Contraception 90 (2014) 502–507

Original research article

Finding the Twitter users who stood with WendyAmanda Jean Stevenson

Population Research Center and Department of Sociology, The University of Texas at Austin, 305 E. 23rd Street, Stop G1800, Austin, TX 78712-1699

Received 19 June 2014; revised 8 July 2014; accepted 16 July 2014

Abstract

Objective: I examine Twitter discussion regarding the Texas omnibus abortion restriction bill before, during and after Wendy Davis’filibuster in summer 2013. This critical moment precipitated wide public discussion of abortion. Digital records allow me to characterize thespatial distribution of participants in Texas and the United States and estimate the proportion of participants who were Texans.Study design: Building a dataset based on all hashtags associated with the bill between June 19th and July 14th, 2013, I use GPS locationsand text descriptions of locations to classify users by county of residence. Mapping tweets from accounts within the continental United Statesby day, I describe the residential composition of the conversation in total and over time. Using indirect estimation, I compute an estimate ofthe number of Texans who participated.Results: About 1.66 million tweets were sent using hashtags associated with the bill from 399,081 user accounts. I estimate counties ofresidence for 160,954 participants (40.3%). An estimated 115,500 participants (29%) were Texans, and Texans sent an estimated 48.8% of alltweets. Tweets were sent from users estimated to live in every region of Texas, including 189 of Texas’ 254 counties. Texans tweeted morethan non-Texans on every day except the filibuster and the day after.Conclusion: The analysis measures real-life responses to proposed abortion restrictions from people across Texas and the United States. Itdemonstrates that Twitter users from across Texas counties opposed HB2 by describing the geographical range of US and Texan abortionrights supporters on Twitter.Implications: The Twitter discussion surrounding Wendy Davis’ filibuster revealed a geographically diverse population of individuals whostrongly oppose abortion restrictions. Texans from across the state were among those who actively voiced opposition. Identifying rightssupporters through online behavior may present a new way of classifying individuals’ orientations regarding abortion rights.© 2014 Elsevier Inc. All rights reserved.

Keywords: Abortion opinion; Social media; Reproductive health; Variability

1. Introduction

On June 25, 2013, Wendy Davis stood on the floor of theTexas Senate for 11 hours to filibuster HB2, an omnibusabortion bill that promised to dramatically decrease thenumber of clinics providing abortion care in Texas, banabortions after 20 weeks “postfertilization,” require allphysicians providing abortion care to have admittingprivileges at a hospital within 30 miles of the facilitywhere they worked, and impose restrictions on the provisionof medication abortion. As she spoke, the Texas Capitolfilled with thousands of supporters and opponents of

E-mail address: stevenam@prc.utexas.edu.

http://dx.doi.org/10.1016/j.contraception.2014.07.0070010-7824/© 2014 Elsevier Inc. All rights reserved.

abortion rights, while 180,000 watched via livestream.Supporters of abortion rights had been rallying in the daysbefore Davis’ filibuster, and they returned day after day tooppose the bill. A hashtag, #StandWithWendy, arose onTwitter as supporters online expressed their outrage with thebill and their support for Davis.

Tweets about the bill and the filibuster represent real-world responses to the proposed restrictions. Johnson-Hanksand coauthors [1] propose a theory of action situatingindividuals’ behaviors within conjunctures, or short-termsets of conditions under which action occurs. Wendy Davis’filibuster and the prospect of HB2’s passage presentedTwitter users in Texas with a conjuncture. How theyresponded to that conjuncture reveals the orientation anddegree of their reaction to the prospect of their very largestate being left with only six or seven abortion clinics.Texans living in or near Austin had the option of marching

503A.J. Stevenson / Contraception 90 (2014) 502–507

on the Capitol to express their opposition to the bill, butthose living far away were likely less able to expressthemselves this way. While tweeting is not equivalent tomarching in the streets, it is something more than privatelyholding an opinion, and thus measuring tweets from theconjuncture during which HB2 was being debated measuresan avenue of action available to all Texans who used Twitter.Taking this interpretation, tweets in support of abortionrights provide data for describing a population of particularlyimpassioned abortion rights supporters without relying onresponses to survey questions. This approach obviatesdesign effects because it measures responses to real-worldevents rather than hypothetical vignettes, but the fact thatTwitter users are substantially different from the generalpopulation means that generalization beyond the descriptionof the discussion is impossible [2–5]. Thus, this analysisdescribes the spatial range of the Twitter conversationaround HB2 and Davis’ filibuster, not the distribution ofabortion opinions in general in Texas.

