Network construction

Our data consist of publication data from the careers of 394 researchers affiliated with the NI between 2004-2021. The analyses discussed here focus specifically on the subset of 208 individuals for whom we had a 4-year window of publications before and 4-year window of publications after the year they joined the NI, with the join year excluded from analysis. The join year was excluded to help account for any publication delay from work performed prior to joining the NI. This overall 9-year window was chosen based on a sensitivity analysis by which we determined that this window maximized the sample size and the duration of publications considered for each researcher before and after joining NI. Consistent with practices in bibliometrics research, where the choice of time window significantly influences a paper’s measured impact [31–33]), we conducted 12 experiments with varying time windows to determine the optimal range for our analysis. These experiments included testing pre- and post-join windows, as well as adjusting the total size of the gap surrounding the NI start date by excluding the join year and varying the gap length. Our time window of 4-years pre- and post-join is consistent with prior studies by Wang and colleagues [32], who found that longer windows (more than 1 year) better capture the citation impact of highly-cited papers. Abramo and colleagues’ [31] analysis of citation impact over an 8-year time window found that while time windows vary across disciplines, windows longer than 2 years provide a more reliable measure of research impact and productivity. The literature also suggests that it is important to have a gap between windows of measurement to account for publication delays [33,34], as the time lag between submission and publication has steadily increased across various disciplines. These studies demonstrate that publication delays tend to be shorter within the natural and physical sciences, such as those predominantly represented in our sample [33]. Hence, we concluded that at a gap length of one year would sufficiently differentiate between pre- and post-join publications, accounting for potential publication delays.

These individuals published 2,887 unique papers with 5,940 unique co-authors in the 4 years before joining the NI, and 4,147 unique papers with 9,160 unique co-authors in the 4 years after they joined the NI. Publication data, including details such as title, type, and journal/conference proceeding name, were collected using the Scopus publication database.

Researchers’ demographic information such as name, job position, year joined institution, and departmental affiliation, were provided through a roster of affiliates managed by the NI. We coded public-facing documents to identify the pronouns used to refer to each researcher, revealing a bias toward masculine pronouns in the sample (147 (70%) he/him, 45 (22%) she/her, and 16 (8%) non-gendered or missing public-facing gendered pronouns). Researchers held primary appointments in 53 unique academic departments at the University. Departmental membership was unevenly distributed such that a few departments contributed many NI members and many other departments contributed a few members. We cannot identify specific departments for privacy reasons, but the most conventional departments represented at the NI were biology departments (including animal, plant, and microbial foci), engineering, chemistry, and computer science disciplines. Less commonly represented departments included the social sciences and humanities. The majority of researchers in the dataset occupied tenure-track appointments in their home departments (71 (34%) Assistant Professor, 41 (20%) Associate Professor, 80 (38%) Full Professor, 16 (8%) other).

Data preprocessing was performed, including data cleaning, author and paper name disambiguation, and the creation of a paper-author network with associated attributes. Subsequent ego-centric network analysis and evaluation involved the extraction of key metrics of focal researchers authorship ego-networks for pre- and post- time periods such as unique paper and co-author counts, and various diversity indices based on departmental affiliations represented.

