Study with the several resources on Docsity
Earn points by helping other students or get them with a premium plan
Prepare for your exams
Study with the several resources on Docsity
Earn points to download
Earn points by helping other students or get them with a premium plan
Community
Ask the community for help and clear up your study doubts
Discover the best universities in your country according to Docsity users
Free resources
Download our free guides on studying techniques, anxiety management strategies, and thesis advice from Docsity tutors
An analysis of the digital divide among Twitter users, comparing their demographics, Internet use patterns, and attitudes to those of non-Twitter users and other Internet users. The study reveals significant differences between Twitter users and other groups, raising questions about the representativeness of Twitter data for social research.
What you will learn
Typology: Lecture notes
1 / 35
This page cannot be seen from the preview
Don't miss anything!
Grant Blank University of Oxford
Please cite as: Blank, Grant. (2016). The digital divide among Twitter users and its implications for social research. Social Science Computer Review. DOI: 10.1177/0894439316671698.
Abstract
Hundreds of papers have been published using Twitter data but few previous papers report the digital divide among Twitter users. British Twitter users are younger, wealthier and better educated than other Internet users, who in turn are younger, wealthier and better educated than the off-line British population. American Twitter users are also younger and wealthier than the rest of the population but they are not better educated. Twitter users are disproportionately members of elites in both countries. Twitter users also differ from other groups in their online activities and their attitudes. These biases and differences have important implications for research based on Twitter data. The unrepresentative characteristics of Twitter users suggest that Twitter data are not suitable for research where representativeness is important such as forecasting elections or gaining insight into attitudes, sentiments or activities of large populations. In general, Twitter data seem to be more suitable for corporate use than for social science research.
Keywords : Social media; Twitter; representativeness; selection bias; elites; Oxford Internet Survey; OxIS; Pew Internet and American Life
States data to show where they are similar and where they differ on demographics, attitudes and Internet use. Finally, I discuss research attempts to use Twitter content to predict behavior. I explore the implications of these results for attempts to use Twitter content as a research tool.
Digital divides
The digital divide has been a major focus of online research; for representative examples see Bonfadelli (2002), van Dijk (2005) or van Deursen & van Dijk (2013). Digital inequality can take many forms that have been explored in the United Kingdom and the United States, as well as elsewhere. In Britain the Oxford Internet Survey (OxIS) has charted 10 years of trends in the online population (Dutton & Blank, 2011, 2013). These studies document that British Internet users have been younger, better educated and wealthier than the off-line population since the earliest wave in 2003. Some differences between the online and off-line populations have disappeared, such as the gender gap which was important in the early 2000s but disappeared by
population, and it is striking that few comparable studies exist of the characteristics of the population of Twitter users.
Prior attempts to estimate demographics
Previous studies of Twitter demographics have been attempts to estimate the demographic characteristics of Twitter users based on tweets and other public data like profiles completed by Twitter users (e.g. Mislove et al 2011; Pennacchiotti & Popescu, 2011; Rao, Yarowsky, Shreevats & Gupta, 2010; Sloan, Morgan, Housley, Burnap, & Williams, 2015; Sloan et al., 2013). These studies attempt to infer characteristics like gender, age, region, race, political orientation or other attributes using machine learning or other computational techniques. They compare their computational classifications to reference datasets that have been hand-coded. This tells them how well their algorithms match hand-coded results; they can obtain a number that measures this kind of accuracy, but it is a weak measure of accuracy. They have few variables on which to compare their results with actual, real-world demographic data on Twitter users, Twitter non-users, or non- users of the Internet. So they can’t tell us much about how Twitter users compare to anyone else. Mislov et al. (2011) and Sloan et al. (2013) make the most methodologically sophisticated attempts to find Twitter demographics.^1 Both are remarkable projects. Using first and last names,
(^1) Mislov et al. (2011) infer geographic location using the self-reported location field in Twitter
user profiles: 75.3% of publicly visible users enter something in that field and Mislov et al. can attach reported locations to latitude and longitude on a US map in 8.8% of the 75.3%. They compare this data to US Census 2000 data at the county level. They infer gender based on first names and race/ethnicity based on last names. They are able to match gender to 64.2% of users and race/ethnicity to 18.5% of users (p. 557). The relatively low percentages are not encouraging.