The theory and method for analyzing social media dataare in their infancy, so I therefore ask simple, basic questionsabout how many people participated, where they lived andwhich side they supported. Whether Texans were among theTwitter users who supported abortion rights in theconversation can help us interpret the outpouring on Twitteras either local resistance or outside resistance to the proposedlaw. Estimating where users lived within Texas couldcombat a perception that the protests at the Capitol reflectedAustin’s liberal population, not a broad base of Texans fromacross the state. While election results indicate thatDemocrats live throughout Texas, not all Democrats supportabortion rights. Thus, I use the Twitter data to investigatewhether or not a geographically diverse population ofTexans participated in the digital resistance to HB2. IfTexans from across the state participated in the discussionand expressed their outrage with the bill, the long-termprospects of abortion access in the state may be moremalleable than if the resistance was concentrated in Austinand outside the state.

2. Materials and methods

2.1. Data sources

I formulated an initial list of hashtags — which are usedon Twitter to build conversations and identify positions —by reviewing Twitter activity during HB2’s proposal, debateand passage. I validated the list through interviews withkey informants (journalists, bloggers and social mediamanagers) and checked it against the Twitter feeds oforganizations and politicians on both sides. Hashtags used atleast five times in reference to the filibuster and HB2 on agiven side were included. The final list included thefollowing: neutral bill name hashtags (#sb1 #hb2, #hb60and #sb5), hashtags used by supporters of abortion rights(#StandWithWendy, #prochoice, #StandWithTXWomen,

#SWTW and #feministarmy) and hashtags used by opponentsof abortion rights (#SitDownWendy and #prolife).

Based on this list of hashtags, I purchased a comprehen-sive dataset of tweets with any of the hashtags from June 19through July 14, 2013, from TweetReach. The datasetincludes both tweets and retweets. The dates covered includeall major events involved in the bill’s passage through theTexas legislature. June 19 was the day before the bill’s firstlarge public hearing, and July 13 was the day the bill finallypassed. Wendy Davis’ filibuster was June 25. The tweet datainclude tweet text, user name, tweet day and time, and othertechnical information.

2.2. Measures and analysis

The tweets themselves do not have locations. In order tobuild the dataset needed to estimate a location for each useraccount, I used the Twitter Application ProgrammingInterface (Twitter REST API v1.1) to collect data on useraccounts. The Twitter API is a feature of the Twitter websitethat allows direct access to some Twitter data from acomputer. For each account whose tweets had GPS data, Icollected 100 tweets from the Twitter REST API v1.1. Forall accounts, I collected location data from user profiles inthe form of text strings. The text strings sometimes includedlatitude and longitude coordinates. Using geocoded textstrings in addition to a sample of GPS encoded tweets, thismethod generates location estimates for about four times asmany Twitter user accounts as typical analyses that relysolely on GPS data [6,7].

I estimated counties of residence in up to two ways, firstusing GPS encoded tweets and second by geocoding textstrings from the user profiles. For user accounts with GPScoded tweets, I determined the county of each of the 100tweets by joining the tweet’s GPS coordinates with ashapefile of the United States. I estimated the user’sresidence as the most frequent county of tweet origin, aslong as more than 50 tweets originated in it. When GPSlocations were outside the United States, I coded them asoutside Texas for the purposes of this analysis. I performedthese processes with Python and the geopy package.