Control group

We required a control group to assess whether joining the NI was the source of any observed effect, rather than other potentially exogenous factors (e.g. changes occuring naturally in researchers’ careers over time, selection effects that influenced researchers with a predisposition for diverse networks to join the NI, etc.). We also knew other important factors driving publication structures in our dataset would likely stem from the disciplines represented at the NI and researchers’ points in their professional career. As such, we sought a control group reproducing the distribution of departments and hiring dates within our NI sample as closely as possible, with the key exception that members of the control group would never join the NI during their careers. We adopted a matched-pairing strategy to select individuals from within the broader university context of the NI to meet these criteria. We first chose a subset of individuals (egos) from our primary dataset to replicate the distribution of departments found in our larger NI sample. From the NI sample, we examined the distribution of unique departments represented among the 208 NI researchers in the sample (n=53) to ensure that the sample for the control group followed a similar distribution. We aimed for a control group size of about 30% of our primary sample. Thus, we selected 60 individuals representing 34 departments to approximate the distribution of departments in the primary dataset. For departments at the tail end of the distribution, we randomly sampled to include some of these departments in the control group, but could not include control representation for each of the tail-departments while still approximating the department distribution of the primary dataset. For each ego in this subset, we identified researchers who joined the same department as the ego at a comparable time and who also possessed a 9-year record of publication in our database surrounding the ego’s start date at the NI (see section on Network Construction for justification of our choice of this 9-year window). Another requirement for inclusion as a control was that these researchers would not eventually join the NI. To offer a point of comparison to our primary dataset, we assigned each of these matched individuals a "start year" of the same year that their referential ego in the primary dataset joined the NI. We repeated the same network construction processes described in the “Network Construction” section above for the control group, and thus created 60x2 ego-networks 4 years pre-start date and 4 years post-start date. This approach allowed us to create matched pairs to investigate the impact of joining the NI on research outcomes while controlling for confounding variables such as time and institutional resources.

Measures

Diversity indices are widely used in many scientific disciplines to quantify the distribution of types within a dataset and are perhaps most strongly associated with the field of ecology, where diversity indices are used to measure in a variety of ways the number and distribution of species in a community. Our previous work [35] reviewed diversity measures from communication, ecology, economics, and bibliometrics, and determined the three main metrics that were directly relevant to team science research: richness, evenness, and Shannon diversity index. In particular, we used the ecological operationalization of diversity as comprising two elements: the number of species in a community [i.e. richness [36] and their relative abundance [i.e. evenness [37]. Building on measures of richness and evenness, the Shannon diversity index [38] measures the evenness of the distribution of species within a system relative to its overall richness. This operationalization largely mirrors the concepts of variety, balance, and disparity discussed in prior works on interdisciplinary science [6,7]

We aimed to operationalize and measure three properties of interdisciplinarity, namely variety, balance, and disparity, [6–8], but due to substantial missing data on departmental affiliations for co-authors not affiliated with the NI, we could only reliably calculate variety. Therefore, in our primary analyses, we use richness (i.e. variety) as the main proxy for diversity of researchers’ collaboration networks, as it captures the breadth of unique connections, and had no missing data as it relies only on counts of papers and co-authors. This measurement approach complements and extends traditional bibliometric measures, which often focus on the variety of citation patterns [7,8,39], or number of disciplines presented [6,40]. While existing approaches measure interdisciplinarity through citation diversity or disciplinary variety at the paper or discipline level, we analyze diversity at the researcher level of analysis.

For descriptive purposes, the first section of our results describes the comparative departmental diversity in ego networks before and after joining NI. We do not compare the departmental diversity of the NI egos with the control group (described below) as departmental affiliation data were exclusively gathered for 394 NI-affiliated researchers in our complete dataset, so our calculation of departmental diversity between egos and alters was restricted to co-authorship collaborations among NI researchers. When interpreting these results, keep in mind that this analysis only examined co-authorship within the NI network (i.e. the 394 researchers in our primary dataset).

Analysis

We conducted generalized linear mixed effects regression (GLMER) modeling with Poisson distribution to determine whether the changes in the richness of each researcher’s numbers of papers and co-authors were significantly different for NI affiliates versus the control group. Our modeling approach was selected after an extensive exploration of the dataset to assess appropriate analytic paradigms. Anonymized data and code used for this project are available at https://github.com/lydinh92/NI_Research. Additional methodological details about data processing and descriptive statistics for our sample are described in our S1 Appendix.

Increase in researcher productivity and diversity of collaboration networks

We drew upon established measures of diversity from ecology (e.g. richness, Shannon evenness) to compare each researcher’s publication networks for the 4 years before and 4 years after joining the NI. Between their pre- and post-join networks, NI researchers had an average increase of 151% in the number of their publications (i.e. paper richness; from 15.6 to 23.7) and a 175% increase in the number of unique co-authors (i.e. co-author richness; from 35.5 to 62.1). After joining, Researchers co-authored with individuals in the NI network from more departments across the Institute’s campus (i.e. departmental richness; from 2.3 to 3.8 departments), and distributed their collaborations more equitably among those departments (i.e. median Shannon evenness increased from 0.791 to 0.895). These statistics suggest that researchers experienced substantial increases in the disciplinary diversity of their collaboration networks after joining the NI. However, because these descriptive results do not disentangle the effects of joining the NI from other potential exogenous factors, next we performed a comparison to a control group.

