Review of an article using bibliometric + qual methods to study sub-discipline collaboration behavior

Mixed methods are always attractive, but many researchers give up because each method typically requires some epistemology which often conflicts with the epistemology of other methods. When mixed methods are done, they are often done in sequence. For example, qualitative work to understand enough about a phenomenon to develop a survey or interviewing survey respondents  to get richer information about their responses. Network methods are neither quantitative* nor qualitative and it's not typical to combine them with qualitative methods - hence my interest in this piece. Of course I'm also interested in collaboration in science.

The authors combine network analysis of the co-authorship network with qualitative interviews with the scientists to look at intergroup collaboration, migrations, and exchange of services or samples.

The citation:

Velden, T., Haque, A., & Lagoze, C. (2010). A new approach to analyzing patterns of collaboration in co-authorship networks: mesoscopic analysis and interpretation Scientometrics DOI: 10.1007/s11192-010-0224-6 (pre-print available at: http://arxiv.org/abs/0911.4761)

As background, the authors note some of the limitations of doing ethnographic studies - rich information about a very small group of people (transferability, but never generalizability) - and of doing large bibliometric studies (mapping a large crowd, but may miss nuances of sub-discipline research area, and also doesn't explain what the network features mean).

The qualitative part of their work is part of a larger ongoing study of chemistry research groups in the US and Europe.  The bibliometric information comes from topical queries in Web of Science. The keywords for the queries were selected to represent three sub-discipline research areas in chemistry. They went back 20 years, and kept the co-authors who had more than one paper in the retrieved set. The result sets were reviewed by participants in the qualitative portion to check that they were on topic.

They extracted the largest component (a component is all of the nodes that are somehow connected to each other), and then did some clustering, some discussing with their participants, calculated a bunch of centrality measures, used a method from Guimera to find "hubs", and they determine if links between the clusters are more of a transfer type or collaboration type depending on their robustness to the removal of one or two author nodes.

Their results:  The extracted clusters matched up with the scientists' immediate research groups plus a few external regular collaboration partners. In some fields, the PI-led groups were almost in a perfect star network shape where as in others, it was still hierarchical, but there was no one node that dominated the cluster.  Some of the clusters only had one or two authors connecting them, where as others had lots of collaborations. The authors asked about these connections, and found that the one or two author connections resulted from visiting professorships, career migration, one-off commissioned work, funded project collaboration of a sub-group leader. The many to many connections resulted from large collaborations on methods and the subject.

Two of the research specialties had mostly large clusters, whereas the third had many more small clusters. The size of the cluster was correlated with the numbers of papers published (not surprising). The field with more small clusters also had a lot more single-hub clusters whereas the other two fields had multiple hubs per cluster. The collaboration type (vs transfer type) connections are more likely to be geographically proximate.

Field A apparently requires large, expensive equipment, so there is some incentive to stronger integration/collaboration. Apparently collaboration isn't funded in the US for field C so there aren't as many in the US as there are in other parts of the world. Field B is some area of synthetic chemistry and fields A and C are some field of physical chemistry - so there do seem to be differences in the ways these authors work. B has more hub and spoke with a PI and his or her lab whereas A & C have more equal distributions with denser connections.

Commentary: Is this combination of methods useful, is the article successful, and did we learn new things from it?

I'm not sure the combination of methods is as new as I originally thought, but they certainly did integrate expertise from their participants at many different stages - and that's healthy. So often we just talk about the discipline level and that's really inadequate when you consider the diversity of research within chemistry, for example.  Elsewhere Velden et al discuss the difference between synthetic type chemists and other types, that seems to hold in this study, too. This paper was mostly about the methods (appropriate for the venue), but it would be nice to see this integrated with some of their other studies (I guess I'll have to wait for the dissertation). It's unfortunate that more detail can't be given on the precise research areas. That information is omitted to protect the privacy of the participants.

* the authors say that network methods are quantitative - I disagree. For one thing, they are about the connections, not about actor attributes. For another, you can't do regular statistics on them because they violate all of the independence of samples, normal distributions things... so any statistics have to be done by bootstrapping.