For all user accounts with non-missing text location data,I geocoded text strings at the county level using the GoogleMaps API v3, which uses gazetteer as well as map data tocode location names like “The Big Apple” as well as namesof cities, counties and states. Location data from user profileswere predominately user-generated text strings like “Austin,TX.” Some users reported their location whimsically, such as“between a rock and a hard place” or “Milky Way.” Somevalid locations were outside the United States and thus hadno county, and some valid locations were identifiably insidethe United States but had insufficient precision to determinea county. I coded these with a binary indicator for outside orinside Texas.

When user accounts had only one residential locationestimate based on GPS data or based on text string data, I

able 1sers and tweets by location

Users Tweets

otal 399,081 100.0% 1,656,252 100.0%

ocation estimates by sourceWith any location estimate 160,954 40.3% 768,140 46.4%

With only textlocation estimate

113,938 28.6% 587,232 35.5%

With only GPSlocation estimate

16,893 4.2% 44,798 2.7%

With text and GPS 30,123 7.5% 136,110 8.2%

504 A.J. Stevenson / Contraception 90 (2014) 502–507

assigned that location as their residence. When estimatesfrom both sources were available, I used the GPS estimate.The two estimates were identical in 79% of accounts withboth estimates.

I generate a map displaying the volume of tweets over theentire period from users estimated to live in each county inthe continental United States and a map of the same data forTexas alone. Maps were generated using SAS® 9.3.Focusing on Texas, I also generate a line chart of thenumber of active user accounts (one or more tweets) withestimated residence in Texas by day and active user accountswith estimated residence outside of Texas by day.

In order to estimate the total number of Texans whoparticipated in the conversation, I applied the proportion ofTexans among all user accountswith valid locations (includinglocations outside the United States) to all user accounts toindirectly estimate the total number of Texans who participat-ed. This calculation relies on the assumption that theproportion of user accounts with missing location data whoresided in Texas was the same as to the proportion of useraccounts with nonmissing location data who resided in Texas.

I estimated orientation toward abortion rights at the level ofthe tweet and the user account. I classified tweets with twobinary indicators: one indicating the presence of any hashtagsassociated with support for abortion rights and one indicatingthe presence of any hashtags associated with opposition toabortion rights. In order to evaluate this classification strategy,I coded three random samples of 100 tweets. One samplecontained tweets with supportive hashtags. The secondcontained tweets with opposing hashtags, and the thirdcontained tweets with exclusively neutral hashtags. I codedall tweets in each sample for support, opposition or neutralitywith respect to abortion rights. Among the tweets withhashtags supportive of abortion rights, 99 of 100 wereidentifiably supportive of rights, while 1 was uninterpretableas supportive or opposed. Similarly, among the sample of 100tweets with hashtags opposed to abortion rights, 100 werehand coded as opposed. In the sample of 100 tweets with onlyneutral hashtags, 87 were supportive of abortion rights, 7 wereneutral, 3 were opposed, and 3 were uninterpretable. Iclassified the user accounts with an indicator of the presenceof hashtags of each type in the full collection of their tweets inthe dataset. Some tweets included emoticons, which couldhave corrupted the text encoding of special characters andmayhave led to errors in classification.

All data from the Twitter API were collected using Pythonand the package Twython. The historical tweets may bepurchased without special approval, and the Twitter API ispublic. This study was determined to be exempt by theInstitutional ReviewBoard at TheUniversity of Texas atAustin.

location estimateMissing location estimate 238,127 59.7% 888,112 53.6%irectly estimated locationsIn Texas 46,614 29.0% 374,993 48.8%Outside Texas 114,340 71.0% 393,147 51.2%direct estimate:TOTAL in Texas

115,500 ~29% 808,500 ~48.8%

3. Results

Table 1 describes the numbers of user accounts and tweetsby estimated location and source of location estimate.

Hashtag-based selection identified 1,656,252 tweets from399,081 unique user accounts. Using residential locationestimates based on text and GPS together, county ofresidence was directly estimated for 160,954 user accounts,or 40.3% of all accounts. Locations were estimated usingGPS data for 11.7% of accounts and using geocoded textdata for 36.1% of accounts. For 7.5% of accounts, bothestimates were available. Of user accounts with directlyestimated residence, 29.0% were estimated to live in Texas.Multiplying this proportion by the total number of useraccounts, an indirect estimate of the total number of Texanswho participated is about 115,500 people. On average, eachTexan who participated tweeted more times than each non-Texan, and thus, Texans are estimated to have sent 808,500tweets, or 48.8% of the total tweets.

