Review of Twitter – A Digital Socioscope
by Yelena Mejova, Ingmar Weber and Michael W. Macy Cambridge University Press © 2015
I recently read an interesting book about Twitter, examining its use as a “digital socioscope”. By this, the writers compare social media to a telescope, microscope, MRI, or any similar technological breakthrough device that reveals the previously unseen, or at least seen through a glass darkly. Social media is a “socioscope”, a device to see the social world, relatively unfiltered. Twitter is an ideal type case, because of its open nature.
The social world is hard to observe and difficult to measure, especially at scale. Interaction networks have mostly been studied at small scales in the past, at the club or village level. Survey data has to focus on (usually) randomly selected individuals, and therefore don't capture the dynamics of social life, in terms of how interactions with others help shape beliefs, opinions, and actions. Everyone is an island.
Twitter produces data that is both enormously macroscopic (millions of tweets daily) and microscopic (each tweet individualized and time-stamped), with interaction data such shares, follows, and the like recorded. Issues such as recall bias or social desirability bias are also reduced or eliminated by the nature of the instrument – i.e. Twitters technical architecture, business practices and user generated norms (e.g. hash-tagging). With smart-phones and tablets, the tweets can even be located in physical space and geographical correlations can be observed (e.g. group events such as demonstrations, organized via Twitter).
The first big application of Twitter to social science was the analyses of networks. Twitter is an excellent laboratory for studying social contagion (tweets going viral), network structure, information diffusion and levels of influence. Some findings have been unexpected; for example the power of “influencers” has probably been overrated. Twitter is also useful for measuring social status via “who gets attention” and how that corresponds to real world outcomes, such as financial power.
Twitter has also proved useful in the study of disease spread – people tweet about their physical and emotional health, as revealed by both content and sentiment analysis.
Twitter's role in such social movements as the Arab Spring and Occupy Wall Street has been much noted, and much studied. Such phenomena are complex and multi-faceted, and thus it is difficult to isolate the effects of social media such as Twitter, but it seems clear that those effects are very significant.
Attempts have been made to not only understand social and political events via Twitter feed analysis, but to use that feed to predict events. The success of these efforts have been mixed, as Twitter users are not randomly drawn samples of the underlying population, and there is a tendency for echo chambers to form. Nonetheless, it is an active area of research, for obvious reasons.
There are challenges with using social media data, and with Twitter in particular. One challenge is that people often don't use their real identities and profile information can be hidden, so it is difficult to get good demographics. Another challenge is teasing out homophily (the tendency for people to share interests) and influence. For example, if someone joins Twitter a few days after his or her friend did, was that due to the friend's influence, or just due to the fact that friends tend to be drawn to similar activities, independently of each other. As with all social research, causality can be tricky. Carefully designed experimental protocols can help, as can multivariate analyses. Simply deciding whether a social tie actually exists between two nodes (people) of the network can be difficult, as can understanding how these affiliations compare to real world relationships. Lastly, it can be a challenge to know how well the Twitter world approximates the off-line world, and therefore how valid generalizations from the former to the latter are.
There are also technical challenges. The ability to tap this data and analyze it often exceeds the skill-sets of most social scientists, being more in the domain of programmers, statisticians and data scientists. So, collaboration is needed among social science experts and information experts to ensure that the right questions are asked of the data, and the right conclusions are drawn from analysis.
In Chapter 1, the book has a rather extensive section on the tools available to access Twitter data via its API. It investigates the pros and cons of some of these. Issues included are:
- The difficulty of pinning down accurate demographic data.
- How to estimate the bias of the Twitter population relative to the overall population.
- Whether the sample data that Twitter normally provides to researchers is biased compared to the entire Twitter data stream.
- The difficult problem of content analysis and the tools that have been developed to aid in that.
- The similarly thorny problem of sentiment analysis.
There is also an extensive list of sources for those who want to use the Twitter API for analysis.
Chapter 2 examines the use of Twitter data to explore public opinion, especially political opinion and election forecasting. Twitter is viewed through the lens of different definitions of “public opinion”, with the main distinction being whether it is more communitarian (shaping debate) or more individualistic (a sounding board). The chapter focuses more on the latter approach, which places Twitter more or less in the tradition of opinion polling.
Many studies have focused on how well sentiment analysis on Twitter correlates with wider public opinion, such as job approval for the U.S. president. There does appear to be some correlation. Using Twitter for election forecasting purposes has had mixed results, to say the least. There are many free parameters (e.g. timing window, choice of parties to follow, choice of key words, operationalization of sentiment analysis), so clearcut conclusions are difficult to draw. Some studies rate Twitter based election forecasting as not too different from polling based forecasting, though small parties' results can be overly optimistic from Twitter data.
