First Monday

Public attention, social media, and the Edward Snowden saga by Yong Jin Park and S. Mo Jang

Although prior research indicated the power of social media in drawing attention to particular issues, little research explored precise patterns by which an issue sustains its salience in mediated social platforms. We dissected the issue-attention cycle of Snowden revelations over a one-year period. We tackled this, using real-time large-scale Twitter data, and found that social media attention varied dramatically over time by the specificities of issue. Volumes of attention associated with — instead of being independent from — traditional media mention of the issues. Directions for future studies as well as implications for using social media for civil rights issues are addressed.


Herd behavior in issue attention
Context and questions
Discussion and practical implications
Potential consequences for civil rights issues




“Being called a traitor by Dick Cheney is the highest honor you can give an American, and the more panicked talk we hear from people like him ... the better off we all are.” — Edward Snowden, 17 June 2013.

Edward Snowden’s revelations about the U.S. National Security Agency (NSA) surveillance program PRISM sparked immense debate, not only about the unprecedented scales of privacy and civil rights violation, but also about the threat to the well-being of citizens in their relationship with the U.S. government (Greenwald and MacAskill, 2013). Snowden’s comment, quoted above, is his belief that continuous public attention and debates on these matters will determine the state of democracy and civil liberty in longevity. Such a concern invites a scrutiny of the role of the emerging digital platforms that host public dialogues on contemporary issues. This also begs the question of digital attention span, i.e., how much the upsurge of interest in a particular issue prevails within the dynamics of digital cornucopia and generates sustained attention, providing political pressure to bring about change (Downs, 1972).

In this study, we propose and test the herd behavior (Banerjee, 1992; Ginneken, 1992) in the context of Twitter conversation — that is, we look for evidence of mass behavior in the phenomenon in which a crowd swiftly gathers around an issue for a given period, but then declines dramatically. By testing herd behavior in social media use, our study aims to examine mass behavior underpinning the issue-attention cycle (see Downs, 1972; Neuman, 1990, 1986). Prior studies have established solid ground for investigating the patterns of Twitter use in issue polarization (e.g., Yardi and boyd, 2010), the participation gap of political dialogue (e.g., Bekafigo and McBride, 2013), and the distribution of emotional sentiment predicting an election outcome (e.g., Tumasjan, et al., 2010). Still, empirical efforts to explore the digital span of attention to a specific issue on Twitter have been scarce. To fill this gap in the literature, we use reference-based large-scale aggregate data sets and dissect more than 300,000 tweets sent out in the U.S. during the one-year span following Edward Snowden’s revelations. We are concerned if the interest in the matter rapidly sags over time despite an initial outrage. Fundamentally, we ask if the opportunities to engage in social media, through greater access to participatory information platforms such as Twitter, may translate into meaningful public attention (Park and Yang, 2017; Shirky, 2011; Webster, 2014).



Herd behavior in issue attention

The capacity of social media to coordinate massive and rapid exchanges — for instance, close to 500 million tweets per day — represents ‘arguably’ the first time in which we can engage in instantaneous issue dialogues irrespective of geography. Do digital tools such as Twitter enhance issue attention? The herd behavior was proposed by early behavioral economists (e.g., Simmel, 1957; Veblen, 1899) who tried to understand sudden shifts in consumption patterns, such as fashion styles or the fluctuation of the stock market. For instance, Veblen (1899) was interested in finding out the process by which one person buys a product just because other people are buying it — that is, a sort of bandwagon effect through which an individual joins the crowd without complete information or any apparent need. Simmel (1957) went a step further to suggest that one’s decision to follow others’ fashion is beyond a form of personal imitation, but more an expression of social equalization in which one desires to be in the same breed as others and achieve a group status.

The analogy of fashion is intriguing. This illustrates the potential motivation for one to join a herd in an effort to eliminate social distance between oneself and a trend that a fashion symbolizes. Traditionally, such assessments have been grounded in the sociological notion of class distinction (Bourdieu, 1984; Simmel, 1957). More recently, consumer literature (e.g., Argo, et al., 2006; Rook, 2006) have suggested that consumers rely on interpersonal cues and imitate others’ purchasing decisions when they decide to purchase a product, such as a new car or clothing. The reality, of course, is complicated. On one hand, the flow of information from a large crowd reduces risk-perception as people tend to converge on similar behaviors prone to fads, but less susceptible to mistakes. On the other hand, this is contrary to rational economic demand according to which individual consumption depends upon the aggregate of complete information and respective needs (Banerjee, 1992; Ginneken, 1992).

Note that the herd behavior concerns collective-crowd conducts at a population level. In this vein, communication and political science scholars have also made the same line of keen observations on mass behavior. A key aspect of earlier studies involves the rise and fall of salience of an issue in public debates as human attention is definitely a scarce resource (Hilgartner and Bosk, 1988). For instance, in such areas as environmental disaster (e.g., Downs, 1972), policy concern of international and domestic affairs (e.g., Peters and Hogwood, 1985), and food safety (e.g., Bodensteiner, 1995), scholars have observed the general tendency by which fashionable issues arise, take center stage, and then gradually fade away.

The insight of Anthony Downs (1972) is noteworthy in this regard because he called for macro-level social dynamics of the public response function. He proposed that each issue of ‘crisis’ has a cycle of issue attention, starting from (1) a pre-problem stage; (2) alarmed discovery; (3) debates for solutions and realization of costs; (4) gradual decline of interest; and, (5) post-problem stage of a prolonged limbo. While the exact applications of each step would vary from issue to issue, the issue-attention cycle provides a time-sensitive contour within which mass behavior, predicted by herd behavior, is to take place in response to real-world stimuli. Put differently, the premise of the attention cycle delineates the conditions under which herd behavior can be applied to fashionable issues over given time. Particularly interesting is Downs’ use of the time intervals that tend to function in congruence with the threshold of attention (Neuman, 1990, 1986). From this, we can assess and predict the progress of herd behavior at different stages of an issue-attention cycle, tracking the dynamics of public discussion over time.