The Pew Internet and American Life project has been the major source of high quality information about social media users in the United States, including Twitter. Although Pew supplies descriptive tables, it does not draw out the implications of its data for the digital divide or for social science. The first demographic breakdowns of Twitter users were published by Duggan and Brenner (2013), and Duggan and Smith (2014) based on 2012 and 2013 Pew surveys of the United States. They report the gender, age, race, education, income and urban-rural status of American Twitter users. This is a considerable step forward, but these reports lack any comparative frame since they include only Twitter users. To understand Twitter users we need comparative data that shows us how Twitter users differ (or not) from other online and off-line populations. Only when we have comparative data can we understand the potential biases of conclusions based solely on Twitter users. As far as I have been able to determine, no one has published the actual comparative demographic breakdown of Twitter users. An exception to this statement is Hargittai and Litt (2012) who presenet comparative demographics for a sample of young Twitter users, age 18-24. Hargittai (2015) presents multivariate results, discussed below, but not comparative demographics. The data needed for comparison are available in the OxIS and the Pew Research Center’s Internet and American Life Project.
Methods
OxIS collects data on British Internet users and non-users. Conducted biennially since 2003, the surveys are random samples of more than 2,000 individuals aged 14 and older in England, Scotland, and Wales. Interviews are conducted face-to-face by an independent survey research company. See Dutton and Blank (2013) for details of the data collection and sample. The analyses below are based on the 480 Twitter users, 1,270 social network site (SNS) users or the 1,610 Internet users out of the full 2013 sample of 2,053 respondents.
The Pew Research Center’s Internet and American Life project conducts regular telephone surveys examining the impact of the Internet. In May 2013 Pew (2013) asked a random sample of the American population age 18 or older about SNS use including Twitter. Analyses of the Pew dataset are based on 341 Twitter users, 1,377 SNS users or 1,912 Internet users (using Pew’s definition of Internet users) out of the full sample of 2,252 respondents (see Table 1). Table 1. British Twitter users compared to other groups Twitter users Non-Twitter SNS users Non-SNS Internet users Off-line population Full sample Number 480 789 341 442 2, Percentage of all Respondents
Percentage of all Internet users
Note : Total SNS users (including Twitter users): 1,270, total Internet users (including Twitter users): 1,610, and the total adds to 2,052 not 2,053 because two SNS users who Don’t Know if they use Twitter are omitted, but rounding the weighted frequencies adds one case. SNS = social network site. We also use nine standard, self-reported demographic variables. We use four education categories: no degree, secondary education degree, further education, and university undergraduate or post-graduate degree. Race is coded as white versus non-white. Place is coded as urban versus rural. Lifestage is a four-category variable: students, employed, unemployed and retired. Marital status has five categories: single, married, living with partner, divorced, widowed. We also include age and binary measures of disability and gender. All of these variables are available in the OxIS dataset; the Pew dataset contains age, education, income, race, lifestage, marital status and gender. To enhance comparability I have attempted to make the demographic categories as similar as possible in the two datasets; I note any differences below. OxIS contains measures of participation in 43 activities that people do on the Internet (The Pew data set does not contain activity measures).The activities cover an extremely wide range, from buying online, to blogging, to making travel plans, to listening to music, to finding out health information, to reading celebrity news or gossip. Each activity was measured using an identical 6-category Likert scale ranging from 0 = never participate to 5 = do several times per
Table 2 contains the same data from the Pew dataset. Notice that American Twitter use is about 8 percentage points less than British Twitter use (15.2% of all respondents compared to 23.4%), although the proportion of Americans who use the Internet is higher. Table 2. American Twitter users compared to other groups Characteristics Twitter users Non-Twitter SNS users Non-SNS Internet users Off-line population Full sample Number 341 1,054 512 340 2, Percentage of allRespondents 15.2 46.9 22.8 15.1 100. Percentage of all Internet users
