Understanding the Political Representativeness of Twitter Users

Introduction: Twitter’s Potential as a Source of Information About Public Opinion and Electoral Preferences

The microblogging service Twitter has become one of the most important social networking sites, with 200 million monthly active users worldwide, ¹ including 18% of all online Americans ² and 15% of online Spaniards. While most of the messages that are sent through this platform are personal updates, its use for political purposes has increased in the past few years. Virtually, all candidates and elected officials have a presence on Twitter; and many users rely on Twitter to stay informed about political events. The content and structure of the political discussion that takes place on this platform, easily accessible through their application programming interface (API), represents a unique opportunity for researchers interested in the study of elections and public opinion.

The increase in the use of social media has led many social scientists to examine whether specific patterns in the stream of tweets might be able to predict real-world outcomes. Asur and Huberman (2010), for example, show how a simple model measuring chatter from Twitter about movies predicts box-office revenues, outperforming market-based predictors. ³ Applying a similar method, Lampos, De Bie, and Cristianini (2010) are able to accurately track the prevalence of influenza-like illnesses in several regions of the United Kingdom. Paul and Dredze (2011) extend this analysis into a broader range of illnesses in the United States, opening a whole new agenda in the field of public health research. In two highly publicized articles, Golder and Macy (2011) and Dodds, Harris, Kloumann, Bliss, and Danforth (2011) study the temporal patterns of happiness of millions of people in real time based on their tweets. An innovative study conducted by Hannak et al. (2012) builds up on this research to show how weather affects aggregated sentiment. Measurements of collective mood states derived from Twitter feeds had already been found to be correlated with stock market indexes by Bollen, Mao, and Zeng (2011). Finally, exploiting the geographic information that Twitter users provide has allowed researchers to estimate the epicenter of earthquakes in Japan (Sakaki, Okazaki, & Matsuo, 2010).

Given the accuracy of these predictions, and the consolidation of Twitter as a source of political information, a battlefield for campaigning, and a public forum of political expression, some researchers have wondered whether “tweets” validly mirror off-line public opinion. “Can we analyze publicly available data to infer population attitudes in the same manner that public opinion pollsters query a population?” (O’Connor, Balasubramanyan, Routledge, & Smith, 2010, p. 122). Were this approach to be successful, its advantages would be obvious: Twitter provides (relatively) easy and free access to millions of public messages in real time and from most countries around the world. Inferring public opinion from social media messages is challenging, but also potentially very rewarding, given the wealth of this information.

The first studies of this kind, in the context of the German legislative elections of 2009 (Tumasjan, Sprenger, Sandner, & Welpe, 2010) and the first 2 years of the Obama presidency in the United States (Cummings, Oh, & Wang, 2010; O’Connor et al., 2010), gave reasons to be optimistic. Tumasjan, Sprenger, Sandner, and Welpe (2010) found that “the mere number of messages [mentioning each German political party] reflect[ed] the election result and even [came] close to traditional electoral polls.” O’Connor, Balasubramanyan, Routledge, and Smith (2010), on the other hand, showed that “a relatively simple sentiment detector based on Twitter data replicate[d] presidential job approval polls. The results highlight the potential of text streams as a substitute for traditional polling.” Similar studies conducted in Singapore (Choy, Cheong, Laik, & Shung, 2011; Skoric, Poor, Achananuparp, Lim, & Jiang, 2012), United Kingdom (Lampos, 2012), Denmark (Sang & Bos, 2012), Spain (Congosto, Fernández, & Moro Egido, 2011), and recent elections in the United States (DiGrazia, McKelvey, Bollen, & Rojas, 2013) have also found that, during the elections, “the Twittersphere represents a rich source of data for gauging public opinion and that the frequency of tweets mentioning names of political parties, political candidates and contested constituencies could be used to make predictions about the share of votes at the national level” (Skoric et al., 2012, p. 2583). With these results, Cummings, Oh, and Wang (2010) even wondered “who needs polls?”