Applied to the context of Twitter, the notion of the herd behavior, as indicated by an aggregate digital attention span, seems to be a reasonable proposition to test. One might argue that Twitter’s low-transactional cost in its easiness to access, read, and write Twitter comments makes it easier to organize attention, to keep up with discussions, and associate oneself with significant trends. But the same rationale could be presented to argue that the low monetary or time cost of obtaining information makes it easier to tag along without attaining a complete grasp of an issue and simply imitate other patterns of news consumption (Webster, 2014).

Unlike the purchases of other consumer goods like cars, houses, or fashion, the cost for sending/receiving tweets is miniscule. Users can easily obtain a Twitter account, follow certain persons, or become associated with dialogue on a particular issue. Instead, the benefit appears greater: gaining the status by following a popular trend, the social currency of talk, and the immediate opportunity to associate with specific individuals. After all, tweets can be simply transferred or retweeted to others as in contagion (e.g., Kramer, et al., 2010), potentially leading people to jump on a bandwagon in the absence of monetary cost and minimal time investment.

At the end, what is at the core of (1) mass psychology of collective-crowd conducts (i.e., herd behavior) and (2) the macro-level attentional dynamics of public response (i.e., Downs’ issue-attention) is ‘attention economy’ (Marazzi, 2008; Goldhaber, 1997). That is, in the abundance of digital information flow, the scarce resource is human attention and its allocation cycle determines how information is valuably consumed and shared. This, to us, would be a key point of integrating the herd behavior and issue-attention in the context of social media. In other words, in order for the crossover to digital media platforms to carry any significant changes, one must demonstrate that the scarcity of attention is disrupted and public dialogues are even extended beyond the limited menu of broadcast media (Hilgartner and Bosk, 1988; Webster, 2014).



Context and questions

Drawing from the propositions of herd behavior and the issue-attention cycle, we propose several investigations with regard to the digital span of issue-attention in the dynamics of Twitter conversation. Our analytical focus is on the actual mentions and expressions of tweets, which are hardly observable through self-reported measures. No single study will determine the complete attentional dynamics of issue dialogues in a complex networked ecology. Our strategy is instead to dissect one particular issue into sub-issue domains and to track their changes over time. This approach is different from that of prior social media studies, which tend to treat a block of time-series data as a single issue. Our approach also differs in that we intend to analyze real-time fluctuations of tweet volume in a natural field setting, not particular characteristics of tweets based on sampled Twitter users.

In fact, the Snowden case presents us with one of the most pivotal political issues involving citizens’ relationships with their government. First, the magnitude of Snowden’s revelations appears to be unprecedented and has instilled a genuine fear of surveillance and a profound sense of crisis. Yet one might argue that with the digital platform, citizens now possess a counter-power to organize their dissents — or even decry — against the secrecy of U.S. government intrusion. We can measure this capacity (or lack of) of voice by observing the distributions of volume of attention to this issue and related sub-issues over a period of time.

Those tweets that enjoy enduring mentions and stimulate attention will have a greater potential to survive in the abundant influx of digital information flow and, ultimately, to exert bottom-up pressure (if not a direct change) on governmental surveillance practices. Here it is important to examine the extent of (in)congruence between traditional news media and social media, i.e., the extent to which increasingly complex, hybrid digital information platforms interact with traditional media in stimulating attention to this matter (Dutta-Bergman, 2004). For the Snowden case, this can be particularly a fruitful area of inquiry because the ‘watchdog’ function of media can be compared with the role of alternative social media activity in drawing attention to an unprecedented scale of government intrusion.

Downs (1972) focused largely on how the traditional media initiate the ‘up and down’ of collective attention, i.e., the herd effect, on an issue. He was also concerned with how constraints on newspaper space or television time makes it difficult to sustain issue attention. We wonder whether social media, particularly Twitter, will differ in its capacity to initiate and sustain digital attention span as an alternative platform, independent of traditional commercial media constraints (Jang, et al., 2016; Neuman, 1990; Shirky, 2011; Webster, 2014). We are particularly interested in the pace and the intensity of Twitter activities at specific time intervals, indicative of the issue-attention cycle in social media. In this vein, the Edward Snowden revelations serve as a lens through which to examine the extent to which social media function to sustain public attention concerning one of the most salient civil rights issues.

Little is known about how new digital ecosystems, with potentially more diverse voices, are related to conventional media systems of information distribution on different, but related, dimensions of one specific issue. In some cases, traditional media outlets may be in competition for story coverage as new social media systems continue to adapt. Or new media platforms, because of the abundance of digital information, may have even greater difficulties than traditional media outlets in generating wide attention to particular types of civil rights issues. In traditional news journalism, there has been much discussion of how a story may have “legs” and its subsequent developments maintain attention. In parallel, it is possible to ask if such dynamics — as the dominance of strategy-oriented or episodic news over hard news (for instance, Edward Snowden the personal story vs. sub-issues such as the policy of NSA surveillance or PRISM program) — would be sustainable in social media (Iyengar, 1991). Peters and Hogwood (1985) also pointed out that there seem to be different patterns of issue cycles depending upon the natures of individual events, as their work examined the policies in the areas of domestic and foreign affairs. Although this research helps us suspect the unique attentional dynamics of sub-issues, there exists no solid empirical ground on which to predict how related-issue subtleties will differ within one issue.