Fig. 1 displays the number of tweets from user accountsestimated to reside in each county in the continental UnitedStates for the 40.3% of user accounts had directly estimatedresidence. It illustrates the fact that Twitter users across theUnited States participated in the conversation. Concentra-tions of tweets in metropolitan areas throughout Texas, theWest coast, the Mid-Atlantic, the Midwest and the coastalNorth East are displayed.

Fig. 2 displays a map of Texas colored by the number oftweets sent from users estimated to live in each county. Itillustrates the fact that the majority of Texas countiescontained at least one participant in the discussion of the bill.Of Texas’ 254 counties, participants were directly estimatedto live in 192. All of Texas’ counties classified asMetropolitan Statistical Areas were estimated to have hadusers who tweeted.

Fig. 3 plots the number of tweets by day for Twitter useraccounts directly estimated to be Texans and Twitter useraccounts directly estimated to reside outside of Texas. Texansoutnumber non-Texans, except on the day of the filibuster andthe day after. Peaks in the plot correspondwith legislative events,hearings and votes. Texans’ tweets peaked on the day of thefilibuster, while tweets from non-Texans peaked the next day.

TU

T

L

D

In

<=5 6-10 11-100 100-200 200-500 500+

Fig. 1. Tweets by county of user's of residence, continental United States.

505A.J. Stevenson / Contraception 90 (2014) 502–507

Based on the high level of agreement between hashtag-basedclassification and hand coding, I accepted hashtag-basedclassification as a proxy for tweet and user orientation. Themajority of tweets and users appear to be supportive ofabortion rights. Specifically, only 2.7% of tweets contained ahashtag associated with opposition to abortion rights, while42.4% of tweets contained a hashtag supportive of abortionrights. Only 0.45% of tweets contained hashtags of bothtypes. Among user accounts, 58.4% tweeted at least oncewith a hashtag supportive of abortion rights, while only 3.6%ever tweeted with a hashtag in opposition to abortion rights.About half of those who ever tweeted with a hashtag inopposition also tweeted with a hashtag in support (1.56%).

<=5 6-10 11-100 100-200 200-500 500+

Fig. 2. Tweets by county of user's residence, Texas.

4. Discussion

The photographs and reports from the June 2013 protestsat the Texas Capitol document thousands of Texans’opposition to HB2, but little has been reported about thevolume and composition of the individuals contributing tothe digital outpouring that accompanied these protests. Thepresent analyses demonstrate that at least 1.66 million tweetswere sent using hashtags associated with the discussion andthat the tweets were largely in support of abortion rights.Further work will more precisely classify the tweets andusers by support for or opposition to abortion rights, but thefinding that only 2.7% of tweets contained hashtags opposedto abortion rights bolsters the interpretation that most of thediscussion surrounding the bill was supportive of abortionrights. The volume of tweets is large considering that thehashtag #YesAllWomen, used by people worldwide todiscuss misogyny, also yielded 1.6 million tweets in its peak4 days of use [8].

By estimating user residences, I find that a large fraction(48.8%) of the tweets were from Texans themselves.Furthermore, users who tweeted came from all over Texas.

0

20,000

40,000

60,000

80,000

100,000

120,000

140,000

160,000

6/19 6/22 6/25 6/28 7/1 7/4 7/7 7/10 7/13

TexansNon-Texans

Fig. 3. Tweets by residence and day.

506 A.J. Stevenson / Contraception 90 (2014) 502–507

Had the outrage on Twitter not included a geographicallydiverse contingent of Texans, the protests last summer couldhave been an artifact of Texas’Capitol being in one of its mostliberal cities. Instead, the findings here point to the existence ofa broad network of Texans who strongly oppose restrictions onabortion access. The presence of relatively ardent (at leastimpassioned enough to speak out on Twitter) rights supportersthroughout Texas rebuts the narrative of Texas’monolithicallyconservative culture. In that narrative, the true Texans aredeeply conservative. While Texas still generally votesRepublican in statewide elections, the present analysisdemonstrates that Texans throughout the state — includingin rural areas — oppose abortion restrictions.

