A Systematic Identification and Analysis of Scientists on Twitter

• The broader-impact interpretation of altmetrics remains empirically ungrounded, and activity attributed to “the public” may instead come from scientists.Identifying scientists and non-scientists is presented as a crucial step toward validating that broader-impact claim.Source: Introduction · p. 2
• Earlier identification efforts were narrow or small-scale and often purposively selected candidates based on prior success in other sources.The examples include selecting highly cited researchers rather than organically finding scientists on Twitter.Source: Introduction · p. 2
• Reliance on bibliographic databases binds studies to traditional citation indicators, introducing bias that overlooks early-career scientists and favors certain disciplines.Source: Introduction · pp. 2–3
• This study offers a first large-scale systematic identification of scientists across disciplines without bibliographic databases, adaptable to other Twitter-listed stakeholders and foundational for studying scholarly communication and altmetrics.Source: Introduction · p. 3

Product-centric perspective

• Only 9.4% of papers indexed by PubMed and Web of Science had Twitter mentions, although this coverage exceeded other social-media metrics except Mendeley.The systematic study covered 1.4 million papers.Source: Background · p. 3
• Twitter coverage varies by discipline: medical and social-science papers, which may appeal to wider publics, are more likely to be covered.Source: Background · p. 3
• A meta-analysis found a negligible correlation between altmetrics and citations (r = 0.003), with results differing by discipline, journal, and time window.Source: Background · p. 3

Producer-centric perspective

• A 2014 Nature survey found that more than 80% of respondents knew of Twitter, but only 13% used it regularly.The survey received more than 3,500 responses.Source: Background · pp. 3–4
• Studies examined 28 scholars sharing papers on Twitter, found many of 672 emergency physicians disconnected from colleagues but a few tightly interconnected, and reported more retweets by biochemists.Source: Background · p. 4
• Earlier studies often manually searched scientists selected outside Twitter, limiting sample size and favoring well-known researchers; a supervised approach relied on DBLP and covered computer science.Source: Background · p. 4

Scientist Occupations

• The authors adopt the 2010 Standard Occupational Classification (SOC) as a practical definition of scientist occupations and augment it with Wikipedia, compiling 322 scientist titles from the two sources (S1 Data).Source: Scientist Occupations · pp. 4–5

List-based Identification of Scientists

• The authors used 8,545 seed users for breadth-first snowball sampling, expanding through members of public lists whose names contained scientist titles.For each queued user, they retrieved matching lists and their members, repeating the process until the queue was empty.Source: List-based Identification of Scientists · p. 6
• The sampling misses scientists who are not listed and may exclude users whose names do not follow the standard format required by the method.The authors only considered users whose names contained spaces, potentially dropping people with non-English names or nonstandard name disclosure.Source: List-based Identification of Scientists · pp. 6–7
• 45,867 users remained after the authors filtered sampled accounts for profile descriptions containing scientist titles.The sampling produced 110,708 users in 4,920 lists before this precision filter.Source: List-based Identification of Scientists · p. 7

Analyzing Scientists

• The analysis collected followers, followings, and up to 3,200 recent statuses for 45,867 identified scientists, totaling 88,412,467 following pairs and 64,449,234 statuses.The dataset underlies the analyses of scientists’ demographics, shared URLs, and interactions.Source: Analyzing Scientists · p. 7

Who are they?

• Using profile descriptions first and list names when needed, the method assigned disciplines to 30,793 users (67.1%) and identified scientists across diverse disciplines.Profiles contained a scientist title for 25,798 users (56.2%); list names contained titles for 24,635 users (53.1%).Source: Who are they? · p. 8
• Social and computer and information scientists are over-represented on Twitter, while mathematical, life, and physical scientists are under-represented relative to the science workforce.This comparison is a rough estimate: the workforce data are U.S.-only, the sample may include users from other countries, and list-based sampling may bias results.Source: Who are they? · p. 9, Table 2 · p. 9
• Among the 71.9% of the sample with identified gender, 38.6% were female and 61.4% male, a ratio of 0.629 versus 0.428 for U.S. scientific authorships.The authors conclude that the gender ratio is less skewed among Twitter scientists than among U.S. scientific authorships, supporting more diverse participation by women.Source: Who are they? · p. 10

What do they share?

• News websites such as The Guardian and The New York Times are numerous; Nature Publishing Group ranks among top domains, major academic publishers appear among top scientific domains, and Science, PNAS, and PLoS ONE attract attention on Twitter.Source: What do they share? · pp. 10–11
• Scientists are more likely to share scientific content from their own disciplines, although some domains, such as nature.com, are popular across disciplines.For example, arxiv.org, a preprint server mainly for physics, is among the top scientific domains shared by scientists in each discipline.Source: What do they share? · p. 12
• Shared scientific domains varied by discipline: aps.org was among the top domains for physicists, acm.org was popular among computer scientists, and the LSE blog among political scientists, economists, and sociologists.Fig. 3 summarizes the top scientific domains by discipline.Source: What do they share? · p. 12, Figure 3 · p. 12
• Scientific URLs formed a small share of scientists’ linked posts across disciplines; biologists, neuroscientists, and ecologists shared more, while the fraction was below 0.2 for nearly all other scientists.The authors characterize sharing links to scientific domains as a minor activity for most scientists on Twitter.Source: What do they share? · p. 13, Figure 4 · p. 13

How do they connect to each other?