Accordingly, we first investigate how Twitter activities about Snowden intensified and became subdued in terms of overall volume over time, and we hypothesize, consistent with our prior premise, that Twitter’s low transactional cost will favor herd behavior, in that the volume of activity will show rapid increases, followed shortly thereafter by rapid declines (H1). This is in line with our prior premise that Twitter’s low transactional costs will favor information in contagion with no sustained attentional focus that may be required to induce change in social institutions. This prediction also derives from the conjecture that information consumption may be more prone to imitation when it comes to controversial issues that would not be intrinsically exciting, such as governmental policies or civil rights issues (Bodensteiner, 1995). Next, our analysis parcels into the sub-issue domains of Twitter expressions about Snowden’s revelations, and looks at whether the Edward Snowden issue, when specified into two different dimensions of surveillance and privacy, exhibit different dynamics of attention flow (RQ1). By sub-issue domains, we mean the specificities of the issue and this is important because we also aim to detect how specific and articulate talks (Neuman, 1986) are different from general topical conversations in networked social media environments. Finally, we examine the relationship between traditional media, and the surge and decline of social media attention (RQ2).




Data collection and procedure

We subscribed to and accessed a digital archive from social media analytic firm Topsy Labs Inc. which has a direct access to and sifts through a full fire hose of more than 400 million daily tweets and all texts tweeted from 6 June 2013 to 5 June 2014 — that is, one full calendar year since Edward Snowden’s revelations. Our data include retweeting of an original 140-character post and replies to someone else’s posted comment and tweet. Importantly, we also accessed the media analytic firm Sysomos, which compiles a digital archive of the news contents of traditional media around the world, including national broadcasts, local newspapers, broadcast Web sites, and print media, such as CNN, Fox News, MSNBC, Wall Street Journal, Los Angeles Times, and New York Times. Sysomos’ servers scan an estimated 100 billion records of news articles and tweets in response to each search phrase we enter, adding two billion news items every week. Thus, there are two units of analysis in our study: (1) the number of tweets; and, (2) news articles for traditional media. For our analysis, we focus on the U.S. and those tweets written in English. We combined the original tweets, retweets, and replies to construct an aggregated volume of Snowden-related tweets.

We normalized the volume of tweets by dividing each day’s total volume by the daily average for the year so that fluctuations can be compared to the average volume (see Neuman, et al., 2014). This is an important step for our purposes because in tweet volumes, there is no natural metric like the currency of dollar, against which the flow is readily standardized at the absolute level (Neuman, 1990). Thus, the appropriative corrective is to use the normalized form so as to standardize the variations across different types of issue-tweets. In combined summary, our data consist of a total baseline volume of 5,303,673 tweets and 676,561 news media article mentions of Edward Snowden and related sub-topics during the one-year period. This means that on an average day, there were about 1,853 mentions in the mass media and 14,530 mentions in tweets that were related to the issue of Edward Snowden.

Boolean search methods

We grabbed the actual tweets of particular mentions, using powerful Boolean search syntax. For instance, we entered a keyword of “NSA AND surveillance” to retrieve all tweets referring to the U.S. government entity in the context of conducting massive surveillance program. For our purpose, the key was to capture thematic subtleties by the issue conversations of different types. This will construct content validity by observing issue-attention of different but related domains, rather than treating attention as having only a single-unified dimension. These refinements result in multiple contingencies of the surveillance and privacy dimensions underlying the Snowden revelations, allowing us to detect overall as well as discrete attention patterns in their unique dynamics.

Our filter process of the Boolean rules (for instance, using operators such as NOT) was better suited to exclude any extraneous contents (i.e., false positive) and to narrow filter ranges in retrieved tweets. This step is different from bots-crawling techniques used by computer scientists in prior work (Helles, 2013). The important merit of our approach is to anchor the thematic emphases qualitatively in our social media measures. Our approach is also different from semantic network analysis (Qin, 2015) in that we start with specific sets of interested references in a deductive process, while inductive reasoning bases the network analysis, which searches for the emergence of associated patterns among words or phrases in retrieved sets of sample texts. Finally, we aim for better generalization as our analysis advances through so-called “big data” — a massive set of tweets, print news articles, and just about every tweeted material available.

Our technique for capturing sub-issue domains can be broken down into a series of steps. First, after looking into Snowden-related tweets, we identified three main entities involving the Snowden revelation-keywords: that is, (1) Snowden (the whistleblower); (2) the NSA (the surveillor); and, (3) PRISM (the surveillance program) [1]. Second, within each of these three sub-issue domains, we dissected the issue attention into two levels of reference specificity:

  1. The issue frame-level, at which tweets are broadly referenced in the dimension of either surveillance or privacy; and,
  2. The issue reference-level, at which tweets are specified with particular reference strings that represent the dimension of either surveillance or privacy.

Table 1 shows the key phrases used on the different issue levels and corresponding summary statistics.


Table 1: Distribution of issue clusters.
Notes: Entries are average daily volumes. Values in parentheses are the raw totals.
Issue reference in Table 1 and subsequent tables denotes specific sets of utterances in two dimensions of privacy and surveillance. Each cell of issue reference contains respective key phrases: protection, liberty, and privacy rights (for privacy cell) and privacy violation, intrusion, and concern (for surveillance cell).
Issue sub-topicSnowdenNSAPRISM
Issue frame70.36
Issue reference15.46


Key identifying phrases

Here the distinction between issue frames and issue references deserves clarification. First, for each issue level, we derived a set of the key identifying phrases unique to that level for the purpose of retrieving the full text of tweet commentaries and traditional media news stories. To warrant the validity of our issue clusters, this process was triangulated by cross-checking randomly-selected tweets in which we discerned different types of issue attention in their respective dimensions. For this, roughly tree percent of all tweets (n=39,777) in the first three months of the time period were selected and examined qualitatively, which cross-validated criterion validity (Lindgren and Lundström, 2011). In addition, we examined the tweets of ‘influencers’ defined by Topsy Lab Inc. as those who have an above-average impact score in having their tweets retweeted by followers (Tufekci, 2013).