The finding that such a small fraction of all tweetssupported abortion restrictions should not be interpreted asevidence of a high prevalence of abortion rights supporters inthe general population, but instead as evidence that socialmedia users who used hashtags to discuss HB2 and thefilibuster were supportive of abortion rights. This is consistentwith Twitter users being younger, better educated, and moreurban than the general population of Internet users [5]. Thedifferences between the Twitter users described here and thegeneral public probably arise jointly from the selection processthat leads people to become Twitter users and the fact thatindividuals who react strongly to restrictions on abortion rightsmay differ substantially from the general population. Whilethis analysis does not measure political opinion generally, thepower of social media to facilitate mass demonstrationtogether with the support for abortion rights among Twitterusers could be welcome news to advocates of abortion rights.Furthermore, the breadth of engagement in the conversationoutside of Texas indicates that social media users across theUnited States and beyond may have a role to play inconversations about local abortion restrictions, an increasinglycommon phenomenon in the states [9].

If abortion opinions are indeed polarizing over time, asmany researchers have argued [10–12], identifying andunderstanding the extremes of support for and opposition toabortion rights may assist social scientists in predicting thefuture of abortion politics and could guide advocates’ effortsto shape that future. The power of understanding theextremes is underscored by evidence that engagement insocial movements about abortion may diffuse extremeideology rather than result from it [13,14]. If this is thecase, at the individual level, even marginal engagement maypredict more fervent views in the future. Analogously, at theaggregate level, the most extreme views of the present maybecome more widespread as polarization increases. Thus, thepopulation of abortion rights supporters in Texas whoresponded to the conjuncture of HB2’s proposal and WendyDavis’ filibuster with public outrage may influence the futureof abortion politics in Texas as a “left flank” of strongsupporters of abortion rights. These individuals couldparticipate in the diffusion of support for abortion rights inTexas. Their energy and passion could be harnessed tofurther discussion, change opinions and mobilize voters.

This study is limited by the fact that it must estimatelocations and by factors associated with the data sources.Twitter users may tweet from nonhome counties (for example,from work), and they may misrepresent or not updateresidence in their profiles. However, if rural users travel tourban areas for work, such travel would bias away fromfinding users in rural areas, making my finding of geographicdiversity conservative. Moreover, this study’s method ofestimating locations is an improvement over relying solely onGPS data since social media users who enable and use GPScoding of their activities have been found to be unrepresen-tative of the total population of socialmedia users [6,15]. Thus,while the estimates here are not perfect, they may be lessbiased than previous work using solely GPS data. Limitationsalso arise from the data and methods. First, because Twitterusers are not representative of the general population, thisstudy solely seeks to describe the locations of Twitter userswho tweeted with the hashtags and cannot be used to describeabortion opinions in the general population. Additionally,selecting tweets on hashtags means that the many tweetswithout hashtags and user accounts that tweeted about HB2and the filibusterwithout ever using a hashtag are omitted fromthis analysis [16]. This bias leads to conservative estimates ofthe number of participants and tweets. It may also mean thatmy sample includes more experienced Twitter users, whichmay contribute to the observation that few tweets or useraccounts are opposed to abortion rights. Hashtag-basedclassification also does not clearly classify the very smallnumber of user accounts and tweets with hashtags of bothtypes. This analysis treats organizations and individuals in thesame way, and it may overestimate the number of individualsparticipating if individuals used more than one Twitteraccount. A final limitation is that error in classification oftweets by support for abortion rights may arise fromundetected irony, retweeting for purposes other than supportor the strategic use of hashtags to insert oneself into the otherside’s conversation. Hand codingwas used to estimate that thiserror leads to themisclassification of less than 5%of all tweets.Unfortunately, special symbols may have obscured a smallproportion of hashtags, which also may have contributed toerrors in classification.