The response to this set of articles arrived in two recent research articles by Metaxas, Mustafaraj, and Gayo-Avello (2011) and Gayo-Avello (2012). These authors warn against “turning social media into another ‘Literary Digest’ poll” and claim that the “predictive power of Twitter regarding elections has been greatly exaggerated.” They illustrate their concerns with analyses of several Senate races in the 2008 and 2010 U.S. Congressional elections and find that electoral predictions applying similar methods as those used by the previous authors do not perform better than chance. These authors also criticize the previous results in this literature. For example, they point out that if Tumasjan et al. (2010) had not restricted their analysis to parties with parliamentary representation, the Pirate Party would have won the 2009 German elections—it was the party with the highest number of mentions in Twitter (Jungherr, Jürgens, & Schoen, 2011). In their view, an accurate prediction can only come through “correctly identifying likely voters and getting an un-biased representative sample of them.” Self-selection biases, overrepresentation of younger, more educated citizens on Twitter, and the simplistic assumptions of the existing sentiment analysis techniques are the three most important methodological challenges to overcome.

On a similar note, it is important to distinguish prediction from causality: The debate about whether Twitter metrics are or not correlated with public opinion metrics is orthogonal to the discussion about the effectiveness of online campaign strategies in vote choice. In fact, it is usually the case that the candidate who is mentioned the most on Twitter is the one who receives more criticism. For this reason, any method developed to measure sentiment should be capable of distinguishing praise from criticism, a positive message from a sarcastic message, and a relevant tweet from just “spam.”

To sum up, this set of research articles evidences some of the challenges that the use of Twitter data presents. The average Internet user is younger, more interested in politics, and comes from a higher socioeconomic background than the average citizen, which raises concerns about external validity (Gong, 2011; Mislove, Lehmann, Ahn, Onnela, & Rosenquist, 2011). Furthermore, the voice of political minorities tend to be underrepresented in the public debate on Twitter, and differences in party strategies regarding their presence in social media can also bias any measure of public opinion that relies on the number and content of tweets. It is therefore necessary to obtain more background information about each individual user, so that it is possible to stratify them and weight public opinion estimates. And even if most studies implicitly assume that Twitter is not a representative sample of the entire population, none of them examines to what extent participants on political discussions on Twitter are self-selected.

That is precisely the purpose of our study. Focusing on the 2011 Spanish legislative election and then on the 2012 U.S. presidential election, we characterize the content and structure of the political conversations about each election that took place on Twitter. These two elections took place in countries with diverse sociopolitical characteristics, and with different adoption rates of Twitter at the time of our analysis, which allows us to examine to what extent the patterns that emerge in our analysis could be generalized to other countries and elections.

We use a unique data set that includes all tweets sent in the 70 days prior to the election with references to the two major candidates that were running, which represents a total of 3 million tweets in Spain, and 62 million tweets in the United States. To analyze inequality in political discussions, we focus on a random sample of 12,000 and 50,000 users, respectively, that participated actively in the political discussion about each election, for which we estimated sociodemographic and political variables of interest.

Our analysis yields four main conclusions. First, we find that political debate on Twitter in both countries has a male name: A majority of users who tweet about politics are men. Second, the geographic distribution of Twitter users approximates that of population, with a slight bias toward urban areas. Third, we find that political discussion on Twitter is highly polarized: Users with clear ideological leaning are much more active and generate a majority of the content. Finally, we find significant differences in the patterns of discussion over our period of analysis: In both countries, campaign events increase or diminish inequality in participation on Twitter. These results have important implications for the use of Twitter data as a source of information about public opinion.