Comparison to control group indicates NI as source of unique diversification effect

We leveraged the NI’s dual-membership appointment model to construct a control group that would allow us to isolate the effects of joining the NI versus other factors known to affect researcher productivity, namely institutional resources and time. We began by selecting 60 focal researchers from our primary dataset approximating the distribution of home departments represented at the NI. From university records, we selected an individual that joined the same home department, at a similar academic rank, at a similar time as each focal individual in the NI. Then, we assigned each control individual the same start date as when their focal individual joined the NI, to facilitate comparison of collaboration networks at similar time points in their careers. Thus, our control group consists of researchers with similar hiring dates, career phases, and departmental affiliations to our NI dataset with a primary difference being that control group researchers did not join the NI. Descriptive statistics for demographics of the control group are available in the S1 Appendix. Limitations in data availability only permitted access to measures of richness for the control group.

For both groups, paper richness and co-author richness increased over time (Fig 1). The control group is similar to the NI sample in terms of their initial research productivity (i.e. the number of papers they published in each time period) and its increase over time (Fig 1a). But, interesting differences emerge when comparing the makeup of each individual’s co-authorship networks (i.e. the number of unique co-authors with whom they published in each time period; Fig 1b).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Differences in mean number of papers (top) and co-authors (bottom) depending on (1) NI membership and (2) time joined NI.

https://doi.org/10.1371/journal.pone.0322943.g001

Two generalized linear mixed effect (GLMER) models were used to compare the two groups in terms of paper richness (Model 1) and co-author richness (Model 2). Both models showed significant effects of NI membership (Model 1 IRR: 1.32, p=0.04; Model 2 IRR: 3.97, p<0.001) and Time (Model 1 IRR: 1.64, p<0.001; Model 2 IRR: 1.44, p<0.001) on paper and co-author richness (Fig 2). The second model also identified a significant interaction effect between Time and NI membership (IRR: 1.22, p=0.001; 95% CI: 1.09, 1.36), whereas the first model did not (IRR: 0.92, p=0.10; CI: 0.61, 1.33).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Regression estimates and incidence rate ratios from GLMER for number of papers (top) and coauthors (bottom).

https://doi.org/10.1371/journal.pone.0322943.g002

The significant model coefficients for NI membership indicate a selection effect wherein researchers who joined the NI exhibited a predisposition to higher productivity with a wider array of collaborators than members of the control group. The positive and significant incidence rates for NI membership in the GLMER predicts that an individual who would eventually join the NI tended to produce 32% more papers and publish with 297% more unique co-authors than an individual from the control group. Furthermore, the models both indicate significant effects for time, suggesting that regardless of NI membership, individuals publish more, and with more unique co-authors in the second time period.

The interaction effects across the two models, however, indicate that joining the NI produced an additional, unique influence on researchers above and beyond these effects. The absence of a significant interaction effect in Model 1 indicates that joining the NI did not result in any additional positive or negative influence on researchers’ paper productivity. The significant interaction coefficient in Model 2, however, indicates that researchers who joined the NI published with an increased diversity of co-authors after joining than would be expected by the other effects in the model. These effects suggest that the NI produced a significant influence on broadening researchers’ collaborations to include a wider variety of co-authors after joining the NI. These effects are particularly notable together: the models suggest that the increased diversity of NI researchers’ co-authorship networks was not simply driven by an increase in productivity, but through a disproportionate diversification of who they publish with after joining the NI.

One counter-explanation for this effect might arise from NI researchers simply publishing articles with substantially larger author lists after joining the NI. This was not the case. A secondary analysis indicated that although the size of co-author lists increased slightly over time across both NI and control groups, this effect was both small (an average increase of less than one co-author per paper for both NI and control groups) and the rate of change was commensurate for both NI and control groups.