• In-degree or in-strength rankings overlapped with PageRank, while k-core leaders differed; Neil deGrasse Tyson ranked first under degree and PageRank in all three networks.The paper examines these rankings across follower, retweet, and mention networks.Source: How do they connect to each other? · p. 15
• Social and life scientists accounted for the largest portions of network centralities, followed by physical scientists, while mathematical and computer scientists occupied very small portions.Across the three networks, k-core distributions were less heterogeneous than the other centralities, and PageRank distributions were similar.Source: How do they connect to each other? · p. 15
• Life and physical scientists had disproportionately larger centralities across all three measures and networks, whereas computer, mathematical, and social scientists showed the opposite pattern.The authors normalized each group’s share of centrality by its share of scientists; values above one indicate disproportionate centrality.Source: How do they connect to each other? · p. 15
• In the follower network, disciplinary assortativity was 0.548 when nodes with unknown discipline were ignored, indicating scientists tended to follow others in their own discipline group.For gender, follower-network assortativity was 0.054, meaning same-gender following was no more frequent than expected by pure chance.Source: How do they connect to each other? · pp. 15–16

Discussion

• The method broadens identification beyond paper-centric sampling, uses occupational terms to include practitioners, and triangulates list- and bio-based views of identity.It favors precision over recall: identified accounts are treated as confidently scientists, although many scientists are missed. Sampling should not rely on external bibliometric data because Twitter scientists may differ in composition and behavior from those databases.Source: Discussion · p. 17
• Self-identified students and faculty comprised 21.9% of the identified population, and practitioners such as psychologists and nutritionists ranked among the five most represented disciplines.Historians, physicists, political and computer scientists, biologists, economists, and sociologists were also widely represented.Source: Discussion · p. 17
• Social scientists were overrepresented on Twitter relative to the scientific workforce, while mathematicians were particularly underrepresented.The discussion notes that Twitter was used across the disciplinary spectrum.Source: Discussion · p. 17
• Of scientists whose gender could be identified, 38.6% were female and 61.4% male, a more equal representation of women than in publication statistics.The authors suggest Twitter scientists may be more gender-balanced than publishing scientists.Source: Discussion · p. 18
• Scientists shared general social-media and news domains as well as scholarly sites, including Nature, Science, and arXiv, but scientific URLs were only a small fraction of tweets.The authors characterize scientists’ tweet content as heterogeneous, with personal and professional boundaries blurred under one handle.Source: Discussion · p. 18
• Social and life scientists dominated the follower, retweet, and mention networks, but social scientists underperformed after centralities were normalized by group size.The authors argue that altmetric indicators should normalize by field because disciplines are unevenly represented on these platforms.Source: Discussion · p. 18
• Scientists tended to follow others in their discipline, and historians were largely isolated in the Twitter network, suggesting disciplinary walls may limit broader access to scholarship.The authors caution that this may constrain the broader impact promised by social-media metrics.Source: Discussion · p. 18
• The approach is blind to unlisted scientists, may skew toward elite communicators, misses private-list members, and may lose users with names lacking spaces, biasing sampling toward English speakers.Unknown list curation may reduce precision, while filtering out profiles without scientist titles favors self-disclosed scientists.Source: Discussion · pp. 18–19

• The study developed a systematic method to identify scientists recognized by other Twitter users through lists or self-identified as scientists in their profiles.Source: Conclusion · p. 19
• Social scientists were over-represented, mathematical and physical scientists under-represented, and women better represented than in scholarly publishing statistics.Source: Conclusion · p. 19
• Only a small portion of URLs shared by identified scientists were science-related.Source: Conclusion · p. 19
• Scientists showed assortative mixing by discipline in follower, retweet, and mention networks.Source: Conclusion · p. 19
• Future work should test machine learning with retweet and mention data, investigate explanations for women's representation, and compare altmetric with bibliometric communities.Source: Conclusion · p. 19

S1 Table. Scientist occupations from 2010 Standard Occupational

• The scientist occupations are classified using the U.S. Department of Labor classification.Source: S1 Table. Scientist occupations from… · p. 20

S4 Table. Top scientists in the follower, retweet, and mention networks

• The S4 table considers scientists in network rankings using in-degree (d←) or in-strength (s←) and PageRank (PR).The section title specifies follower, retweet, and mention networks.Source: S4 Table. Top scientists in the… · p. 20

Getting initial seed users

• The authors collected 285,760,507 unique users from a Gardenhose dataset covering about 10% of public tweets, then obtained attributes for 2,436,889 users.After filtering, they selected 8,545 seed users whose attributes included “science” and a scientist title.Source: S1 Text · p. 24

Academic rank

• Academic rank was inferred from profile-description keywords; when multiple categories appeared, the first was selected, and professors disclose professional information on Twitter.Source: S1 Text · pp. 24–25

Community structure

• Infomap identified 343 follower-network communities with more than 10 nodes, which the authors labeled using the five most frequent profile-description words.They counted individual words in community members’ profile descriptions to characterize the communities.Source: S1 Text · p. 25
• Scientists’ follower communities appear discipline-based, with scientists following others in their own scholarly communities.Ecologists and biologists are tightly connected; astronomers and physicists, and political scientists, economists, and sociologists, also form connected groups.Source: S1 Text · p. 25