Tweet filer strategies for the issue frame level include two broad categories of “surveillance” and “privacy”, as defined by Boolean keyword search rules to exclude mutually each other from the respective dimension. For the issue reference-level, the keyword strings of “privacy violation”, “privacy intrusion”, and “privacy concern” were added and coded as specific references to the “surveillance” dimension, whereas the strings of “privacy protection”, “civil liberty”, and “privacy rights” were added for the “privacy” dimension. Reference-level strings reflect the ways in which one can articulate each dimension of the issue. For instance, the proactive assessment using the terms like rights, protection, and liberty constitutes “privacy” (civil rights frame), whereas the negative or worrisome appraisal using violation, intrusion, and concern represents “surveillance” (government surveillance frame).

These strings are not an exclusive list, but as explained by Neuman, et al. (2014), this approach provides key identifying phrases unique to the public dialogue on Snowden, as construed in terms of privacy and surveillance. In other words, we used the parameters of these additional keyword strings to capture the potential range of the references in terms of a dimensional difference. Note the discrete distinction in the dialectic between surveillance (i.e., being public; open; and violated) and privacy (i.e., being private; closed; and protected) (Park, 2015a, 2015b, 2011), as we classified the issue-clusters into the two dimensions of the civil rights debate underlying the Edward Snowden case. Based on each of these criteria, our filter strings scanned the entire dataset during the one-year span for the corresponding volume of each day. Accordingly, we had a 2 (frame and reference levels of specificity) x 2 (surveillance and privacy dimensions) issue clusters in each of the three sub-issue domains (see Table 1 for summary statistics). The words or phrases contained in hashtags are included in our issue clusters. The hashtag itself is not a unit of our analysis because we were not interested in identifying particular patterns of organized nodes in connection, as in studies that regard hashtags as framing devices (e.g., Hemphill, et al., 2013).

Overall, the logic is to move from the categories of the broad frames (issue frame-level) to the narrower filters (issue reference-level) that allow us to discern specificities of the issue (see Goffman, 1974). In this regard, one of the main goals in this study is to examine (1) the different frames and (2) the discrete references of tweets, with a particular focus on the contents of actual conversations. This is important for our purpose, because cognitive attention often remains subject to the potential bias embedded in word frames and references (Entman, 2004). For instance, the dominant frames of how stories are told can promote unique dimensions of the issue, while diluting alternative collective attention. Furthermore, specific sets of utterances may be even more highly sensitive and distill certain issue dimensions that otherwise would not stand out (Goffman, 1974). Here we also intend to establish a well-grounded rule for future studies, in that the selection of social media measures can be cross-checked based on both empirical and logical reasoning.




To examine the distribution of attention to the Snowden revelations and related sub-issues over the one-year time span, we mapped out a global overview of the flow of tweet volumes over time. Figure 1 shows the fluctuation of each of the three sub-issues of Snowden, NSA, and PRISM from 6 June 2013 to 5 June 2014, with the related events shown in the daily time sequence. The inflection is straightforward, lending strong support to H1. For instance, during the one-year span, the daily volume ranged from 0 up to 117,912 tweet mentions (that is, the large attention gap between the highest and the lowest points of any sub-issues). The patterns of time-variation display the hikes of tweet volumes remained concentrated in the first 30 days after the initial Snowden revelation. Furthermore, the acmes of hypes rapidly waned within the five days after 6 June 2013.


Attentional trends, Edward Snowden/NSA/PRISM for one year
Figure 1: Attentional trends: Edward Snowden/NSA/PRISM for one year.
Note: Data are the raw volumes (y, unit in 1000) prior to the normalization for each day (x).
Note: Larger version of figure available here.


Reviewing the patterns as illustrated in Figure 1, however, one also comes to appreciate the subtle dynamics of the three sub-issue domains. First, the tweet mention of Edward Snowden (M = 2,267.67, SD = 4,473.11, Max = 38,539) declined rapidly, but rebounded slightly at the end of the one-year span, while the NSA tweets, the largest in daily volume (M = 12,071.57, SD = 13,018.51, Max = 117,912), also swiftly disappeared in a similar pattern of a wave. PRISM, on the other hand, appeared to have its own life span (M = 191.35, SD = 711.74, Max = 7,857.00) as its initial spike was followed by a flat line of attention that never bounced back.

Table 2 displays the average volumes and ranges of issue attention when parceled into discrete types of issue clusters across the four-quarter intervals (RQ1). The year-long pattern displays a trend in which the tweets during the first quarter accounted for a disproportionately large share of the total volumes in all issue clusters. For instance, the issue clusters of PRISM in the first quarter (just 25 percent of the 365 days captured in our dataset) account for about three times the number of Twitter mentions each of the following three quarters. We ran simple mean t-tests between the average volume in the first quarter and the average yearly volume.