Data and Method. Twitter in the 2012 Legislative Election in Spain and the 2013 U.S. Presidential Election

Data Set

To analyze the structure of content of political discussion on Twitter about our two study cases, we captured all tweets mentioning the two most important candidates in each election, Rajoy and Rubalcaba, and Romney and Obama, over the 70 days prior to the date of the election (November 20, 2011, and November 6, 2012). ⁴ During this period, we systematically captured this set of tweets, at regular intervals. ⁵ In particular, in the first case, we included all tweets that mentioned at least one of the following four key words: “rubalcaba,” “rajoy,” “conrubalcaba,” and “marianorajoy.” ⁶ This represents a total of 2.8 million tweets sent by 375,000 unique users. For the U.S. presidential election, we included all tweets that mentioned at least one of the following four key words: “Obama,” “Romney,” “Barack Obama,” “Mittromney.” In this case, our data set contains a total of 62 million tweets and 8.8 million unique users.

As it has been shown before, production of tweets by users is profoundly unequal: Not all participants in the global discussion contribute to the same extent, and the differences are similar across our two case studies. While 47% of Spanish users and 48% of U.S. users published only one tweet mentioning one of the candidates over our 70 days of study, others—94 in Spain and 4,873 in the United States—tweeted more than 1,000 times.

Similarly, while the 1% most active users published 34.8% of all tweets in Spain (and 44.6% of all tweets in the United States), and the top 10% generated 70.9% of all tweets (70.6% in the United States), the remaining 90% sent only around 29% of all tweets. This distribution corresponds to a Gini coefficient of 0.764 in Spain and 0.755 in the United States, which denote a high concentration of the publication of tweets in a very limited number of users.

Given the magnitude of our data set, and with the purpose of providing a closer look at political discussion about these two elections, most of our analysis in this article focuses on a random sample of 12,000 active users ⁷ in Spain and 50,000 active users in the United States. At the end of our period of analysis, we captured the list of friends and followers of each of these users, as well as their description and location, as provided in their personal profile on Twitter. As we show subsequently, this information will allows us to characterize their activity on this online networking site. Figure 1 shows how, even within these two groups of active users, there is a high degree of inequality in participation: For example, we see that the bottom 50% of least active users sent only around 10–15% of the total of tweets.

OPEN IN VIEWER

Figure 1.

Lorenz curve. Inequality in the number of tweets.

Independent Variables

The indicators we presented in the previous section vary at the individual and aggregate levels as a function of a long series of factors, many of them difficult to measure empirically. In our analysis, this variation becomes particularly relevant when it is systematically associated with other variables that can affect citizens’ electoral behavior. When that is indeed the case, all electoral predictions that do not control in some way (either by weighting or calibrating based on the characteristics of the sample of individuals included in the analysis) will be systematically biased. In this article, we will focus on four of such variables: gender, geographic location, ideology, and the moment of the electoral campaign.

Gender

We have inferred the gender of each user in our random sample of 62,000 individuals based on the name they report in the “full name” field using a Naïve Bayes classifier (Bird et al., 2009). Our training data set was a list of names labeled by gender in each country. In Spain, our source is the National Institute of Statistics; in the United States, we rely on a large data set of anonymized names available in the Random Names R library (Betebenner, 2012). In both cases, the accuracy of our classifier is higher than 70%. ⁹

This variable is relevant because gender has a significant effect on political behavior, even after controlling for other confounding factors (see, e.g., Chaney, Alvarez, & Nagler, 1998; Kellstedt et al., 2010; Martínez & Calvo, 2010). Similarly, in the specific context of Twitter, different comparative studies have found that women are underrepresented (Bode, Hanna, Sayre, Yang, & Shah, 2011; Mislove et al., 2011; Parmelee & Richard, 2012). As we will show with greater detail in the following section, this appears to be the case also in Spain and the United States: We estimate that only between 35% and 40% of users who tweet about politics are women.

Geographic location

In our analysis, we also examine the geographic location of Twitter users. Several studies suggest that Twitter users are located predominantly in urban areas (see, e.g., Mislove et al., 2011). However, there is some controversy about this finding, since several authors (Hale, Gaffney, & Graham, 2013; Hecht, Hong, Suh, & Chi, 2011) argue that most Twitter users cannot be located accurately.

