Exploring co-authorship networks in Statistics: methodological issues and empirical results

Several studies recognized scientific collaboration as a fundamental element in knowledge advancement since it facilitates interactions, exchanges, sharing methods, techniques and new ideas among scientists. Thanks to the availability of bibliographic archives, co-authorship networks have been analyzed in various fields as a proxy of scholars' collaborative behaviors. The present contribution aims at analyzing co-authorship networks in Statistics by defining groups of collaborative scientists. Two main methodological issues in the definition of collaboration data are discussed referring to the heterogeneity of the bibliographic archives available to derive co-authorship networks, and the disambiguation problem to obtain a correct identification of authors’ papers. Within this scenario, the focus here is in comparing different community detection algorithms to discover collaborative groups, and in analyzing the changes in the collaboration structure over time. Two main research questions motivate the interest: i) Does the presence of clusters related to authors' characteristics, localizations and field affiliations is confirmed by recognizing groups with community detection algorithms?; and ii) Does the stability of the composition of research groups and of collaboration behaviors can be affected by research assessment exercises to evaluate the activities of researchers and their research products? To this end, empirical results on Italian academic statisticians and their co-authorship relationships will be provided. Bibliographic data on authors and papers are extracted from online repositories by using web scraping techniques. Co-authorship patterns are examined before and after the first Italian research assessment exercise, carried out in 2013 by the National Agency for the Evaluation of the University and Research system. The networks are then described in order to underline changes in groups' collaboration structure by comparing results from community detection algorithms