Table 2: Average volumes and ranges of issue clusters by issue frame and reference.
Notes: Data are normalized for each day. Range is the difference between the highest and the lowest volumes of each quarter.
 Issue frame Issue reference
 PrivacySurveillance PrivacySurveillance
 RangeVolumeRangeVolume RangeVolumeRangeVolume
First quarter         
Snowden18.601.6544.351.88 17.141.7036.021.83
NSA37.182.0313.732.18 14.741.9238.582.15
PRISM29.283.3846.883.63 41.333.1490.403.43
Second quarter         
Snowden3.310.432.940.58 4.370.407.150.15
NSA4.810.802.040.64 5.310.652.350.30
PRISM13.930.493.350.19 22.320.7612.160.36
Third quarter         
Snowden24.131.104.560.60 7.070.4717.891.15
NSA7.820.807.580.82 35.511.1625.401.01
PRISM2.680.101.750.09 3.830.087.600.11
Fourth quarter         
Snowden12.160.8014.910.92 35.181.4136.400.84
NSA1.630.351.020.34 1.920.2521.320.52
PRISM0.320.020.940.07 0.950.017.600.09
The year         
Snowden24.1570.3644.35195.50 35.1815.4636.400.45
NSA37.31276.0613.901045.75 35.5432.1738.584.38
PRISM29.2821.2446.8849.47 41.331.6790.400.52


Table 3 shows significant and negative t-values at the .001 level in all issue clusters, displaying the lopsided concentrations of issue-tweets in the first quarter. Interestingly, no clear dimensional difference between privacy and surveillance was found at either the issue frame or issue reference level, although at the frame-level the surveillance dimension was inclined to generate a larger proportion of daily volumes, indicating that understandings of Snowden-related issues may be more broadly anchored in the frame of government surveillance than in the specific articulations of civil rights debates over privacy.


Table 3: First attention peak and yearly average attention.
Notes: ***Significance level at .001.
Entries are t-values. Mean differences between the average volume in the first quarter and the yearly average were tested in the general linear model using one-sample t-test.
 Issue frameIssue reference
Snowden-17.60 ***-43.35 ***-105.07 ***-48.43 ***
NSA-5.86 ***-12.97 ***-41.62 ***-48.14 ***
PRISM-28.28 ***-45.89 ***-112.44 ***-86.80 ***


When examined in this way, the issue clusters at the reference-level call for particular considerations. First, unlike the generic frame-level, the reference-level clusters had low overall volumes, indicating that highly-specific references to Snowden-related events were confined within a small number of tweets. Second, there was a pattern of bouncing back of tweet fluctuation, in which issue clusters in the first and fourth quarters reached similar ranges of attention [2]. Third, in this regard, the bouncing back of the Snowden and NSA issue clusters (36.02 — 36.40, Snowden; 38.58 — 21.32, NSA) at the (surveillance) reference-level were noteworthy as these were in contrast to the frame-level attention where no such pattern existed.

Overall, the observed distributions indicate two characteristics:

  1. A stable pattern that is common across issue clusters (that is, a flat line of attention in the second and third quarters); and,
  2. The volatility of fluctuations as manifested in the discrete dimensions of surveillance (that is, similar hikes in the first and fourth quarters).

RQ2 asked how the surge and decline of tweets were related to the traditional media mentions (retrieved using the same Boolean rules as in tweets) (M = 765.05, SD = 899.96, for Snowden) (M = 897.18, SD = 884.62, for the NSA) (M = 191.35, SD = 711.74, for PRISM, for media mentions). Table 4 shows the correlations between media mentions and each of the 12 issue clusters. In all issue clusters, we found significant correlations. Furthermore, the strengths of the average correlations show that the fluctuations of tweets were in tight sync with the volume of media mentions (average r = 0.51 and 0.67, for the issue frames of privacy and surveillance; 0.56, 0.37, issue references of privacy and surveillance, respectively). Although this is not indicative of causal directionality, the findings point to the potentially mutual influences in that tweet flows rarely function independently of traditional media. There was no clear difference in the patterns of correlations across sub-issue domains.


Table 4: Correlations of issue attention: Tweet and traditional media.
Note: ***Significance level at .001.
Media mention SnowdenNSAPRISM Average correlation
Issue frame      
Privacy 0.23 ***0.48 ***0.83 *** 0.51
Surveillance 0.41 ***0.82 ***0.79 *** 0.67
Issue reference      
Privacy 0.31 ***0.69 ***0.70 *** 0.56
Surveillance 0.38 ***0.19 ***0.56 *** 0.37




Discussion and practical implications

This study investigated the distribution of issue attention by looking at the Edward Snowden case as a lens to demonstrate the function of social media in sustaining digital attention. The premise of our analysis was an observation of herd behavior (Banerjee, 1992; Ginneken, 1992) — that is, the collective behavior of joining a massive flow of tweets at a transient attention cycle (Downs, 1972). Fundamentally, the key aspect of our analysis concerns the role of social media platforms in sustaining the salience of an issue in public debates, given the scarcity of human attention in an increasing number of information choices (Neuman, 1990, 1986; Webster, 2014). In sum, the case of Snowden revelations epitomizes our concern over whether social media, such as Twitter, can function as an alternative platform to spur sustained attention and debate regarding important matters of civil liberty.

We found evidence in support of herd behavior. However, the results of our analysis also suggest that the mass behavior was subtler than one might suspect, as the issue clusters exhibited discrete patterns within their own unique dynamics. Thus, simply concluding that there was herd behavior in Twitter activities would downplay contextual nuances as to how issue conversations of different types were initiated and progressed, despite the overwhelming predominance of attention at a very early stage of an issue. Importantly, the flows of tweet volumes and media mention were found highly correlated in sync, displaying how deeply Twitter conversations may be embedded in the intricate progression of traditional media.

One of this study’s main findings concerning the massive exodus of issue attention at the start of the Snowden-NSA revelations deserves attention. First, this finding indicates that social media discussion may not necessarily go through all phases of a traditional issue attention cycle. It appears that a social media attention cycle bypasses substantial issue development stages of ‘debates for solutions’ and ‘gradual decline of interest’ at an extremely fleeting rate [3]. The disproportionately large share of total volumes in the first quarter shows that what remains salient in the social media issue cycle may be the alarmed discovery alone — that is, the first stage after the issue is known. Related, one of the significant revelations in our data is the apparent fact that Snowden as the person seems to have overwhelmed the real story, that is the revelations about the NSA and related issues of civil liberties and privacy.