Our results also point in this direction. Using the Yahoo Maps API, we tried to locate each user in our random sample of 62,000 individuals using the “location” field in Twitter users’ profiles. In Spain, we found that 39% of them do not provide enough information to determine the province in which they live. This proportion increases to 47% if we try to place users in cities. In the United States, we find similar results: It is possible to identify the state from which users are tweeting in 84% of the cases and the county for 67% of them.

With this caveat, our analysis of the geographic distribution by province or state of the 7,307 Spanish Twitter users and 42,060 U.S. Twitter users that we were able to geolocate shows that this sample is not systematically biased with respect to the geographic distribution of the whole population. ¹⁰ As we show in Figures 2 and 3, the proportions of Twitter users and citizens living in each province/state are very similar (Pearson’s correlation coefficient between the two variables is .866 with N = 52 in Spain, and 0.959 with N = 50 in the United States), with the single exception of Madrid, with 30% of Twitter users but only 13.7% of the total population.

OPEN IN VIEWER

Figure 2.

Geographic distribution of Twitter users in Spain and the United States.

OPEN IN VIEWER

Figure 3.

Volume of tweets in our data set, by day, over our period of analysis.

Ideology/Party identification ¹¹

The political science literature that studies electoral behavior has shown the importance of ideology and party identification in anchoring the vote in general and in the two cases of study in particular (Campbell, Converse, Miller, & Stokes, 1960; Torcal & Medina, 2002). An analysis of the representativeness of Twitter users thus requires an understanding of the extent to which voters that identify with different parties or ideological orientations are present on Twitter, and whether they participate at similar levels.

To examine these issues, we implement a Bayesian item-response model that allows us to infer the orientation and intensity of ideological identification of each user as a function of the political orientation that is predominant in their network of “friends” (those that they decided to follow on Twitter). Hence, in the model, we assume that ideology is a latent variable that can be inferred from the distribution of political actors that each user is following, and it is based on the assumption that users prefer to follow politicians whose ideological viewpoint is similar to theirs. One of the main advantages of the statistical model producing the results is that it naturally incorporates the possibility that some accounts are less discriminative than others for learning the ideology of a given user by reason of the high profile or visibility of the politicians behind them. A full description of the method can be found in Barberá (2014).

Timing of the electoral campaign

Finally, we also analyze how different metrics evolve throughout the electoral campaign. Timing is important for two reasons. First, it has been shown that voters with low interests in politics decide whom to vote for during the electoral campaign. In Spain, for example, the postelection survey conducted by the Center of Sociological Research in 2011 showed that 28.2% of voters made that decision in the last 2 weeks before the election. Second, in both the elections, there is an important asymmetry in the online strategies of the two main parties. While Rubalcaba opened his Twitter account in August 2011 and sent personal message since then, Rajoy didn’t start until the end of September of that year. As a result, the Partido Popular (the main right-wing party) concentrated their online activity during the electoral campaign, while the Partido Socialista Obrero Español (the main left-wing party) had an active presence on Twitter since earlier in the year. On a similar fashion, Obama had more than 15 million followers on Twitter since the beginning of 2012, while Romney did not reach one million followers until a few weeks before the election.

As we show in Figure 3, the volume of tweets included in our data set largely varies over the electoral campaign in both countries. Furthermore, we find that the intensity of the debate increases during specific political events, such as the candidate debates, the dissolution of the terrorist group Euskadi Ta Askatasuna, or the attack on the American embassy in Libya.

Results

As indicated previously, one of the most remarkable results in our analysis relates to the deep inequality in the proportion of men and women that are participating in the political discussion on Twitter. As shown in Table 1, a large majority (60–65%) of the users who were active during our period of study are men and we can clearly reject the hypothesis of an equal representation by gender in both Spain and the United States. Similar results have been reported in previous studies (Bode et al., 2011; Mislove et al., 2011; Parmelee & Bichard, 2011) that also extend this finding to other social networks and contexts. ¹²

OPEN IN VIEWER

Table 1.

