https://orcid.org/0000-0002-8441-3915
![Sample our Geography journals, sign in here to start your FREE access for 14 days]({"type":"image" , "src" : "/sda/5937/TOC-Subject-Sample-2021-Geography.jpg"})
ABSTRACT
This paper studies productivity spillovers from colleagues and co-authors simultaneously in computer science accounting for both quantity and quality of peers. For the identification of colleague spillovers, we model colleagues’ mobility behaviours and use the estimated staying colleagues to instrument for the actual colleagues. For the identification of co-author spillovers, we exploit the characteristics of (estimated staying) colleagues of a scientist’s non-colleague co-authors to construct instruments. Our results provide strong evidence that co-authors generate significant spillover effects, while colleagues generally do not. One additional co-author on average increases a scientist’s productivity by 2%. Further analyses suggest that co-author spillovers function through the sharing of ideas and knowledge, and researchers with more co-authors are more interdisciplinary.
ACKNOWLEDGEMENTS
The authors gratefully acknowledge the insightful comments and suggestions offered by the editor and three anonymous referees. We are thankful to Bruce Weinberg and Lung-fei Lee for helpful discussions. The errors and omissions are the authors’ own.
DISCLOSURE STATEMENT
No potential conflict of interest was reported by the author(s).
Notes
1 See Borjas and Doran (Citation2015) for exception.
2 Spillover effects and peer effects are used interchangeably in this paper.
3 For example, the author name disambiguation can combine articles by John Smith and J. Smith. Or it can split articles written by two authors whose names are both John Smith but are different in other dimensions.
4 This does not invalidate the relevance requirement of instruments when we use co-authors’ colleagues as instruments for co-authors, because a scientist’s group of colleagues also contain some of his or her co-authors.
5 Kim et al. (Citation2009) also report weakening and disappearing of local spillover effects since the 1980s and 1990s for elite universities.
6 Even earlier publications are also available, but the coverage in early years is not good.
7 ASJC codes are subject categories assigned to each article by subject experts based on its content. For more on ASJC, see https://service.elsevier.com/app/answers/detail/a_id/15181/supporthub/scopus/.
8 Genni is available at http://abel.lis.illinois.edu/cgi-bin/genni/search.cgi.
9 Ethnea maps people into 26 ethnicity groups: English, Hispanic, Chinese, German, Japanese, French, Italian, Slav, Indian, Arab, Korean, Vietnamese, Nordic, Dutch, Turkish, Israel, Greek, African, Hungarian, Thai, Romanian, Baltic, Indonesian, Caribbean, Mongolian and Polynesian. Ethnea data are available at http://abel.lis.illinois.edu/cgi-bin/ethnea/search.py.
10 The data are available at https://www.nsf.gov/statistics/herd/pub_data.cfm.
11 Including both federal and non-federal expenditure.
12 This could be due to false lumping of two or more scientists’ publications.
13 We use five-year forward citations. The results are not sensitive to using an alternative measure such as three-year forward citations.
14 Data in earlier or later years may be used to construct some of the variables. For example, data after 2011 may be used to calculate citations. Data in years before 1996 may be needed to construct a researcher’s career age.
15 ASJC subjects are coded at multiple levels. We use the two-digit ASJC codes, which include: Agricultural and Biological Sciences, Arts and Humanities, Biochemistry, Genetics and Molecular Biology, Business, Management and Accounting, Chemical Engineering, Chemistry, Computer Science, Decision Sciences, Earth and Planetary Sciences, Economics, Econometrics and Finance, Energy, Engineering, Environmental Science, Immunology and Microbiology, Materials Science, Mathematics, Medicine, Neuroscience, Nursing, Pharmacology, Toxicology and Pharmaceutics, Physics and Astronomy, Psychology, Social Sciences, Veterinary, Dentistry, Health Professions, and Multidisciplinary. We tabulate the frequencies of all two-digit ASJC codes that show up in a researcher’s publications and pick the subject that appears most frequently as the scientist’s main research field.
16 For example, if two researchers co-wrote a paper 10 years ago and never collaborate again, we do not consider them as co-authors in the current period. If they collaborate again within the last four years, then they are included in each other’s current co-author networks.
17 The number of co-authors presented here only considers a computer scientist’s co-authors in the field of computer science and does not include graduate students or co-authors in other disciplines. This is why the number of co-authors in seems to be low. Actually, computer scientists often have co-authors from other fields such as economics, sociology, physics and medicine.
18 We define the career age as the number of years since first publication. Data in all available years, instead of just the sample periods, are used to calculate career age.
19 Gender and ethnicity are not included when individual fixed effects are added.
20 In order to control for more flexible age patterns, we tried a variety of alternative functional forms, including a quadratic function and a full set of career–age dummies, but the results are literally unchanged.
21 Yet, it is worthwhile to note that the standard errors are relatively large because the variations in this variable drop considerably after controlling for department fixed effects.
22 This instrument strategy only accounts for the endogeneity issue associated with the year-to-year change in a scientist’s colleague network. However, any endogeneity associated with the initial group of colleagues is removed by the combination of individual fixed effects and department fixed effects.
23 Our method is similar to that of Staudt (Citation2015).
24 We also tried a different cutoff of 5%. The results are largely the same.
25 In some cases, it is difficult to find out the exact year when scientists move to different institutions if they have gaps in publishing years. We define the year of moving as the year when we observe a change in a researcher’s affiliation.
26 However, we do not claim this as a causal relationship since endogeneity issues are not addressed.
27 Especially for intra-institutional co-authors (Aman, Citation2016).
28 The impact of a journal is measured by its quantile in the distribution of the average number of citations received by an article published in the journal in a year. We looked at both three- and five-year forward citations. The results are similar.
29 There is a total of 334 four-digit ASJC codes.
30 Considering that the average productivity varies across the three subsamples, even in percentage terms, computer scientists from the higher group still experience larger spillovers.
31 For example, see https://grants.nih.gov/grants/guide/pa-files/PAR-17-244.html.