Note : Total SNS users (including Twitter users) = 1,377, total Internet users (including Twitter users) = 1,912, and four respondents who refused to answer the SNS question have been omitted. SNS = social network site. Demographic comparisons of Twitter users to other groups are in Table 3 for both Great Britain and the United States. I first talk about the British data, followed by comparisons to the American data. The columns in both of these tables correspond to the columns in Tables 1 and 2. The nine demographic variables are ordered approximately from the strongest effects to the weakest. I discuss each demographic variable in the British part of Table 3 in succession, beginning with age. Age differences are large. Twitter users are more likely to be young by about 26 percentage points: 30% of Twitter users are between age 18-24 compared to 4.3% of the off-line population. Comparatively few Twitter users are over age 55, whereas over 70% of the off-line population is over age 55. Looking at education, Twitter users are less likely to have no educational qualifications by 60 percentage points compared to off-line respondents. Twitter users are more likely to have graduated from college. Thirty-five percent of Twitter users have at least one higher education degree, about seven percentage points more than non-Twitter users and 29 percentage points more than people who are off-line. For marital status, Twitter users are about 17 percentage points more likely to be single compared to SNS users who don’t use Twitter and about 15 percentage points less likely to be married. Among the other marital status categories,
Twitter users are more likely to be living with a partner but less likely to be divorced, separated or widowed than the other categories. In terms of lifestage, Twitter users are eight percentage points more likely to be students than non-Twitter SNS users and 17 percentage points more likely than the off-line population. Twitter users are also more likely to be employed, and they tend not to be unemployed or retired. Income also shows large differences. Twitter users are about five percentage points more likely to have incomes of £50,000/year or more compared to non-Twitter SNS users and about seven percentage points more likely to have incomes of between £40,000/year and £50,000/year. Compared to the off-line population they are over 50 percentage points less likely to earn less than £12,500 per year. Twitter users are about 10 percentage points more likely to be white than those who are not online. They are slightly more likely to be male than off-line people. Twitter users are somewhat less likely to be disabled than non-Twitter SNS users by about six percentage points. They are about 11 percentage points less likely to be disabled than off-line respondents.
Race Gender Disability Urban Note : Disability and Urban-Rural are not available in the Pew dataset.
The American data in Table 3 show patterns that are sometimes similar and sometimes different than the British. The age portion of the American table looks very similar to the British table. The modal users are the youngest respondents and Twitter use declines monotonically with age. In the education categories, in both countries highly educated people are more likely to be Twitter users. The modal Twitter user has some college in the US but is a secondary school graduate in Britain. The best explanation is that this reflects differences in the meaning of “further education” and “some college” in the two countries. In other words, it is the difference in the definition of the category. For marital status, the modal category in both countries is single, followed by married. The other differences reflect differences in the American and British populations more than differences in the way people of different marital status’s use Twitter. In lifestage we see a large difference: Only about 3% of American students use Twitter compared to about 18% of British students. The lack of American student use is striking. Almost three-quarters of American employed respondents use Twitter compared to less than two-thirds of British employed persons. Income is another area where the two countries differ. Among the lowest income category, Americans show 30 percentage points difference between Twitter users and Internet non-users; low income British show over 50 percentage points difference. Among high income Americans, 30 percentage points separate Twitter users from Internet non-users, compared to about 10 percentage points in Britain. For race, nonwhites are 13 percentage points more likely to use Twitter in the US compared to nine percentage points in Britain. Finally, gender doesn’t seem to matter in the U.S. data. In summary, the big demographic differences between Twitter users and other groups are that, in both countries, Twitter users are more likely to be younger, better educated, students or employed, single, and wealthier. Specifically they are younger than other SNS users, who are in turn younger than other Internet users, who are younger than non-users. This monotonic relationship shows up in most of the strong relationships in the table. It shows up in all education