Patterns of Behavior on Twitter, by Gender.

Spain	Variable	Men	Women	All	t-Stat
	Number of users	6,192	3,389	9,581
		[65%]	[35%]	[100%]	29.9**
	Mean of tweets	19.6	21.4	20.2	1.9
	Mean of retweets	8.8	9.5	9.0	1.2
	Mean of replies	3.6	3.9	3.7	1.1
United States	Variable	Men	Women	All	t-stat
	Number of users	25,452	17,385	42,837
		[60%]	[40%]	[100%]	39.6**
	Mean of tweets	74.3	91.0	81.1	5.5**
	Mean of retweets	38.1	55.7	45.2	8.7**
	Mean of replies	7.1	6.7	6.9	0.6

Note. These results are based on the analysis of a random sample of size 12,000 and 50,000 active users in the conversation on Twitter about the elections in Spain and the United States. Gender of the users was identified based on the username, although the procedure did not allow to infer the gender of 1,773 users (15%) in Spain and 7,163 (14%) in the United States.

Significance (equality of means or proportions): *10%. **5%.

However, regardless of the descriptive distribution of users, in this article, we are interested in the extent to which gender is a variable affecting behavior on the platform. Table 1 shows that women tend to be more active and they also show a level of sophistication slightly higher than men, as suggested by the number of retweets and @-replies. However, the size of those differences is not large enough to compensate for the general underrepresentation of women on Twitter. In fact, in the case of Spain, the difference between the two groups is not statistically significant.

Our next variable of interest is the geographical distribution of Twitter users. We see in Figure 2 that there are no clear differences in the distribution of the general population and the distribution of active Twitter users at least when we aggregate data at the province level in Spain or the state level in the United States: More populated provinces/states also show a higher number of users in a roughly similar proportion, with perhaps the only exception of Madrid. This result is robust to other metrics, like the total volume of tweets (Figure 4) with a very high correlation (Pearson’s ρ = .846 with N = 52 in Spain, and Pearson’s ρ = .977 with N = 50 in the United States). Therefore, it seems clear that Twitter users are a representative sample of the geographical diversity of both countries.

OPEN IN VIEWER

Figure 4.

Geographical distribution of tweets sent during the electoral campaign.

But geography may affect behavior through a different dimension. In particular, the urban/rural cleavage is usually found to be one of the structural factors of political competition (Lipset & Rokkan, 1967) and in fact, the kind of habitat appears as one of the most solid regularities in the prediction of political behavior. More specifically, both in Spain and in the United States, large urban centers are usually found to be associated with left-wing/liberal electoral outcomes, and it is a commonplace to point out that voters in rural areas are more likely to vote for conservative parties than those living in large cities.

In order to analyze whether there are differences by kind of habitat, we focus on the town/“municipio” from which each Twitter user has sent messages to the platform. We split the sample in two groups: individuals living in cities with populations larger than 500,000 in Spain or 1,000,000 in the United States and the rest of the users. Results are shown in Table 2. Although we do not have data for our full sample, ¹³ our data suggest that localities of smaller size are clearly underrepresented when political issues are debated online. Although less than 17% of the Spanish population lives in one of the six larger cities (Madrid, Barcelona, Valencia, Sevilla, Zaragoza and Málaga), the users in our sample who live in one of them represent around 32% of the total volume of tweets. Similarly, 7% of the U.S. population lives in the nine most populated cities, but tweets sent from them represent around 21% of the total.

OPEN IN VIEWER

Table 2.

Patterns of Behavior on Twitter, by Size of Habitat.

Spain	Variables	Cities	Other	All	t-Stat
	Number of users	2,007	4,288	6,295
		[32%]	[68%]	[100%]
	Mean of tweets	20.4	21.8	21.3	1.1
	Mean of retweets	8.9	9.8	9.6	1.2
	Mean of replies	3.7	3.8	3.7	0.1
United States	Variable	Cities	Other	All	t-stat
	Number of users	7,042	26,300	33,342
		[21%]	[79%]	[100%]
	Mean of tweets	72.3	68.0	81.1	1.2
	Mean of retweets	38.2	37.3	45.2	0.3
	Mean of replies	5.4	5.7	5.6	0.30

Note. These results are based on a random sample of size 12,000 and 50,000 active users in the conversation on Twitter about the elections in Spain and the United States. The location of the tweets was recovered using the information provided by each user, using the geolocation application programming interface (API) of Yahoo!, although this procedure did not allows to infer the location of 5,705 users (47%) in Spain and 16,658 users (33%) in the United States.

Significance (equality of means): *10%. **5%.

If we focus on different metrics, we do not find significant differences between these two groups in our population: The average number of tweets, retweets, and “@-replies” is similar for users located in big cities and the rest for both countries. As a consequence, city size is an important source of inequality, as the smaller number of users located outside large cities is not compensated by a larger number of tweets.

We now present our results in relation to ideology as an explanatory factor. As indicated previously, this variable has been estimated using an iterative method based on the political identification dominant in the network of friends for each user (Barberá, 2014). This procedure generates a score for each individual in a continuous scale with mean 0 and standard deviation 1, in which smaller values are associated with higher intensity in the political self-identification with the political left (liberalism), and higher values are associated with the political right (conservatism). In the results shown subsequently, we restrict our analysis to the hard core of each ideological orientation in Twitter—that is, to users in our sample whose values in our scale are statistically different from zero.

Results are shown in Table 3. First, and with respect to the distribution of users by ideological orientation, we see that users are divided in similar proportions between left and right in both countries. Second, we find that the activity of users shows a remarkable ideological bias with respect to the number of tweets, retweets, and “@-replies” published during our period of analysis. In fact, the total participation of users in the extremes of the ideological scale is higher than for users that we could classify as belonging to the center. To put it differently, this result suggests that the political discussion in Twitter is mainly driven by citizens with extreme values in the ideological scale, a situation that certainly favors the level of polarization of the political discussion on Twitter. This result is consistent with the growing literature on the political consequences of the Internet that claims that these new communication platforms are creating polarized “echo chamber” environments where individuals are exposed to political information that reinforces their previous beliefs (Adamic & Glance, 2005; Prior, 2007; Sunstein, 2001).

OPEN IN VIEWER

Table 3.

Patterns of Behavior in Twitter, by Ideology.

Spain	Variable	Right	Left	All	t-Stat
	Number of users	2,895	2,626	12,000
		[25%]	[23%]	[100%]
	Mean of tweets	26.4	26.6	20.1	0.1
	Mean of retweets	10.9	13.0	9.0	2.2**
	Mean of replies	4.7	5.1	3.8	1.0
United States	Variable	Right	Left	All	t-Stat
	Number of users	12.513	12.293	50.000
		[25%]	[25%]	[100%]
	Mean of tweets	145.8	95.0	84.0	10,0**
	Mean of retweets	81.4	58.8	46.3	6.6**
	Mean of replies	12.5	7.5	7.0	5.0**

Note. These results are based on a random sample of size 12,000 and 50,000 active users in the conversation on Twitter about the elections in Spain and the United States. Party identification was recovered applying an item-response model to the network of friends of the users (Barberá, 2014).

Significance (equality of means between left and right): *10%, **5%. All differences between those identifying with left and right are statistically significant at the 5% level.

When dividing the sample between right-wing and left-wing users (or conservatives and liberals), we also find other relevant asymmetries. For instance, although the total number of tweets sent by each group in the Spanish case is roughly similar, conservative users retweet at a considerably higher rate than liberal ones. This kind of behavior suggests that the political discussion among right-wing voters is more hierarchical and is more structured, at least if we consider that most tweets from this group come from a relatively reduced number of users that are echoed by individuals politically more closely related to them.