Theoretically, the results provide insight into the function of herd behavior in social media platforms. In other words, the collective behavior of Twitter use may tend to reach a tipping point so quickly and easily that the massive transfer of the conversation initiated by tweets may not necessarily carry substantial influence on individual cognition (Banerjee, 1992; Ginneken, 1992). Instead, herd behavior appears to work in such a way that information choices in Twitter gravitate toward an issue that has already gained attention. This also adds to evidence that the benefits of Twitter’s low transactional cost result in not only the massive initiation of issue conversation and information distribution, but also its swift termination, while foregoing critical stages of the issue-attention cycle.

When broken into issue clusters, our findings concerning the discrete patterns of attention are noteworthy. What is critical is the bouncing back of reference-level attention on Snowden and the NSA at the end of the one-year span because it displays the divergent patterns by which a few highly specific references had an enduring power to generate discussion. First, this is different from generic frame-level issue clusters in which massive issue attention was quick to dissipate. Second, more importantly, the finding raises a possibility that digital social media platforms, such as Twitter, may be better at promoting extremely few, but articulated (more specified), issue attention. In other words, there appears to be a volatile period during which attention can resurface and be articulated in specific references, potentially progressing into substantial dialogues. Note the contrast with the mass fleeting of tweets in the early stage of the Snowden case. The attentional dynamic in newly emerging social media platforms is in fact non-linear and highly stratified, where a few articulated tweets may generate different gains in attention threshold and resist possible herd behavior.

Collectively, this finding suggests an issue-attention gap between the contagious stage of issue discovery and the stage of bouncing back of a few highly-specific issue references. Scholars (such as Brynjolfsson, et al., 2011) elsewhere have proposed a long-tail distribution in which a heavy concentration occurs at the head of the distribution, whereas a large share of the distribution remains dispersed in the long tail [4]. An interesting comparison can be made here. In our case, the overall pattern of distribution is similar, but also distinctive in that the digital attention cycle exhibits a pattern of volatility (a tail dispersion of few references) as well as a long-tail stability (excessive head concentration in the first quarter of aggregate-collective attention dropping to the flat line of no attention). On the one hand, technological progress reflected in the abundance of social media platforms seems to eliminate a certain process of the issue cycle. On the other hand, the sustenance of a small number of highly-specified references implies that these special issue ranges tend to remain relatively immune to the imitation of some fashionable tweets, whereas the massive attention stays homogeneously concentrated, brief, and generically framed.

We capture this in Figure 2, which depicts the reverse-power log of the highly skewed distribution of attention at the surveillance frame-level. The trend is salient in that it is far from a normal distribution. That is, as much as 45 percent of the total tweets mentioning Snowden at the generic frame-level fall into just a two-week span. Put it differently, a ratio between the volume of tweet and the number of days taken is 0.08, with about four percent of the entire year of 2013–2014 accounting for little less than half of the total attention paid to the Snowden surveillance issue. A particularly unique element of our study is that it can identify these distinctive characteristics through real-time population-level data and depict the lopsided distribution of the issue-cycle at the macro-level of social networked environments.


Snowden surveillance frame and its tweet concentration
Figure 2: Snowden surveillance frame and its tweet concentration.
Notes: Data are normalized (tweet volume y) for each day (x) prior to analysis. Volumes are stacked up by rank order.


The findings concerning high correlations between tweets and traditional news media mentions suggest that the function of social media may not be unique or bypass influence from traditional media in drawing attention to the matters of civil liberty and government intrusion [5]. This result is consistent with the prior observation by Neuman, et al. (2014) who found the attentional dynamics of traditional and social media are correlated in bi-directional influence. This is an important insight, given there is no commercial constraint of advertising support or distribution cost in Twitter information consumption. Instead, the possible mutual influences between the tweets and traditional news media mentions allow us to speculate that the traditional media may heavily influence or even contribute to the excessively-skewed concentration of tweets at the beginning of the Snowden case.

It is important to note that we ‘know’ the Edward Snowden saga originated from traditional media sources because the Guardian broke the story along with other mass media channels like the Washington Post. In other words, while we do not have the ways to discern the precise temporality or directionality of mutual influences between tweets and traditional news (as shown in the study by Neuman, et al., 2014), the resilient power of mass media cannot be discounted. Here one might rightfully argue for its powerful agenda-setting function at the very beginning of Edward Snowden case, whether tweets or mass media are better at sustaining public attention at a specific moment for particular sets of stories.

The point here is that despite few articulate tweets, social media alone may not appear to bring fundamental changes into the nature of public attention. This also makes sense, given the complementarity between consumptions of traditional and new media, as people seeking information in specific areas tend to conduct similar information search across different media platforms and outlets (cf., Dutta-Bergman, 2004; Gottfried and Shearer, 2016).



Potential consequences for civil rights issues

The concern addressed in our study is not new. In the tradition of marketing research, scholars have long been concerned possible ways to extend the consumption cycle and understand consumers’ purchase decisions. What is new is that we have imported their insights on collective behavior to identify the attentional dynamics of a civil rights issue in newly-emerging social media platforms. Our investigations led us to conclude the limited opportunities as well as new challenges social media offer. Twitter’s instantaneous exposure of the U.S. government’s surveillance activities had generated massive discussions, yet in exchange for fleeting attention. At the core lies the utility of social media platforms, notably, Twitter, in bringing meaningful dialogues to civil rights issues and generating political pressures to change unwarranted government surveillance. Our analysis presented no compelling evidence.