However, Twitter's adversarial nature and non-representativeness, compared to the overall population, raise doubts about whether it can reliably be used to determine the public mood. Machine generated tweets may also be gaming the system. Sentiment analysis is particularly tricky in political tweets, as irony and sarcasm abound in this area, so machine algorithms are often widely off the mark. Political tweets are also more prone to false, misleading, or manipulative content, another area which is problematical for sentiment algorithms.
Chapter 3 looks at the use of Twitter to determine various socioeconomic phenomena, such as unemployment rates, consumer confidence, social moods, investor sentiments, and activity in financial markets. For obvious reasons, much research has been directed at using Twitter (and other social media, especially Google search metrics) in correlational and predictive studies of the stock market. A number of studies have found correlations between sentiment on Twitter and the activity o the stock market at various levels; calmer sentiment on Twitter seems to correlate with an up-trending market. Of particular interest to readers of this blog, (Gruhl et al. (2005)) show that on-line chatter volume can predict book sales. Similar claims were made for movie box-office receipts.
Analysis of the level of tweets that mention subjects such as “job loss” have been found to correlate well with official unemployment measures, such as unemployment insurance claims, in the U.S.. Similar findings have been reported for consumer confidence measures; both social media sentiment and official consumer confidence measures seem to be measuring the same underlying economic factors, though there is mixed evidence whether or not one is a more time-sensitive measure than the other. As noted above, much detailed work has been done on how measures of “social mood” via Twitter correlate with the Dow Jones Industrial average – calmness is a good sign, fear is a bad sign. Indicators of the ratio of “bullish” to “bearish” tweets also appear to be predictive of the stock market, with high “bearish” ratios especially predictive of sell-offs. Network analysis (e.g. the number of edges that a particular node has, which corresponds to followers on Twitter) also seems to have some complex predictive power for the stock market.
It is noted that the use of Twitter and other social media in these areas should complement, rather than replace traditional measures. Many of these are exploratory, atheoretical and post-hoc studies; these are fine for efforts to “beat the market”, but social science should have a deeper theoretical grounding.
Chapter 4 continues somewhat in this vein, though focusing on whether the general sentiment of tweets from a given area are a reflection of that area's overall sense of social well-being or “hyper-local happiness”. The intent is to complement pure economic measures such as GDP to gain a fuller picture of how communities and societies are doing.
Happiness, well-being and satisfaction with life has been tracked by survey data. However, those mechanisms are expensive, among other problems. Social media, such as Twitter, present the opportunity to capture sentiment and content of “well-being” at scale and inexpensively. Text analysis (content and sentiment), at the census tract level, has been performed on tweets in both the U.S. and U.K.. Words related to pro-social activities tend to relate to positive community well-being, while areas with higher levels of disengagement words related to more negative community well-being. This data has been checked for time trends – not surprisingly, happiness peaks on weekends and bottoms out on Monday and Tuesday. Traumatic events, whether political, economic or climatic tended to be negative days for Twitter content and sentiment.
A study in London, in the U.K., showed that “deprived” areas tweets tended to have lower normalized sentiment scores (less happy, if you like). Conversely, advantaged areas had higher normalized sentiment scores. The correlation coefficient was 0.35, quite high for social science work of this type. Content differed as well. For example, users in more deprived areas tweeted more about celebrity gossip, while users in advantaged areas tweeted about vacations. Deprived areas used more Blackberries, advantaged areas used more iPhones. Regression analysis showed that the general content of tweets was quite predictive of the area's socioeconomic status.
Chapter 5 shows how Twitter data can be used for public health purposes. Firstly, social media can be a source of epidemic intelligence, helping to track the spread of a disease in space and time. Secondly, it can be a source of health information, though its uncontrolled nature raises problems about the reliability of that information. Thirdly, it can track sentiment about public health concerns, such as vaccination campaigns.
In the case of the 2009 Swine Flu, it was found that tweets reporting flu correlated highly with actual flu cases presenting in the health system in the U.S. and U.K.. In fact “Twitter outbreaks” also correlated highly with actual cases in the following week, indicating that Twitter was useful for predicting spikes in flu. During this outbreak, the use of Twitter as a public health dissemination tool was also studied. Twitter was mostly used to link to official sources or media sources, for which it was an effective tool. In terms of sentiment analysis, it was found that in the early stages of the Swine Flu vaccination program, tweets tended to be negative. The sentiment shifted to positive as the vaccines became available and as vaccination coverage increased.
Finally, Chapter 6 demonstrates how Twitter has been used during natural disasters. One important way that Twitter has been found to be useful, is in actually identifying developing situations and providing helpful information to responders. This can be either through human recognition of relevant tweets, or sophisticated machine learning algorithms tailored to the task. There are challenges: among these are the sheer volume of tweets, problems with accurately determining the locations of tweets, and issues regarding judging the reliability and honesty of tweets.
Twitter has also been used by officials to disseminate information during emergencies, and by citizens to request assistance.
And here is a very appropriate xkcd comic about twitter and earthquakes: