, collected April 2015-March 2016, Figure S4: Geographic distribution of percent tweets using Hispanic-related terms that are positive, collected April 2015-March 2016, Figure S5: Geographic distribution of percent tweets using Asian-related terms that are positive, collected April 2015-March 2016, Table
Data sources: 1,249,653 geolocated tweets from the 49 contiguous United States collected between April 2015-March 2016.

Table 2 .
Content analysis themes of 2015-2016 US tweets with illustrative Examples, 2015-2016.Mexican flag in Az. they need to be arrested.
FoodGod bless Mexicans and their delicious food <33,343 [url] Mexican food and family can definitely turn a day around We are getting Chinese food for lunch.I repeat.We are getting Chinese food for lunch.Im not in the business of keep a n*gga that aint tryna be kept When "*" is present, it was inserted by the study team and not part of the original text.Data sources: 1,249,653 geolocated tweets from the 49 contiguous United States collected between April 2015-March 2016.