In the United States, we also find significant differences between ideological groups, but at a deeper level. The users we classify as conservatives tend to be more active in general, with a higher number of tweets coming from this side of the ideological spectrum. This result is consistent with the findings reported in Conover, Gonçalves, Flammini, and Menczer (2012), who showed that Twitter users closer to the Republican party have a more polarized behavior that those leaning toward the Democrats.

Figure 5 complements this analysis of the different patterns of behavior arising from party identification by exploiting the fact that our measurement of ideological leaning not only allows us to classify individuals but also locates them in a continuous scale. In the figure, the color saturation of each hexagon represents the number of users whose ideology and number of tweets are located in that region of the graphic. The image that arises confirms our previous conclusions. Again, we find that individuals with more extreme values in the ideological scale post a higher number of tweets. Therefore, it is more than reasonable to expect that the political discussion on Twitter is more polarized than it would have been had the number of tweets been more homogenously spread among users.

OPEN IN VIEWER

Figure 5.

Number of tweets sent during the period of analysis, by ideology.

Finally, our last dimension of analysis is the relation between participation and the moment of the political campaign. In this case, we are interested in the inequality in participation at the aggregate level. Thus, we returned to the full data set and we calculated the Gini index for the number of users and tweets mentioning each of the two main candidates in each election. In the Gini index, higher values are associated with higher levels of inequality and it thus allows us to study the extent to which the discussion related to candidates is limited to a few users or, on the contrary, is open to more users with a similar level of participation. Moreover, a longitudinal analysis allows us to observe how this index changes as we move into the political campaign. One would expect that, as Election Day approaches, political promises and campaign events should increase the volume of messages, be that because the most involved users increase their participation or because new users join the discussion about current events.

Figure 6 presents the results of our analysis. The main finding is that inequality in participation changes substantially throughout the campaign as a consequence of the main political events (indicated by dashed lines). In the Spanish case, we find that these events increase inequality or, to put it differently, they increase the rate of participation of the most active users. The picture is more complicated for the United States and the effect of those events on the distribution of participation among users seems to depend on the nature of the event. On one hand, electoral debates attract a higher number of participants, which makes the distribution of tweets more homogeneous. On the other hand, events such as the attack against the U.S. embassy in Libya show the opposite effect, most likely because that particular event generated an intense response from voters from the Republican party. ¹⁴

OPEN IN VIEWER

Figure 6.

Evolution of the inequality in the participation on Twitter.

Hence, the evolution of inequality throughout the campaign suggests that any analysis based on Twitter data not accounting for the changes in behavior due to these exogenous events risks introducing relevant biases arising from the unequal participation of different types of users.

Conclusion

We started this article by underlining the opportunities offered by Twitter for the analysis of public opinion: Messages are exchanged by numerous users in a public forum and they may contain valuable information about individual preferences and reactions to different political events in an environment that is fully accessible to the researcher. However, the validity of the generalizations that one can make from that potential data source is conditional on our ability to overcome the limitations arising from the fact that participation, at least in the political discussion in this platform, is not homogeneously distributed among users.

Our study about the use of Twitter in the 2011 legislative elections of Spain and the 2012 presidential elections of the United States analyzes some of these possible shortcomings. Users participating in the political discussion were mostly men, living in urban areas, and with strong ideological preferences. Even more, our results remark the deep inequality that exists in the behavior of the groups overrepresented in Twitter, namely, that the followers of political parties are more active in the political discussion than the rest. We also found similar patterns along ideological lines, with conservative users being more active than liberal ones.

Beyond these specific results relative to our cases of study, our analysis has relevant, more general implications for the use of Twitter as a tool for electoral prediction. An analysis based on the volume of tweets or their “sentiment” will inherently be biased if it does not account for the factors indicated previously. In this sense, our analysis should not be considered as a criticism, but as an opportunity. The same way public opinion research significantly improved after taking more seriously the biases that arise from missing data, sample selection, and social desirability, we also expect researchers using information based on tweets to be able to improve their estimates by taking advantage of a deeper understanding of the patterns of inequality that arise in the political discussion in the Internet.