These findings make us doubt claims that massive bottom-up tweets will enable civil rights debates differently from traditional news media, which have been criticized for undermining important social issues due to bottom-line considerations. What remains to be seen is the role of a few influential elite-tweets in marshalling collective responses to civil rights and privacy violations, beyond a brief stint of Snowden hype. These tweets may be likely to come from opinion leaders, such as journalists, whistleblowers, civic organization and advocacy leaders, and educators (Katz and Lazarsfeld, 1955). In this regard, the distinction between ‘journalist-driven’ and ‘event-driven’ stories is useful and important. For instance, the Snowden story, largely driven by investigative reporting (much to the credit of an individual journalist, Glenn Greenwald), may have better places than stories about catastrophic events or terrorism. It is quite possible that newly-emerging digital media systems will increasingly consist of numerous information actors whose role is beyond conventional news relay. Accordingly, we do not entirely eliminate our cautious optimism that the power of various information actors can renew wider circulation and offset the poor digital attention span for a specific issue.

Our recommendation is to spur the flow of information from a few opinion leaders and learn how to integrate their articulate issue tweets into prolonged conversations and debates. The explosion of emerging social media platforms will make opinion leaders even more important because the collective attention needs help to find sufficient interest on important civil rights issues, such as privacy and surveillance, environments, or policies. One of the limitations in this work is that we could not precisely follow the entire and complex flow of Edward Snowden’s revelations. In fact, leaks were published across the full length of the study period and beyond, and numerous expert-based and digital rights organizations were involved in a broader range of related issues. PRISM, one of the three key issues, was only one of many revelations by Snowden, and the short discussion span on this particular program may be therefore understandable. Despite continued new revelations by Snowden, the decline of discussions about Snowden and NSA should be an important attentional dynamic that deserves further investigation.

In this vein, we will need individual-level experimental studies to understand the precise mechanism by which interpersonal dynamics play out for few influential Twitter users and their followers. An important inquiry is how the different social positions of social media users may affect the contagion of attention to some unfashionable civil rights issues. Related, our study cannot tell the potential influence of Twitter users’ socio-demographic backgrounds (Bourdieu, 1984; Park, 2015a, 2015b, 2011). For instance, it will be valuable to learn to what extent age or gender affects the types of issue-tweets that individuals generate and whether such differences are related to the ability to trigger broader interest in a given issue. End of article


About the authors

Yong Jin Park is Professor at the Cathy Hughes School of Communications at Howard University.
E-mail: yongjin [dot] park [at] howard [dot] edu

S. Mo Jang is Assistant Professor in the School of Journalism and Mass Communications at the University of South Carolina.
E-mail: mo7788 [at] gmail [dot] com



The authors express their gratitude to Dr. W. Russ Neuman for his guidance in earlier drafts; First Monday’s anonymous reviewers for their generous comments; and Roland J. Park for his assistance. The first author also thanks colleagues at the Cathy Hugh School of Communications at Howard University.



1. It is reasonable to suspect that NSA and PRISM might be somewhat related; however, we distinguish these two to discern a set of utterances in specific conversational contexts.

2. Here the range (the difference between the highest and lowest volumes of each quarter) indicates the breadth of volume fluctuations. This has important implications for displaying the volatility of attention at specific time intervals.

3. Fluctuations of mass media coverage remained similar to those of tweets. Notably, at the issue-frame level, the volumes of Snowden media mentions (either dimension of privacy or surveillance) displayed a pattern of lopsided concentration at the outset, although the magnitudes were smaller than tweets (0.511, for surveillance; 0.307, for privacy, in the ratio between first quarter and yearly total). This raises two interesting possibilities. First, disrupted by social media outlets, traditional media in digital environments no longer follow the conventional attention cycle, as proposed by Downs (1972). Second, investigative reporting, like the Snowden case, may be prone to herd behavior, for the investigation itself depends on few elite journalists. We speculate these because the Snowden issue deviates from other conventional news stories like policy debates or disastrous events, which might follow Down’s public attention cycle more closely.

4. Their emphasis was on a rather optimistic interpretation of the tail end where niches can be small in volume, but unlimited in type. This assessment has been criticized for failing to recognize the other end of the head concentration.

5. This does not necessarily contradict the study by Qin (2015), who found differences in the news frame between social media and legacy news. The robust correlations in this study’s findings concern the temporal flow of volumes, whereas Qin’s analysis focused on network patterns of connected nodes.



J. Argo, K. White, and D. Dahl. 2006. “Social comparison theory and deception in the interpersonal exchange of consumption information,” Journal of Consumer Research, volume 33, number 1, pp. 99–108.
doi:, accessed 19 July 2017.

A. Banerjee, 1992. “A simple model of herd behavior,” Quarterly Journal of Economics, volume 107, number 3, pp. 797–817.
doi:, accessed 19 July 2017.

M. Bekafigo and A. McBride, 2013. “Who tweets about politics? Political participation of Twitter users during the 2011 gubernatorial elections,” Social Science Computer Review, volume 31, number 5, pp. 625–643.
doi:, accessed 19 July 2017.

C. Bodensteiner, 1995. “Predicting public and media attention span for social issues,” Public Relations Quarterly, volume 40, number 2, pp. 14–19.

P. Bourdieu, 1984. Distinction: A social critique of the judgment of taste. Translated by R. Nice. Cambridge, Mass.: Harvard University Press.

E. Brynjolfsson, Y. Hu, and D. Simester, 2011. “Goodbye Pareto principle, hello long tail: The effect of search costs on the concentration of product sales,” Management Science, volume 57, number 8, pp. 1,373–1,386.
doi:, accessed 19 July 2017.