This analysis only scratches the surface of what ispossible to learn from these data. Further work shouldinvestigate the correlates of persistent engagement inpolitical action among these users, the network diffusion ofengagement and spatial variation in the ratio of supporters toopponents of abortion rights. These efforts will requirecollaborations across reproductive justice, sociology, com-puter science and communications. The findings of suchwork could inform advocacy and help predict the futuredirection of reproductive health care access in the states.

Acknowledgments

The author is grateful to Jonathan Rosenberg for expertprogramming assistance and to Joseph Potter, Kristine

507A.J. Stevenson / Contraception 90 (2014) 502–507

Hopkins, Celia Hubert, David McClendon and threeanonymous reviewers for insights and feedback on previousversions. This work was supported in part by the EuniceKennedy Shriver National Institute of Child Health andHuman Development grant for Infrastructure for PopulationResearch at the University of Texas at Austin, Grant 5 R24HD042849, NICHD.

References

[1] Johnson-Hanks JA, Bachrach CA, Morgan SP, Kohler H-P. Under-standing family change and variation. Towards a theory of conjecturalaction. In: Stillwell J, editor. Understanding population trends andprocesses, vol. 5. Dordrecht, The Netherlands: Springer; 2011.

[2] Bumpass LL. The measurement of public opinion on abortion: theeffects of survey design. Fam Plan Perspect 1997;29(4):177–80.

[3] Tourangeau R, Rasinski KA. Cognitive processes underlying contexteffects in attitude measurement. Psychol Bull 1988;103(3):299–314.

[4] Zigerell LJ, Rice HM. Intense Ambivalence. In: McGraw KM, &Aldrich John H, editors. Improving public opinion surveys: interdis-ciplinary innovation and the American national election studies.Princeton University Press; 2011, pp. 303–22.

[5] Brenner J, Smith A. 72% of online adults are social networking siteusers. Washington, D.C: Pew Research Center; 2013.

[6] HechtB, StephensM.A tale of cities: urbanbiases in volunteeredgeographicinformation. ICWSM 2014. Menlo Park, CA: AAAI Press; 2014.

[7] Takhteyev Y, Gruzd A, Wellman B. Geography of Twitter networks.Soc Networks 2012;34(1):73–81.

[8] Hofherr J. #Yesallwomen takes Twitter by storm: a rundown of thenumbers. [cited 2014 June 18, 2014]; Available from , http://www.boston.com/news/nation/2014/05/27/yesallwomen-takes-twitter-storm-rundown-the-numbers/dtjOWwQvol6upAwOY2hAUK/story.html2014.

[9] Nash E, Gold RB, Rowan A, Rathburn G, VierboomY. Laws affectingreproductive health and rights: 2013 state policy review. NY:NY: TheGuttmacher Institute; 2013.

[10] Bolzendahl CI, Myers DJ. Feminist attitudes and support for genderequality: opinion change in women and men, 1974–1998. Soc Forces2004;83(2):759–89.

[11] Hout M. Abortion politics in the United States, 1972–1994: fromsingle issue to ideology. Gend Issues 1999;17(2):3–4.

[12] Mouw T, Sobel Michael E. Culture wars and opinion polarization: thecase of abortion. Am J Sociol 2001;106(4):913–43.

[13] Munson ZW. The making of pro-life activists: how social movementmobilization works. University of Chicago Press; 2009.

[14] Yardi S, Boyd D. Dynamic debates: an analysis of group polarizationover time on Twitter. Bull Sci Technol Soc 2010;30(5):316–27.

[15] Leetaru KH, Wang S, Cao G, Padmanabhan A, Shook E. Mapping theglobal Twitter heartbeat: the geographyofTwitter. FirstMonday2013;18(5).

[16] Liu Y, Kliman-Silver C, Mislove A. The tweets they are a-changin’:evolution of Twitter users and behavior. Proceedings of the EighthInternational AAAI Conference on Weblogs and Social Media. AnnArbor, MI: AAAI Press; 2014.