The Structure of a Social Science Collaboration Network : Disciplinary Cohesion from 1963 to 1999

s, I include an indicator for how completely Sociological Abstracts indexes the journals where people publish. Coverage is indicated at three levels: complete (100% of the articles in that journal are indexed), priority (more than 50% of the articles are indexed), and selective (less than 50% are indexed). I construct the collaboration network by assigning an edge between any two people who wrote a paper together, regardless of the how often they have coauthored. Figure 2 demonM A R C Z A K B U IN E SS SE R V IC E S, IN C . 51856-2041 • AX 51856-0109 • -M AIL: rm azak@ r.com S E C O N D P R O O F STRUCTURE OF SOCIAL SCIENCE COLLABORATION NETWORK—–219 #1471-ASR 69:2 filename:69204-moody 09 A cutoff of 15 appearances of a name was used to distinguish common from uncommon. 10 Prior work on large collaboration networks has not attempted to identify these types of errors. An alternative source for name cleaning would be to use author aff iliation. Unfortunately, Sociological Abstracts, only lists affiliation for the first author, and authors may have multiple affiliations (such as a research center and an academic department). While such corrections are important to help ensure accurate measures, the general graph features examined here do not differ significantly if I use the corrected versus the non-corrected data. 11 This results in a probability score based on the proportion of people with a given first name who are male. Since not all first names can be matched, using this measure results in missing data for 32% of cases, likely predominantly among rare and foreign names. The substantive model results for the other variables do not differ with the inclusion of the gender measure. Tables not using gender are available from the author upon request. 12 The sociological abstract area categories are likely not substantively ideal, being subject to both errors of misclassification and internal heterogeneity. However, they remain the only tractable information on substantive area. While each record contains keywords describing content, the sheer number of such words (over 8000 unique keywords) would require some sort of categorization routine, the development of which is not transparent. 13 This is also an imperfect measure, since while all quantitative papers include tables some non-quantitative papers also include textual tables, so this measure over-estimates the number of quantitative papers. At the aggregate level used here, the small number of nonquantitative papers with tables washes out relative to differences across specialties. strates how the networks are constructed from the authorship data.14 The top panel of figure 2 is a schematic representation of data as given in Sociological Abstracts, with authors (squares) connected to the papers (circles) they write. The data include single authored papers (persons A,B,C and D) as well as those with more authors. The structure on the top right of figure 2 represents a large connected set of authors, each of whom has coauthored with someone who has coauthored with someone else. The bottom panel of figure 2 provides the resulting collaboration network. Those who have written only single authored papers do not participate in the collaboration network, but can be represented as structural isolates. Pairs of people who have only coauthored with each other are represented as isolated dyads {EF, GH}. The largest connected component is the maximal set of people who are connected by a chain of any length to each other. The large structure at the bottom right of figure 2 is the largest connected component. Nested within this component is a bicomponent (circled). While a component requires only a single traceable path between each actor, a bicomponent requires that there be at least two node-independent paths connecting every pair of actors in the network. Simmel (1950) argued that the necessary condition for a group is that a supra-individual body remains even if a person leaves. Bicomponents meet this criterion, since the group remains connected even if a single person is deleted (Moody and White 2003). This conception scales, as tricomponents (3-node independent paths), 4-components, and higher order k-components identify increasingly cohesive subgroups in a network. The degree distribution of the network is used to test the preferential attachment model. M A R C Z A K B U IN E SS SE R V IC E S, IN C . 51856-2041 • AX 51856-0109 • -M AIL: rm azak@ r.com S E C O N D P R O O F 220—–AMERICAN SOCIOLOGICAL REVIEW #1471-ASR 69:2 filename:69204-moody Figure 2. Constructing Collaboration Networks 14 Thanks to a reviewer for suggesting this figure. An actor’s degree is the number of unique people they are directly connected to, in this context the number of unique collaborators. The degree distribution for the example network is given in the lower right of figure 2. Geodesic paths define network distance, as the number of intermediaries on the shortest path connected two nodes in a network. So, for example, nodes L and S are 3 steps apart.