A. Downs, 1972. “Up and down with ecology — the ‘issue-attention cycle’,” Public Interest, volume 28, pp. 38–50.

M. Dutta-Bergman, 2004. “Complementarity in consumption of news types across traditional and new media,” Journal of Broadcasting & Electronic Media, volume 48, number 1, pp. 41–60.
doi:, accessed 19 July 2017.

R. Entman, 2004. Projections of power: Framing news, public opinion, and U.S. foreign policy. Chicago: University of Chicago Press.

J. van Ginneken, 1992. Crowds, psychology, and politics, 1871–1899. New York: Cambridge University Press.

E. Goffman, 1974. Frame analysis: An essay on the organization of experience. Cambridge, Mass.: Harvard University Press.

M. Goldhaber, 1997. “The attention economy and the Net,” First Monday, volume 2, number 4, at, accessed 19 July 2017.
doi:, accessed 19 July 2017.

J. Gottfried and E. Shearer, 2016. “News use across social media platforms 2016,” Pew Research Center (26 May), at, accessed 19 July 2017.

G. Greenwald and E. MacAskill, 2013. “NSA prism program taps in to user data of Apple, Google and others,” Guardian (7 June), at, accessed 19 July 2017.

R. Helles, 2013. “The big head and the long tail: An illustration of explanatory strategies for big data Internet studies,” First Monday, volume 18, number 10, at, accessed 19 July 2017.
doi:, accessed 19 July 2017.

L. Hemphill, A. Culotta, and M. Heston, 2013. “Framing in social media: How the US Congress uses Twitter hashtags to frame political issues,” SSRN (29 August), at, accessed 19 July 2017.

S. Hilgartner and C. Bosk, 1988. “The rise and fall of social problems: A public arenas model,” American Journal of Sociology, volume 94, number 1, pp. 53–78.
doi:, accessed 19 July 2017.

S. Iyengar, 1991. Is anyone responsible? How television frames political issues. Chicago: University of Chicago Press.

S. Jang, Y. Park, and H. Lee, 2016. “Round-trip agenda setting: Tracking the intermedia process over time in the ice bucket challenge,” Journalism (30 August).
doi:, accessed 19 July 2017.

E. Katz and P. Lazarsfeld, 1955. Personal influence: The part played by people in the flow of mass communications. Glencoe, Ill.: Free Press.

C. Marazzi, 2008. Capital and language: From the new economy to the war economy. Translated by G. Conti. Los Angeles, Calif.: Semiotext(e).

W. Neuman, 1990. “The threshold of public attention,” Public Opinion Quarterly, volume 54, number 2, pp. 159–176.
doi:, accessed 19 July 2017.

W. Neuman, 1986. The paradox of mass politics: Knowledge and opinion in the American electorate. Cambridge, Mass.: Harvard University Press.

W> Neuman, L. Guggenheim, S. Jang, and S. Bae, 2014. “The dynamics of public attention: Agenda setting theory meets big data,” Journal of Communication, volume 64, number 2, pp. 193–214.
doi:, accessed 19 July 2017.

Y. Park, 2015a. “Do men and women differ in privacy? Gendered privacy and (in)equality in the Internet,” Computers in Human Behavior, volume 50, pp. 252–258.
doi:, accessed 19 July 2017.

Y. Park, 2015b. “Explicating net diversity in trend assessment,” Communication Research (2 September).
doi:, accessed 19 July 2017.

Y. Park, 2011. “Provision of Internet privacy and market conditions: An empirical analysis,” Telecommunications Policy, volume 35, number 7, pp. 650–662.
doi:, accessed 19 July 2017.

Y. Park and G. Yang, 2017. “Personal network on the Internet: How the socially marginalized stay marginalized in personal network diversity and multiplicity,” Telematics and Informatics, volume 34, number 1, pp. 1–10.
doi:, accessed 19 July 2017.

B. Peters and B. Hogwood, 1985. “In search of the issue-attention cycle,” Journal of Politics, volume 47, number 1, pp. 238–253.
doi:, accessed 19 July 2017.

J. Qin, 2015. “Hero on Twitter, traitor on news: How social media and legacy media frame Snowden,” International Journal of Press/Politics, volume 20, number 2, pp. 166–184.
doi:, accessed 19 July 2017.

L. Rook, 2006. “An economic psychological approach to herd behavior,” Journal of Economic Issues, volume 40, number 1, pp. 75–95.
doi:, accessed 19 July 2017.

G. Simmel, 1957. “Fashion,” American Journal of Sociology, volume 62, number 6, pp. 541–558.
doi:, accessed 19 July 2017.

Z. Tufekci, 2013. “‘Not this one’: Social movements, the attention economy, and microcelebrity networked activism,” American Behavioral Scientist, volume 57, number 7, pp. 848–870.
doi:, accessed 19 July 2017.

T. Veblen, 1899. The theory of the leisure class; An economic study in the evolution of institutions. New York: Macmillan.

J. Webster, 2014. The marketplace of attention: How audiences take shape in a digital age. Cambridge, Mass.: MIT Press.

S. Yardi and d. boyd, 2010. “Dynamic debates: An analysis of group polarization over time on Twitter,” Bulletin of Science, Technology & Society, volume 30, number 5, pp. 316–327.
doi:, accessed 19 July 2017.


Editorial history

Received 10 April 2017; revised 16 June 2017; revised 18 July 2017; accepted 19 July 2017.

Creative Commons License
This paper is in the Public Domain.

Public attention, social media, and the Edward Snowden saga
by Yong Jin Park and S. Mo Jang.
First Monday, Volume 22, Number 8 - 7 August 2017