than whites. Hargittai’s (2015, Table 1) multivariate analysis of Twitter using the same Pew dataset produces similar results, although difference in model specification make detailed comparisons difficult. Table 4: Logistic Regressions Predicting Twitter Use British and American Data British Odds Ratios American Odds Ratios Age 0.94*** Age 0.97*** Household income £12.5-£20,000 0.83 $20-$30,000 0. £20-£30,000£30-£40,000 0.801.47 $30-$40,000$40-$50,000 1.151. £40-£50,000 4.12*** $50-$75,000 1. £50,000 or more 2.81** $75,000 or more 2.11** Education Secondary graduateFurther education 2.331.80 High school graduateSome college 0.880. Higher ed. degree 2.27 College degree 1. Lifestage Employed 1.87 Employed 0. RetiredUnemployed 3.221.29 RetiredUnemployed 0.290.28 Marital status Married 0.83 Married 0.57* Living with partner 0.91 Living with partner 0. Divorced/SeparatedWidowed 1.140.61 Divorced/SeparatedWidow 0.890. Ethnicity 1.27 Race 1.56* Gender 0.88 Gender 1. Disability 1.07 (not available) Urban-ruralConstant 1.241.26 (not available)Constant 2. N 1,298 1, McFadden's R² 0.16 0. Note white, men, not disabled, rural. American: < $20,000, less than high school,: Omitted categories. British: < £12,500, no qualifications, student, single, student, single, white, men. Two variables, disability and urban-rural, are not available in the Pew dataset.
race. The differences in education, lifestage and race seem important and worth thinking about. In terms of who uses Twitter, these countries have different patterns of use. The preceding pages describe differences between the Twitter users and others. What is the impact of these differences? The next two sections explore this question, comparing activities and attitudes of Twitter users to others.
Activities of Twitter users
Once people are online, they act in a variety of ways. Do Twitter users act differently from other users? Figure 1 compares Twitter users to non-users on 10 types of activities. From this point onward, we can no longer compare British and American Twitter use because this Pew dataset doesn’t have any of the activity variables; these tables use only OxIS data. The striking result is how much more Twitter users do. In every activity, Twitter users are 12-38 percentage points more likely to participate than non-Twitter users. This gap extends across the entire range of the 10 activities, from the most frequent activity—socializing—to the least frequent—vice. The smallest differences are information seeking (12 percentage points) and vice (16 percentage points). The largest differences are in entertainment use of the Internet (38 percentage points), blogging (34 percentage points) and creative production (30 percentage points). It is notable that the biggest differences tend to be in the most challenging online activities; the blogging category includes not only blogging but maintaining a personal website, and creative production includes posting videos and posting anything the respondent considers ‘creative’. The entertainment category is an exception. It is important to reiterate that Twitter use is not included in any of the 10 activity scales. Considering all the errors inherent in self-reported survey variables, this figure shows surprisingly large differences between Twitter users and non-users.
Attitudes of Twitter users
Finally we can look at attitudes of Twitter users. There are some 48 attitude variables in the 2013 wave of OxIS, most of which show statistically significant differences between Twitter users and non-users. Figure 4 contains four variables selected to illustrate the range of variables on which there are differences. Twitter users are between 13 and 26 percentage points more
likely to agree that the Internet keeps them in touch with people, saves time and helps them escape. They are seven percentage points less likely to agree that the Internet is frustrating to work with. The overall pattern is that Twitter users have a much more positive attitude toward all aspects of the Internet than non-users.
The digital divide is strong among Twitter users: Twitter users are unlike other groups in many ways. Furthermore, Twitter users in the UK and the US have different characteristics. What does this mean? We explore the implications of these differences for research using Twitter.
Discussion
The implications of the digital divide in Twitter use are different than for most digital divide research. Digital divide research has usually been concerned that large elements of the population are missing the benefits of being online. Since off-line people are more likely to be poor, uneducated, or elderly, the lack of access to Internet benefits reinforces their marginal
